def test(data, weights=None, batch_size=32, imgsz=640, conf_thres=-1, # not for NMS iou_thres=0.25, # not for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling save_hybrid=False, # for hybrid auto-labelling save_conf=False, # save auto-label confidences plots=True, wandb_logger=None, compute_loss=None, half_precision=True, # dsv is_coco=False, max_by_class=True, opt=None): print_size, print_batches = 640, 3 log_errors = -1 if opt is not None: print_size = opt.print_size print_batches = opt.print_batches max_by_class = opt.max_by_class log_errors = opt.log_errors ct = ast.literal_eval(opt.ct) if len(ct) == 0: ct = None else: ct = None if print_batches < 0: print_batches = 1000 # Initialize/load model and set device training = model is not None if training: # i.e. called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() # device = select_device(opt.device, batch_size=batch_size) device = select_device(opt.device, batch_size=2) # Directories save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir # Load model model = attempt_load(weights, map_location=device) # load FP32 model gs = max(int(model.stride.max()), 32) # grid size (max stride) imgsz = check_img_size(imgsz, s=gs) # check img_size # Half half = device.type != 'cpu' and half_precision # half precision only supported on CUDA if half: model.half() # Configure model.eval() if isinstance(data, str): is_coco = data.endswith('coco.yaml') with open(data) as f: data = yaml.safe_load(f) check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes # iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 iouv = torch.arange(iou_thres, 1, 0.05).to(device) # iou_thres : 0.95 : 0.05 niou = iouv.numel() # Logging log_imgs = 0 if wandb_logger and wandb_logger.wandb: log_imgs = min(wandb_logger.log_imgs, 100) # Dataloader if not training: if device.type != 'cpu': model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters()))) # run once task = opt.task if opt.task in ('train', 'val', 'test') else 'val' # path to train/val/test images dataloader = create_dataloader(data[task], imgsz, batch_size, gs, opt, pad=0.5, rect=True, training=False, prefix=colorstr(f'{task}: '))[0] seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = data['names'] # names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} names_dict = {k: v for k, v in enumerate(names)} coco91class = coco80_to_coco91_class() fmt = '%{}s'.format(2 + max([len(s) for s in names])) s = (fmt + '%12s' * 7) % ('Class', 'Images', 'Targets', 'P', 'R', 'F1', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1, mf1 = 0., 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] error_log = [] for batch_i, (imgs, targets, paths, shapes) in enumerate(dataloader): imgs = imgs.to(device, non_blocking=True) imgs = imgs.half() if half else imgs.float() # uint8 to fp16/32 imgs /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = imgs.shape # batch size, channels, height, width # Run model t = time_synchronized() inf_out, train_out = model(imgs, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if compute_loss: loss += compute_loss([x.float() for x in train_out], targets)[1][:3] # box, obj, cls # Run NMS targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling t = time_synchronized() output = non_max_suppression(inf_out, labels=lb, multi_label=False, agnostic=True) # , conf_thres=conf_thres, iou_thres=iou_thres) t1 += time_synchronized() - t # pfunc(f'test_batch_size=={len(output)}') # Statistics per image idx = [] fn,fp = 0,0 for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] # Dims of all target boxes (in pixels) target_dims = targets[targets[:, 0] == si, -2:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path = Path(paths[si]) seen += 1 if len(pred) == 0: idx.append(None) if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Filter out-of-frame predictions (in padding) gain, img_box = shapes[si][1][0][0], None if gain == 1: img_box = torch.zeros([1, 4]) img_box[0, :2] = torch.FloatTensor(shapes[si][1][1]) img_box[0, 2:] = torch.FloatTensor(shapes[si][0]) + torch.FloatTensor(shapes[si][1][1]) img_box = img_box.to(device) io2s = box_io2(img_box, pred[:, :4]) k = (io2s > 0.95).nonzero(as_tuple=True)[1] pred = pred[k, :] idx.append(k) else: idx.append(None) # Predictions if single_cls: pred[:, 5] = 0 predn = pred.clone() scale_coords(imgs[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Dims of all pred boxes (in pixels) pred_target_dims = torch.zeros(pred.shape[0], 4, dtype=torch.float32, device=device) pred_target_dims[:, :2] = pred[:, 2:4] - pred[:, :2] # Append to text file if save_txt: gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0]] # normalization gain whwh for *xyxy, conf, cls in predn.tolist(): xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: f.write(('%g ' * len(line)).rstrip() % line + '\n') # W&B logging - Media Panel Plots if len(wandb_images) < log_imgs and wandb_logger.current_epoch > 0: # Check for test operation if wandb_logger.current_epoch % wandb_logger.bbox_interval == 0: box_data = [{"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, "class_id": int(cls), "box_caption": "%s %.3f" % (names[int(cls)], conf), "scores": {"class_score": conf}, "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()] boxes = {"predictions": {"box_data": box_data, "class_labels": names_dict}} # inference-space wandb_images.append(wandb_logger.wandb.Image(imgs[si], boxes=boxes, caption=path.name)) wandb_logger.log_training_progress(predn, path, names) if wandb_logger and wandb_logger.wandb_run else None # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int(path.stem) if path.stem.isnumeric() else path.stem box = xyxy2xywh(predn[:, :4]) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({'image_id': image_id, 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5)}) # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) scale_coords(imgs[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix.process_batch(predn, torch.cat((labels[:, 0:1], tbox), 1)) # Per target class for cls in torch.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero(as_tuple=False).view(-1) # target indices pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1) # prediction indices # Search for detections if pi.shape[0]: # Prediction to target ious ious, i = box_iou(predn[pi, :4], tbox[ti]).max(1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti[i[j]] # detected target... index into target array.... if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn k = pi[j] # index into pred array..... pred_target_dims[k, 2:] = target_dims[d] if len(detected) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls, pred_target_dims.cpu(), target_dims.cpu())) # FP/FN counts (per image) if log_errors > -1: corr = correct.cpu().numpy() idx_conf = pred[:, 4].cpu() > conf_thres if idx_conf.cpu().numpy().mean() < 1: corr = corr[idx_conf] tp = corr[:, 0].sum() fp, fn = len(corr) - tp, len(tcls) - tp if fp + fn > log_errors: error_log.append(path.stem) # Plot images if plots and batch_i < print_batches: prefix = Path(paths[0]).stem if batch_size == 1 else f'test_batch{batch_i}' f1 = save_dir / f'{prefix}_labels.jpg' # labels thread1 = Thread(target=plot_images, args=(imgs, targets, paths, f1, names, print_size), daemon=True) f2 = save_dir / f'{prefix}_pred.jpg' # predictions thread2 = Thread(target=plot_images, args=(imgs, output_to_target(output, idx), paths, f2, names, print_size), daemon=True) thread1.start() thread1.join() thread2.start() thread2.join() ################################## # show FP and FN detections... if log_errors > -1 and batch_size == 1: # fn boxes... if fn > 0: idx_fn = np.ones(len(targets)) for d in detected: idx_fn[d.item()] = 0 idx_fn = np.nonzero(idx_fn)[0] f = save_dir / f'{prefix}_fn.jpg' # labels thread = Thread(target=plot_images, args=(imgs, targets[idx_fn], paths, f, names, print_size, 16, True), daemon=True) thread.start() thread.join() # fp boxes... if fp > 0: idx_fp = ~corr[:, 0] f = save_dir / f'{prefix}_fp.jpg' # labels thread = Thread(target=plot_images, args=(imgs, output_to_target(output, idx, idx_fp, idx_conf), paths, f, names, print_size, 16, True), daemon=True) thread.start() thread.join() # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy conf_best = -1 if len(stats) and stats[0].any(): mp, mr, map50, map, mf1, ap_class, conf_best, nt, (p, r, ap50, ap, f1, cc) = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names, ct=ct, max_by_class=max_by_class, conf_thres=conf_thres) else: nt = torch.zeros(1) # Print results pfunc('------------------------------------------- Validation Set -----------------------------------------------') pfunc(s) pf = fmt + '%12.3g' * 7 # print format # Print results per class if (verbose or nc < 50) and nc > 1 and len(stats): for i, c in enumerate(ap_class): pfunc(pf % (names[c], seen, nt[c], p[i], r[i], f1[i], ap50[i], ap[i])) # log # Print averages if nc > 1: pfunc('') pfunc(pf % ('AVG', seen, nt.sum(), p.mean(), r.mean(), f1.mean(), ap50.mean(), ap.mean())) # unweighted average ss = 'WEIGHTED AVG' if nc > 1 else names[0] pfunc(pf % (ss, seen, nt.sum(), mp, mr, mf1, map50, map)) # weighted average (if nc>1) if conf_best > -1: pfunc('\nOptimal Confidence Threshold: {0:0.3f}'.format(conf_best)) # log if max_by_class: pfunc('Optimal Confidence Thresholds (Per-Class): {}'.format(list(cc.round(3)))) # log # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=names) if wandb_logger and wandb_logger.wandb: val_batches = [wandb_logger.wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg'))] wandb_logger.log({"Validation": val_batches}) if wandb_images: wandb_logger.log({"Bounding Box Debugger/Images": wandb_images}) # Save JSON if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights anno_json = '../coco/annotations/instances_val2017.json' # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.img_files] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[:2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print(f'pycocotools unable to run: {e}') # Return results model.float() # for training if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {save_dir}{s}") # if len(error_log)>0: # fn = f'{save_dir}/error_log.txt' # save_list(error_log, fn) maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, mf1, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def train(hyp, opt, device, tb_writer=None): logger.info( colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) save_dir, epochs, batch_size, total_batch_size, weights, rank = \ Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Directories wdir = save_dir / 'weights' wdir.mkdir(parents=True, exist_ok=True) # make dir last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = save_dir / 'results.txt' # Save run settings with open(save_dir / 'hyp.yaml', 'w') as f: yaml.safe_dump(hyp, f, sort_keys=False) with open(save_dir / 'opt.yaml', 'w') as f: yaml.safe_dump(vars(opt), f, sort_keys=False) # Configure plots = not opt.evolve # create plots cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.safe_load(f) # data dict # Logging- Doing this before checking the dataset. Might update data_dict loggers = {'wandb': None} # loggers dict if rank in [-1, 0]: opt.hyp = hyp # add hyperparameters run_id = torch.load(weights).get('wandb_id') if weights.endswith( '.pt') and os.path.isfile(weights) else None wandb_logger = WandbLogger(opt, save_dir.stem, run_id, data_dict) loggers['wandb'] = wandb_logger.wandb data_dict = wandb_logger.data_dict if wandb_logger.wandb: weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp # WandbLogger might update weights, epochs if resuming nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes names = ['item'] if opt.single_cls and len( data_dict['names']) != 1 else data_dict['names'] # class names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check is_coco = opt.data.endswith('coco.yaml') and nc == 80 # COCO dataset # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): weights = attempt_download( weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create exclude = [ 'anchor' ] if (opt.cfg or hyp.get('anchors')) and not opt.resume else [ ] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] # Freeze freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay logger.info(f"Scaled weight_decay = {hyp['weight_decay']}") pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR if opt.linear_lr: lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp[ 'lrf'] # linear else: lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # EMA ema = ModelEMA(model) if rank in [-1, 0] else None # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # EMA if ema and ckpt.get('ema'): ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) ema.updates = ckpt['updates'] # Results if ckpt.get('training_results') is not None: results_file.write_text( ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = max(int(model.stride.max()), 32) # grid size (max stride) nl = model.model[ -1].nl # number of detection layers (used for scaling hyp['obj']) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: ')) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # Process 0 if rank in [-1, 0]: testloader = create_dataloader( test_path, imgsz_test, batch_size * 2, gs, opt, # testloader hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, pad=0.5, prefix=colorstr('val: '))[0] if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) if plots: plot_labels(labels, names, save_dir, loggers) if tb_writer: tb_writer.add_histogram('classes', c, 0) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) model.half().float() # pre-reduce anchor precision # DDP mode if cuda and rank != -1: model = DDP( model, device_ids=[opt.local_rank], output_device=opt.local_rank, # nn.MultiheadAttention incompatibility with DDP https://github.com/pytorch/pytorch/issues/26698 find_unused_parameters=any( isinstance(layer, nn.MultiheadAttention) for layer in model.modules())) # Model parameters hyp['box'] *= 3. / nl # scale to layers hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers hyp['label_smoothing'] = opt.label_smoothing model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) model.class_weights = labels_to_class_weights( dataset.labels, nc).to(device) * nc # attach class weights model.names = names # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) compute_loss = ComputeLoss(model) # init loss class logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n' f'Using {dataloader.num_workers} dataloader workers\n' f'Logging results to {save_dir}\n' f'Starting training for {epochs} epochs...') for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices if rank in [-1, 0]: cw = model.class_weights.cpu().numpy() * ( 1 - maps)**2 / nc # class weights iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights dataset.indices = random.choices( range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx # Broadcast if DDP if rank != -1: indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'labels', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device)) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode if opt.quad: loss *= 4. # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot if plots and ni < 3: f = save_dir / f'train_batch{ni}.jpg' # filename Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start() if tb_writer and ni == 0: with warnings.catch_warnings(): warnings.simplefilter( 'ignore') # suppress jit trace warning tb_writer.add_graph( torch.jit.trace(de_parallel(model), imgs, strict=False), []) # graph elif plots and ni == 10 and wandb_logger.wandb: wandb_logger.log({ 'Mosaics': [ wandb_logger.wandb.Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg') if x.exists() ] }) # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP ema.update_attr(model, include=[ 'yaml', 'nc', 'hyp', 'gr', 'names', 'stride', 'class_weights' ]) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP wandb_logger.current_epoch = epoch + 1 results, maps, times = test.test(data_dict, batch_size=batch_size * 2, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, save_json=is_coco and final_epoch, verbose=nc < 50 and final_epoch, plots=plots and final_epoch, wandb_logger=wandb_logger, compute_loss=compute_loss, is_coco=is_coco) # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # append metrics, val_loss # Log tags = [ 'train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2' ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): if tb_writer: tb_writer.add_scalar(tag, x, epoch) # tensorboard if wandb_logger.wandb: wandb_logger.log({tag: x}) # W&B # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi wandb_logger.end_epoch(best_result=best_fitness == fi) # Save model if (not opt.nosave) or (final_epoch and not opt.evolve): # if save ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': results_file.read_text(), 'model': deepcopy(de_parallel(model)).half(), 'ema': deepcopy(ema.ema).half(), 'updates': ema.updates, 'optimizer': optimizer.state_dict(), 'wandb_id': wandb_logger.wandb_run.id if wandb_logger.wandb else None } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) if wandb_logger.wandb: if ((epoch + 1) % opt.save_period == 0 and not final_epoch) and opt.save_period != -1: wandb_logger.log_model(last.parent, opt, epoch, fi, best_model=best_fitness == fi) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: logger.info( f'{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.\n' ) if plots: plot_results(save_dir=save_dir) # save as results.png if wandb_logger.wandb: files = [ 'results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')] ] wandb_logger.log({ "Results": [ wandb_logger.wandb.Image(str(save_dir / f), caption=f) for f in files if (save_dir / f).exists() ] }) if not opt.evolve: if is_coco: # COCO dataset for m in [last, best ] if best.exists() else [last]: # speed, mAP tests results, _, _ = test.test(opt.data, batch_size=batch_size * 2, imgsz=imgsz_test, conf_thres=0.001, iou_thres=0.7, model=attempt_load( m, device).half(), single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, save_json=True, plots=False, is_coco=is_coco) # Strip optimizers for f in last, best: if f.exists(): strip_optimizer(f) # strip optimizers if wandb_logger.wandb: # Log the stripped model wandb_logger.wandb.log_artifact( str(best if best.exists() else last), type='model', name='run_' + wandb_logger.wandb_run.id + '_model', aliases=['latest', 'best', 'stripped']) wandb_logger.finish_run() else: dist.destroy_process_group() torch.cuda.empty_cache() return results
def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info(f'Hyperparameters {hyp}') save_dir, epochs, batch_size, total_batch_size, weights, rank = \ Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Directories wdir = save_dir / 'weights' wdir.mkdir(parents=True, exist_ok=True) # make dir last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = save_dir / 'results.txt' # Save run settings with open(save_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(save_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure plots = not opt.evolve # create plots cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int(data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round(hyp['anchors']) # force autoanchor model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Logging if wandb and wandb.run is None: opt.hyp = hyp # add hyperparameters wandb_run = wandb.init(config=opt, resume="allow", project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, name=save_dir.stem, id=ckpt.get('wandb_id') if 'ckpt' in locals() else None) loggers = {'wandb': wandb} # loggers dict # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (weights, epochs) if epochs < start_epoch: logger.info('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size] # verify imgsz are gs-multiples # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # EMA ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, image_weights=opt.image_weights) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1) # Process 0 if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt, # testloader hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, pad=0.5)[0] if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) if plots: Thread(target=plot_labels, args=(labels, save_dir, loggers), daemon=True).start() if tb_writer: tb_writer.add_histogram('classes', c, 0) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights model.names = names # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0) # P, R, [email protected], [email protected], val_loss(box, obj, cls) with open('tmp/val_loss.txt', 'r', encoding='utf-8') as f: val_loss = float(f.read()) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) logger.info('Image sizes %g train, %g test\n' 'Using %g dataloader workers\nLogging results to %s\n' 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs)) # Early stopping early_stopping = EarlyStopping(patience=3, verbose=True) for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices if rank in [-1, 0]: cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx # Broadcast if DDP if rank != -1: indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float() / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max(1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) if 'momentum' in x: x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_items = compute_loss(pred, targets.to(device), model) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ( '%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot if plots and ni < 3: f = save_dir / f'train_batch{ni}.jpg' # filename Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start() # if tb_writer: # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard elif plots and ni == 3 and wandb: wandb.log({"Mosaics": [wandb.Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg')]}) # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema: ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test(opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, plots=plots and final_epoch, log_imgs=opt.log_imgs if wandb else 0) # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Log tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2'] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): if tb_writer: tb_writer.add_scalar(tag, x, epoch) # tensorboard if wandb: wandb.log({tag: x}) # W&B # Update best mAP fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = {'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict(), 'wandb_id': wandb_run.id if wandb else None} # Save last, best and delete torch.save(ckpt, last) torch.save(ckpt, 'tmp/best.pt') if best_fitness == fi: with open('tmp/mAP.txt', 'w', encoding='utf-8') as f: f.write(str('%.3f'%results[-4])) torch.save(ckpt, best) # torch.save(ckpt, 'tmp/best.pt') # Early stopping print('val_loss:', round(sum(results[-3:]), 3)) if sum(results[-3:]) < val_loss: # torch.save(ckpt, best) # torch.save(ckpt, 'tmp/best.pt') val_loss = sum(results[-3:]) with open('tmp/val_loss.txt', 'w', encoding='utf-8') as f: f.write(str(val_loss)) early_stopping(round(sum(results[-3:]), 3), ckpt) if early_stopping.early_stop: print("Early stopping") torch.cuda.empty_cache() break del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers for f in [last, best]: if f.exists(): # is *.pt strip_optimizer(f) # strip optimizer os.system('gsutil cp %s gs://%s/weights' % (f, opt.bucket)) if opt.bucket else None # upload # Plots if plots: plot_results(save_dir=save_dir) # save as results.png if wandb: files = ['results.png', 'precision_recall_curve.png', 'confusion_matrix.png'] wandb.log({"Results": [wandb.Image(str(save_dir / f), caption=f) for f in files if (save_dir / f).exists()]}) logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) # Test best.pt if opt.data.endswith('coco.yaml') and nc == 80: # if COCO results, _, _ = test.test(opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=attempt_load(best if best.exists() else last, device).half(), single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, save_json=True, # use pycocotools plots=False) else: dist.destroy_process_group() wandb.run.finish() if wandb and wandb.run else None torch.cuda.empty_cache() return results
def test( data, weights=None, batch_size=16, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir='', merge=False, save_txt=False): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() device = select_device(opt.device, batch_size=batch_size) merge, save_txt = opt.merge, opt.save_txt # use Merge NMS, save *.txt labels if save_txt: out = Path('inference/output') if os.path.exists(out): shutil.rmtree(out) # delete output folder os.makedirs(out) # make new output folder # Remove previous for f in glob.glob(str(Path(save_dir) / 'test_batch*.jpg')): os.remove(f) # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Dataloader if not training: img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img _ = model(img.half() if half else img ) if device.type != 'cpu' else None # run once path = data['test'] if opt.task == 'test' else data[ 'val'] # path to val/test images dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, hyp=None, augment=False, cache=True, pad=0.5, rect=True)[0] seen = 0 names = model.names if hasattr(model, 'names') else model.module.names coco91class = coco80_to_coco91_class() s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class = [], [], [], [] evaluator = COCOEvaluator(root=DATA_ROOT, model_name=opt.weights.replace('.pt', '')) for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width whwh = torch.Tensor([width, height, width, height]).to(device) # Disable gradients with torch.no_grad(): # Run model t = time_synchronized() inf_out, train_out = model( img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if training: # if model has loss hyperparameters loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # GIoU, obj, cls # Run NMS t = time_synchronized() output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, merge=merge) t1 += time_synchronized() - t # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class seen += 1 if pred is None: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Append to text file if save_txt: gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0 ]] # normalization gain whwh x = pred.clone() x[:, :4] = scale_coords(img[si].shape[1:], x[:, :4], shapes[si][0], shapes[si][1]) # to original for *xyxy, conf, cls in x: xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh with open(str(out / Path(paths[si]).stem) + '.txt', 'a') as f: f.write( ('%g ' * 5 + '\n') % (cls, *xywh)) # label format # Clip boxes to image bounds clip_coords(pred, (height, width)) # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = Path(paths[si]).stem box = pred[:, :4].clone() # xyxy scale_coords(img[si].shape[1:], box, shapes[si][0], shapes[si][1]) # to original shape box = xyxy2xywh(box) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): result = { 'image_id': int(image_id) if image_id.isnumeric() else image_id, 'category_id': coco91class[int(p[5])], 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5) } jdict.append(result) #evaluator.add([result]) #if evaluator.cache_exists: # break # # Assign all predictions as incorrect # correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) # if nl: # detected = [] # target indices # tcls_tensor = labels[:, 0] # # # target boxes # tbox = xywh2xyxy(labels[:, 1:5]) * whwh # # # Per target class # for cls in torch.unique(tcls_tensor): # ti = (cls == tcls_tensor).nonzero(as_tuple=False).view(-1) # prediction indices # pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1) # target indices # # # Search for detections # if pi.shape[0]: # # Prediction to target ious # ious, i = box_iou(pred[pi, :4], tbox[ti]).max(1) # best ious, indices # # # Append detections # detected_set = set() # for j in (ious > iouv[0]).nonzero(as_tuple=False): # d = ti[i[j]] # detected target # if d.item() not in detected_set: # detected_set.add(d.item()) # detected.append(d) # correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn # if len(detected) == nl: # all targets already located in image # break # # # Append statistics (correct, conf, pcls, tcls) # stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # # Plot images # if batch_i < 1: # f = Path(save_dir) / ('test_batch%g_gt.jpg' % batch_i) # filename # plot_images(img, targets, paths, str(f), names) # ground truth # f = Path(save_dir) / ('test_batch%g_pred.jpg' % batch_i) # plot_images(img, output_to_target(output, width, height), paths, str(f), names) # predictions evaluator.add(jdict) evaluator.save()
def test( data, weights=None, batch_size=16, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling plots=True): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() device = select_device(opt.device, batch_size=batch_size) save_txt = opt.save_txt # save *.txt labels if save_txt: out = Path('inference/output') if os.path.exists(out): shutil.rmtree(out) # delete output folder os.makedirs(out) # make new output folder # Remove previous for f in glob.glob(str(save_dir / 'test_batch*.jpg')): os.remove(f) # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Dataloader if not training: img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img _ = model(img.half() if half else img ) if device.type != 'cpu' else None # run once path = data['test'] if opt.task == 'test' else data[ 'val'] # path to val/test images dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, hyp=None, augment=False, cache=False, pad=0.5, rect=True)[0] seen = 0 names = model.names if hasattr(model, 'names') else model.module.names coco91class = coco80_to_coco91_class() s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class = [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width whwh = torch.Tensor([width, height, width, height]).to(device) # Disable gradients with torch.no_grad(): # Run model t = time_synchronized() inf_out, train_out = model( img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if training: # if model has loss hyperparameters loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # GIoU, obj, cls # Run NMS t = time_synchronized() output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres) t1 += time_synchronized() - t # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class seen += 1 if pred is None: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Append to text file if save_txt: gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0 ]] # normalization gain whwh x = pred.clone() x[:, :4] = scale_coords(img[si].shape[1:], x[:, :4], shapes[si][0], shapes[si][1]) # to original for *xyxy, conf, cls in x: xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh with open(str(out / Path(paths[si]).stem) + '.txt', 'a') as f: f.write( ('%g ' * 5 + '\n') % (cls, *xywh)) # label format # Clip boxes to image bounds clip_coords(pred, (height, width)) # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = Path(paths[si]).stem box = pred[:, :4].clone() # xyxy scale_coords(img[si].shape[1:], box, shapes[si][0], shapes[si][1]) # to original shape box = xyxy2xywh(box) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({ 'image_id': int(image_id) if image_id.isnumeric() else image_id, 'category_id': coco91class[int(p[5])], 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5) }) # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) * whwh # Per target class for cls in torch.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero(as_tuple=False).view( -1) # prediction indices pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view( -1) # target indices # Search for detections if pi.shape[0]: # Prediction to target ious ious, i = box_iou(pred[pi, :4], tbox[ti]).max( 1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct[ pi[j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append( (correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # Plot images if plots and batch_i < 1: f = save_dir / ('test_batch%g_gt.jpg' % batch_i) # filename plot_images(img, targets, paths, str(f), names) # ground truth f = save_dir / ('test_batch%g_pred.jpg' % batch_i) plot_images(img, output_to_target(output, width, height), paths, str(f), names) # predictions # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, fname=save_dir / 'precision-recall_curve.png') p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean( 1) # [P, R, [email protected], [email protected]:0.95] mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%12.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if verbose and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print( 'Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Save JSON if save_json and len(jdict): f = 'detections_val2017_%s_results.json' % \ (weights.split(os.sep)[-1].replace('.pt', '') if isinstance(weights, str) else '') # filename print('\nCOCO mAP with pycocotools... saving %s...' % f) with open(f, 'w') as file: json.dump(jdict, file) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval imgIds = [int(Path(x).stem) for x in dataloader.dataset.img_files] cocoGt = COCO( glob.glob('../coco/annotations/instances_val*.json') [0]) # initialize COCO ground truth api cocoDt = cocoGt.loadRes(f) # initialize COCO pred api cocoEval = COCOeval(cocoGt, cocoDt, 'bbox') cocoEval.params.imgIds = imgIds # image IDs to evaluate cocoEval.evaluate() cocoEval.accumulate() cocoEval.summarize() map, map50 = cocoEval.stats[: 2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print('ERROR: pycocotools unable to run: %s' % e) # Return results model.float() # for training maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def test(cfg = None, data = None, weights=None, batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling save_hybrid=False, # for hybrid auto-labelling save_conf=False, # save auto-label confidences plots=True): # Initialize/load model and set device training = model is not None if not training: # called by train.py # called directly set_logging() # Directories save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir # Load model model = Model(cfg) model.load(weights) model = model.fuse() imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Configure model.eval() is_coco = data.endswith('coco.yaml') # is COCO dataset with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes iouv = jt.linspace(0.5, 0.95, 10) # iou vector for [email protected]:0.95 niou = iouv.numel() # Dataloader if not training: img = jt.zeros((1, 3, imgsz, imgsz)) # init img path = data['test'] if opt.task == 'test' else data['val'] # path to val/test images dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True, prefix=colorstr('test: ' if opt.task == 'test' else 'val: ')) seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} coco91class = coco80_to_coco91_class() s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = jt.zeros((3,)) jdict, stats, ap, ap_class = [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): img = img.float32() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets nb, _, height, width = img.shape # batch size, channels, height, width with jt.no_grad(): # Run model t = time_synchronized() inf_out, train_out = model(img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if training: loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # box, obj, cls # Run NMS targets[:, 2:] *= jt.array([width, height, width, height]) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling t = time_synchronized() output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb) t1 += time_synchronized() - t # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path = Path(paths[si]) seen += 1 if len(pred) == 0: if nl: stats.append((jt.zeros((0, niou), dtype="bool"), jt.array([]), jt.array([]), tcls)) continue # Predictions predn = pred.clone() predn[:, :4] = scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Append to text file if save_txt: gn = jt.array(shapes[si][0])[jt.array([1, 0, 1, 0])] # normalization gain whwh for *xyxy, conf, cls in predn.tolist(): xywh = (xyxy2xywh(jt.array(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: f.write(('%g ' * len(line)).rstrip() % line + '\n') # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int(path.stem) if path.stem.isnumeric() else path.stem box = xyxy2xywh(predn[:, :4]) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({'image_id': image_id, 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5)}) # Assign all predictions as incorrect correct = jt.zeros((pred.shape[0], niou), dtype="bool") if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) tbox = scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix.process_batch(predn, jt.contrib.concat((labels[:, 0:1], tbox), 1)) # Per target class for cls in jt.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero().view(-1) # prediction indices pi = (cls == pred[:, 5]).nonzero().view(-1) # target indices # Search for detections if pi.shape[0]: # Prediction to target ious i ,ious = box_iou(predn[pi, :4], tbox[ti]).argmax(1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(): d = ti[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn if len(detected) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append((correct.numpy(), pred[:, 4].numpy(), pred[:, 5].numpy(), tcls)) # Plot images if plots and batch_i < 3: f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start() f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(img, output_to_target(output), paths, f, names), daemon=True).start() # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) ap50, ap = ap[:, 0], ap.mean(1) # [email protected], [email protected]:0.95 mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = np.zeros((1,)) # Print results pf = '%20s' + '%12.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if (verbose or (nc <= 20 and not training)) and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) # Save JSON if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights anno_json = '../coco/annotations/instances_val2017.json' # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.img_files] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[:2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print(f'pycocotools unable to run: {e}') # Return results if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {save_dir}{s}") maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.numpy() / len(dataloader)).tolist()), maps, t
def test( data, weights=None, batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling save_hybrid=False, # for hybrid auto-labelling save_conf=False, # save auto-label confidences plots=True, log_imgs=0, # number of logged images compute_loss=None): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() device = select_device(opt.device, batch_size=batch_size) # Directories save_dir = Path( increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir( parents=True, exist_ok=True) # make dir # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() is_coco = data.endswith('coco.yaml') # is COCO dataset with open(data) as f: data = yaml.load(f, Loader=yaml.SafeLoader) # model dict check_dataset(data) # check nc1 = 1 if single_cls else int(data['nc1']) # number of classes # edit nc2 = 1 if single_cls else int(data['nc2']) # number of classes # edit iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Logging log_imgs, wandb = min(log_imgs, 100), None # ceil try: import wandb # Weights & Biases except ImportError: log_imgs = 0 # Dataloader if not training: if device.type != 'cpu': model( torch.zeros(1, 3, imgsz, imgsz).to(device).type_as( next(model.parameters()))) # run once path = data['test'] if opt.task == 'test' else data[ 'val'] # path to val/test images dataloader = create_dataloader( path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True, prefix=colorstr('test: ' if opt.task == 'test' else 'val: '))[0] seen = 0 confusion_matrix1 = ConfusionMatrix(nc=nc1) confusion_matrix2 = ConfusionMatrix(nc=nc2) # edit names1 = { k: v for k, v in enumerate( model.names1 if hasattr(model, 'names1') else model.module.names1) } # edit names2 = { k: v for k, v in enumerate( model.names2 if hasattr(model, 'names2') else model.module.names2) } # edit coco91class = coco80_to_coco91_class() s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') t0, t1 = 0., 0. p_1, r_1, f1_1, mp_1, mr_1, map50_1, map_1 = 0., 0., 0., 0., 0., 0., 0. p_2, r_2, f1_2, mp_2, mr_2, map50_2, map_2 = 0., 0., 0., 0., 0., 0., 0. # edit loss = torch.zeros(4, device=device) # edit jdict, stats1, stats2, ap_1, ap_2, ap50_1, ap50_2, ap_class_1, ap_class_2, wandb_images =\ [], [], [], [], [], [], [], [], [], [] # edit # targets: img_id, cls1, cls2, xywh # edit for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): # print("targets 105:", targets.shape, targets) # todo img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width with torch.no_grad(): # Run model t = time_synchronized() inf_out, train_out = model( img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if compute_loss: # print("new_loss:", new_loss.shape) # todo loss += compute_loss( [x.float() for x in train_out], targets)[1][:4] # box, obj, cls1, cls2 # edit # Run NMS targets[:, 3:] *= torch.Tensor([width, height, width, height ]).to(device) # to pixels # edit lb = [targets[targets[:, 0] == i, 1:] for i in range(nb) ] if save_hybrid else [] # for autolabelling t = time_synchronized() output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb, nc1=nc1, nc2=nc2) # edit t1 += time_synchronized() - t # Statistics per image for si, pred in enumerate( output): # pred: xyxy, conf1, cls1, conf2, cls2 # edit labels = targets[targets[:, 0] == si, 1:] # labels: cls1, cls2, xywh # edit # print("labels 130:", targets.shape, pred.shape, labels.shape, targets, labels) # todo nl = len(labels) tcls1 = labels[:, 0].tolist() if nl else [] # target class 1 tcls2 = labels[:, 1].tolist() if nl else [] # target class 2 # edit path = Path(paths[si]) seen += 1 if len(pred) == 0: if nl: stats1.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls1)) stats2.append( (torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls2)) # edit continue # Predictions predn = pred.clone() scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Append to text file if save_txt: gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0 ]] # normalization gain whwh for *xyxy, conf1, cls1, conf2, cls2 in predn.tolist(): xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh line = (cls1, cls2, *xywh, conf1, conf2) if save_conf else (cls1, cls2, *xywh ) # label format # edit with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: f.write(('%g ' * len(line)).rstrip() % line + '\n') # W&B logging # edit if plots and len(wandb_images) < log_imgs: box_data = [{ "position": { "minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3] }, "class_id": int(cls1), "box_caption": "%s (%s) %.3f" % (names1[cls1], names2[cls2], conf1), "scores": { "class_score": conf1 }, "domain": "pixel" } for *xyxy, conf1, cls1, conf2, cls2 in pred.tolist()] boxes = { "predictions": { "box_data": box_data, "class_labels": names1 } } # inference-space wandb_images.append( wandb.Image(img[si], boxes=boxes, caption=path.name)) box_data = [{ "position": { "minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3] }, "class_id": int(cls2), "box_caption": "(%s) %s %.3f" % (names1[cls1], names2[cls2], conf2), "scores": { "class_score": conf2 }, "domain": "pixel" } for *xyxy, conf1, cls1, conf2, cls2 in pred.tolist()] boxes = { "predictions": { "box_data": box_data, "class_labels": names2 } } # inference-space wandb_images.append( wandb.Image(img[si], boxes=boxes, caption=path.name)) # Append to pycocotools JSON dictionary if save_json: # todo ~ # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int( path.stem) if path.stem.isnumeric() else path.stem box = xyxy2xywh(predn[:, :4]) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({ 'image_id': image_id, 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5) }) # Assign all predictions as incorrect correct1 = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) correct2 = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) # edit if nl: detected = [] # target indices tcls_tensor_1 = labels[:, 0] tcls_tensor_2 = labels[:, 1] # edit # target boxes tbox = xywh2xyxy(labels[:, 2:6]) # edit scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix1.process_batch(predn, torch.cat( (labels[:, 0:1], tbox), 1)) # edit confusion_matrix2.process_batch(predn, torch.cat( (labels[:, 0:1], tbox), 1)) # edit # Per target class unique_classes = torch.unique(torch.stack( (tcls_tensor_1, tcls_tensor_2)), dim=1).T # print("unique_classes", unique_classes.shape, unique_classes) # todo # print("tcls tensors", tcls_tensor_1, tcls_tensor_2) # todo for cls1, cls2 in unique_classes: ti1 = (cls1 == tcls_tensor_1).nonzero(as_tuple=False).view( -1) # prediction indices pi1 = (cls1 == pred[:, 5]).nonzero(as_tuple=False).view( -1) # target indices ti2 = (cls2 == tcls_tensor_2).nonzero(as_tuple=False).view( -1) # prediction indices #edit pi2 = (cls2 == pred[:, 7]).nonzero(as_tuple=False).view( -1) # target indices #edit # print("ti/pi", ti1, pi1, ti2, pi2) # todo # Search for detections if pi1.shape[0]: # Prediction to target ious ious, i = box_iou(predn[pi1, :4], tbox[ti1]).max( 1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti1[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct1[pi1[ j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break if pi2.shape[0]: # edit # Prediction to target ious ious, i = box_iou(predn[pi2, :4], tbox[ti2]).max( 1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti2[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct2[pi2[ j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break # Append statistics (correct, conf, predicted class, target class) stats1.append( (correct1.cpu(), pred[:, 4].cpu(), pred[:, 6].cpu(), tcls1)) stats2.append( (correct2.cpu(), pred[:, 5].cpu(), pred[:, 7].cpu(), tcls2)) # edit # Plot images if plots and batch_i < 3: f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(img, targets, paths, f, names1, names2), daemon=True).start() # edit f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(img, output_to_target(output), paths, f, names1, names2), daemon=True).start() # edit # Compute statistics stats1 = [np.concatenate(x, 0) for x in zip(*stats1)] # to numpy stats2 = [np.concatenate(x, 0) for x in zip(*stats2)] # to numpy # edit # s1 = [np.count_nonzero(ba) for ba in stats1[0]] # todo # s2 = [np.count_nonzero(ba) for ba in stats2[0]] # todo # print("stats 1:", np.sum(s1), stats1[1:]) # todo # print("stats 2:", np.sum(s2), stats2[1:]) # todo if len(stats1) and stats1[0].any(): p_1, r_1, ap_1, f1_1, ap_class_1 = ap_per_class(*stats1, plot=plots, save_dir=save_dir, names=names1, suffix='_1') ap50_1, ap_1 = ap_1[:, 0], ap_1.mean(1) # [email protected], [email protected]:0.95 mp_1, mr_1, map50_1, map_1 = p_1.mean(), r_1.mean(), ap50_1.mean( ), ap_1.mean() nt_1 = np.bincount(stats1[3].astype(np.int64), minlength=nc1) # number of targets per class else: nt_1 = torch.zeros(1) if len(stats2) and stats2[0].any(): p_2, r_2, ap_2, f1_2, ap_class_2 = ap_per_class(*stats2, plot=plots, save_dir=save_dir, names=names2, suffix='_2') # edit ap50_2, ap_2 = ap_2[:, 0], ap_2.mean(1) # [email protected], [email protected]:0.95 # edit mp_2, mr_2, map50_2, map_2 = p_2.mean(), r_2.mean(), ap50_2.mean( ), ap_2.mean() # edit nt_2 = np.bincount(stats2[3].astype( np.int64), minlength=nc2) # number of targets per class # edit else: nt_2 = torch.zeros(1) # edit # Print results pf = '%20s' + '%12.3g' * 6 # print format print(pf % ('all_1', seen, nt_1.sum(), mp_1, mr_1, map50_1, map_1)) print(pf % ('all_2', seen, nt_2.sum(), mp_2, mr_2, map50_2, map_2)) # edit # Print results per class if (verbose or (nc1 < 50 and not training)) and nc1 > 1 and len(stats1): for i, c in enumerate(ap_class_1): print( pf % (names1[c], seen, nt_1[c], p_1[i], r_1[i], ap50_1[i], ap_1[i])) if (verbose or (nc2 < 50 and not training)) and nc2 > 1 and len(stats2): # edit for i, c in enumerate(ap_class_2): print( pf % (names2[c], seen, nt_2[c], p_2[i], r_2[i], ap50_2[i], ap_2[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print( 'Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Plots if plots: confusion_matrix1.plot(save_dir=save_dir, names=list(names1.values()), suffix='_1') confusion_matrix2.plot(save_dir=save_dir, names=list(names2.values()), suffix='_2') # edit if wandb and wandb.run: val_batches = [ wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg')) ] wandb.log({ "Images": wandb_images, "Validation": val_batches }, commit=False) # Save JSON if save_json and len(jdict): # todo ~ w = Path(weights[0] if isinstance(weights, list) else weights ).stem if weights is not None else '' # weights anno_json = '../coco/annotations/instances_val2017.json' # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [ int(Path(x).stem) for x in dataloader.dataset.img_files ] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[: 2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print(f'pycocotools unable to run: {e}') # Return results if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {save_dir}{s}") model.float() # for training maps_1 = np.zeros(nc1) + map_1 for i, c in enumerate(ap_class_1): maps_1[c] = ap_1[i] maps_2 = np.zeros(nc2) + map_2 for i, c in enumerate(ap_class_2): maps_2[c] = ap_2[i] res_loss = (loss.cpu() / len(dataloader)).tolist() return maps_1, maps_2, t,\ (mp_1, mr_1, map50_1, map_1, *res_loss),\ (mp_2, mr_2, map50_2, map_2, *res_loss)
def psuedolabel(data, weights=None, batch_size=16, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS augment=False, pseudo_threshold=.4, model=None, dataloader=None, merge=False): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly device = utils.torch_utils.select_device(opt.device, batch_size=batch_size) merge = opt.merge # use Merge NMS, save *.txt labels # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict # Dataloader if not training: img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once path = data['extra'] if opt.task == 'extra' else data['val'] # path to val/test images dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, hyp=None, augment=False, cache=False, pad=0.5, rect=True)[0] count = 1 uncount = 0 boundary_error = 0 for batch_i, (img, targets, paths, shapes) in tqdm(enumerate(dataloader), desc="PseudoLabel", mininterval=0.01): img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 nb, _, height, width = img.shape # batch size, channels, height, width # Disable gradients with torch.no_grad(): # Run model inf_out, _ = model(img, augment=augment) # inference and training outputs # Prediction output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, merge=merge) plabels, boundary_error = psuedolabel_generation(output, width, height, pseudo_threshold, boundary_error) if len(plabels) > 0: for i in plabels[:,0].astype('int'): idx = np.where(plabels[:,0] == i)[0] save_labels = plabels[idx] labels = save_labels[:,1:-1] file_name = paths[i].replace('images/extra', 'labels/pseudo').replace('jpg','txt').replace('JPG','txt').replace('png','txt').replace('PNG','txt') if (np.sum(np.isnan(labels)) == 0) and (np.sum(np.isinf(labels)) == 0): np.savetxt(file_name, labels, delimiter=' ',fmt=['%d','%4f','%4f','%4f','%4f']) image = Image.open(paths[i]) image.save(paths[i].replace('extra', 'pseudo')) count += 1 image.close() else: print(file_name) uncount += 1 print(f'Completed generating {count} pseudo labels.') print(f'Eliminated {uncount} images.') print(f'Boundary Error: {boundary_error} objects')
def train(hyp, opt, device, tb_writer=None): # 控制台打印日志 logger.info(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = log_dir / 'weights' # weights directory os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = str(log_dir / 'results.txt') # weights:权重文件(预训练的);rank:全局进程; epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' # 初始化随机种子(numpy,random,torch的) init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict ''' 执行逻辑:如果是0号进程进来会直接从torch_distributed_zero_first返回,然后执行check_dataset, 执行完check_dataset后会进入torch_distributed_zero_first函数从yield之下开始执行barrier函数暂停等到所有线程都到这个函数再继续执行, 如果不是0号线程进入那么就会执行barrier函数等待,等到所有进程都进入此函数的时候解除barrier继续执行 解除barrier的方法就是等所有进程都执行barrier函数的时候就会解除 ''' # 核实数据 with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') # 加载预训练的模型参数 if pretrained: # 下载与训练数据 with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round( hyp['anchors']) # force autoanchor # 载入输入的配置或者是加载的pretrained的配置,ch=3是输入channel model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [ ] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 # 只加载在预训练的模型和当前模型中都有的组件的参数,这要求与训练的模型和当前模型的shape要相等 # intersect_dicts的左右就是将与训练的模型参数和当前的模型参数进行比较,取shape一致的那些参数(shape不一样的是没法运用在当前模型的) state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect # 非严格模式加载参数 model.load_state_dict(state_dict, strict=False) # load logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze # 冻结某几层,finetune可以用 freeze = [ '', ] # parameter names to freeze (full or partial) if any(freeze): for k, v in model.named_parameters(): if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer # 将batch_size和64比较,当64不是batch_size的整数倍的时候,权重做相应的调整 nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_parameters(): v.requires_grad = True if '.bias' in k: pg2.append(v) # biases elif '.weight' in k and '.bn' not in k: pg1.append(v) # apply weight decay else: pg0.append(v) # all else if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) # 添加卷积权重参数和biases参数,其中biases不需要权重衰减,这里的params只能是这个名字,执行完add_param_group后optimizer的数据就是一个list中有三个值 optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR # 权重衰减策略 lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[ 'lrf']) + hyp['lrf'] # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}' ) # save previous weights if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) # 得到处理(就是让图片大小能被网格大小整除)后的图像和测试图像的大小 imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode # 分布式(单机多GPU) if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: # 同步batchnorm,如果不同步的话每个GPU上的batchnorm都会使用当前GPU上数据的方差和均值,那几个GPU虽然训练的是同一个batch的数据,值却是不一样的 model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # Exponential moving average 指数移动平均 # 给予近期数据更高的权重,就是说对于参数,我们给予最近的几次的参数更高的权重,其假设就是最近几次的参数是在最优处抖动,所以最近几次的参数权重就给高点 ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode # 分布式(多机器多GPU) if cuda and rank != -1: # local_rank指定的是当前进程使用的是哪块GPU,local_rank表示的就是GPU序号 model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # Process 0 if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers)[0] # testloader if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) plot_labels(labels, save_dir=log_dir) if tb_writer: # tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384 tb_writer.add_histogram('classes', c, 0) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights model.names = names # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) logger.info('Image sizes %g train, %g test\n' 'Using %g dataloader workers\nLogging results to %s\n' 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs)) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices if rank in [-1, 0]: cw = model.class_weights.cpu().numpy() * ( 1 - maps)**2 # class weights iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights dataset.indices = random.choices( range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx # Broadcast if DDP if rank != -1: indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot if ni < 3: f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename result = plot_images(images=imgs, targets=targets, paths=paths, fname=f) if tb_writer and result is not None: tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard # end batch ------------------------------------------------------------------------------------------------ # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema: ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=log_dir, plots=epoch == 0 or final_epoch) # plot first and last # Write with open(results_file, 'a') as f: f.write( s + '%10.4g' * 7 % results + '\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Tensorboard if tb_writer: tags = [ 'train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2' ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): tb_writer.add_scalar(tag, x, epoch) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict() } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers n = opt.name if opt.name.isnumeric() else '' fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt' for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename if str(f2).endswith('.pt'): # is *.pt strip_optimizer(f2) # strip optimizer os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish if not opt.evolve: plot_results(save_dir=log_dir) # save as results.png logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() return results
def main(string): #set parameter collapse_check_list =[148,149,150] model_path = os.path.join(base_path,'yolov5l_19_0.988_0.764.pt') model_path2 = os.path.join(base_path,'yolov5l_34_0.988_0.794.pt') model_path3 = os.path.join(base_path,'yolov5l_34_0.989_0.769.pt') json_path = os.path.join(base_path,'t1_res_U0000000279.json') img_size = 640 confidence_threshold = 0.175 #0.1, 0.2 iou_threshold = 0.4 #0.4, 0.5 tracking_iou = 0.65 device_choice = '' fps_collapse_rate = 150 batch_size = 20 single_cls = True device = select_device(device_choice) half = device.type != 'cpu' # half precision only supported on CUDA model = attempt_load(model_path, map_location=device) # load FP32 model model2 = attempt_load(model_path2, map_location=device) # load FP32 model model3 = attempt_load(model_path3, map_location=device) # load FP32 model img_size = check_img_size(img_size, s=model.stride.max()) # check img_size model.half() model2.half() model3.half() img = torch.zeros((1, 3, img_size, img_size), device=device) # init img _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once _ = model2(img.half() if half else img) if device.type != 'cpu' else None # run once _ = model3(img.half() if half else img) if device.type != 'cpu' else None # run once with open(json_path,'w') as f: json_file = dict() json_file['annotations'] = [] json.dump(json_file,f) f.close() for folder_path in os.listdir(string): if os.path.splitext(folder_path)[-1] == '.cache': continue folder_path = os.path.join(string,folder_path) info_of_video = [] Answer_Manager_list = [] dataloader = create_dataloader(folder_path, img_size, batch_size, model.stride.max(),single_cls, pad=0.5, rect=False)[0] for batch_i, (img, _, paths, shapes) in enumerate(dataloader): img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 nb, _, height, width = img.shape # batch size, channels, height, width with torch.no_grad(): # Run model inf_out, train_out = model(img, augment=True) # inference and training outputs inf_out2, train_out2 = model2(img,augment=True) inf_out3, train_out3 = model3(img,augment=True) inf_out = torch.cat([inf_out,inf_out2,inf_out3],1) for index,_ in enumerate(inf_out): sort,idx = torch.sort(_[:,4],0,descending=True) inf_out[index] = _[idx] # Run NMS output = non_max_suppression(inf_out, conf_thres=confidence_threshold, iou_thres=iou_threshold) for si, pred in enumerate(output): path = Path(paths[si]) pred_boxes = np.asarray([]) pred_scores = np.asarray([]) if len(pred): pred[:, :4] = scale_coords(img[si].shape[1:], pred[:, :4], shapes[si][0], shapes[si][1]) pred_boxes = np.asarray(pred[:,:4].tolist()) pred_scores = np.asarray(pred[:,4].tolist()) delete_index = np.where(np.logical_or(abs(pred_boxes[:,0]-pred_boxes[:,2])<= 32,abs(pred_boxes[:,1]-pred_boxes[:,3])<=32)) pred_scores = np.delete(pred_scores,delete_index,axis=0) pred_boxes = np.delete(pred_boxes,delete_index,axis=0) info_of_frame = dict() info_of_frame['file_name'] = os.path.basename(str(path)) info_of_frame['boxes'] = pred_boxes.astype('int') info_of_frame['scores'] = pred_scores info_of_video.append(info_of_frame) #tracking for index, info in enumerate(info_of_video): info_boxes = info['boxes'] info_scores = info['scores'] info_of_video[index]['boxes'] = [] info_of_video[index]['scores'] = [] #정답 Manager 유무 확인 if len(Answer_Manager_list) != 0: #pred 유무 확인 if len(info_scores) != 0: delete_index_list = [] for idx,answer in enumerate(Answer_Manager_list): delete_index, iou = iou_matching(answer['box'],info_boxes) if iou >= tracking_iou: Answer_Manager_list[idx]['box'] = info_boxes[delete_index] Answer_Manager_list[idx]['score'] = info_scores[delete_index] Answer_Manager_list[idx]['stack'] += 1 info_of_video[index]['boxes'].append(info_boxes[delete_index].tolist()) info_of_video[index]['scores'].append(info_scores[delete_index].tolist()) # del info_boxes[delete_index] # del info_scores[delete_index] info_boxes = np.delete(info_boxes,[delete_index],axis=0) info_scores = np.delete(info_scores,[delete_index],axis=0) else: if answer['stack'] >= fps_collapse_rate: delete_index_list.append(idx) else: Answer_Manager_list[idx]['stack'] += 1 info_of_video[index]['boxes'].append(answer['box'].tolist()) info_of_video[index]['scores'].append(answer['score'].tolist()) Answer_Manager_list = np.delete(Answer_Manager_list,delete_index_list) Answer_Manager_list = Answer_Manager_list.tolist() if index < len(info_of_video)-fps_collapse_rate: for info_box, info_score in zip(info_boxes[:],info_scores): for check_val in collapse_check_list: _, iou = iou_matching(info_box,np.array(info_of_video[index+check_val]['boxes'])) if iou >= tracking_iou: answer_dict = dict() answer_dict['box'] = info_box answer_dict['score'] = info_score answer_dict['stack'] = 1 Answer_Manager_list.append(answer_dict) info_of_video[index]['boxes'].append(info_box.tolist()) info_of_video[index]['scores'].append(info_score.tolist()) break else: delete_index_list = [] for idx,answer in enumerate(Answer_Manager_list): if answer['stack']>=fps_collapse_rate: delete_index_list.append(idx) else: Answer_Manager_list[idx]['stack'] += 1 info_of_video[index]['boxes'].append(answer['box'].tolist()) info_of_video[index]['scores'].append(answer['score'].tolist()) Answer_Manager_list = np.delete(Answer_Manager_list,delete_index_list) Answer_Manager_list = Answer_Manager_list.tolist() else: #pred 유무 확인 if len(info_scores) != 0: if index < len(info_of_video)-fps_collapse_rate: for info_box, info_score in zip(info_boxes[:],info_scores): for check_val in collapse_check_list: _, iou = iou_matching(info_box,np.array(info_of_video[index+check_val]['boxes'])) if iou >= tracking_iou: answer_dict = dict() answer_dict['box'] = info_box answer_dict['score'] = info_score answer_dict['stack'] = 1 Answer_Manager_list.append(answer_dict) info_of_video[index]['boxes'].append(info_box.tolist()) info_of_video[index]['scores'].append(info_score.tolist()) break with open(json_path,'r') as f: json_data = json.load(f) f.close() for _ in info_of_video: detection_num = len(_['scores']) # if detection_num == 0: # continue json_d = dict() json_d['file_name'] = _['file_name'] json_d['box'] = [] for index in range(detection_num): j_d = dict() j_d['position'] = _['boxes'][index] j_d['confidence_score'] = str(_['scores'][index]) json_d['box'].append(j_d) json_data['annotations'].append(json_d) with open(json_path,'w') as f: json.dump(json_data,f) f.close()
def train(hyp, opt, device): save_dir, epochs, batch_size, total_batch_size, weights, rank = \ Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank do_semi = opt.do_semi # Directories wdir = save_dir / 'weights' wdir.mkdir(parents=True, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = save_dir / 'results.txt' # Save run settings with open(save_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(save_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure plots = not opt.evolve #create plots cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.SafeLoader) nc = 1 if opt.single_cls else int(data_dict['nc']) #number of classes names = ['item'] if opt.single_cls and len( data_dict['names']) != 1 else data_dict['names'] assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) #load checkpoint model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) #create exclude = [ 'anchor' ] if (opt.cfg or hyp.get('anchors')) and not opt.resume else [ ] #exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) #intersect model.load_state_dict(state_dict, strict=False) #load else: model = Model(opt.cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) with torch_distributed_zero_first(rank): check_dataset(data_dict) #check train_path = data_dict['train'] test_path = data_dict['val'] # Optimizer nbs = 64 accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply dacay if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust betal to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) del pg0, pg1, pg2 if opt.linear_lr: lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp[ 'lrf'] # linear else: lf = one_cycle(1, hyp['lrf'], epochs) scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # EMA ema = ModelEMA(model) if rank in [-1, 0] else None # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # EMA if ema and ckpt.get('ema'): ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) ema.updates = ckpt['updates'] # Results if ckpt.get('training_results') is not None: results_file.write_text( ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weight, epochs) if epochs < start_epoch: epochs += ckpt['epoch'] del ckpt, state_dict # Image sizes gs = max(int(model.stride.max()), 32) # grid size (max stride) nl = model.model[ -1].nl # number of detection layer (used for scaling hyp['obj]) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to( device) # Trainloader if do_semi: dataloader, dataset, unlabeldataloader = create_dataloader( train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: '), do_semi=opt.do_semi) else: dataloader, dataset = create_dataloader( train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: '), do_semi=opt.do_semi) # Train teacher model mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # process 0 if rank in [-1, 0]: testloader = create_dataloader( test_path, imgsz_test, batch_size * 2, gs, opt, # testloader hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, pad=0.5, prefix=colorstr('val: '), do_semi=False)[0] if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) model.half().float() # pre-reduce anchor precision # DDP mode if cuda and rank != 1: model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank, find_unused_parameters=any( isinstance(layer, nn.MultiheadAttention) for layer in model.modules())) # Model parameters hyp['box'] *= 3. / nl # scale to layers hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers hyp['label_smoothing'] = opt.label_smoothing model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) model.class_weights = labels_to_class_weights( dataset.labels, nc).to(device) * nc # attach class weights model.names = names # Train teacher model --> burn in t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) compute_loss = ComputeLoss(model) # init loss class burnin_epochs = epochs / 2 # burn in for epoch in range(start_epoch, burnin_epochs): # epoch------------------------- model.train() nb = len(dataloader) mloss = torch.zeros(4, device=device) # mean loss if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) optimizer.zero_grad() for i, (imgs, targets, paths, _) in pbar: ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warm up if ni <= [0, nw]: xi = [0, nw] accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size].round())) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_item = compute_loss( pred, targets.to(device)) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between device in DDP mode if opt.quad: loss *= 4. # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) scaler.update() optimizer.zero_grad() if ema: ema.update(model) # print if rank in [-1, 0]: mloss = (mloss * i + loss_item) / (i + 1) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP ema.update_attr(model, include=[ 'yaml', 'nc', 'hyp', 'gr', 'names', 'stride', 'class_weights' ]) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test(data_dict, batch_size=batch_size * 2, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, verbose=nc < 50 and final_epoch, plots=plots and final_epoch, compute_loss=compute_loss) fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, mAP@50, [email protected]] if fi > best_fitness: best_fitness = fi if (not opt.nosave) or (final_epoch and not opt.evolve): # if save ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': results_file.read_text(), 'model': deepcopy(model.module if is_parallel(model) else model).half(), 'ema': deepcopy(ema.ema).half(), 'updates': ema.updates, 'optimizer': optimizer.state_dict() } if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------- # end warm up # get persudo label # STAC # first apply weak augmentation on unlabeled dataset then use teacher net to predict the persudo labels # Then apply strong augmentation on unlabeled dataset, use student net to get the logists and compute the unlabeled loss. model.eval() img = [] target = [] Path = [] imgsz = opt.img_size for idx, batch in tqdm(enumerate(unlabeldataloader), total=len(unlabeldataloader)): imgs0, _, path, _ = batch # from uint8 to float16 with torch.no_grad(): pred = model(imgs0.to(device, non_blocking=True).float() / 255.0)[0] gn = torch.tensor(imgs0.shape)[[3, 2, 3, 2]] pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms) for index, pre in enumerate(pred): predict_number = len(pre) if predict_number == 0: continue Class = pre[:, 5].view(predict_number, 1).cpu() XYWH = (xyxy2xywh(pre[:, :4])).cpu() XYWH /= gn pre = torch.cat((torch.zeros(predict_number, 1), Class, XYWH), dim=1) img.append(imgs0[index]) target.append(pre) Path.append(path[index]) unlabeldataset = semiDataset(img, target, Path) del img, targets, Path model.train()
def train_on_large_batch(classes_to_update, train_path, model, device, logger, valid_path=None, imgsz=416, imgsz_test=416, gs=None, opt=None, hyp=None, nc=6, log_dir=None, tb_writer=None, names=None, optimizer=None, extMem=None, best_fitness=None, use_ext_mem=True, epochs=None): lf = lambda x: ((1 + math.cos(x * math.pi / opt.epochs)) / 2) * (1 - hyp[ 'lrf']) + hyp['lrf'] scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) nbs = 64 # nominal batch size results_file = str(log_dir / 'results.txt') wdir = log_dir / 'weights' # weights dfirectory # os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' epochs = opt.epochs if epochs == None else epochs batch_size, total_batch_size = opt.batch_size, opt.total_batch_size external_files_path = extMem.get_memory_file() path_to_add = train_path.copy() if use_ext_mem: # train_p = [train_path, external_files_path] train_path.append(external_files_path) # else: # train_p = train_path print(f'train_path{train_path}') dataloader, dataset = create_dataloader( train_path, imgsz, batch_size, gs, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, workers=opt.workers, ) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) if valid_path is not None: testloader = create_dataloader( valid_path, # '/media/ivan/share/core50_350_1f/test.txt', imgsz_test, total_batch_size, gs, hyp=hyp, augment=False, cache=opt.cache_images and not opt.notest, rect=False, rank=-1, world_size=opt.world_size, workers=opt.workers)[0] # testloader # if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.IntTensor(labels[:, 0]) # classes plot_labels(labels, save_dir=log_dir) print(torch.bincount(c)) # if tb_writer: # # tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384 # tb_writer.add_histogram('classes', c, core_batch) # Anchors # if not opt.noautoanchor: # check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights model.names = names model.imgsize = imgsz # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) start_epoch = 0 scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=True) logger.info('Image sizes %g train, %g test\n' 'Using %g dataloader workers\nLogging results to %s\n' 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs)) # update number of epochs to iterative training # x_train, y_train = dataset.get_all_data() for epoch in tqdm( range(start_epoch, epochs) ): # epoch ------------------------------------------------------------------ model.train() mloss = torch.zeros(4, device=device) # mean losses logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) # x_train_splitted = torch.split(x_train, 4) # y_train_splitted = torch.split(y_train, 4) # pbar = enumerate(zip(x_train_splitted, y_train_splitted)) pbar = enumerate(dataloader) pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) for i, ( imgs, targets, _, _ ) in pbar: # batch ------------------------------------------------------------- # imgs = x_train[i * batch_size:(i + 1) * batch_size] # targets = y_train[i * batch_size:(i + 1) * batch_size] # # # preprocess tensor to proper form # # img, label = zip(imgs, targets) # transposed # for i, l in enumerate(targets): # l[:, 0] = i # add target image index for build_targets() # # imgs = torch.stack(imgs) # targets = torch.cat(targets) ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # # Multi-scale # if opt.multi_scale: # sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size # sf = sz / max(imgs.shape[2:]) # scale factor # if sf != 1: # ns = [math.ceil(x * sf / gs) * gs for x in # imgs.shape[2:]] # new shape (stretched to gs-multiple) # imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=True): pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() # if ema: # ema.update(model) # Print mloss = (mloss * i + loss_items) / (i + 1) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # end batch ------------------------------------------------------------------------------------------------ # Scheduler scheduler.step() # mAP print('train:') results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, # model=ema.ema, model=model, single_cls=opt.single_cls, dataloader=dataloader, save_dir=log_dir, # plots=epoch == 0, # plot first and last log_imgs=0, verbose=True, nc=nc) if (epoch % 5 == 0) and valid_path is not None: print(valid_path) print('valid: (is broken)') test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, # model=ema.ema, model=model, dataloader=testloader, save_dir=log_dir, # plots=epoch == 0, # plot first and last log_imgs=0, verbose=True, nc=nc) # wandb.log({'per class/AP per class': maps}) # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Log # tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss', # train loss # 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', # 'val/giou_loss', 'val/obj_loss', 'val/cls_loss', # val loss # 'x/lr0', 'x/lr1', 'x/lr2'] # params # for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): # if tb_writer: # tb_writer.add_scalar(tag, x, epoch) # tensorboard # if wandb: # wandb.log({tag: x}) # W&B # Strip optimizers n = opt.name if opt.name.isnumeric() else '' fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt' for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename if str(f2).endswith('.pt'): # is *.pt strip_optimizer(f2) # strip optimizer os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish plot_results(save_dir=log_dir) # save as results.png logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi # Save model save = not opt.nosave last = wdir / 'last.pt' if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), # 'model': ema.ema, 'model': model, 'optimizer': optimizer.state_dict(), } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training extMem.update_memory(path_to_add, update_iters=classes_to_update) return
def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = log_dir / 'weights' # weights dfirectory os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = str(log_dir / 'results.txt') epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs_init, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round( hyp['anchors']) # force autoanchor model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [ ] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[ 'lrf']) + hyp['lrf'] # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Logging if wandb and wandb.run is None: id = ckpt.get('wandb_id') if 'ckpt' in locals() else None wandb_run = wandb.init(config=opt, resume="allow", project="YOLOv5", name=os.path.basename(log_dir), id=id) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}' ) # save previous weights if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) all_test_dataloader = create_dataloader( test_path, # '/media/ivan/share/core50_350_1f/test.txt', imgsz_test, total_batch_size, gs, hyp=hyp, augment=False, # cache=opt.cache_images and not opt.notest, rect=False, rank=-1, world_size=opt.world_size, workers=opt.workers)[0] root = '/media/ivan/share/core50_350_1f/batches/' paths = os.listdir(root) train_paths = [] valid_paths = [] for p in paths: if 'train' in p: train_paths.append(root + p) elif 'val' in p: valid_paths.append(root + p) else: print(p) extMem = externalMemory(size=200) print(f'external memory file: {extMem.get_memory_file()}') train_paths = [ '/media/ivan/share/demoset/train_4.txt', '/media/ivan/share/demoset/train_2.txt' ] valid_paths = [ '/media/ivan/share/demoset/valid.txt', '/media/ivan/share/demoset/valid.txt' ] # prepare_params(hyp, opt, device, tb_writer=None, wandb=None) for core_batch in range(2): print(f'------------CORE50 itertaion №:{core_batch}------------') train_on_large_batch(core_batch, train_paths[core_batch], valid_paths[core_batch], model, device, logger, imgsz=imgsz, imgsz_test=imgsz_test, gs=gs, opt=opt, hyp=hyp, nc=nc, log_dir=log_dir, tb_writer=tb_writer, names=names, optimizer=optimizer, extMem=extMem, scheduler=scheduler, lf=lf, best_fitness=best_fitness, wandb_run=wandb_run) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=model, single_cls=opt.single_cls, dataloader=all_test_dataloader, save_dir=log_dir / 'images' / str(core_batch), # plots=epoch == 0 or final_epoch, # plot first and last log_imgs=opt.log_imgs, verbose=True) # wandb.log({'per class/AP per class All': maps[0]}) # tb_writer.add_scalar('per class/AP per class All', maps[0]) # Log tags = [ # train loss 'test/precision', 'test/recall', 'test/mAP_0.5', 'test/mAP_0.5:0.95', 'test/giou_loss', 'test/obj_loss', 'test/cls_loss' ] # params for x, tag in zip(list(results), tags): if tb_writer: tb_writer.add_scalar(tag, x, core_batch) # tensorboard if wandb: wandb.log({tag: x}) # W&B return results
def test_prune(cfg, data, opt, weights=None, batch_size=16, img_size=416, iou_thres=0.5, conf_thres=0.001, nms_thres=0.5, save_json=False, model=None, model_name="darknet"): # Initialize/load model and set device if model is None: device = torch_utils.select_device(opt.device) verbose = True # Initialize model if (model_name == "darknet"): model = Darknet(cfg, img_size).to(device) else: model = torch.load(weights, map_location=device)['model'].float() # Load weights attempt_download(weights) if weights.endswith('.pt'): # pytorch format model.load_state_dict( torch.load(weights, map_location=device)['model']) else: # darknet format _ = load_darknet_weights(model, weights) if torch.cuda.device_count() > 1: model = nn.DataParallel(model) else: device = next(model.parameters()).device # get model device verbose = False # Configure run data = parse_data_cfg(data) nc = int(data['nc']) # number of classes test_path = data['val'] # path to test images # names = load_classes(data['names']) # class names names = data['names'][2:-2].split("', '") model.eval() # Dataloader # with torch_distributed_zero_first(-1): # dataset = LoadImagesAndLabels(test_path, img_size, batch_size) # batch_size = min(batch_size, len(dataset)) # dataloader = DataLoader(dataset, # batch_size=batch_size, # num_workers=min([os.cpu_count(), batch_size, 16]), # pin_memory=True, # collate_fn=LoadImagesAndLabels.collate_fn) opt.single_cls = False dataloader = create_dataloader(test_path, img_size, batch_size, 32, opt, hyp=None, augment=False, cache=False, pad=0.5, rect=True)[0] seen = 0 coco91class = coco80_to_coco91_class() s = ('%20s' + '%10s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP', 'F1') p, r, f1, mp, mr, map, mf1 = 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3) jdict, stats, ap, ap_class = [], [], [], [] for batch_i, (imgs, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): imgs = torch.div(imgs, 255.0) targets = targets.to(device) imgs = imgs.to(device) _, _, height, width = imgs.shape # batch size, channels, height, width # Plot images with bounding boxes if batch_i == 0 and not os.path.exists('test_batch0.jpg'): plot_images(imgs=imgs, targets=targets, paths=paths, fname='test_batch0.jpg') # Run model inf_out, train_out = model(imgs) # inference and training outputs # Compute loss if hasattr(model, 'hyp'): # if model has loss hyperparameters loss += compute_loss(train_out, targets, model)[1][:3].cpu() # GIoU, obj, cls # Run NMS output = non_max_suppression(inf_out, conf_thres=conf_thres, nms_thres=nms_thres) # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class seen += 1 if pred is None: if nl: stats.append(([], torch.Tensor(), torch.Tensor(), tcls)) continue # Append to text file # with open('test.txt', 'a') as file: # [file.write('%11.5g' * 7 % tuple(x) + '\n') for x in pred] # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int(Path(paths[si]).stem.split('_')[-1]) box = pred[:, :4].clone() # xyxy scale_coords(imgs[si].shape[1:], box, shapes[si]) # to original shape box = xyxy2xywh(box) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for di, d in enumerate(pred): jdict.append({ 'image_id': image_id, 'category_id': coco91class[int(d[6])], 'bbox': [floatn(x, 3) for x in box[di]], 'score': floatn(d[4], 5) }) # Clip boxes to image bounds clip_coords(pred, (height, width)) # Assign all predictions as incorrect correct = [0] * len(pred) if nl: detected = [] tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) tbox[:, [0, 2]] *= width tbox[:, [1, 3]] *= height # Search for correct predictions for i, (*pbox, pconf, pcls_conf, pcls) in enumerate(pred): # Break if all targets already located in image if len(detected) == nl: break # Continue if predicted class not among image classes if pcls.item() not in tcls: continue # Best iou, index between pred and targets m = (pcls == tcls_tensor).nonzero().view(-1) iou, bi = bbox_iou(pbox, tbox[m]).max(0) # If iou > threshold and class is correct mark as correct if iou > iou_thres and m[ bi] not in detected: # and pcls == tcls[bi]: correct[i] = 1 detected.append(m[bi]) # Append statistics (correct, conf, pcls, tcls) stats.append((correct, pred[:, 4].cpu(), pred[:, 6].cpu(), tcls)) # Compute statistics stats = [np.concatenate(x, 0) for x in list(zip(*stats))] # to numpy if len(stats): p, r, ap, f1, ap_class = ap_per_class(*stats) mp, mr, map, mf1 = p.mean(), r.mean(), ap.mean(), f1.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%10.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map, mf1)) # Print results per class if verbose and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap[i], f1[i])) # Save JSON if save_json and map and len(jdict): try: imgIds = [ int(Path(x).stem.split('_')[-1]) for x in dataset.img_files ] with open('results.json', 'w') as file: json.dump(jdict, file) from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb cocoGt = COCO('../coco/annotations/instances_val2014.json' ) # initialize COCO ground truth api cocoDt = cocoGt.loadRes('results.json') # initialize COCO pred api cocoEval = COCOeval(cocoGt, cocoDt, 'bbox') cocoEval.params.imgIds = imgIds # [:32] # only evaluate these images cocoEval.evaluate() cocoEval.accumulate() cocoEval.summarize() map = cocoEval.stats[1] # update mAP to pycocotools mAP except: print( 'WARNING: missing dependency pycocotools from requirements.txt. Can not compute official COCO mAP.' ) # Return results maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map, mf1, *(loss / len(dataloader)).tolist()), maps
def test( data, weights=None, batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling save_hybrid=False, # for hybrid auto-labelling save_conf=False, # save auto-label confidences plots=True, log_imgs=0): # number of logged images # Initialize/load model and set device # 判断是否在训练时调用test,如果是则获取训练时的设备 training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() device = select_device(opt.device, batch_size=batch_size) # Directories save_dir = Path( increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir( parents=True, exist_ok=True) # make dir # Load model model = attempt_load(weights, map_location=device) # load FP32 model # 检查输入图片分辨率是否能被模型的最大步长(默认32)整除 imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half # 如果设备不是cpu并且gpu数目为1,则将模型由Float32转为Float16,提高前向传播的速度 half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # to FP16 # Configure # 加载数据配置信息 model.eval() is_coco = data.endswith('coco.yaml') # is COCO dataset with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes # 设置iou阈值,从0.5~0.95,每间隔0.05取一次 iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 # iou个数 niou = iouv.numel() # Logging log_imgs, wandb = min(log_imgs, 100), None # ceil try: import wandb # Weights & Biases except ImportError: log_imgs = 0 # Dataloader if not training: # 创建一个全0数组测试一下前向传播是否正常运行 img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img _ = model(img.half() if half else img ) if device.type != 'cpu' else None # run once # 获取图片路径 path = data['test'] if opt.task == 'test' else data[ 'val'] # path to val/test images # 创建dataloader # 注意这里rect参数为True,yolov5的测试评估是基于矩形推理的 dataloader = create_dataloader( path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True, prefix=colorstr('test: ' if opt.task == 'test' else 'val: '))[0] seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) # 获取类别的名字 names = { k: v for k, v in enumerate( model.names if hasattr(model, 'names') else model.module.names) } """ 获取coco数据集的类别索引 这里要说明一下,coco数据集有80个类别(索引范围应该为0~79), 但是他的索引却属于0~90(笔者是通过查看coco数据测试集的json文件发现的,具体原因不知) coco80_to_coco91_class()就是为了与上述索引对应起来,返回一个范围在0~90的索引数组 """ coco91class = coco80_to_coco91_class() # 设置tqdm进度条的显示信息 s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') # 初始化指标,时间 p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. # 初始化测试集的损失 loss = torch.zeros(3, device=device) # 初始化json文件的字典,统计信息,ap jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): img = img.to(device, non_blocking=True) # 图片也由Float32->Float16 img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width with torch.no_grad(): # Run model """ time_synchronized()函数里面进行了torch.cuda.synchronize(),再返回的time.time() torch.cuda.synchronize()等待gpu上完成所有的工作 总的来说就是这样测试时间会更准确 """ t = time_synchronized() # inf_out为预测结果, train_out训练结果 inf_out, train_out = model( img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss # 如果是在训练时进行的test,则通过训练结果计算并返回测试集的GIoU, obj, cls损失 if training: loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # box, obj, cls # Run NMS targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb) ] if save_hybrid else [] # for autolabelling t = time_synchronized() """ non_max_suppression进行非极大值抑制; conf_thres为置信度阈值,iou_thres为iou阈值 merge为是否合并框 """ output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb) t1 += time_synchronized() - t # Statistics per image # 为每一张图片做统计, 写入预测信息到txt文件, 生成json文件字典, 统计tp等 for si, pred in enumerate(output): # 获取第si张图片的标签信息, 包括class,x,y,w,h # targets[:, 0]为标签属于哪一张图片的编号 labels = targets[targets[:, 0] == si, 1:] nl = len(labels) # 获取标签类别 tcls = labels[:, 0].tolist() if nl else [] # target class path = Path(paths[si]) # 统计测试图片数量 seen += 1 # 如果预测为空,则添加空的信息到stats里 if len(pred) == 0: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Predictions predn = pred.clone() scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Append to text file # 保存预测结果为txt文件 if save_txt: # 获得对应图片的长和宽 gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0 ]] # normalization gain whwh for *xyxy, conf, cls in predn.tolist(): # xyxy格式->xywh, 并对坐标进行归一化处理 xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: f.write(('%g ' * len(line)).rstrip() % line + '\n') # W&B logging if plots and len(wandb_images) < log_imgs: box_data = [{ "position": { "minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3] }, "class_id": int(cls), "box_caption": "%s %.3f" % (names[cls], conf), "scores": { "class_score": conf }, "domain": "pixel" } for *xyxy, conf, cls in pred.tolist()] boxes = { "predictions": { "box_data": box_data, "class_labels": names } } # inference-space wandb_images.append( wandb.Image(img[si], boxes=boxes, caption=path.name)) # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int( path.stem) if path.stem.isnumeric() else path.stem box = xyxy2xywh(predn[:, :4]) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({ 'image_id': image_id, 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5) }) # Assign all predictions as incorrect # 初始化预测评定,niou为iou阈值的个数 correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: # detected用来存放已检测到的目标 detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes # 获得xyxy格式的框并乘以wh tbox = xywh2xyxy(labels[:, 1:5]) # 将预测框的坐标调整到基于其原本长宽的坐标 scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix.process_batch( pred, torch.cat((labels[:, 0:1], tbox), 1)) # Per target class # 对图片中的每个类单独处理 for cls in torch.unique(tcls_tensor): # 标签框该类别的索引 ti = (cls == tcls_tensor).nonzero(as_tuple=False).view( -1) # prediction indices # 预测框该类别的索引 pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view( -1) # target indices # Search for detections if pi.shape[0]: """ Prediction to target ious box_iou计算预测框与标签框的iou值,max(1)选出最大的ious值,i为对应的索引 pred shape[N, 4] tbox shape[M, 4] box_iou shape[N, M] ious shape[N, 1] i shape[N, 1], i里的值属于0~M """ ious, i = box_iou(predn[pi, :4], tbox[ti]).max( 1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): # 获得检测到的目标 d = ti[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) # iouv为以0.05为步长 0.5到0.95的序列 # 获得不同iou阈值下的true positive correct[ pi[j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) # 每张图片的结果统计到stats里 stats.append( (correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # Plot images # 画出第1个batch的图片的ground truth和预测框并保存 if plots and batch_i < 3: f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start() f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(img, output_to_target(output), paths, f, names), daemon=True).start() # Compute statistics # 将stats列表的信息拼接到一起 stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): # 根据上面得到的tp等信息计算指标 # 精准度TP/TP+FP,召回率TP/P,map,f1分数,类别 p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean( 1) # [P, R, [email protected], [email protected]:0.95] mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() # nt是一个列表,测试集每个类别有多少个标签框 nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%12.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if (verbose or (nc <= 20 and not training)) and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print( 'Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) if wandb and wandb.run: wandb.log({"Images": wandb_images}) wandb.log({ "Validation": [ wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg')) ] }) # Save JSON # 采用之前保存的json格式预测结果,通过cocoapi评估指标 # 需要注意的是 测试集的标签也需要转成coco的json格式 if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights ).stem if weights is not None else '' # weights anno_json = '../coco/annotations/instances_val2017.json' # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [ int(Path(x).stem) for x in dataloader.dataset.img_files ] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[: 2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print(f'pycocotools unable to run: {e}') # Return results if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {save_dir}{s}") model.float() # for training maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def test( data, weights=None, batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling save_hybrid=False, # for hybrid auto-labelling save_conf=False, # save auto-label confidences plots=False, log_imgs=0, # number of logged images compute_loss=None): # Initialize/load model and set device logger = setup_logger('Test', './') write_info(logger, True) training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() device = select_device(opt.device, batch_size=batch_size) # Directories save_dir = Path( increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run # (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Half half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Logging log_imgs, wandb = min(log_imgs, 100), None # ceil try: import wandb # Weights & Biases except ImportError: log_imgs = 0 # Dataloader if not training: img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img _ = model(img.half() if half else img ) if device.type != 'cpu' else None # run once path = data['test'] if opt.task == 'test' else data[ 'val'] # path to val/test images dataloader = create_dataloader( path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True, prefix=colorstr('test: ' if opt.task == 'test' else 'val: '))[0] # write_imglist(path, logger) seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = { k: v for k, v in enumerate( model.names if hasattr(model, 'names') else model.module.names) } p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(dataloader): stats_perimg = [] img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width with torch.no_grad(): # Run model t = time_synchronized() inf_out, train_out = model( img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if compute_loss: loss += compute_loss([x.float() for x in train_out], targets)[1][:3] # box, obj, cls # Run NMS targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb) ] if save_hybrid else [] # for autolabelling t = time_synchronized() output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb) t1 += time_synchronized() - t # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path = Path(paths[si]) seen += 1 if len(pred) == 0: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) stats_perimg.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Predictions predn = pred.clone() scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix.process_batch( pred, torch.cat((labels[:, 0:1], tbox), 1)) # Per target class for cls in torch.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero(as_tuple=False).view( -1) # prediction indices pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view( -1) # target indices # Search for detections if pi.shape[0]: # Prediction to target ious ious, i = box_iou(predn[pi, :4], tbox[ti]).max( 1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct[ pi[j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append( (correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) stats_perimg.append( (correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # # f1 per image # mf1_perimg = 0 # mp_perimg = 0 # mr_perimg = 0 # stats_perimg = [np.concatenate(x, 0) for x in zip(*stats_perimg)] # if len(stats_perimg) and stats_perimg[0].any(): # p_perimg, r_perimg, _, f1_perimg, ap_class_perimg = ap_per_class(*stats_perimg, plot=plots, save_dir=save_dir, names=names) # p_perimg, r_perimg, f1_perimg = p_perimg[:, 0], r_perimg[:, 0], f1_perimg[:, 0] # [P, R, [email protected], [email protected]:0.95] # nt_perimg = np.bincount(stats_perimg[3].astype(np.int64), minlength=nc) # number of targets per class # for ind, c in enumerate(ap_class_perimg): # mf1_perimg += f1_perimg[ind] * nt_perimg[c] # mp_perimg += p_perimg[ind] * nt_perimg[c] # mr_perimg += r_perimg[ind] * nt_perimg[c] # mf1_perimg = mf1_perimg/nt_perimg.sum() # mp_perimg = mp_perimg / nt_perimg.sum() # mr_perimg = mr_perimg / nt_perimg.sum() # # logger.info('[{}] {} [F1 score:{:4f} (Prec: {:4f}, Rec: {:4f})]'.format(str(batch_i + 1), paths[0].split('/')[-1], mf1_perimg, mp_perimg, mr_perimg)) # save GPS log # Logger_System('./xmls', './Logger', output, paths, names) # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) p, r, ap50, ap, f1 = p[:, 0], r[:, 0], ap[:, 0], ap.mean( 1), f1[:, 0] # [P, R, [email protected], [email protected]:0.95] mp, mr, map50, map, mf1 = p.mean(), r.mean(), ap50.mean(), ap.mean( ), f1.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) meanf1 = 0 meanp = 0 meanr = 0 meanap = 0 # Print results per class if (verbose or (nc <= 20)) and nc > 1 and len(stats): for i, c in enumerate(ap_class): meanf1 += f1[i] * nt[c] meanp += p[i] * nt[c] meanr += r[i] * nt[c] meanap += ap50[i] * nt[c] meanf1 = meanf1 / nt.sum() meanp = meanp / nt.sum() meanr = meanr / nt.sum() meanap = meanap / nt.sum() logger.info('[Final] F1 score:{:4f} (Prec: {:4f}, Rec: {:4f})\n'.format( meanf1, meanp, meanr)) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print( 'Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Return results model.float() # for training maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] write_info(logger, False) return (meanf1, mp, mr, map50, meanap, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info(f"Hyperparameters {hyp}") save_dir, epochs, batch_size, total_batch_size, weights, rank = ( Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank, ) # Directories wdir = save_dir / "weights" wdir.mkdir(parents=True, exist_ok=True) # make dir last = wdir / "last.pt" best = wdir / "best.pt" results_file = save_dir / "results.txt" # Save run settings with open(save_dir / "hyp.yaml", "w") as f: yaml.dump(hyp, f, sort_keys=False) with open(save_dir / "opt.yaml", "w") as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure plots = not opt.evolve # create plots cuda = device.type != "cpu" init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict["train"] test_path = data_dict["val"] nc, names = ( (1, ["item"]) if opt.single_cls else (int(data_dict["nc"]), data_dict["names"]) ) # number classes, names assert len(names) == nc, "%g names found for nc=%g dataset in %s" % ( len(names), nc, opt.data, ) # check # Model pretrained = weights.endswith(".pt") if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get("anchors"): ckpt["model"].yaml["anchors"] = round(hyp["anchors"]) # force autoanchor model = Model(opt.cfg or ckpt["model"].yaml, ch=3, nc=nc).to(device) # create exclude = ["anchor"] if opt.cfg or hyp.get("anchors") else [] # exclude keys state_dict = ckpt["model"].float().state_dict() # to FP32 state_dict = intersect_dicts( state_dict, model.state_dict(), exclude=exclude ) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info( "Transferred %g/%g items from %s" % (len(state_dict), len(model.state_dict()), weights) ) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print("freezing %s" % k) v.requires_grad = False # Optimizer nbs = 64 # nominal batch size accumulate = max( round(nbs / total_batch_size), 1 ) # accumulate loss before optimizing hyp["weight_decay"] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, "bias") and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, "weight") and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay if opt.adam: optimizer = optim.Adam( pg0, lr=hyp["lr0"], betas=(hyp["momentum"], 0.999) ) # adjust beta1 to momentum else: optimizer = optim.SGD( pg0, lr=hyp["lr0"], momentum=hyp["momentum"], nesterov=True ) optimizer.add_param_group( {"params": pg1, "weight_decay": hyp["weight_decay"]} ) # add pg1 with weight_decay optimizer.add_param_group({"params": pg2}) # add pg2 (biases) logger.info( "Optimizer groups: %g .bias, %g conv.weight, %g other" % (len(pg2), len(pg1), len(pg0)) ) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = ( lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp["lrf"]) + hyp["lrf"] ) # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Logging if wandb and wandb.run is None: opt.hyp = hyp # add hyperparameters wandb_run = wandb.init( config=opt, resume="allow", project="YOLOv5" if opt.project == "runs/train" else Path(opt.project).stem, name=save_dir.stem, id=ckpt.get("wandb_id") if "ckpt" in locals() else None, ) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt["optimizer"] is not None: optimizer.load_state_dict(ckpt["optimizer"]) best_fitness = ckpt["best_fitness"] # Results if ckpt.get("training_results") is not None: with open(results_file, "w") as file: file.write(ckpt["training_results"]) # write results.txt # Epochs start_epoch = ckpt["epoch"] + 1 if opt.resume: assert ( start_epoch > 0 ), "%s training to %g epochs is finished, nothing to resume." % ( weights, epochs, ) if epochs < start_epoch: logger.info( "%s has been trained for %g epochs. Fine-tuning for %g additional epochs." % (weights, ckpt["epoch"], epochs) ) epochs += ckpt["epoch"] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) imgsz, imgsz_test = [ check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info("Using SyncBatchNorm()") # EMA ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) # Trainloader dataloader, dataset = create_dataloader( train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, ) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert ( mlc < nc ), "Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g" % ( mlc, nc, opt.data, nc - 1, ) # Process 0 if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader( test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, )[ 0 ] # testloader if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) if plots: plot_labels(labels, save_dir=save_dir) if tb_writer: tb_writer.add_histogram("classes", c, 0) if wandb: wandb.log( { "Labels": [ wandb.Image(str(x), caption=x.name) for x in save_dir.glob("*labels*.png") ] } ) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp["anchor_t"], imgsz=imgsz) # Model parameters hyp["cls"] *= nc / 80.0 # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device ) # attach class weights model.names = names # Start training t0 = time.time() nw = max( round(hyp["warmup_epochs"] * nb), 1000 ) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0) # P, R, [email protected], [email protected], val_loss(box, obj, cls) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) logger.info( "Image sizes %g train, %g test\n" "Using %g dataloader workers\nLogging results to %s\n" "Starting training for %g epochs..." % (imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs) ) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices if rank in [-1, 0]: cw = ( model.class_weights.cpu().numpy() * (1 - maps) ** 2 ) # class weights iw = labels_to_image_weights( dataset.labels, nc=nc, class_weights=cw ) # image weights dataset.indices = random.choices( range(dataset.n), weights=iw, k=dataset.n ) # rand weighted idx # Broadcast if DDP if rank != -1: indices = ( torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n) ).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ("\n" + "%10s" * 8) % ("Epoch", "gpu_mem", "box", "obj", "cls", "total", "targets", "img_size") ) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _, ) in ( pbar ): # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = ( imgs.to(device, non_blocking=True).float() / 255.0 ) # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round() ) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x["lr"] = np.interp( ni, xi, [ hyp["warmup_bias_lr"] if j == 2 else 0.0, x["initial_lr"] * lf(epoch), ], ) if "momentum" in x: x["momentum"] = np.interp( ni, xi, [hyp["warmup_momentum"], hyp["momentum"]] ) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [ math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate( imgs, size=ns, mode="bilinear", align_corners=False ) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device), model ) # loss scaled by batch_size if rank != -1: loss *= ( opt.world_size ) # gradient averaged between devices in DDP mode # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1) # update mean losses mem = "%.3gG" % ( torch.cuda.memory_reserved() / 1e9 if torch.cuda.is_available() else 0 ) # (GB) s = ("%10s" * 2 + "%10.4g" * 6) % ( "%g/%g" % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1], ) pbar.set_description(s) # Plot if plots and ni < 3: f = save_dir / f"train_batch{ni}.jpg" # filename plot_images(images=imgs, targets=targets, paths=paths, fname=f) # if tb_writer: # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard elif plots and ni == 3 and wandb: wandb.log( { "Mosaics": [ wandb.Image(str(x), caption=x.name) for x in save_dir.glob("train*.jpg") ] } ) # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler lr = [x["lr"] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema: ema.update_attr( model, include=["yaml", "nc", "hyp", "gr", "names", "stride"] ) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, plots=plots and final_epoch, log_imgs=opt.log_imgs if wandb else 0, ) # Write with open(results_file, "a") as f: f.write( s + "%10.4g" * 7 % results + "\n" ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) if len(opt.name) and opt.bucket: os.system( "gsutil cp %s gs://%s/results/results%s.txt" % (results_file, opt.bucket, opt.name) ) # Log tags = [ "train/box_loss", "train/obj_loss", "train/cls_loss", # train loss "metrics/precision", "metrics/recall", "metrics/mAP_0.5", "metrics/mAP_0.5:0.95", "val/box_loss", "val/obj_loss", "val/cls_loss", # val loss "x/lr0", "x/lr1", "x/lr2", ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): if tb_writer: tb_writer.add_scalar(tag, x, epoch) # tensorboard if wandb: wandb.log({tag: x}) # W&B # Update best mAP fi = fitness( np.array(results).reshape(1, -1) ) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, "r") as f: # create checkpoint ckpt = { "epoch": epoch, "best_fitness": best_fitness, "training_results": f.read(), "model": ema.ema, "optimizer": None if final_epoch else optimizer.state_dict(), "wandb_id": wandb_run.id if wandb else None, } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers n = opt.name if opt.name.isnumeric() else "" fresults, flast, fbest = ( save_dir / f"results{n}.txt", wdir / f"last{n}.pt", wdir / f"best{n}.pt", ) for f1, f2 in zip( [wdir / "last.pt", wdir / "best.pt", results_file], [flast, fbest, fresults] ): if f1.exists(): os.rename(f1, f2) # rename if str(f2).endswith(".pt"): # is *.pt strip_optimizer(f2) # strip optimizer os.system( "gsutil cp %s gs://%s/weights" % (f2, opt.bucket) ) if opt.bucket else None # upload # Finish if plots: plot_results(save_dir=save_dir) # save as results.png if wandb: files = [ "results.png", "precision_recall_curve.png", "confusion_matrix.png", ] wandb.log( { "Results": [ wandb.Image(str(save_dir / f), caption=f) for f in files if (save_dir / f).exists() ] } ) logger.info( "%g epochs completed in %.3f hours.\n" % (epoch - start_epoch + 1, (time.time() - t0) / 3600) ) else: dist.destroy_process_group() wandb.run.finish() if wandb and wandb.run else None torch.cuda.empty_cache() return results
def run( data, weights=None, # model.pt path(s) batch_size=32, # batch size imgsz=640, # inference size (pixels) conf_thres=0.001, # confidence threshold iou_thres=0.6, # NMS IoU threshold task='val', # train, val, test, speed or study device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu workers=8, # max dataloader workers (per RANK in DDP mode) single_cls=False, # treat as single-class dataset augment=False, # augmented inference verbose=False, # verbose output save_txt=False, # save results to *.txt save_hybrid=False, # save label+prediction hybrid results to *.txt save_conf=False, # save confidences in --save-txt labels save_json=False, # save a COCO-JSON results file project=ROOT / 'runs/val', # save to project/name name='exp', # save to project/name exist_ok=False, # existing project/name ok, do not increment half=True, # use FP16 half-precision inference dnn=False, # use OpenCV DNN for ONNX inference model=None, dataloader=None, save_dir=Path(''), plots=True, callbacks=Callbacks(), compute_loss=None, ): # Initialize/load model and set device training = model is not None if training: # called by train.py device, pt, jit, engine = next(model.parameters()).device, True, False, False # get model device, PyTorch model half &= device.type != 'cpu' # half precision only supported on CUDA model.half() if half else model.float() else: # called directly device = select_device(device, batch_size=batch_size) # Directories save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir # Load model model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half) stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine imgsz = check_img_size(imgsz, s=stride) # check image size half = model.fp16 # FP16 supported on limited backends with CUDA if engine: batch_size = model.batch_size else: device = model.device if not (pt or jit): batch_size = 1 # export.py models default to batch-size 1 LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models') # Data data = check_dataset(data) # check # Configure model.eval() cuda = device.type != 'cpu' is_coco = isinstance(data.get('val'), str) and data['val'].endswith('coco/val2017.txt') # COCO dataset nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10, device=device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Dataloader if not training: if pt and not single_cls: # check --weights are trained on --data ncm = model.model.yaml['nc'] assert ncm == nc, f'{weights[0]} ({ncm} classes) trained on different --data than what you passed ({nc} ' \ f'classes). Pass correct combination of --weights and --data that are trained together.' model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz)) # warmup pad = 0.0 if task in ('speed', 'benchmark') else 0.5 rect = False if task == 'benchmark' else pt # square inference for benchmarks task = task if task in ('train', 'val', 'test') else 'val' # path to train/val/test images dataloader = create_dataloader(data[task], imgsz, batch_size, stride, single_cls, pad=pad, rect=rect, workers=workers, prefix=colorstr(f'{task}: '))[0] seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} class_map = coco80_to_coco91_class() if is_coco else list(range(1000)) s = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', '[email protected]', '[email protected]:.95') dt, p, r, f1, mp, mr, map50, map = [0.0, 0.0, 0.0], 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class = [], [], [], [] callbacks.run('on_val_start') pbar = tqdm(dataloader, desc=s, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar for batch_i, (im, targets, paths, shapes) in enumerate(pbar): callbacks.run('on_val_batch_start') t1 = time_sync() if cuda: im = im.to(device, non_blocking=True) targets = targets.to(device) im = im.half() if half else im.float() # uint8 to fp16/32 im /= 255 # 0 - 255 to 0.0 - 1.0 nb, _, height, width = im.shape # batch size, channels, height, width t2 = time_sync() dt[0] += t2 - t1 # Inference out, train_out = model(im) if training else model(im, augment=augment, val=True) # inference, loss outputs dt[1] += time_sync() - t2 # Loss if compute_loss: loss += compute_loss([x.float() for x in train_out], targets)[1] # box, obj, cls # NMS targets[:, 2:] *= torch.tensor((width, height, width, height), device=device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling t3 = time_sync() out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls) dt[2] += time_sync() - t3 # Metrics for si, pred in enumerate(out): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path, shape = Path(paths[si]), shapes[si][0] seen += 1 if len(pred) == 0: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Predictions if single_cls: pred[:, 5] = 0 predn = pred.clone() scale_coords(im[si].shape[1:], predn[:, :4], shape, shapes[si][1]) # native-space pred # Evaluate if nl: tbox = xywh2xyxy(labels[:, 1:5]) # target boxes scale_coords(im[si].shape[1:], tbox, shape, shapes[si][1]) # native-space labels labelsn = torch.cat((labels[:, 0:1], tbox), 1) # native-space labels correct = process_batch(predn, labelsn, iouv) if plots: confusion_matrix.process_batch(predn, labelsn) else: correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool) stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # (correct, conf, pcls, tcls) # Save/log if save_txt: save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / (path.stem + '.txt')) if save_json: save_one_json(predn, jdict, path, class_map) # append to COCO-JSON dictionary callbacks.run('on_val_image_end', pred, predn, path, names, im[si]) # Plot images if plots and batch_i < 3: f = save_dir / f'val_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(im, targets, paths, f, names), daemon=True).start() f = save_dir / f'val_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(im, output_to_target(out), paths, f, names), daemon=True).start() callbacks.run('on_val_batch_end') # Compute metrics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) ap50, ap = ap[:, 0], ap.mean(1) # [email protected], [email protected]:0.95 mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%11i' * 2 + '%11.3g' * 4 # print format LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats): for i, c in enumerate(ap_class): LOGGER.info(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in dt) # speeds per image if not training: shape = (batch_size, 3, imgsz, imgsz) LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) callbacks.run('on_val_end') # Save JSON if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json') # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...') with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb check_requirements(['pycocotools']) from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[:2] # update results ([email protected]:0.95, [email protected]) except Exception as e: LOGGER.info(f'pycocotools unable to run: {e}') # Return results model.float() # for training if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def test(data, weights=None, batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir=Path(''), # for saving images save_txt=False, # for auto-labelling save_conf=False, plots=True, log_imgs=0, refine=False): # number of logged images # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly set_logging() device = select_device(opt.device, batch_size=batch_size) save_txt = opt.save_txt # save *.txt labels # Directories save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half # half = device.type != 'cpu' # half precision only supported on CUDA half = False if half: model.half() # Configure model.eval() is_coco = data.endswith('coco.yaml') # is COCO dataset with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict check_dataset(data) # check nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Logging log_imgs, wandb = min(log_imgs, 100), None # ceil try: import wandb # Weights & Biases except ImportError: log_imgs = 0 # Dataloader if not training: img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once path = data['test'] if opt.task == 'test' else data['val'] # path to val/test images dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True)[0] seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} coco91class = coco80_to_coco91_class() s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width with torch.no_grad(): # Run model t = time_synchronized() (inf_out, _),feature = model(img, refine=True) # inference and training outputs # if len(boxes)!=res.shape[0]: # new_res = [] # for bbox in boxes: # len_bb = bbox.shape[0] # print(res[c:c+len_bb,...].view(1,-1,res.shape[-1]).shape) # new_res.append(res[c:c+len_bb,...].view(1,-1,res.shape[-1])) # c = c+len_bb # res = torch.cat(new_res,0) # Compute loss # if training: # loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # box, obj, cls # res,boxes = model.detector_(inf_out,feature) # model.refine_net = model.refine_net.to(device) # res = model.refine_net(res,boxes) # print(res.shape) # inf_out=torch.cat((inf_out,res),1) # Run NMS targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_txt else [] # for autolabelling t = time_synchronized() if refine: res,boxes = model.detector_(inf_out,feature) model.refine_net = model.refine_net.to(device) res = model.refine_net(res,boxes,eval=True) t0 += time_synchronized() - t output = non_max_suppression(res, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb) c = 0 if len(boxes)!=res.shape[0]: new_res = [] for bbox in boxes: len_bb = bbox.shape[0] # if len_bb==0: # len_bb=1 if len_bb>1: # print(torch.cat(output[c:c+len_bb],0).shape) k__ = [] for i in range(len_bb): k__.append(output[c+i].to(device)) a = torch.cat(k__,0) elif len_bb==1: a = output[c] else: a = torch.zeros(0,6).to(device) new_res.append(a) c = c+len_bb output = new_res else: output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb) t1 += time_synchronized() - t # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path = Path(paths[si]) seen += 1 if len(pred) == 0: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Predictions predn = pred.clone() scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Append to text file if save_txt: gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0]] # normalization gain whwh for *xyxy, conf, cls in predn.tolist(): xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: f.write(('%g ' * len(line)).rstrip() % line + '\n') # W&B logging if plots and len(wandb_images) < log_imgs: box_data = [{"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, "class_id": int(cls), "box_caption": "%s %.3f" % (names[cls], conf), "scores": {"class_score": conf}, "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()] boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space wandb_images.append(wandb.Image(img[si], boxes=boxes, caption=path.name)) # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int(path.stem) if path.stem.isnumeric() else path.stem box = xyxy2xywh(predn[:, :4]) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({'image_id': image_id, 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5)}) # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix.process_batch(pred, torch.cat((labels[:, 0:1], tbox), 1)) # Per target class for cls in torch.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero(as_tuple=False).view(-1) # prediction indices pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1) # target indices # Search for detections if pi.shape[0]: # Prediction to target ious ious, i = box_iou(predn[pi, :4], tbox[ti]).max(1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn if len(detected) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # Plot images if plots and batch_i < 3: f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start() f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(img, output_to_target(output), paths, f, names), daemon=True).start() # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1) # [P, R, [email protected], [email protected]:0.95] mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%12.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if verbose and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple if not training: print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) if wandb and wandb.run: wandb.log({"Images": wandb_images}) wandb.log({"Validation": [wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg'))]}) # Save JSON if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights anno_json = glob.glob('../coco/annotations/instances_val*.json')[0] # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.img_files] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[:2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print('ERROR: pycocotools unable to run: %s' % e) # Return results if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {save_dir}{s}") model.float() # for training maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = log_dir / 'weights' # weights dfirectory os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = str(log_dir / 'results.txt') epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs_init, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round( hyp['anchors']) # force autoanchor model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [ ] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False if opt.reg_lambda != 0: # the regularization is based on Synaptic Intelligence as described in the # paper. ewcData is a list of two elements (best parametes, importance) # while synData is a dictionary with all the trajectory data needed by SI model.ewcData, model.synData = create_syn_data(model) # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[ 'lrf']) + hyp['lrf'] # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Logging if wandb and wandb.run is None: id = ckpt.get('wandb_id') if 'ckpt' in locals() else None wandb_run = wandb.init(config=opt, resume="allow", project="YOLOv5", name=os.path.basename(log_dir), id=id) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}' ) # save previous weights if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) all_test_dataloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, n_batch=-1)[0] root = 'G:/projects/core50_350_1f/batches/' paths = os.listdir(root) train_paths = [] valid_paths = [] for p in paths: if 'train' in p: train_paths.append(root + p) elif 'val' in p: valid_paths.append(root + p) else: print(p) # external_memory = ext_memory() extMem = externalMemory() for core_batch in range(11): # Trainloader if opt.reg_lambda != 0: init_batch(model, model.ewcData, model.synData) print(f'------------CORE50 itertaion №:{core_batch}------------') external_files_path = extMem.file if core_batch > 0: train_path = [train_paths[core_batch], external_files_path] else: train_path = train_paths[core_batch] extMem.update_memory(train_paths[core_batch], update_iters=10 if core_batch == 0 else 1) dataloader, dataset = create_dataloader( train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, ) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) testloader = create_dataloader(valid_paths[core_batch], imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers)[0] # testloader #if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.IntTensor(labels[:, 0]) # classes plot_labels(labels, save_dir=log_dir) print(torch.bincount(c)) if tb_writer: # tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384 tb_writer.add_histogram('classes', c, core_batch) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights model.names = names # Start training t0 = time.time() nw = max( round(hyp['warmup_epochs'] * nb), 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) logger.info( 'Image sizes %g train, %g test\n' 'Using %g dataloader workers\nLogging results to %s\n' 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs)) # update number of epochs to iterative training if core_batch != 0: epochs = opt.epochs_iter # x_train, y_train = dataset.get_all_data() for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() mloss = torch.zeros(4, device=device) # mean losses logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) # x_train_splitted = torch.split(x_train, 4) # y_train_splitted = torch.split(y_train, 4) # pbar = enumerate(zip(x_train_splitted, y_train_splitted)) pbar = enumerate(dataloader) pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, _, _ ) in pbar: # batch ------------------------------------------------------------- # imgs = x_train[i * batch_size:(i + 1) * batch_size] # targets = y_train[i * batch_size:(i + 1) * batch_size] # # # preprocess tensor to proper form # # img, label = zip(imgs, targets) # transposed # for i, l in enumerate(targets): # l[:, 0] = i # add target image index for build_targets() # # imgs = torch.stack(imgs) # targets = torch.cat(targets) ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 if opt.reg_lambda != 0: pre_update(model, model.synData) # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # # Multi-scale # if opt.multi_scale: # sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size # sf = sz / max(imgs.shape[2:]) # scale factor # if sf != 1: # ns = [math.ceil(x * sf / gs) * gs for x in # imgs.shape[2:]] # new shape (stretched to gs-multiple) # imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step if opt.reg_lambda != 0: post_update(model, model.synData) scaler.update() optimizer.zero_grad() # if ema: # ema.update(model) # Print mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # end batch ------------------------------------------------------------------------------------------------ # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # mAP results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, # model=ema.ema, model=model, single_cls=opt.single_cls, dataloader=testloader, save_dir=log_dir, plots=epoch == 0, # plot first and last log_imgs=opt.log_imgs) # wandb.log({'per class/AP per class': maps}) # Write with open(results_file, 'a') as f: f.write( s + '%10.4g' * 7 % results + '\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Log tags = [ 'train/giou_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2' ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): if tb_writer: tb_writer.add_scalar(tag, x, epoch) # tensorboard if wandb: wandb.log({tag: x}) # W&B # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi # Save model save = not opt.nosave if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), # 'model': ema.ema, 'model': model, 'optimizer': optimizer.state_dict(), 'wandb_id': wandb_run.id if wandb else None } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training #consolidate_weights(model, cur_class) if opt.reg_lambda != 0: update_ewc_data(model, model.ewcData, model.synData, 0.001, 1) if rank in [-1, 0]: # Strip optimizers n = opt.name if opt.name.isnumeric() else '' fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt' for f1, f2 in zip( [wdir / 'last.pt', wdir / 'best.pt', results_file], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename if str(f2).endswith('.pt'): # is *.pt strip_optimizer(f2) # strip optimizer os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish plot_results(save_dir=log_dir) # save as results.png logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, #model=ema.ema, model=model, single_cls=opt.single_cls, dataloader=all_test_dataloader, save_dir=log_dir, #plots=epoch == 0 or final_epoch, # plot first and last log_imgs=opt.log_imgs, verbose=True) #wandb.log({'per class/AP per class All': maps[0]}) #tb_writer.add_scalar('per class/AP per class All', maps[0]) # Log tags = [ # train loss 'test/precision', 'test/recall', 'test/mAP_0.5', 'test/mAP_0.5:0.95', 'test/giou_loss', 'test/obj_loss', 'test/cls_loss' ] # params for x, tag in zip(list(results), tags): if tb_writer: tb_writer.add_scalar(tag, x, core_batch) # tensorboard if wandb: wandb.log({tag: x}) # W&B return results
def run( data, weights=None, # model.pt path(s) batch_size=32, # batch size imgsz=640, # inference size (pixels) conf_thres=0.001, # confidence threshold iou_thres=0.6, # NMS IoU threshold task='val', # train, val, test, speed or study device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu single_cls=False, # treat as single-class dataset augment=False, # augmented inference verbose=False, # verbose output save_txt=False, # save results to *.txt save_hybrid=False, # save label+prediction hybrid results to *.txt save_conf=False, # save confidences in --save-txt labels save_json=False, # save a cocoapi-compatible JSON results file project='runs/test', # save to project/name name='exp', # save to project/name exist_ok=False, # existing project/name ok, do not increment half=True, # use FP16 half-precision inference model=None, dataloader=None, save_dir=Path(''), plots=True, wandb_logger=None, compute_loss=None, ): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly device = select_device(device, batch_size=batch_size) # Directories save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir( parents=True, exist_ok=True) # make dir # Load model model = attempt_load(weights, map_location=device) # load FP32 model gs = max(int(model.stride.max()), 32) # grid size (max stride) imgsz = check_img_size(imgsz, s=gs) # check image size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Half half &= device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() if isinstance(data, str): with open(data) as f: data = yaml.safe_load(f) check_dataset(data) # check is_coco = type(data['val']) is str and data['val'].endswith( 'coco/val2017.txt') # COCO dataset nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Logging log_imgs = 0 if wandb_logger and wandb_logger.wandb: log_imgs = min(wandb_logger.log_imgs, 100) # Dataloader if not training: if device.type != 'cpu': model( torch.zeros(1, 3, imgsz, imgsz).to(device).type_as( next(model.parameters()))) # run once task = task if task in ( 'train', 'val', 'test') else 'val' # path to train/val/test images dataloader = create_dataloader(data[task], imgsz, batch_size, gs, single_cls, pad=0.5, rect=True, prefix=colorstr(f'{task}: '))[0] seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = { k: v for k, v in enumerate( model.names if hasattr(model, 'names') else model.module.names) } coco91class = coco80_to_coco91_class() s = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1, t2 = 0., 0., 0., 0., 0., 0., 0., 0., 0., 0. loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): t_ = time_synchronized() img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width t = time_synchronized() t0 += t - t_ # Run model out, train_out = model( img, augment=augment) # inference and training outputs t1 += time_synchronized() - t # Compute loss if compute_loss: loss += compute_loss([x.float() for x in train_out], targets)[1][:3] # box, obj, cls # Run NMS targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling t = time_synchronized() out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls) t2 += time_synchronized() - t # Statistics per image for si, pred in enumerate(out): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path = Path(paths[si]) seen += 1 if len(pred) == 0: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Predictions if single_cls: pred[:, 5] = 0 predn = pred.clone() scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred # Append to text file if save_txt: gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0 ]] # normalization gain whwh for *xyxy, conf, cls in predn.tolist(): xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: f.write(('%g ' * len(line)).rstrip() % line + '\n') # W&B logging - Media Panel plots if len( wandb_images ) < log_imgs and wandb_logger.current_epoch > 0: # Check for test operation if wandb_logger.current_epoch % wandb_logger.bbox_interval == 0: box_data = [{ "position": { "minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3] }, "class_id": int(cls), "box_caption": "%s %.3f" % (names[cls], conf), "scores": { "class_score": conf }, "domain": "pixel" } for *xyxy, conf, cls in pred.tolist()] boxes = { "predictions": { "box_data": box_data, "class_labels": names } } # inference-space wandb_images.append( wandb_logger.wandb.Image(img[si], boxes=boxes, caption=path.name)) wandb_logger.log_training_progress( predn, path, names) if wandb_logger and wandb_logger.wandb_run else None # Append to pycocotools JSON dictionary if save_json: # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... image_id = int( path.stem) if path.stem.isnumeric() else path.stem box = xyxy2xywh(predn[:, :4]) # xywh box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner for p, b in zip(pred.tolist(), box.tolist()): jdict.append({ 'image_id': image_id, 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5) }) # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 1:5]) scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels if plots: confusion_matrix.process_batch( predn, torch.cat((labels[:, 0:1], tbox), 1)) # Per target class for cls in torch.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero(as_tuple=False).view( -1) # target indices pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view( -1) # prediction indices # Search for detections if pi.shape[0]: # Prediction to target ious ious, i = box_iou(predn[pi, :4], tbox[ti]).max( 1) # best ious, indices # Append detections detected_set = set() for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti[i[j]] # detected target if d.item() not in detected_set: detected_set.add(d.item()) detected.append(d) correct[ pi[j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append( (correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # Plot images if plots and batch_i < 3: f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start() f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(img, output_to_target(out), paths, f, names), daemon=True).start() # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) ap50, ap = ap[:, 0], ap.mean(1) # [email protected], [email protected]:0.95 mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%11i' * 2 + '%11.3g' * 4 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t2)) # speeds per image if not training: shape = (batch_size, 3, imgsz, imgsz) print( f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) if wandb_logger and wandb_logger.wandb: val_batches = [ wandb_logger.wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg')) ] wandb_logger.log({"Validation": val_batches}) if wandb_images: wandb_logger.log({"Bounding Box Debugger/Images": wandb_images}) # Save JSON if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights ).stem if weights is not None else '' # weights anno_json = '../coco/annotations/instances_val2017.json' # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb check_requirements(['pycocotools']) from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [ int(Path(x).stem) for x in dataloader.dataset.img_files ] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[: 2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print(f'pycocotools unable to run: {e}') # Return results model.float() # for training if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {save_dir}{s}") maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def train(hyp, opt, device, tb_writer=None): logger.info(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = log_dir / 'weights' # weights directory os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = str(log_dir / 'results.txt') epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint # added by jiangrong if not opt.resume: ckpt['epoch'] = -1 if opt.nas: model = NasModel(opt.cfg, ch=3, nc=nc, nas=opt.nas, nas_stage=opt.nas_stage).to(device) # create else: model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create exclude = ['anchor'] if opt.cfg else [] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: if opt.nas: model = NasModel(opt.cfg, ch=3, nc=nc, nas=opt.nas, nas_stage=opt.nas_stage).to(device) # create if opt.nas_stage == 3: # TODO, Remapping with BN Statistics on Width-level model.re_organize_middle_weights() else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create if opt.nas and opt.nas_stage > 0: from models.experimental import attempt_load """ P R [email protected] 0.535 0.835 0.742 python test.py \ --weights /workspace/yolov5-v3/yolov5/runs/exp122/weights/best.pt \ --data ./data/baiguang.yaml \ --device 1 \ --conf-thres 0.2 """ teacher_model = attempt_load( "/workspace/yolov5-v3/yolov5/runs/exp259/weights/best.pt", map_location='cuda:1') teacher_model.eval() # Freeze freeze = [ '', ] # parameter names to freeze (full or partial) if any(freeze): for k, v in model.named_parameters(): if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_parameters(): v.requires_grad = True if '.bias' in k: pg2.append(v) # biases elif '.weight' in k and '.bn' not in k: pg1.append(v) # apply weight decay else: pg0.append(v) # all else if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: (( (1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.8 + 0.2 # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None and not opt.nas > 0: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}' ) # save previous weights if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode # TheModel = model if cuda and rank == -1 and torch.cuda.device_count() > 1 and not ( opt.nas and opt.nas_stage > 0): # https://pytorch.org/docs/stable/generated/torch.nn.DataParallel.html # >>> net = torch.nn.DataParallel(model, device_ids=[0, 1, 2]) # >>> output = net(input_var) # input_var can be on any device, including CPU model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # Exponential moving average ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=(opt.local_rank)) # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # Testloader if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader( test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers)[0] # only runs on process 0 # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights model.names = names # Class frequency if rank in [-1, 0]: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # model._initialize_biases(cf.to(device)) plot_labels(labels, save_dir=log_dir) if tb_writer: # tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384 tb_writer.add_histogram('classes', c, 0) # Check anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Start training t0 = time.time() nw = max(3 * nb, 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = ( 0, 0, 0, 0, 0, 0, 0 ) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification' scheduler.last_epoch = start_epoch - 1 # do not move # scaler = amp.GradScaler(enabled=cuda) logger.info('Image sizes %g train, %g test' % (imgsz, imgsz_test)) logger.info('Using %g dataloader workers' % dataloader.num_workers) logger.info('Starting training for %g epochs...' % epochs) # torch.autograd.set_detect_anomaly(True) plot_csum = 0 for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if dataset.image_weights: # Generate indices if rank in [-1, 0]: w = model.class_weights.cpu().numpy() * ( 1 - maps)**2 # class weights image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w) dataset.indices = random.choices( range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx # Broadcast if DDP if rank != -1: indices = torch.zeros([dataset.n], dtype=torch.int) if rank == 0: indices[:] = torch.tensor(dataset.indices, dtype=torch.int) dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- # print(type(targets), targets.size()) # [[_,classid(start from 0), x,y,w,h (0-1)]] # print('---> targets: ', targets) ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp( ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) if 'momentum' in x: x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward ###### jiangrong, turn off mixed precision ########## # with amp.autocast(enabled=cuda): if 1 == 1: pred = model(imgs) # forward, format x(bs,3,20,20,80+1+4) loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode # z= [] # for i in range(TheModel._modules['model'][-1].nl): # bs, _, ny, nx, _ = pred[i].shape # if TheModel._modules['model'][-1].grid[i].shape[2:4] != pred[i].shape[2:4]: # TheModel._modules['model'][-1].grid[i] = TheModel._modules['model'][-1]._make_grid(nx, ny).to(pred[i].device) # y = pred[i].sigmoid() # y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + TheModel._modules['model'][-1].grid[i].to(pred[i].device)) * TheModel._modules['model'][-1].stride[i] # xy # y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * TheModel._modules['model'][-1].anchor_grid[i] # wh # z.append(y.view(bs, -1, TheModel._modules['model'][-1].no)) # inf_out = torch.cat(z, 1) # teacher_pred = non_max_suppression(inf_out, conf_thres=0.2, iou_thres=0.6, merge=False) # assert len(teacher_pred) == imgs.size()[0] # for i, (det, plot_img) in enumerate(zip(teacher_pred, imgs.detach().cpu().numpy())): # plot_img = np.transpose(plot_img, (1,2,0)) # plot_img = np.uint8(plot_img * 255.0) # plot_csum += 1 # cv2.imwrite('./tmp/{}.jpg'.format(plot_csum), plot_img) # plot_img = cv2.imread('./tmp/{}.jpg'.format(plot_csum)) # for tgt in targets.detach().cpu().numpy(): # _, tgt_class_id, c_x, c_y, c_w, c_h = tgt # c_x, c_y, c_w, c_h = float(c_x), float(c_y), float(c_w), float(c_h) # c_x, c_y, c_w, c_h = c_x * plot_img.shape[1], c_y * plot_img.shape[0], c_w * plot_img.shape[1], c_h * plot_img.shape[0] # cv2.rectangle(plot_img, (int(c_x - c_w / 2), int(c_y - c_h / 2)), (int(c_x + c_w / 2), int(c_y + c_h / 2)), (0,0,255), 2) # print('===> ', int(c_x - c_w / 2), int(c_y - c_h / 2), int(c_x + c_w / 2), int(c_y + c_h / 2), tgt_class_id) # if det is not None: # det = det.detach().cpu().numpy() # for each_b in det: # pass # cv2.rectangle(plot_img, (int(each_b[0]), int(each_b[1])), (int(each_b[2]), int(each_b[3])), (255,0,0), 2) # print('---> ', int(each_b[0]), int(each_b[1]), int(each_b[2]), int(each_b[3]), float(each_b[4]), int(each_b[5])) # cv2.imwrite('./tmp/{}.jpg'.format(plot_csum), plot_img) if opt.nas and opt.nas_stage > 0: teacher_imgs = imgs.to('cuda:1') with torch.no_grad(): inf_out, _ = teacher_model(teacher_imgs) # forward # filter by obj confidence 0.05 teacher_pred = non_max_suppression_teacher( inf_out, conf_thres=0.05, iou_thres=0.6, merge=False ) # (x1, y1, x2, y2, conf, cls) in resized image size teacher_targets = teacher2targets(teacher_pred, teacher_imgs) # print('---> teacher_pred', teacher_pred) # print('---> targets', targets) # print('---> teacher_targets', teacher_targets) # TODO: apply soft label loss teacher_loss, teacher_loss_items = compute_teacher_loss( pred, teacher_targets.to(device), model) # loss scaled by batch_size # print("===> origin loss", loss, loss_items) # print("===> teacher loss", teacher_loss, teacher_loss_items) teacher_loss_scale = 2.0 loss += teacher_loss * teacher_loss_scale loss_items += teacher_loss_items * teacher_loss_scale ########## the targets and teacher predictions are matched, but they both can not be restored to the image, need TODO!! ########### # assert len(teacher_pred) == imgs.size()[0] # for i, (det, plot_img) in enumerate(zip(teacher_pred, imgs.detach().cpu().numpy())): # plot_img = np.transpose(plot_img, (1,2,0)) # plot_img = np.uint8(plot_img * 255.0) # plot_csum += 1 # cv2.imwrite('./tmp/{}.jpg'.format(plot_csum), plot_img) # plot_img = cv2.imread('./tmp/{}.jpg'.format(plot_csum)) # for tgt in targets.detach().cpu().numpy(): # _, tgt_class_id, c_x, c_y, c_w, c_h = tgt # c_x, c_y, c_w, c_h = float(c_x), float(c_y), float(c_w), float(c_h) # c_x, c_y, c_w, c_h = c_x * plot_img.shape[1], c_y * plot_img.shape[0], c_w * plot_img.shape[1], c_h * plot_img.shape[0] # cv2.rectangle(plot_img, (int(c_x - c_w / 2), int(c_y - c_h / 2)), (int(c_x + c_w / 2), int(c_y + c_h / 2)), (0,0,255), 2) # print('===> ', int(c_x - c_w / 2), int(c_y - c_h / 2), int(c_x + c_w / 2), int(c_y + c_h / 2), tgt_class_id) # if det is not None: # det = det.detach().cpu().numpy() # for each_b in det: # pass # cv2.rectangle(plot_img, (int(each_b[0]), int(each_b[1])), (int(each_b[2]), int(each_b[3])), (255,0,0), 2) # print('---> ', int(each_b[0]), int(each_b[1]), int(each_b[2]), int(each_b[3]), float(each_b[4]), int(each_b[5])) # cv2.imwrite('./tmp/{}.jpg'.format(plot_csum), plot_img) # Backward # scaler.scale(loss).backward() loss.backward() # Optimize if ni % accumulate == 0: # scaler.step(optimizer) # optimizer.step # scaler.update() optimizer.step() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot if ni < 3: f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename result = plot_images(images=imgs, targets=targets, paths=paths, fname=f) if tb_writer and result is not None: tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard # end batch ------------------------------------------------------------------------------------------------ # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema: ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP if opt.nas: # only evaluate the super network ema.ema.nas_stage = 0 results, maps, times = test.test(opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=log_dir) if opt.nas: ema.ema.nas_stage = opt.nas_stage # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Tensorboard if tb_writer: tags = [ 'train/giou_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2' ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): tb_writer.add_scalar(tag, x, epoch) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1] if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict() } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.name fresults, flast, fbest = 'results%s.txt' % n, wdir / f'last{n}.pt', wdir / f'best{n}.pt' for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', 'results.txt'], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename if str(f2).endswith('.pt'): # is *.pt strip_optimizer(f2) # strip optimizer os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish if not opt.evolve: plot_results(save_dir=log_dir) # save as results.png logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() return results
def train( hyp, # path/to/hyp.yaml or hyp dictionary opt, device, callbacks): save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \ Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \ opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze # Directories w = save_dir / 'weights' # weights dir (w.parent if evolve else w).mkdir(parents=True, exist_ok=True) # make dir last, best = w / 'last.pt', w / 'best.pt' # Hyperparameters if isinstance(hyp, str): with open(hyp, errors='ignore') as f: hyp = yaml.safe_load(f) # load hyps dict LOGGER.info( colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) # Save run settings if not evolve: with open(save_dir / 'hyp.yaml', 'w') as f: yaml.safe_dump(hyp, f, sort_keys=False) with open(save_dir / 'opt.yaml', 'w') as f: yaml.safe_dump(vars(opt), f, sort_keys=False) # Loggers data_dict = None if RANK in [-1, 0]: loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance if loggers.wandb: data_dict = loggers.wandb.data_dict if resume: weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size # Register actions for k in methods(loggers): callbacks.register_action(k, callback=getattr(loggers, k)) # Config plots = not evolve # create plots cuda = device.type != 'cpu' init_seeds(1 + RANK) with torch_distributed_zero_first(LOCAL_RANK): data_dict = data_dict or check_dataset(data) # check if None train_path, val_path = data_dict['train'], data_dict['val'] nc = 1 if single_cls else int(data_dict['nc']) # number of classes names = ['item'] if single_cls and len( data_dict['names']) != 1 else data_dict['names'] # class names assert len( names ) == nc, f'{len(names)} names found for nc={nc} dataset in {data}' # check is_coco = isinstance(val_path, str) and val_path.endswith( 'coco/val2017.txt') # COCO dataset # Model check_suffix(weights, '.pt') # check weights pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(LOCAL_RANK): weights = attempt_download( weights) # download if not found locally ckpt = torch.load(weights, map_location='cpu' ) # load checkpoint to CPU to avoid CUDA memory leak model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create exclude = [ 'anchor' ] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys csd = ckpt['model'].float().state_dict( ) # checkpoint state_dict as FP32 csd = intersect_dicts(csd, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(csd, strict=False) # load LOGGER.info( f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}' ) # report else: model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create # Freeze freeze = [ f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0])) ] # layers to freeze for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): LOGGER.info(f'freezing {k}') v.requires_grad = False # Image size gs = max(int(model.stride.max()), 32) # grid size (max stride) imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple # Batch size if RANK == -1 and batch_size == -1: # single-GPU only, estimate best batch size batch_size = check_train_batch_size(model, imgsz) loggers.on_params_update({"batch_size": batch_size}) # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay LOGGER.info(f"Scaled weight_decay = {hyp['weight_decay']}") g0, g1, g2 = [], [], [] # optimizer parameter groups for v in model.modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias g2.append(v.bias) if isinstance(v, nn.BatchNorm2d): # weight (no decay) g0.append(v.weight) elif hasattr(v, 'weight') and isinstance( v.weight, nn.Parameter): # weight (with decay) g1.append(v.weight) if opt.optimizer == 'Adam': optimizer = Adam(g0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum elif opt.optimizer == 'AdamW': optimizer = AdamW(g0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = SGD(g0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': g1, 'weight_decay': hyp['weight_decay'] }) # add g1 with weight_decay optimizer.add_param_group({'params': g2}) # add g2 (biases) LOGGER.info( f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups " f"{len(g0)} weight (no decay), {len(g1)} weight, {len(g2)} bias") del g0, g1, g2 # Scheduler if opt.cos_lr: lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] else: lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf' ] # linear scheduler = lr_scheduler.LambdaLR( optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # EMA ema = ModelEMA(model) if RANK in [-1, 0] else None # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # EMA if ema and ckpt.get('ema'): ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) ema.updates = ckpt['updates'] # Epochs start_epoch = ckpt['epoch'] + 1 if resume: assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.' if epochs < start_epoch: LOGGER.info( f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs." ) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, csd # DP mode if cuda and RANK == -1 and torch.cuda.device_count() > 1: LOGGER.warning( 'WARNING: DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n' 'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.' ) model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and RANK != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) LOGGER.info('Using SyncBatchNorm()') # Trainloader train_loader, dataset = create_dataloader( train_path, imgsz, batch_size // WORLD_SIZE, gs, single_cls, hyp=hyp, augment=True, cache=None if opt.cache == 'val' else opt.cache, rect=opt.rect, rank=LOCAL_RANK, workers=workers, image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: '), shuffle=True) mlc = int(np.concatenate(dataset.labels, 0)[:, 0].max()) # max label class nb = len(train_loader) # number of batches assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}' # Process 0 if RANK in [-1, 0]: val_loader = create_dataloader(val_path, imgsz, batch_size // WORLD_SIZE * 2, gs, single_cls, hyp=hyp, cache=None if noval else opt.cache, rect=True, rank=-1, workers=workers * 2, pad=0.5, prefix=colorstr('val: '))[0] if not resume: labels = np.concatenate(dataset.labels, 0) # c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) if plots: plot_labels(labels, names, save_dir) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) model.half().float() # pre-reduce anchor precision callbacks.run('on_pretrain_routine_end') # DDP mode if cuda and RANK != -1: model = DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK) # Model attributes nl = de_parallel( model).model[-1].nl # number of detection layers (to scale hyps) hyp['box'] *= 3 / nl # scale to layers hyp['cls'] *= nc / 80 * 3 / nl # scale to classes and layers hyp['obj'] *= (imgsz / 640)**2 * 3 / nl # scale to image size and layers hyp['label_smoothing'] = opt.label_smoothing model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.class_weights = labels_to_class_weights( dataset.labels, nc).to(device) * nc # attach class weights model.names = names # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 100) # number of warmup iterations, max(3 epochs, 100 iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training last_opt_step = -1 maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) stopper = EarlyStopping(patience=opt.patience) compute_loss = ComputeLoss(model) # init loss class LOGGER.info( f'Image sizes {imgsz} train, {imgsz} val\n' f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n' f"Logging results to {colorstr('bold', save_dir)}\n" f'Starting training for {epochs} epochs...') for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional, single-GPU only) if opt.image_weights: cw = model.class_weights.cpu().numpy() * ( 1 - maps)**2 / nc # class weights iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx # Update mosaic border (optional) # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(3, device=device) # mean losses if RANK != -1: train_loader.sampler.set_epoch(epoch) pbar = enumerate(train_loader) LOGGER.info( ('\n' + '%10s' * 7) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'labels', 'img_size')) if RANK in [-1, 0]: pbar = tqdm( pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device)) # loss scaled by batch_size if RANK != -1: loss *= WORLD_SIZE # gradient averaged between devices in DDP mode if opt.quad: loss *= 4. # Backward scaler.scale(loss).backward() # Optimize if ni - last_opt_step >= accumulate: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) last_opt_step = ni # Log if RANK in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB) pbar.set_description(('%10s' * 2 + '%10.4g' * 5) % (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1])) callbacks.run('on_train_batch_end', ni, model, imgs, targets, paths, plots, opt.sync_bn) if callbacks.stop_training: return # end batch ------------------------------------------------------------------------------------------------ # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for loggers scheduler.step() if RANK in [-1, 0]: # mAP callbacks.run('on_train_epoch_end', epoch=epoch) ema.update_attr(model, include=[ 'yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights' ]) final_epoch = (epoch + 1 == epochs) or stopper.possible_stop if not noval or final_epoch: # Calculate mAP results, maps, _ = val.run(data_dict, batch_size=batch_size // WORLD_SIZE * 2, imgsz=imgsz, model=ema.ema, single_cls=single_cls, dataloader=val_loader, save_dir=save_dir, plots=False, callbacks=callbacks, compute_loss=compute_loss) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi log_vals = list(mloss) + list(results) + lr callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi) # Save model if (not nosave) or (final_epoch and not evolve): # if save ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'model': deepcopy(de_parallel(model)).half(), 'ema': deepcopy(ema.ema).half(), 'updates': ema.updates, 'optimizer': optimizer.state_dict(), 'wandb_id': loggers.wandb.wandb_run.id if loggers.wandb else None, 'date': datetime.now().isoformat() } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) if (epoch > 0) and (opt.save_period > 0) and (epoch % opt.save_period == 0): torch.save(ckpt, w / f'epoch{epoch}.pt') del ckpt callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi) # Stop Single-GPU if RANK == -1 and stopper(epoch=epoch, fitness=fi): break # Stop DDP TODO: known issues shttps://github.com/ultralytics/yolov5/pull/4576 # stop = stopper(epoch=epoch, fitness=fi) # if RANK == 0: # dist.broadcast_object_list([stop], 0) # broadcast 'stop' to all ranks # Stop DPP # with torch_distributed_zero_first(RANK): # if stop: # break # must break all DDP ranks # end epoch ---------------------------------------------------------------------------------------------------- # end training ----------------------------------------------------------------------------------------------------- if RANK in [-1, 0]: LOGGER.info( f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.' ) for f in last, best: if f.exists(): strip_optimizer(f) # strip optimizers if f is best: LOGGER.info(f'\nValidating {f}...') results, _, _ = val.run( data_dict, batch_size=batch_size // WORLD_SIZE * 2, imgsz=imgsz, model=attempt_load(f, device).half(), iou_thres=0.65 if is_coco else 0.60, # best pycocotools results at 0.65 single_cls=single_cls, dataloader=val_loader, save_dir=save_dir, save_json=is_coco, verbose=True, plots=True, callbacks=callbacks, compute_loss=compute_loss) # val best model with plots if is_coco: callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi) callbacks.run('on_train_end', last, best, plots, epoch, results) LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}") torch.cuda.empty_cache() return results
def train(hyp, opt, device, tb_writer=None): logger.info(f'Hyperparameters {hyp}') print("tb_writer.log_dir: ", tb_writer.log_dir) #resumed的时候可以从这里恢复 log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = log_dir / 'weights' # weights directory os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' print("log_dir: ", log_dir) print("wdir: ", wdir) print("last: ", last) print("best: ", best) results_file = str(log_dir / 'results.txt') epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank print("epochs: ", epochs) print("batch_size: ", batch_size) print("total_batch_size: ", total_batch_size) print("weights: ", weights) print("rank: ", rank) # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' print("cuda: ", cuda) init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict print("data_dict: ", data_dict) with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') print("pretrained: ", pretrained) if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint """ckpt中,['epoch'], ['best_fitness'], ['training_results'], ['model'], ['optimizer']""" # print("ckpt: ", ckpt) print( "ckpt: ['epoch'], ['best_fitness'], ['training_results'], ['optimizer']: ", ckpt['epoch'], ckpt['best_fitness'], ckpt['training_results'], ckpt['optimizer']) # print("ckpt['model']: ", ckpt['model']) # print("ckpt['model'].model: ", ckpt['model'].model) # print("ckpt['model'].state_dict(): ", ckpt['model'].state_dict()) print("ckpt['model'].save: ", ckpt['model'].save) print("ckpt['model'].yaml: ", ckpt['model'].yaml) print("hyp.get('anchors'): ", hyp.get('anchors')) if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round( hyp['anchors']) # force autoanchor print("opt.cfg: ", opt.cfg) # create, 都为真则取or前面的, 即以opt.cfg中的内容(eg:yolov5s.yaml, yolov5x.yaml...)为主,其次是ckpt['model'].yaml model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) exclude = (['anchor'] if opt.cfg or hyp.get('anchors') else [] ) # exclude keys, 以opt.cfg中的anchor为主,其次是hyp中的 print("exclude: ", exclude) state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load # print("state_dict: ", state_dict) logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze freeze = [ '', ] # parameter names to freeze (full or partial) if any(freeze): for k, v in model.named_parameters(): if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer """当模型梯度累积了(nbs/total_batch_size)次之后,再更新一次模型参数,变相的扩大了batch_size""" nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_parameters(): v.requires_grad = True if '.bias' in k: pg2.append(v) # biases elif '.weight' in k and '.bn' not in k: pg1.append(v) # apply weight decay else: pg0.append(v) # all else #optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # 设置学习率衰减,这里为余弦退火方式进行衰减 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp[ 'lrf']) + hyp['lrf'] # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) #每个epochs具有一个不同的学习率 # plot_lr_scheduler(optimizer, scheduler, epochs) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}' ) # save previous weights if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(max(model.stride)) # grid size (max stride) # 获取模型总步长和模型输入图片分辨率 imgsz, imgsz_test = [ check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples #检查输入图片分辨率是gs=32的整数倍 print("imgsz: ", imgsz) print("imgsz_test: ", imgsz_test) # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: print( "DP mode..............................................................................." ) model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # Exponential moving average ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers) print("dataloader: ", dataloader) print("dataset: ", dataset) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) nb = len(dataloader) # number of batches print("nb: ", nb) # Process 0 if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates print("ema.updates: ", ema.updates) testloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers)[0] # testloader if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) plot_labels(labels, save_dir=log_dir) if tb_writer: # tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384 tb_writer.add_histogram('classes', c, 0) # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights model.names = names #print("model: ", model) # Start training t0 = time.time() n_warmup = max( round(hyp['warmup_epochs'] * nb), 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # n_warmup = min(n_warmup, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training print("n_warmup: ", n_warmup) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) #混合精度梯度放大模块 maps = np.zeros(nc) # mAP per class results = ( 0, 0, 0, 0, 0, 0, 0 ) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification' logger.info( 'Image sizes %g train, %g test\nUsing %g dataloader workers\nLogging results to %s\n' 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs)) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices if rank in [-1, 0]: cw = model.class_weights.cpu().numpy() * ( 1 - maps)**2 # class weights iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights dataset.indices = random.choices( range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx # Broadcast if DDP if rank != -1: indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start batch) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= n_warmup: # model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou) accumulate = max( 1, np.interp(ni, [0, n_warmup], [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, [0, n_warmup], [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, [0, n_warmup], [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): #自动混合精度 pred = model(imgs) # forward loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode # Backward scaler.scale(loss).backward() # Optimize """每accumulate个batch时更新一次, 在n_warmup之内时,accumulate从1逐渐增大到4""" if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot if ni < 3: f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename result = plot_images(images=imgs, targets=targets, paths=paths, fname=f) if tb_writer and result is not None: tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard # end batch ------------------------------------------------------------------------------------------------ # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema: ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP if final_epoch: # replot predictions [ os.remove(x) for x in glob.glob( str(log_dir / 'test_batch*_pred.jpg')) if os.path.exists(x) ] results, maps, times = evaluate.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=log_dir) # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Tensorboard if tb_writer: tags = [ 'train/giou_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2' ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): tb_writer.add_scalar(tag, x, epoch) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1] if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict() } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers n = opt.name if opt.name.isnumeric() else '' fresults, flast, fbest = log_dir / f'results{n}.txt', wdir / f'last{n}.pt', wdir / f'best{n}.pt' for f1, f2 in zip([wdir / 'last.pt', wdir / 'best.pt', results_file], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename if str(f2).endswith('.pt'): # is *.pt strip_optimizer(f2) # strip optimizer os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish if not opt.evolve: plot_results(save_dir=log_dir) # save as results.png logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() return results
def run( data, weights=None, # model.pt path(s) batch_size=32, # batch size imgsz=640, # inference size (pixels) conf_thres=0.001, # confidence threshold iou_thres=0.6, # NMS IoU threshold task='val', # train, val, test, speed or study device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu single_cls=False, # treat as single-class dataset augment=False, # augmented inference verbose=False, # verbose output save_txt=False, # save results to *.txt save_hybrid=False, # save label+prediction hybrid results to *.txt save_conf=False, # save confidences in --save-txt labels save_json=False, # save a COCO-JSON results file project=ROOT / 'runs/val', # save to project/name name='exp', # save to project/name exist_ok=False, # existing project/name ok, do not increment half=True, # use FP16 half-precision inference model=None, dataloader=None, save_dir=Path(''), plots=True, callbacks=Callbacks(), compute_loss=None, ): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly device = select_device(device, batch_size=batch_size) # Directories save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run (save_dir / 'labels' if save_txt else save_dir).mkdir( parents=True, exist_ok=True) # make dir # Load model check_suffix(weights, '.pt') model = attempt_load(weights, map_location=device) # load FP32 model gs = max(int(model.stride.max()), 32) # grid size (max stride) imgsz = check_img_size(imgsz, s=gs) # check image size # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 # if device.type != 'cpu' and torch.cuda.device_count() > 1: # model = nn.DataParallel(model) # Data data = check_dataset(data) # check # Half half &= device.type != 'cpu' # half precision only supported on CUDA model.half() if half else model.float() # Configure model.eval() is_coco = isinstance(data.get('val'), str) and data['val'].endswith( 'coco/val2017.txt') # COCO dataset nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Dataloader if not training: if device.type != 'cpu': model( torch.zeros(1, 3, imgsz, imgsz).to(device).type_as( next(model.parameters()))) # run once task = task if task in ( 'train', 'val', 'test') else 'val' # path to train/val/test images dataloader = create_dataloader(data[task], imgsz, batch_size, gs, single_cls, pad=0.5, rect=True, prefix=colorstr(f'{task}: '))[0] seen = 0 confusion_matrix = ConfusionMatrix(nc=nc) names = { k: v for k, v in enumerate( model.names if hasattr(model, 'names') else model.module.names) } class_map = coco80_to_coco91_class() if is_coco else list(range(1000)) s = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', '[email protected]', '[email protected]:.95') dt, p, r, f1, mp, mr, map50, map = [0.0, 0.0, 0.0], 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 loss = torch.zeros(3, device=device) jdict, stats, ap, ap_class = [], [], [], [] for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): t1 = time_sync() img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width t2 = time_sync() dt[0] += t2 - t1 # Run model out, train_out = model( img, augment=augment) # inference and training outputs dt[1] += time_sync() - t2 # Compute loss if compute_loss: loss += compute_loss([x.float() for x in train_out], targets)[1] # box, obj, cls # Run NMS targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling t3 = time_sync() out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls) dt[2] += time_sync() - t3 # Statistics per image for si, pred in enumerate(out): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class path, shape = Path(paths[si]), shapes[si][0] seen += 1 if len(pred) == 0: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Predictions if single_cls: pred[:, 5] = 0 predn = pred.clone() scale_coords(img[si].shape[1:], predn[:, :4], shape, shapes[si][1]) # native-space pred # Evaluate if nl: tbox = xywh2xyxy(labels[:, 1:5]) # target boxes scale_coords(img[si].shape[1:], tbox, shape, shapes[si][1]) # native-space labels labelsn = torch.cat((labels[:, 0:1], tbox), 1) # native-space labels correct = process_batch(predn, labelsn, iouv) if plots: confusion_matrix.process_batch(predn, labelsn) else: correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool) stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # (correct, conf, pcls, tcls) # Save/log if save_txt: save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / (path.stem + '.txt')) if save_json: save_one_json(predn, jdict, path, class_map) # append to COCO-JSON dictionary callbacks.run('on_val_image_end', pred, predn, path, names, img[si]) # Plot images if plots and batch_i < 3: f = save_dir / f'val_batch{batch_i}_labels.jpg' # labels Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start() f = save_dir / f'val_batch{batch_i}_pred.jpg' # predictions Thread(target=plot_images, args=(img, output_to_target(out), paths, f, names), daemon=True).start() # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) ap50, ap = ap[:, 0], ap.mean(1) # [email protected], [email protected]:0.95 mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%11i' * 2 + '%11.3g' * 4 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in dt) # speeds per image if not training: shape = (batch_size, 3, imgsz, imgsz) print( f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t) # Plots if plots: confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) callbacks.run('on_val_end') # Save JSON if save_json and len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights ).stem if weights is not None else '' # weights anno_json = str( Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json') # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print(f'\nEvaluating pycocotools mAP... saving {pred_json}...') with open(pred_json, 'w') as f: json.dump(jdict, f) try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb check_requirements(['pycocotools']) from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') if is_coco: eval.params.imgIds = [ int(Path(x).stem) for x in dataloader.dataset.img_files ] # image IDs to evaluate eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[: 2] # update results ([email protected]:0.95, [email protected]) except Exception as e: print(f'pycocotools unable to run: {e}') # Return results model.float() # for training if not training: s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' print(f"Results saved to {colorstr('bold', save_dir)}{s}") maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def test( data, weights=None, batch_size=16, imgsz=640, conf_thres=0.3, iou_thres=0.5, # for NMS save_json=False, single_cls=False, augment=False, verbose=False, model=None, dataloader=None, save_dir='', merge=False, emb_dim=256, save_txt=False): # Initialize/load model and set device training = model is not None if training: # called by train.py device = next(model.parameters()).device # get model device else: # called directly device = select_device(opt.device, batch_size=batch_size) merge, save_txt = opt.merge, opt.save_txt # use Merge NMS, save *.txt labels if save_txt: out = Path('inference/output') if os.path.exists(out): shutil.rmtree(out) # delete output folder os.makedirs(out) # make new output folder # Remove previous for f in glob.glob(str(Path(save_dir) / 'test_batch*.jpg')): os.remove(f) # Load model model = attempt_load(weights, map_location=device) # load FP32 model imgsz = [check_img_size(x, model.stride.max()) for x in imgsz] # Half half = device.type != 'cpu' # half precision only supported on CUDA if half: model.half() # Configure model.eval() with open(data) as f: data = yaml.load(f, Loader=yaml.FullLoader) # model dict nc = 1 if single_cls else int(data['nc']) # number of classes iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for [email protected]:0.95 niou = iouv.numel() # Dataloader if not training: if len(imgsz) == 1: img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img else: img = torch.zeros((1, 3, imgsz[1], imgsz[0]), device=device) _ = model(img.half() if half else img ) if device.type != 'cpu' else None # run once root = data['root'] path = data['test'] if opt.task == 'test' else data[ 'test_emb'] # path to val/test images dataloader = create_dataloader(root, path, imgsz, batch_size, model.stride.max(), opt, hyp=None, augment=False, cache=False, pad=0.5, rect=False)[0] seen = 0 names = model.names if hasattr(model, 'names') else model.module.names s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', '[email protected]', '[email protected]:.95') p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. jdict, stats, ap, ap_class = [], [], [], [] loss = torch.zeros(4, device=device) for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): img = img.to(device, non_blocking=True) img = img.half() if half else img.float() # uint8 to fp16/32 img /= 255.0 # 0 - 255 to 0.0 - 1.0 targets = targets.to(device) nb, _, height, width = img.shape # batch size, channels, height, width whwh = torch.Tensor([width, height, width, height]).to(device) # Disable gradients with torch.no_grad(): # Run model t = time_synchronized() inf_out, train_out_p, train_out_pemb = model( img, augment=augment) # inference and training outputs t0 += time_synchronized() - t # Compute loss if training: # if model has loss hyperparameters loss += compute_loss([x.float() for x in train_out_p], [x.float() for x in train_out_pemb], targets, model)[1][:4] # GIoU, obj, cls, lid # Run NMS t = time_synchronized() output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, merge=merge, emb_dim=emb_dim) t1 += time_synchronized() - t ''' images = letterbox(cv2.imread(paths[1]), [608,1088], auto=False, scaleup=False)[0] d = output[1] if d is None: continue for i in range(len(d)): cv2.rectangle(images, (int(d[i][0]), int(d[i][1])), (int(d[i][2]), int(d[i][3])), (0, 0, 255), 2) cv2.imshow("image", images) cv2.waitKey(0) ''' # Statistics per image for si, pred in enumerate(output): labels = targets[targets[:, 0] == si, 1:] nl = len(labels) tcls = labels[:, 0].tolist() if nl else [] # target class seen += 1 if pred is None: if nl: stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) continue # Clip boxes to image bounds clip_coords(pred, (height, width)) # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) if nl: detected = [] # target indices tcls_tensor = labels[:, 0] # target boxes tbox = xywh2xyxy(labels[:, 2:6]) * whwh # Per target class for cls in torch.unique(tcls_tensor): ti = (cls == tcls_tensor).nonzero(as_tuple=False).view( -1) # prediction indices pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view( -1) # target indices # Search for detections if pi.shape[0]: # Prediction to target ious ious, i = box_iou(pred[pi, :4], tbox[ti]).max( 1) # best ious, indices # Append detections for j in (ious > iouv[0]).nonzero(as_tuple=False): d = ti[i[j]] # detected target if d not in detected: detected.append(d) correct[ pi[j]] = ious[j] > iouv # iou_thres is 1xn if len( detected ) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) stats.append( (correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) # Plot images if batch_i < 1: f = Path(save_dir) / ('test_batch%g_gt.jpg' % batch_i) # filename plot_images(img, targets, paths, str(f), names) # ground truth f = Path(save_dir) / ('test_batch%g_pred.jpg' % batch_i) plot_test_images(img, output_to_target(output, width, height), paths, str(f), names) # predictions # Compute statistics stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy if len(stats) and stats[0].any(): p, r, ap, f1, ap_class = ap_per_class(*stats) p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean( 1) # [P, R, [email protected], [email protected]:0.95] mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class else: nt = torch.zeros(1) # Print results pf = '%20s' + '%12.3g' * 6 # print format print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) # Print results per class if verbose and nc > 1 and len(stats): for i, c in enumerate(ap_class): print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) # Print speeds t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz[0], imgsz[1], batch_size) if not training: print( 'Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) # Return results model.float() # for training maps = np.zeros(nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info( colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) print(f'Hyperparameters {hyp}') """ 训练日志包括:权重、tensorboard文件、超参数hyp、设置的训练参数opt(也就是epochs,batch_size等),result.txt result.txt包括: 占GPU内存、训练集的GIOU loss, objectness loss, classification loss, 总loss, targets的数量, 输入图片分辨率, 准确率TP/(TP+FP),召回率TP/P ; 测试集的mAP50, [email protected]:0.95, GIOU loss, objectness loss, classification loss. 还会保存batch<3的ground truth """ # 获取保存路径、总轮次、批次、总批次(涉及到分布式训练)、权重、进程序号(主要用于分布式训练) save_dir, epochs, batch_size, total_batch_size, weights, rank = \ Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Directories wdir = save_dir / 'weights' wdir.mkdir(parents=True, exist_ok=True) # make dir last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = save_dir / 'results.txt' # Save run settings # 保存hyp和opt with open(save_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(save_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure plots = not opt.evolve # create plots cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict # torch_distributed_zero_first同步所有进程 # check_dataset检查数据集,如果没找到数据集则下载数据集(仅适用于项目中自带的yaml文件数据集) with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc = 1 if opt.single_cls else int(data_dict['nc']) # number of classes names = ['item'] if opt.single_cls and len( data_dict['names']) != 1 else data_dict['names'] # class names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') if pretrained: # 加载模型,从google云盘中自动下载模型 # 但通常会下载失败,建议提前下载下来放进weights目录 with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally # 加载检查点 ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round( hyp['anchors']) # force autoanchor """ 这里模型创建,可通过opt.cfg,也可通过ckpt['model'].yaml 这里的区别在于是否是resume,resume时会将opt.cfg设为空,则按照ckpt['model'].yaml创建模型 这也影响着下面是否除去anchor的key(也就是不加载anchor),如果resume则不加载anchor 主要是因为保存的模型会保存anchors,有时候用户自定义了anchor之后,再resume,则原来基于coco数据集的anchor就会覆盖自己设定的anchor, 参考https://github.com/ultralytics/yolov5/issues/459 所以下面设置了intersect_dicts,该函数就是忽略掉exclude """ model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [ ] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load # 显示加载预训练权重的的键值对和创建模型的键值对 # 如果设置了resume,则会少加载两个键值对(anchors,anchor_grid) logger.info( 'Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze """ 冻结模型层,设置冻结层名字即可 具体可以查看https://github.com/ultralytics/yolov5/issues/679 但作者不鼓励冻结层,因为他的实验当中显示冻结层不能获得更好的性能,参照:https://github.com/ultralytics/yolov5/pull/707 并且作者为了使得优化参数分组可以正常进行,在下面将所有参数的requires_grad设为了True 其实这里只是给一个freeze的示例 """ freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer """ nbs为模拟的batch_size; 就比如默认的话上面设置的opt.batch_size为16,这个nbs就为64, 也就是模型梯度累积了64/16=4(accumulate)次之后 再更新一次模型,变相的扩大了batch_size """ nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing # 根据accumulate设置权重衰减系数 hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay logger.info(f"Scaled weight_decay = {hyp['weight_decay']}") # 将模型分成三组(weight、bn, bias, 其他所有参数)优化 pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay # 选用优化器,并设置pg0组的优化方式 if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) # 设置weight、bn的优化方式 optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay # 设置biases的优化方式 optimizer.add_param_group({'params': pg2}) # add pg2 (biases) # 打印优化信息 logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # 设置cosine调度器,定义学习率衰减学习率衰减,这里为余弦退火方式进行衰减 # 就是根据以下公式lf,epoch和超参数hyp['lrf']进行衰减 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR # lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] # cosine lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Logging if rank in [-1, 0] and wandb and wandb.run is None: opt.hyp = hyp # add hyperparameters wandb_run = wandb.init( config=opt, resume="allow", project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, name=save_dir.stem, id=ckpt.get('wandb_id') if 'ckpt' in locals() else None) loggers = {'wandb': wandb} # loggers dict # EMA # 在深度学习中,经常会使用EMA(指数移动平均)这个方法对模型的参数做滑动平均,以求提高测试指标并增加模型鲁棒,如果GPU进程数大于1,则不创建 # Exponential moving average ema = ModelEMA(model) if rank in [-1, 0] else None # Resume # best_fitness是以[0.0, 0.0, 0.1, 0.9]为系数并乘以[精确度, 召回率, [email protected], [email protected]:0.95]再求和所得 # 根据best_fitness来保存best.pt start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer # 加载优化器与best_fitness if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # EMA if ema and ckpt.get('ema'): ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) ema.updates = ckpt['updates'] # Results # 加载训练结果result.txt if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs 加载训练的轮次 start_epoch = ckpt['epoch'] + 1 """ 如果resume,则备份权重 尽管目前resume能够近似100%成功的起作用了,参照:https://github.com/ultralytics/yolov5/pull/756 但为了防止resume时出现其他问题,把之前的权重覆盖了,所以这里进行备份,参照:https://github.com/ultralytics/yolov5/pull/765 """ if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % ( weights, epochs) """ 如果新设置epochs小于加载的epoch, 则视新设置的epochs为需要再训练的轮次数而不再是总的轮次数 """ if epochs < start_epoch: logger.info( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # 获取模型最大步长和模型输入图片分辨率 gs = int(model.stride.max()) # grid size (max stride) nl = model.model[ -1].nl # number of detection layers (used for scaling hyp['obj']) # 检查训练和测试图片分辨率确保能够整除总步长gs imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode # 分布式训练,参照:https://github.com/ultralytics/yolov5/issues/475 # DataParallel模式,仅支持单机多卡 # rank为进程编号, 这里应该设置为rank=-1则使用DataParallel模式 # rank=-1且gpu数量=1时,不会进行分布式 if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm # 使用跨卡同步BN if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # DDP mode # 如果rank不等于-1,则使用DistributedDataParallel模式 # local_rank为gpu编号,rank为进程,例如rank=3,local_rank=0 表示第 3 个进程内的第 1 块 GPU。 if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: ')) """ 获取标签中最大的类别值,并于类别数作比较 如果小于类别数则表示有问题 """ mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # Process 0 if rank in [-1, 0]: # 更新ema模型的updates参数,保持ema的平滑性 ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader( test_path, imgsz_test, total_batch_size, gs, opt, # testloader hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, pad=0.5, prefix=colorstr('val: '))[0] if not opt.resume: # 将所有样本的标签拼接到一起shape为(total, 1),统计后做可视化 labels = np.concatenate(dataset.labels, 0) # 获得所有样本的类别 c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) if plots: # 根据上面的统计对所有样本的类别,中心点xy位置,长宽wh做可视化 plot_labels(labels, save_dir, loggers) if tb_writer: tb_writer.add_histogram('classes', c, 0) # Check anchors """ 计算默认锚点anchor与数据集标签框的长宽比值 标签的长h宽w与anchor的长h_a宽w_a的比值, 即h/h_a, w/w_a都要在(1/hyp['anchor_t'], hyp['anchor_t'])是可以接受的 如果标签框满足上面条件的数量小于总数的99%,则根据k-mean算法聚类新的锚点anchor """ if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Model parameters # 根据自己数据集的类别数设置分类损失的系数 hyp['box'] *= 3. / nl # scale to layers hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers hyp['obj'] *= (imgsz / 640)**2 * 3. / nl # scale to image size and layers model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) # 根据labels初始化图片采样权重 model.class_weights = labels_to_class_weights( dataset.labels, nc).to(device) * nc # attach class weights model.names = names """ 设置giou的值在objectness loss中做标签的系数, 使用代码如下 tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype) 这里model.gr=1,也就是说完全使用标签框与预测框的giou值来作为该预测框的objectness标签 """ # Start training t0 = time.time() # 获取热身训练的迭代次数 nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = (0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls) """ 设置学习率衰减所进行到的轮次, 目的是打断训练后,--resume接着训练也能正常的衔接之前的训练进行学习率衰减 """ scheduler.last_epoch = start_epoch - 1 # do not move # 通过torch1.6自带的api设置混合精度训练 scaler = amp.GradScaler(enabled=cuda) """ 打印训练和测试输入图片分辨率 加载图片时调用的cpu进程数 从哪个epoch开始训练 """ logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n' f'Using {dataloader.num_workers} dataloader workers\n' f'Logging results to {save_dir}\n' f'Starting training for {epochs} epochs...') # 加载图片权重(可选),定义进度条,设置偏差Burn-in,使用多尺度,前向传播,损失函数,反向传播,优化器,打印进度条,保存训练参数至tensorboard,计算mAP,保存结果到results.txt,保存模型(最好和最后) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices """ 如果设置进行图片采样策略, 则根据前面初始化的图片采样权重model.class_weights以及maps配合每张图片包含的类别数 通过random.choices生成图片索引indices从而进行采样 """ if rank in [-1, 0]: cw = model.class_weights.cpu().numpy() * ( 1 - maps)**2 / nc # class weights iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights # 类平衡采样 dataset.indices = random.choices( range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx # Broadcast if DDP # 如果是DDP模式,则广播采样策略 if rank != -1: indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders # 初始化训练时打印的平均损失信息 mloss = torch.zeros(4, device=device) # mean losses if rank != -1: # DDP模式下打乱数据, ddp.sampler的随机采样数据是基于epoch+seed作为随机种子, # 每次epoch不同,随机种子就不同 dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar tqdm 创建进度条,方便训练时 信息的展示 optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup """ 热身训练(前nw次迭代) 在前nw次迭代中,根据以下方式选取accumulate和学习率 """ if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 """ bias的学习率从0.1下降到基准学习率lr*lf(epoch),其他的参数学习率从0增加到lr*lf(epoch) lf为上面设置的余弦退火的衰减函数 """ x['lr'] = np.interp(ni, xi, [ hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch) ]) if 'momentum' in x: x['momentum'] = np.interp( ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale # 设置多尺度训练,从imgsz * 0.5, imgsz * 1.5 + gs随机选取尺寸 if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # 混合精度 with amp.autocast(enabled=cuda): pred = model(imgs) # forward # 计算损失,包括分类损失,objectness损失,框的回归损失 # loss为总损失值,loss_items为一个元组,包含分类损失,objectness损失,框的回归损失和总损失 if (IS_Debug()): #loss, loss_items = compute_loss(pred, targets.to(device), model, imgs) # loss scaled by batch_size loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size else: loss, loss_items = compute_loss( pred, targets.to(device), model) # loss scaled by batch_size if rank != -1: # 平均不同gpu之间的梯度 loss *= opt.world_size # gradient averaged between devices in DDP mode if opt.quad: loss *= 4. # Backward scaler.scale(loss).backward() # Optimize # 模型反向传播accumulate次之后再根据累积的梯度更新一次参数 if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema is not None: ema.update(model) # Print if rank in [-1, 0]: # 打印显存,进行的轮次,损失,target的数量和图片的size等信息 mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot # 将前三次迭代batch的标签框在图片上画出来并保存 if plots and ni < 3: f = save_dir / f'train_batch{ni}.jpg' # filename Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start() # if tb_writer: # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard elif plots and ni == 3 and wandb: wandb.log({ "Mosaics": [ wandb.Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg') ] }) # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler # 进行学习率衰减 lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP # 更新EMA的属性 # 添加include的属性 if ema: ema.update_attr(model, include=[ 'yaml', 'nc', 'hyp', 'gr', 'names', 'stride', 'class_weights' ]) # 判断该epoch是否为最后一轮 final_epoch = epoch + 1 == epochs # 对测试集进行测试,计算mAP等指标 # 测试时使用的是EMA模型 if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, plots=plots and final_epoch, log_imgs=opt.log_imgs if wandb else 0) # Write with open(results_file, 'a') as f: f.write( s + '%10.4g' * 7 % results + '\n') # P, R, [email protected], [email protected], val_loss(box, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Log tags = [ 'train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss 'x/lr0', 'x/lr1', 'x/lr2' ] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): if tb_writer: tb_writer.add_scalar(tag, x, epoch) # tensorboard if wandb: wandb.log({tag: x}) # W&B # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # weighted combination of [P, R, [email protected], [email protected]] if fi > best_fitness: best_fitness = fi # Save model """ 保存模型,还保存了epoch,results,optimizer等信息, optimizer将不会在最后一轮完成后保存 model保存的是EMA的模型 """ save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict(), 'wandb_id': wandb_run.id if wandb else None } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers """ 模型训练完后,strip_optimizer函数将optimizer从ckpt中去除; 并且对模型进行model.half(), 将Float32的模型->Float16, 可以减少模型大小,提高inference速度 """ final = best if best.exists() else last # final model for f in [last, best]: if f.exists(): strip_optimizer(f) # strip optimizers if opt.bucket: os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload # Plots if plots: # 可视化results.txt文件 plot_results(save_dir=save_dir) # save as results.png if wandb: files = [ 'results.png', 'precision_recall_curve.png', 'confusion_matrix.png' ] wandb.log({ "Results": [ wandb.Image(str(save_dir / f), caption=f) for f in files if (save_dir / f).exists() ] }) if opt.log_artifacts: wandb.log_artifact(artifact_or_path=str(final), type='model', name=save_dir.stem) # Test best.pt logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) if opt.data.endswith('coco.yaml') and nc == 80: # if COCO for conf, iou, save_json in ([0.25, 0.45, False], [0.001, 0.65, True]): # speed, mAP tests results, _, _ = test.test(opt.data, batch_size=total_batch_size, imgsz=imgsz_test, conf_thres=conf, iou_thres=iou, model=attempt_load(final, device).half(), single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, save_json=save_json, plots=False) else: dist.destroy_process_group() # 释放显存 wandb.run.finish() if wandb and wandb.run else None torch.cuda.empty_cache() return results
def train(hyp, opt, device, tb_writer=None): print(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = str(log_dir / 'weights') + os.sep # weights directory os.makedirs(wdir, exist_ok=True) last = wdir + 'last.pt' best = wdir + 'best.pt' results_file = str(log_dir / 'results.txt') epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # TODO: Use DDP logging. Only the first process is allowed to log. # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict root_path = data_dict['root'] train_path = data_dict['train'] test_emb_path = data_dict['test_emb'] test_path = data_dict['test'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Remove previous results if rank in [-1, 0]: for f in glob.glob('*_batch*.jpg') + glob.glob(results_file): os.remove(f) # Create model model = Model(opt.cfg, nc=nc).to(device) # Image sizes gs = int(max(model.stride)) # grid size (max stride) imgsz = imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # Optimizer nbs = 64 # nominal batch size # default DDP implementation is slow for accumulation according to: https://pytorch.org/docs/stable/notes/ddp.html # all-reduce operation is carried out during loss.backward(). # Thus, there would be redundant all-reduce communications in a accumulation procedure, # which means, the result is still right but the training speed gets slower. # TODO: If acceleration is needed, there is an implementation of allreduce_post_accumulation # in https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/BERT/run_pretraining.py accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_parameters(): if v.requires_grad: if '.bias' in k: pg2.append(v) # biases elif '.weight' in k and '.bn' not in k: pg1.append(v) # apply weight decay else: pg0.append(v) # all else if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: (( (1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.8 + 0.2 # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # Load Model with torch_distributed_zero_first(rank): attempt_download(weights) start_epoch, best_fitness = 0, 0.0 if weights.endswith('.pt'): # pytorch format ckpt = torch.load(weights, map_location=device) # load checkpoint # load model try: exclude = ['anchor'] # exclude keys ckpt['model'] = { k: v for k, v in ckpt['model'].float().state_dict().items() if k in model.state_dict() and not any(x in k for x in exclude) and model.state_dict()[k].shape == v.shape } model.load_state_dict(ckpt['model'], strict=False) print('Transferred %g/%g items from %s' % (len(ckpt['model']), len(model.state_dict()), weights)) except KeyError as e: s = "%s is not compatible with %s. This may be due to model differences or %s may be out of date. " \ "Please delete or update %s and try again, or use --weights '' to train from scratch." \ % (weights, opt.cfg, weights, weights) raise KeyError(s) from e # load optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # load results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # epochs start_epoch = ckpt['epoch'] + 1 if epochs < start_epoch: print( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) print('Using SyncBatchNorm()') # Exponential moving average ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=(opt.local_rank)) # Trainloader dataloader, dataset = create_dataloader(root_path, train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, local_rank=rank, world_size=opt.world_size) # Testloader testloader = create_dataloader(root_path, test_path, imgsz_test, total_batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images, rect=True, local_rank=-1, world_size=opt.world_size)[0] mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights model.names = names # model.nID = dataset.nID # Class frequency if rank in [-1, 0]: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes plot_labels(labels, save_dir=log_dir) if tb_writer: # tb_writer.add_hparams(hyp, {}) # causes duplicate https://github.com/ultralytics/yolov5/pull/384 tb_writer.add_histogram('classes', c, 0) # Check anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Start training t0 = time.time() nw = max(3 * nb, 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = ( 0, 0, 0, 0, 0, 0, 0 ) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification' scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) if rank in [0, -1]: print('Image sizes {} train, {} test'.format(str(imgsz), str(imgsz_test))) print('Using %g dataloader workers' % dataloader.num_workers) print('Starting training for %g epochs...' % epochs) # torch.autograd.set_detect_anomaly(True) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() mloss = torch.zeros(5, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) if rank in [-1, 0]: print(('\n' + '%10s' * 8 + '%13s') % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'lid', 'total', 'targets', 'img_size')) pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp( ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) if 'momentum' in x: x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']]) # Multi-scale if opt.multi_scale and random.random() < 0.5: candidate_shapes = [[608, 1088], [480, 864], [320, 576], [512, 960], [384, 640]] curr_shapes = candidate_shapes[random.randint(0, 4)] imgs = F.interpolate(imgs, size=curr_shapes, mode='bilinear', align_corners=False) # Autocast with amp.autocast(enabled=cuda): # Forward pred_detect, pred_emb = model(imgs) # Loss loss, loss_items = compute_loss(pred_detect, pred_emb, targets.to(device), model) # scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema is not None: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6 + "%10s") % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], ' [%g,%g]' % (imgs.shape[-1], imgs.shape[-2])) pbar.set_description(s) # Plot if ni < 3: f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename result = plot_images(images=imgs, targets=targets, paths=paths, fname=f) if tb_writer and result is not None: tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard # end batch ------------------------------------------------------------------------------------------------ # Scheduler scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema is not None: ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP results, maps, times = test.test( opt.data, batch_size=total_batch_size, imgsz=imgsz_test, save_json=final_epoch and opt.data.endswith(os.sep + 'coco.yaml'), model=ema.ema.module if hasattr(ema.ema, 'module') else ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=log_dir, emb_dim=model.module.emb_dim) # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 8 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Tensorboard if tb_writer: tags = [ 'train/giou_loss', 'train/obj_loss', 'train/cls_loss', 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss', 'val/cls_loss' ] for x, tag in zip(list(mloss[:-1]) + list(results), tags): tb_writer.add_scalar(tag, x, epoch) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1] if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema.module if hasattr(ema, 'module') else ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict() } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.name fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename ispt = f2.endswith('.pt') # is *.pt strip_optimizer(f2) if ispt else None # strip optimizer os.system('gsutil cp %s gs://%s/weights' % ( f2, opt.bucket)) if opt.bucket and ispt else None # upload # Finish if not opt.evolve: plot_results(save_dir=log_dir) # save as results.png print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() return results
def train(hyp, opt, device, tb_writer=None): print(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory wdir = str(log_dir / 'weights') + os.sep # weights directory os.makedirs(wdir, exist_ok=True) last = wdir + 'last.pt' best = wdir + 'best.pt' results_file = str(log_dir / 'results.txt') epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # TODO: Use DDP logging. Only the first process is allowed to log. # Save run settings with open(log_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(log_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict train_path = data_dict['train'] test_path = data_dict['val'] nc, names = (1, ['item']) if opt.single_cls else (int( data_dict['nc']), data_dict['names']) # number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint model = Darknet(opt.cfg).to(device) # create state_dict = { k: v for k, v in ckpt['model'].items() if model.state_dict()[k].numel() == v.numel() } model.load_state_dict(state_dict, strict=False) print('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Darknet(opt.cfg).to(device) # create # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in dict(model.named_parameters()).items(): if '.bias' in k: pg2.append(v) # biases elif 'Conv2d.weight' in k: pg1.append(v) # apply weight_decay else: pg0.append(v) # all else if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) print('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR lf = lambda x: (( (1 + math.cos(x * math.pi / epochs)) / 2)**1.0) * 0.8 + 0.2 # cosine scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if epochs < start_epoch: print( '%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = 32 # grid size (max stride) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size ] # verify imgsz are gs-multiples # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) print('Using SyncBatchNorm()') # Exponential moving average ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=(opt.local_rank)) # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, local_rank=rank, world_size=opt.world_size) mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class nb = len(dataloader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, opt.data, nc - 1) # Testloader if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates *** # local_rank is set to -1. Because only the first process is expected to do evaluation. testloader = create_dataloader(test_path, imgsz_test, batch_size, gs, opt, hyp=hyp, augment=False, cache=opt.cache_images, rect=True, local_rank=-1, world_size=opt.world_size)[0] # Model parameters hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights model.names = names # Class frequency if rank in [-1, 0]: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes # cf = torch.bincount(c.long(), minlength=nc) + 1. # model._initialize_biases(cf.to(device)) plot_labels(labels, save_dir=log_dir) if tb_writer: tb_writer.add_histogram('classes', c, 0) # Check anchors #if not opt.noautoanchor: # check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Start training t0 = time.time() nw = max(3 * nb, 1e3) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class results = ( 0, 0, 0, 0, 0, 0, 0 ) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification' scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) if rank in [0, -1]: print('Image sizes %g train, %g test' % (imgsz, imgsz_test)) print('Using %g dataloader workers' % dataloader.num_workers) print('Starting training for %g epochs...' % epochs) # torch.autograd.set_detect_anomaly(True) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if dataset.image_weights: # Generate indices if rank in [-1, 0]: w = model.class_weights.cpu().numpy() * ( 1 - maps)**2 # class weights image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w) dataset.indices = random.choices( range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx # Broadcast if DDP if rank != -1: indices = torch.zeros([dataset.n], dtype=torch.int) if rank == 0: indices[:] = torch.from_tensor(dataset.indices, dtype=torch.int) dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(4, device=device) # mean losses if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) if rank in [-1, 0]: print( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size')) pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float( ) / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou) accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp( ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) if 'momentum' in x: x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:] ] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Autocast with amp.autocast(enabled=cuda): # Forward pred = model(imgs) # Loss loss, loss_items = compute_loss(pred, targets.to(device), model) # scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode # if not torch.isfinite(loss): # print('WARNING: non-finite loss, ending training ', loss_items) # return results # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema is not None: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1 ) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) pbar.set_description(s) # Plot if ni < 3: f = str(log_dir / ('train_batch%g.jpg' % ni)) # filename result = plot_images(images=imgs, targets=targets, paths=paths, fname=f) if tb_writer and result is not None: tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard # end batch ------------------------------------------------------------------------------------------------ # Scheduler scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema is not None: ema.update_attr(model) final_epoch = epoch + 1 == epochs save_FLAG = False # if not opt.notest or final_epoch: # Calculate mAP if (epoch + 1) % opt.test_step == 0 or final_epoch: save_FLAG = True results, maps, times = test.test( opt.data, batch_size=batch_size, imgsz=imgsz_test, save_json=final_epoch and opt.data.endswith(os.sep + 'coco.yaml'), model=ema.ema.module if hasattr(ema.ema, 'module') else ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=log_dir) # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Tensorboard if tb_writer: tags = [ 'train/giou_loss', 'train/obj_loss', 'train/cls_loss', 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/giou_loss', 'val/obj_loss', 'val/cls_loss' ] for x, tag in zip(list(mloss[:-1]) + list(results), tags): tb_writer.add_scalar(tag, x, epoch) # Update best mAP fi = fitness(np.array(results).reshape( 1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1] if fi > best_fitness: best_fitness = fi # Save model # save = (not opt.nosave) or (final_epoch and not opt.evolve) save = (save_FLAG) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = { 'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema.module.state_dict() if hasattr(ema, 'module') else ema.ema.state_dict(), 'optimizer': None if final_epoch else optimizer.state_dict() } # Save last, best and delete torch.save(ckpt, last) if epoch >= (epochs - 5): torch.save(ckpt, last.replace('.pt', '_{:03d}.pt'.format(epoch))) if (best_fitness == fi) and not final_epoch: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers n = ('_' if len(opt.name) and not opt.name.isnumeric() else '') + opt.name fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n for f1, f2 in zip([wdir + 'last.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]): if os.path.exists(f1): os.rename(f1, f2) # rename ispt = f2.endswith('.pt') # is *.pt strip_optimizer(f2) if ispt else None # strip optimizer os.system('gsutil cp %s gs://%s/weights' % ( f2, opt.bucket)) if opt.bucket and ispt else None # upload # Finish if not opt.evolve: plot_results(save_dir=log_dir) # save as results.png print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() return results
def train(hyp, opt, device, tb_writer=None, wandb=None): logger.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) save_dir, epochs, batch_size, total_batch_size, weights, rank = \ Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # Directories wdir = save_dir / 'weights' wdir.mkdir(parents=True, exist_ok=True) # make dir last = wdir / 'last.pt' best = wdir / 'best.pt' results_file = save_dir / 'results.txt' # Save run settings with open(save_dir / 'hyp.yaml', 'w') as f: yaml.dump(hyp, f, sort_keys=False) with open(save_dir / 'opt.yaml', 'w') as f: yaml.dump(vars(opt), f, sort_keys=False) # Configure plots = not opt.evolve # create plots cuda = device.type != 'cpu' init_seeds(2 + rank) with open(opt.data) as f: data_dict = yaml.load(f, Loader=yaml.SafeLoader) # data dict with torch_distributed_zero_first(rank): check_dataset(data_dict) # check train_path = data_dict['train'] test_path = data_dict['val'] nc1 = 1 if opt.single_cls else int(data_dict['nc1']) # number of classes names1 = ['item'] if opt.single_cls and len(data_dict['names1']) != 1 else data_dict['names1'] # class names assert len(names1) == nc1, '%g names found for nc=%g dataset in %s' % (len(names1), nc1, opt.data) # check nc2 = 1 if opt.single_cls else int(data_dict['nc2']) # number of classes # edit names2 = ['item'] if opt.single_cls and len(data_dict['names2']) != 1 else data_dict['names2'] # class names # edit assert len(names2) == nc2, '%g names found for nc=%g dataset in %s' % (len(names2), nc2, opt.data) # check # edit # Model pretrained = weights.endswith('.pt') if pretrained: with torch_distributed_zero_first(rank): attempt_download(weights) # download if not found locally ckpt = torch.load(weights, map_location=device) # load checkpoint if hyp.get('anchors'): ckpt['model'].yaml['anchors'] = round(hyp['anchors']) # force autoanchor model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc1=nc1, nc2=nc2).to(device) # create # edit exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [] # exclude keys state_dict = ckpt['model'].float().state_dict() # to FP32 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect model.load_state_dict(state_dict, strict=False) # load logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report else: model = Model(opt.cfg, ch=3, nc1=nc1, nc2=nc2).to(device) # create # Freeze freeze = [] # parameter names to freeze (full or partial) for k, v in model.named_parameters(): v.requires_grad = True # train all layers if any(x in k for x in freeze): print('freezing %s' % k) v.requires_grad = False # Optimizer nbs = 64 # nominal batch size accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay logger.info(f"Scaled weight_decay = {hyp['weight_decay']}") pg0, pg1, pg2 = [], [], [] # optimizer parameter groups for k, v in model.named_modules(): if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): pg2.append(v.bias) # biases if isinstance(v, nn.BatchNorm2d): pg0.append(v.weight) # no decay elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): pg1.append(v.weight) # apply decay if opt.adam: optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum else: optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay optimizer.add_param_group({'params': pg2}) # add pg2 (biases) logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) del pg0, pg1, pg2 # Scheduler https://arxiv.org/pdf/1812.01187.pdf # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR if opt.linear_lr: lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear else: lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Logging if rank in [-1, 0] and wandb and wandb.run is None: opt.hyp = hyp # add hyperparameters wandb_run = wandb.init(config=opt, resume="allow", project=opt.wandb if opt.project == 'runs/train' else Path(opt.project).stem, name=save_dir.stem, id=ckpt.get('wandb_id') if 'ckpt' in locals() else None) loggers = {'wandb': wandb} # loggers dict # Resume start_epoch, best_fitness = 0, 0.0 if pretrained: # Optimizer if ckpt['optimizer'] is not None: optimizer.load_state_dict(ckpt['optimizer']) best_fitness = ckpt['best_fitness'] # Results if ckpt.get('training_results') is not None: with open(results_file, 'w') as file: file.write(ckpt['training_results']) # write results.txt # Epochs start_epoch = ckpt['epoch'] + 1 if opt.resume: assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (weights, epochs) if epochs < start_epoch: logger.info('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % (weights, ckpt['epoch'], epochs)) epochs += ckpt['epoch'] # finetune additional epochs del ckpt, state_dict # Image sizes gs = int(model.stride.max()) # grid size (max stride) nl = model.model[-1].nl # number of detection layers (used for scaling hyp['obj']) imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size] # verify imgsz are gs-multiples # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm if opt.sync_bn and cuda and rank != -1: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) logger.info('Using SyncBatchNorm()') # EMA ema = ModelEMA(model) if rank in [-1, 0] else None # DDP mode if cuda and rank != -1: model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) # Trainloader (labels: cls1, cls2, xyxy) dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, world_size=opt.world_size, workers=opt.workers, image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: ')) nb = len(dataloader) # number of batches mlc1 = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class 1 # edit assert mlc1 < nc1, 'Label class 1 %g exceeds nc1=%g in %s. Possible class 1 labels are 0-%g'\ % (mlc1, nc1, opt.data, nc1 - 1) # edit mlc2 = np.concatenate(dataset.labels, 0)[:, 1].max() # max label class 2 # edit assert mlc2 < nc2, 'Label class 2 %g exceeds nc2=%g in %s. Possible class 2 labels are 0-%g'\ % (mlc2, nc2, opt.data, nc2 - 1) # edit # Process 0 if rank in [-1, 0]: ema.updates = start_epoch * nb // accumulate # set EMA updates testloader = create_dataloader(test_path, imgsz_test, batch_size * 2, gs, opt, # testloader hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, rank=-1, world_size=opt.world_size, workers=opt.workers, pad=0.5, prefix=colorstr('val: '))[0] if not opt.resume: labels = np.concatenate(dataset.labels, 0) c1 = torch.tensor(labels[:, 0]) # classes 1 c2 = torch.tensor(labels[:, 1]) # classes 2 # edit # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency # model._initialize_biases(cf.to(device)) if plots: plot_labels(labels[:, 0], labels[:, 2:], save_dir, loggers, suffix="_1") plot_labels(labels[:, 1], labels[:, 2:], save_dir, loggers, suffix="_2") # edit if tb_writer: tb_writer.add_histogram('classes 1', c1, 0) tb_writer.add_histogram('classes 1', c2, 0) # edit # Anchors if not opt.noautoanchor: check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) # Model parameters hyp['box'] *= 3. / nl # scale to layers hyp['cls1'] = hyp['cls'] * nc1 / 80. * 3. / nl # scale to classes and layers # edit hyp['cls2'] = hyp['cls'] * nc2 / 80. * 3. / nl # scale to classes and layers # edit hyp['obj'] *= (imgsz / 640) ** 2 * 3. / nl # scale to image size and layers model.nc1 = nc1 # attach number of classes to model # edit model.nc2 = nc2 # attach number of classes to model # edit model.hyp = hyp # attach hyperparameters to model model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) model.class1_weights, model.class2_weights = labels_to_class_weights(dataset.labels, nc1, nc2) # attach class weights # edit model.class1_weights = model.class1_weights.to(device) * nc1 # edit model.class2_weights = model.class1_weights.to(device) * nc2 # edit model.names1 = names1 # edit model.names2 = names2 # edit # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps1 = np.zeros(nc1) # mAP per class 1 # edit maps2 = np.zeros(nc2) # mAP per class 2 # edit results1 = (0, 0, 0, 0, 0, 0, 0, 0) # P_1, R_1, [email protected]_1, [email protected]_1, val_loss(box, obj, cls1, cls2) # edit results2 = (0, 0, 0, 0, 0, 0, 0, 0) # P_2, R_2, [email protected]_2, [email protected]_2, val_loss(box, obj, cls1, cls2) # edit scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) compute_loss = ComputeLoss(model) # init loss class logger.info(f'Image sizes {imgsz} train, {imgsz_test} test\n' f'Using {dataloader.num_workers} dataloader workers\n' f'Logging results to {save_dir}\n' f'Starting training for {epochs} epochs...') for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ model.train() # Update image weights (optional) if opt.image_weights: # Generate indices if rank in [-1, 0]: cw1 = model.class1_weights.cpu().numpy() * (1 - maps1) ** 2 / nc1 # class weights # edit cw2 = model.class2_weights.cpu().numpy() * (1 - maps2) ** 2 / nc2 # class weights # edit iw1 = labels_to_image_weights(dataset.labels, nc=nc1, class_weights=cw1) # image weights # edit iw2 = labels_to_image_weights(dataset.labels, nc=nc2, class_weights=cw2) # image weights # edit dataset.indices1 = random.choices(range(dataset.n), weights=iw1, k=dataset.n) # rand weighted idx # edit dataset.indices2 = random.choices(range(dataset.n), weights=iw2, k=dataset.n) # rand weighted idx # edit # Broadcast if DDP if rank != -1: indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() dist.broadcast(indices, 0) if rank != 0: dataset.indices = indices.cpu().numpy() # Update mosaic border # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) # dataset.mosaic_border = [b - imgsz, -b] # height, width borders mloss = torch.zeros(5, device=device) # mean losses # edit if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info(('\n' + '%10s' * 9) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls1', 'cls2', 'total', 'targets', 'img_size')) # edit if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- # print(i, targets) # todo ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float() / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 # Warmup if ni <= nw: xi = [0, nw] # x interp # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) accumulate = max(1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) if 'momentum' in x: x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) # Multi-scale if opt.multi_scale: sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size sf = sz / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward with amp.autocast(enabled=cuda): pred = model(imgs) # forward # loss: loss * batch_size; loss_items: lbox, lobj, lcls1, lcls2, loss loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size if rank != -1: loss *= opt.world_size # gradient averaged between devices in DDP mode if opt.quad: loss *= 4. # Backward scaler.scale(loss).backward() # Optimize if ni % accumulate == 0: scaler.step(optimizer) # optimizer.step scaler.update() optimizer.zero_grad() if ema: ema.update(model) # Print if rank in [-1, 0]: mloss = (mloss * i + loss_items) / (i + 1) # update mean losses mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) s = ('%10s' * 2 + '%10.4g' * 7) %\ ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) # edit pbar.set_description(s) # Plot if plots and ni < 3: f = save_dir / f'train_batch{ni}.jpg' # filename Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start() # if tb_writer: # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) # tb_writer.add_graph(model, imgs) # add model to tensorboard elif plots and ni == 10 and wandb: wandb.log({"Mosaics": [wandb.Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg') if x.exists()]}, commit=False) # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # mAP if ema: ema.update_attr(model, include=['yaml', 'nc1', 'nc2', 'hyp', 'gr', 'names1', 'names2', 'stride', 'class_weights']) # edit final_epoch = epoch + 1 == epochs if not opt.notest or final_epoch: # Calculate mAP maps1, maps2, times, results1, results2 = test.test(opt.data, batch_size=batch_size * 2, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, verbose=nc1 < 50 and nc2 < 50 and final_epoch, # edit plots=plots and final_epoch, log_imgs=opt.log_imgs if wandb else 0, compute_loss=compute_loss) # Write # print("before write results:", results1, results2) # todo with open(results_file, 'a') as f: # epoch/epochs, mem, train_loss(box, obj, cls1, cls2, loss), targets.shape[0], imgs.shape[-1] # edit # P_1, R_1, [email protected]_1, [email protected]_1, P_2, R_2, [email protected]_2, [email protected]_2, val_loss(box, obj, cls1, cls2) f.write(s + '%10.4g' * 4 % results1[:4] + '%10.4g' * 8 % results2 + '\n') # edit if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # Log tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss_1', 'train/cls_loss_2', # train loss # edit 'metrics/precision_1', 'metrics/recall_1', 'metrics/mAP_0.5_1', 'metrics/mAP_0.5:0.95_1', 'metrics/precision_2', 'metrics/recall_2', 'metrics/mAP_0.5_2', 'metrics/mAP_0.5:0.95_2', # edit 'val/box_loss', 'val/obj_loss', 'val/cls_loss_1', 'val/cls_loss_2', # val loss # edit 'x/lr0', 'x/lr1', 'x/lr2'] # params for x, tag in zip(list(mloss[:-1]) + list(results1)[:4] + list(results2) + lr, tags): # edit if tb_writer: tb_writer.add_scalar(tag, x, epoch) # tensorboard if wandb: wandb.log({tag: x}, step=epoch, commit=tag == tags[-1]) # W&B # Update best mAP # weighted combination of [P, R, [email protected], [email protected]] for class 1 and class 2 fi = fitness(np.array(results1).reshape(1, -1), np.array(results2).reshape(1, -1)) # edit if fi > best_fitness: best_fitness = fi # Save model save = (not opt.nosave) or (final_epoch and not opt.evolve) if save: with open(results_file, 'r') as f: # create checkpoint ckpt = {'epoch': epoch, 'best_fitness': best_fitness, 'training_results': f.read(), 'model': ema.ema, 'optimizer': None if final_epoch else optimizer.state_dict(), 'wandb_id': wandb_run.id if wandb else None} # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training if rank in [-1, 0]: # Strip optimizers final = best if best.exists() else last # final model for f in [last, best]: if f.exists(): strip_optimizer(f) # strip optimizers if opt.bucket: os.system(f'gsutil cp {final} gs://{opt.bucket}/weights') # upload # Plots if plots: plot_results(save_dir=save_dir) # save as results.png if wandb: files = ['results.png', 'confusion_matrix_1.png', 'confusion_matrix_2.png', *[f'{x}_curve_{y}.png' for x in ('F1', 'PR', 'P', 'R') for y in ('1', '2')]] # edit wandb.log({"Results": [wandb.Image(str(save_dir / f), caption=f) for f in files if (save_dir / f).exists()]}) if opt.log_artifacts: wandb.log_artifact(artifact_or_path=str(final), type='model', name=save_dir.stem) # Test best.pt logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) if opt.data.endswith('coco.yaml') and nc1 == 80: # if COCO for conf, iou, save_json in ([0.25, 0.45, False], [0.001, 0.65, True]): # speed, mAP tests maps1, maps2, times, results1, results2 = test.test(opt.data, batch_size=batch_size * 2, imgsz=imgsz_test, conf_thres=conf, iou_thres=iou, model=attempt_load(final, device).half(), single_cls=opt.single_cls, dataloader=testloader, save_dir=save_dir, save_json=save_json, plots=False) else: dist.destroy_process_group() wandb.run.finish() if wandb and wandb.run else None torch.cuda.empty_cache() return results1, results2