def train(hyp, opt, device, tb_writer=None): logger.info( colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) print( 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) shutil.copyfile(os.path.basename(__file__), os.path.join(str(save_dir), os.path.basename(__file__))) # 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 is_coco = opt.data.endswith('coco.yaml') # 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, Path(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 # Model pretrained = weights.endswith('.pt') if 'yolov5' in opt.cfg: model = V5Centernet(opt.cfg, num_classes=nc, pretrained=weights, device=device).to(device) else: model = V5Dual(opt.cfg, num_classes=nc, pretrained=weights, device=device).to(device) # model = FlexibleModel(model_config=opt.cfg).to(device) if pretrained: # weights = torch.load(pretrained) # self.model.load_state_dict(weights) exclude = [] # exclude keys ckpt = torch.load(weights, map_location=device) # load checkpoint 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('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report del ckpt, state_dict # segLoss = FocalLoss() segLoss = nn.MSELoss() 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']}") print(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 = MADGRAD(pg0, lr=hyp['lr0'], momentum=hyp['momentum']) 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))) 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 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: if 0: # 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)) 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 if hasattr(model, 'model'): gs = max(int(model.model.stride.max()), 32) # grid size (max stride) try: nl = model.model.model[-1].nl except: nl = model.model.detection.nl # number of detection layers (used for scaling hyp['obj']) else: gs = max(int(model.stride.max()), 32) # grid size (max stride) nl = model.detection.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()') print('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] 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) # 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...') print(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...') criterion = torch.nn.BCEWithLogitsLoss() 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' * 11) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'labels', 'img_size', 'hm loss', 'lo_loss', 'loss')) print(('\n' + '%10s' * 11) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'labels', 'img_size', 'hm loss', 'lo_loss', 'loss')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, hms, targets, paths, _, logt ) 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 hms = hms.to(device, non_blocking=True).float() # print(hms.shape, imgs.shape) # if i>10: # break # print('output_layer====output_layer: ',torch.max(hms)) # 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, seg_out, logits = 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. # print(seg_out.shape) seg_out = _sigmoid(seg_out) # seg_out = seg_out[:,0,:,:] hms = torch.unsqueeze(hms, 1) # print(hms.shape, seg_out.shape) # print(targets.shape, imgs.shape, logits.shape, paths.shape) # logits = torch.clip(logits, -11, 11) # print(logits) logit_loss = criterion(logits, logt.to(device)) hm_loss = segLoss(seg_out, hms) # print(hm_loss) ratio = sigmoid_rampup(epoch, int(40 * epochs / 50)) ratio = 10 * (1 - ratio) loss = 1 * loss + ratio * hm_loss + 0.5 * logit_loss # 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' * 9) % ( '%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1], hm_loss, logit_loss, loss) 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, hms, 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_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, 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 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_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( model.module if is_parallel(model) else 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]: # Plots 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() ] }) # Test best.pt logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) print('%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 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 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 if wandb_logger.wandb and not opt.evolve: # Log the stripped model wandb_logger.wandb.log_artifact( str(final), type='model', name='run_' + wandb_logger.wandb_run.id + '_model', aliases=['last', '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): 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 ## hyps : command line hyperparameters (overwrites hyp.yaml) hyps = None try: if opt.hyps is not None and len(opt.hyps) > 0: ## hyps should evaluate to a python dict() hyps = ast.literal_eval(opt.hyps) ## add hyps to hyp (overwrite)... for k, v in hyps.items(): hyp[k] = v except: pfunc(f'ERROR: problem parsing hyps (hyperparameter string): {hyps}') # Print swagger job json string... if opt.job_str: pfunc(f'swagger job submitted:\n{opt.job_str}\n') # 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 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) if hyp.get('freeze'): ## freeze backbone layers? N = int(hyp['freeze']) + 1 freeze = ['model.%s.' % x for x in range(N)] logger.info('Freezing first {} layers of network'.format(N)) 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('freezing %s' % k) v.requires_grad = False # ## create separate testing model # test_model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create # for k, v in test_model.named_parameters(): # v.requires_grad = False # freeze all layers # 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 lr_epochs, init_epochs = epochs, 0 if hyp.get('init_epochs'): init_epochs = hyp['init_epochs'] lr_epochs += init_epochs if opt.linear_lr: lf = lambda x: (1 - x / (lr_epochs - 1)) * (1.0 - hyp['lrf']) + hyp[ 'lrf'] # linear else: lf = one_cycle(1, hyp['lrf'], lr_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 # Fix 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 # Fix Crop size if hyp.get('crop') and hyp['crop'] > 0: hyp['crop'] = check_img_size(hyp['crop'], gs) imgsz_test = hyp['crop'] # DP mode if cuda and rank == -1 and torch.cuda.device_count() > 1: pfunc('DOING DATA PARALLEL 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()') # if rank in [-1, 0]: pfunc( f'RANK={rank} opt.world_size={opt.world_size} dist.get_rank()={dist.get_rank()} dist.get_world_size()={dist.get_world_size()}' ) # Trainloader trainloader, dataset = create_dataloader( train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, # cache='disk', cache_efficient_sampling=True, 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(trainloader) # 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 # test_batch_size = batch_size test_batch_size = 1 ## so test_batch_size-per-GPU = 1 (needed for DDP validation) testloader = create_dataloader(test_path, imgsz_test, test_batch_size, gs, opt, hyp=hyp, cache=opt.cache_images and not opt.notest, cache_efficient_sampling=True, drop_last=False, shuffle=False, rect=True, training=False, rank=rank, world_size=opt.world_size, workers=opt.workers, pad=0.5, prefix=colorstr('val: '))[0] # Process 0 new_best_model = False if rank in [-1, 0]: # orig_testloader = create_dataloader(test_path, imgsz_test, test_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, # # lazy_caching=True, # pad=0.5, prefix=colorstr('val: '))[0] if not opt.resume: labels = np.concatenate(dataset.labels, 0) c = torch.tensor(labels[:, 0]) # classes 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 = t1 = 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 ###### resuming training run..... ######################################### if init_epochs > 0: nw = -1 logger.info( 'Stepping lr_scheduler forward {} epochs...'.format(init_epochs)) for i in range(init_epochs - 1): scheduler.step() ## check initial model performance.... # if init_epochs>0 and rank in [-1, 0]: # test.test(opt.data, # batch_size=test_batch_size, # imgsz=imgsz_test, # model=ema.ema, # single_cls=opt.single_cls, # dataloader=orig_testloader, # save_dir=save_dir, # verbose=True, # plots=False, # log_imgs=opt.log_imgs if wandb else 0, # compute_loss=compute_loss) ########################################################################### pfunc(f'Image sizes {imgsz} train, {imgsz_test} test\n' f'Using {trainloader.num_workers} trainloader 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: trainloader.sampler.set_epoch(epoch) pbar = enumerate(trainloader) if rank in [-1, 0]: t1 = time.time() num_img = 0 steps = list(range(100, 0, -2)) pfunc( '==========================================================================================================' ) pfunc(f'Epoch {epoch+1}/{epochs}') optimizer.zero_grad() # logging.StreamHandler.terminator = "" 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 ## simple progress indicator.... if rank in [-1, 0]: prog = int(np.ceil(100 * (i + 1) / nb)) while len(steps) > 0 and prog >= steps[-1]: step = steps.pop() # pfunc('.') if step % 10 == 0: pfunc(f' {step}%') # gpu_stats() # 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]: td = time.time() - t1 num_img += imgs.shape[0] imgs_sec = (num_img / td) * opt.world_size 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) # Plot if plots and ni < 5: 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_graph(torch.jit.trace(de_parallel(model), imgs, strict=False), []) # model graph # # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) 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() # logging.StreamHandler.terminator = "\n" if rank in [-1, 0]: pfunc( (' ' + '%10s' * 3) % ('total_min', 'gpu_mem', 'imgs_sec')) pfunc((' ' + '%10.2f' + '%10s' + '%10.4g') % (((time.time() - t1) / 60), mem, imgs_sec)) t1 = time.time() final_epoch = epoch + 1 == epochs ################################################################################## ## DDP VALIDATION.... # results = (mp, mr, mf1, map50, map)#, *(loss.cpu() / len(dataloader)).tolist()) try: results = test_ddp( opt, de_parallel(model), testloader, rank, device, names, ) if rank in [-1, 0]: pfunc(f'Validation Time: {(time.time()-t1)/60:0.2f} min') # Logging tags = [ 'train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss 'metrics/precision', 'metrics/recall', 'metrics/F1', '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 fitness # fitness = weighted combination of [P, R, F1, [email protected], [email protected]:0.95] fi = fitness(np.array(results).reshape(1, -1)) 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: pfunc('Saving best model!') torch.save(ckpt, best) new_best_model = True best_model_msg = f'Best Model: Epoch {epoch+1}, mF1={results[2]:0.3f}, [email protected]:0.95={results[4]:0.3f}' 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 # Upload best model to s3 if (epoch + 1) % 10 == 0 and epoch > 15: if new_best_model: strip_optimizer(best) upload_model(opt) new_best_model = False # print best model so far pfunc(best_model_msg) # Upload output log to s3 upload_log(opt) ## END DDP VALIDATION except Exception as e: pfunc('Validation failed.') ################################################################ # end epoch ---------------------------------------------------------------------------------------------------- # end training ===================================================================================================== if rank in [-1, 0]: # ## Plots # 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()]}) # ## Test best.pt # pfunc('%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 m in [best] if best.exists() else [last]: # speed, mAP tests # results, _, _ = test.test(data_dict, # batch_size=test_batch_size, # imgsz=imgsz_test, # model=attempt_load(m, device),#.half(), # single_cls=opt.single_cls, # dataloader=orig_testloader, # save_dir=save_dir, # # verbose=nc < 50 and final_epoch, # # plots=plots and final_epoch, # wandb_logger=wandb_logger, # plots=False, # # compute_loss=compute_loss, # ) # ## 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 # if wandb_logger.wandb and not opt.evolve: # Log the stripped model # wandb_logger.wandb.log_artifact(str(final), 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, # path/to/hyp.yaml or hyp dictionary opt, device, ): save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, = \ 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 # Directories save_dir = Path(save_dir) 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' # Hyperparameters if isinstance(hyp, str): with open(hyp) 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 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 evolve # create plots cuda = device.type != 'cpu' init_seeds(1 + RANK) with open(data) as f: data_dict = yaml.safe_load(f) # data dict # Loggers loggers = {'wandb': None, 'tb': None} # loggers dict if RANK in [-1, 0]: # TensorBoard if not evolve: prefix = colorstr('tensorboard: ') LOGGER.info( f"{prefix}Start with 'tensorboard --logdir {opt.project}', view at http://localhost:6006/" ) loggers['tb'] = SummaryWriter(str(save_dir)) # W&B opt.hyp = hyp # add hyperparameters run_id = torch.load(weights).get('wandb_id') if weights.endswith( '.pt') and os.path.isfile(weights) else None run_id = run_id if opt.resume else None # start fresh run if transfer learning wandb_logger = WandbLogger(opt, save_dir.stem, run_id, data_dict) loggers['wandb'] = wandb_logger.wandb if loggers['wandb']: data_dict = wandb_logger.data_dict weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp # may update weights, epochs if resuming 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, '%g names found for nc=%g dataset in %s' % ( len(names), nc, data) # check is_coco = 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(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 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(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'] val_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 / 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']}") 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 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 = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple # DP mode if cuda and RANK == -1 and torch.cuda.device_count() > 1: logging.warning( 'DP not recommended, instead 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: raise Exception( 'can not train with --sync-bn, known issue https://github.com/ultralytics/yolov5/issues/3998' ) 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=opt.cache_images, rect=opt.rect, rank=RANK, workers=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(train_loader) # number of batches assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % ( mlc, nc, data, 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=opt.cache_images and not noval, rect=True, rank=-1, workers=workers, 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, loggers) # 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=[LOCAL_RANK], output_device=LOCAL_RANK) # 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 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) compute_loss = ComputeLoss(model) # init loss class LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n' f'Using {train_loader.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: train_loader.sampler.set_epoch(epoch) pbar = enumerate(train_loader) 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 / 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 *= 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 # 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 loggers['tb'] and ni == 0: # TensorBoard with warnings.catch_warnings(): warnings.simplefilter( 'ignore') # suppress jit trace warning loggers['tb'].add_graph( torch.jit.trace(de_parallel(model), imgs[0:1], strict=False), []) elif plots and ni == 10 and loggers['wandb']: wandb_logger.log({ 'Mosaics': [ loggers['wandb'].Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg') if x.exists() ] }) # end batch ------------------------------------------------------------------------------------------------ # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for loggers 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 noval or final_epoch: # Calculate mAP wandb_logger.current_epoch = epoch + 1 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, 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) # 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 loggers['tb']: loggers['tb'].add_scalar(tag, x, epoch) # TensorBoard if loggers['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 nosave) or (final_epoch and not 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 loggers['wandb'] else None } # Save last, best and delete torch.save(ckpt, last) if best_fitness == fi: torch.save(ckpt, best) if loggers['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 loggers['wandb']: files = [ 'results.png', 'confusion_matrix.png', *[f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R')] ] wandb_logger.log({ "Results": [ loggers['wandb'].Image(str(save_dir / f), caption=f) for f in files if (save_dir / f).exists() ] }) if not evolve: if is_coco: # COCO dataset for m in [last, best ] if best.exists() else [last]: # speed, mAP tests results, _, _ = val.run( data_dict, batch_size=batch_size // WORLD_SIZE * 2, imgsz=imgsz, model=attempt_load(m, device).half(), single_cls=single_cls, dataloader=val_loader, save_dir=save_dir, save_json=True, plots=False) # Strip optimizers for f in last, best: if f.exists(): strip_optimizer(f) # strip optimizers if loggers['wandb']: # Log the stripped model loggers['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() torch.cuda.empty_cache() return results
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), # 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 is_coco = opt.data.endswith("coco.yaml") # 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 # 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 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 # -------------- SageMaker の train と val ------------- # #train_path = data_dict["train"] #test_path = data_dict["val"] data_path = opt.data_dir train_path = data_path + "/train2017.txt" test_path = data_path + "/val2017.txt" # 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 ## ここで dataset 作ってる、 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: ## 変更ポイント3、ラップ 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"] *= 3.0 / nl # scale to layers hyp["cls"] *= nc / 80.0 * 3.0 / nl # scale to classes and layers hyp["obj"] *= (imgsz / 640)**2 * 3.0 / 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 ------------------------------------------------------------------ print(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) print(f'batch: {ni}') 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 *= 4.0 # 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_graph( # torch.jit.trace(model, imgs, strict=False), [] # ) # add model graph # # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) 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 # 変更ポイント5 の dist.get_rank()==0 はここで反映されている 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, 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( model.module if is_parallel(model) else 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]: # Plots 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() ] }) # 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 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 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 if wandb_logger.wandb and not opt.evolve: # Log the stripped model wandb_logger.wandb.log_artifact( str(final), type="model", name="run_" + wandb_logger.wandb_run.id + "_model", aliases=["last", "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): 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 is_coco = opt.data.endswith('coco.yaml') # 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 # modal stage model modal_stage_model = None if opt.modal_stage_model is not None and opt.modal_stage_model != "": modal_ckpt = torch.load(opt.modal_stage_model, map_location=device) modal_stage_model = Model(modal_ckpt['model'].yaml, ch=3, nc=nc).to(device) state_dict = modal_ckpt['model'].float().state_dict() state_dict = intersect_dicts(state_dict, modal_stage_model.state_dict(), exclude=[]) modal_stage_model.load_state_dict(state_dict, strict=False) logger.info('Transferred %g/%g items from %s for modal stage model' % (len(state_dict), len(modal_stage_model.state_dict()), opt.modal_stage_model)) modal_stage_model.eval() # Model if modal_stage_model is not None: input_ch = 3 + nc else: input_ch = 3 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=input_ch, 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=input_ch, 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'] val_paths = dict() if type(data_dict['val']) == str: val_paths[MASTER_VAL_NAME] = data_dict['val'] elif len(data_dict['val']) == 1: val_paths[MASTER_VAL_NAME] = data_dict['val'] else: if MASTER_VAL_NAME not in data_dict['val']: raise ValueError(f"When you use multiple validation sets, one MUST be named '{MASTER_VAL_NAME}'. This is the val set" + f" we will use for early stopping/model selection. Your data yaml file ({opt.data}) does NOT " + "conform to this requirement. Please fix it.") for k, v in data_dict['val'].items(): # in this case, the yaml file has several datasets under the key 'val' val_paths[k] = v # 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: ')) if dataset.single_labelset: mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class else: mlc = np.concatenate([dataset.labels[i]['amodal'] for i in range(len(dataset.labels))], 0)[:, 0].max() # max label class, as taken from *amodal* labels 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]: val_loaders = dict() for k, v in val_paths.items(): val_loaders[k] = create_dataloader(v, imgsz_test, batch_size * 2, gs, opt, 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(f"val/{k}: "))[0] if not opt.resume: if dataset.single_labelset: labels = np.concatenate(dataset.labels, 0) else: # use the class distribution from the amodal labels for this generated histogram labels = np.concatenate([dataset.labels[i]['amodal'] for i in range(len(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 # init tensorboardx tags train_tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss'] lr_tags = ['x/lr0', 'x/lr1', 'x/lr2'] 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: if not dataset.single_labelset: raise NotImplementedError("We don't support image weighting in the modal/amodal label case yet, but it can be added.") # 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 """ In the modal/amodal use case, 'targets' is a dictionary of tensors, so you should choose targets['modal'] or targets['amodal'] depending on your needs. """ # TODO: change this logic when ready! if isinstance(targets, dict): targets = targets['amodal'] if modal_stage_model is not None: with torch.no_grad(): img_shape = (imgs.shape[2], imgs.shape[3]) boxes, _ = modal_stage_model.forward(imgs) pixel_map = predicted_bboxes_to_pixel_map(boxes, img_shape) imgs = torch.cat([imgs, pixel_map], dim=1) # 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) % ( '%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(torch.jit.trace(model, imgs, strict=False), []) # add model 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()]}) # SEND RESULTS TO TBX if ni % nb != 0 and ni % opt.tbx_report_train_every_n_batches == 1: # ni % nb != 0 condition to avoid double-logging anything, ... == 1 condition to log information (almost) right away for x, tag in zip(mloss[:-1], train_tags): tb_writer.add_scalar(tag, x, ni) # init tensorboardx 'lr' list lr = [x['lr'] for x in optimizer.param_groups] # gather learning rates for x, tag in zip(lr, lr_tags): tb_writer.add_scalar(tag, x, ni) # TODO: if we want to update val loss more frequently, do it here # 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 for val_loader_name in val_loaders.keys(): temp_results, temp_maps, temp_times = test.test(data_dict, batch_size=batch_size * 2, imgsz=imgsz_test, model=ema.ema, single_cls=opt.single_cls, dataloader=val_loaders[val_loader_name], save_dir=save_dir, verbose=nc < 50 and final_epoch, plots=plots and final_epoch, wandb_logger=wandb_logger, compute_loss=compute_loss, is_coco=is_coco, modal_stage_model=modal_stage_model) if val_loader_name != MASTER_VAL_NAME: # Log tbx metrics for all non-master validation sets tbx_tags = ['precision', 'recall', 'mAP_0.5', 'mAP_0.5:0.95', 'box_loss', 'obj_loss', 'cls_loss'] for x, tag in zip(list(temp_results), tbx_tags): if tb_writer: tb_writer.add_scalar(f"{val_loader_name}/{tag}", x, nb * (epoch + 1)) if wandb_logger.wandb: wandb_logger.log({f"{val_loader_name}/{tag}": x}) # W&B else: results = temp_results maps = temp_maps times = temp_times # Log all_tbx_tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss # IMPORTANT: when looking at tbx results, the metrics under 'metrics' are calculated on the MASTER VAL SET. 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', f"{MASTER_VAL_NAME}/box_loss", f"{MASTER_VAL_NAME}/obj_loss", f"{MASTER_VAL_NAME}/cls_loss", # val loss 'x/lr0', 'x/lr1', 'x/lr2'] # params for x, tag in zip(list(mloss[:-1]) + list(results) + lr, all_tbx_tags): # TBX UPDATES if tb_writer: tb_writer.add_scalar(tag, x, nb * (epoch + 1)) # tensorboard if wandb_logger.wandb: wandb_logger.log({tag: x}) # W&B # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # append metrics, val_loss if len(opt.name) and opt.bucket: os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) # 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(model.module if is_parallel(model) else 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]: # Plots 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()]}) # 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 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=val_loaders[MASTER_VAL_NAME], save_dir=save_dir, save_json=True, plots=False, is_coco=is_coco, modal_stage_model=modal_stage_model) # 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 if wandb_logger.wandb and not opt.evolve: # Log the stripped model wandb_logger.wandb.log_artifact(str(final), type='model', name='run_' + wandb_logger.wandb_run.id + '_model', aliases=['last', 'best', 'stripped']) wandb_logger.finish_run() else: dist.destroy_process_group() torch.cuda.empty_cache() return results