def train(hyp, opt, device, tb_writer=None): logger.info(f'Hyperparameters {hyp}') """ 获取记录训练日志的路径: 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 """ # 如果设置进化算法则不会传入tb_writer(则为None),设置一个evolve文件夹作为日志目录 log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory # 设置生成文件的保存路径 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') # 获取轮次、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) 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 # 保存hyp和opt 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 # 获取数据路径 cuda = device.type != 'cpu' # 设置随机种子 # 需要在每一个进程设置相同的随机种子,以便所有模型权重都初始化为相同的值,即确保神经网络每次初始化都相同 init_seeds(2 + rank) # 加载数据配置信息 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 # 获取训练集、测试集图片路径 train_path = data_dict['train'] test_path = data_dict['val'] # 获取类别数量和类别名字 # 如果设置了opt.single_cls则为一类 nc, names = (1, ['item']) if opt.single_cls else ( int(data_dict['nc']), data_dict['names']) # 保存data.yaml中的number classes, names assert len(names) == nc, '%g names found for nc=%g dataset in %s' % ( len(names), nc, opt.data) # check # Model # 判断weights字符串是否以'.pt'为结尾。若是,则说明本次训练需要预训练模型 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 ckpt = torch.load(weights, map_location=device) # load checkpoint 导入权重文件 """ 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,则加载权重中保存的anchor来继续训练; 主要是预训练权重里面保存了默认coco数据集对应的anchor, 如果用户自定义了anchor,再加载预训练权重进行训练,会覆盖掉用户自定义的anchor; 所以这里主要是设定一个,如果加载预训练权重进行训练的话,就去除掉权重中的anchor,采用用户自定义的; 如果是resume的话,就是不去除anchor,就权重和anchor一起加载, 接着训练; 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中的键对应的值 """ ''' ckpt: {'epoch': -1, 'best_fitness': array([ 0.49124]), 'training_results': None, 'model': Model( ... ) 'optimizer': None } ''' if hyp.get('anchors'): # 用户自定义的anchors优先级大于权重文件中自带的anchors ckpt['model'].yaml['anchors'] = round( hyp['anchors']) # force autoanchor # 创建并初始化yolo模型 model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create ''' model = Model( (model): Sequential( (0): Focus(...) ... (24): Detect(...) ) ) ''' # 如果opt.cfg存在,或重新设置了'anchors',则将预训练权重文件中的'anchors'参数清除,使用用户自定义的‘anchors’信息 exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [ ] # exclude keys # state_dict变量存放训练过程中需要学习的权重和偏执系数,state_dict 是一个python的字典格式,以字典的格式存储,然后以字典的格式被加载,而且只加载key匹配的项 # 将ckpt中的‘model’中的”可训练“的每一层的参数建立映射关系(如 'conv1.weight': 数值...)存在state_dict中 state_dict = ckpt['model'].float().state_dict() # to FP32 # 加载除了与exclude以外,所有与key匹配的项的参数 即将权重文件中的参数导入对应层中 state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect # 将最终模型参数导入yolo模型 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: # 不进行预训练,则直接创建并初始化yolo模型 model = Model(opt.cfg, ch=3, nc=nc).to(device) # create # Freeze #freeze = ['', ] # parameter names to freeze (full or partial) freeze = ['model.%s.' % x for x in range(10) ] # 冻结带有'model.0.'-'model.9.'的所有参数 即冻结0-9层的backbone 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人为模拟的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 pg0, pg1, pg2 = [], [], [] # optimizer parameter groups # 将模型分成三组(w权重参数(非bn层), bias, 其他所有参数)优化 for k, v in model.named_parameters(): # named_parameters:网络层的名字和参数的迭代器 ''' (0): Focus( (conv): Conv( (conv): Conv2d(12, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn): BatchNorm2d(80, eps=0.001, momentum=0.03, affine=True, track_running_stats=True) (act): Hardswish() ) ) k: 网络层可训练参数的名字所属 如: model.0.conv.conv.weight 或 model.0.conv.bn.weight 或 model.0.conv.bn.bias (Focus层举例) v: 对应网络层的具体参数 如:对应model.0.conv.conv.weight的 size为(80,12,3,3)的参数数据 即 卷积核的数量为80,深度为12,size为3×3 ''' v.requires_grad = True # 设置当前参数在训练时保留梯度信息 if '.bias' in k: pg2.append(v) # biases (所有的偏置参数) elif '.weight' in k and '.bn' not in k: pg1.append(v) # apply weight decay (非bn层的权重参数w) else: pg0.append(v) # all else (网络层的其他参数) # 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) # 设置权重参数weights(非bn层)的优化方式 optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['weight_decay'] }) # add pg1 with weight_decay # 设置偏置参数bias的优化方式 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 # 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 匿名余弦退火函数 scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Resume # 初始化开始训练的epoch和最好的结果 # 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'] # 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 # 加载上次断点模型中训练的轮次,并在此基础上继续训练 start_epoch = ckpt['epoch'] + 1 # 如果使用断点重训的同时发现 start_epoch= 0,则说明上次训练正常结束,不存在断点 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 # 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 # Image sizes # 获取模型总步长和模型输入图片分辨率 gs = int(max(model.stride)) # grid size (max stride) # 检查输入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,不支持混合精度训练 # rank为进程编号, 这里应该设置为rank=-1则使用DataParallel模式 # 如果 当前运行设备为gpu 且 进程编号=-1 且gpu数量大于1时 才会进行分布式训练 ,将model对象放入DataParallel容器即可进行分布式训练 if cuda and rank == -1 and torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # SyncBatchNorm # 实现多GPU之间的BatchNorm 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 : YOLOv5优化策略之一 EMA + SGD可提高模型鲁棒性 为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建 ''' ema = ModelEMA(model) if rank in [-1, 0] else None # 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 # class dataloader 和 dataset . 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) # 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) ''' dataloader和testloader不同之处在于: 1. testloader:没有数据增强,rect=True(大概是测试图片保留了原图的长宽比) 2. dataloader:数据增强,保留了矩形框训练。 ''' # Process 0 if rank in [-1, 0]: # local_rank is set to -1. Because only the first process is expected to do evaluation. # testloader 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 # 根据自己数据集的类别数设置分类损失的系数 hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset # 设置类别数,超参数 model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model """ 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 """ 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(图像类别所占比例高的采样频率低) model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights # 获取类别的名字 model.names = names # Start training t0 = time.time() # 获取warm-up训练的迭代次数 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 # 初始化mAP和results maps = np.zeros(nc) # mAP per class results = ( 0, 0, 0, 0, 0, 0, 0, 0 ) # P, R, [email protected], [email protected], val_loss(box, obj, cls, angleloss) """ 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( '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)) # 训练 for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ # model设置为训练模式,其中training属性表示BatchNorm与Dropout层在训练阶段和测试阶段中采取的策略不同,通过判断training值来决定前向传播策略 model.train() # Update image weights (optional) # 加载图片权重(可选) 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 # 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 # 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(5, 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' * 9) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'angle', 'total', 'targets', 'img_size')) if rank in [-1, 0]: # tqdm 创建进度条,方便训练时 信息的展示 pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------ ''' i: batch_index, 第i个batch imgs : torch.Size([batch_size, 3, resized_height, resized_weight]) targets : torch.Size = (该batch中的目标数量, [该image属于该batch的第几个图片, class, xywh, θ]) paths : List['img1_path','img2_path',......,'img-1_path'] len(paths)=batch_size shapes : size= batch_size, 不进行mosaic时进行矩形训练时才有值 ''' # ni计算迭代的次数iteration 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 """ warmup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 """ # 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 # 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) # 采用上采样下采样函数interpolate完成imgs尺寸的转变,模式设置为双线性插值 imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) # Forward # 前向传播 with amp.autocast(enabled=cuda): ''' 训练时返回x x list: [small_forward, medium_forward, large_forward] eg:small_forward.size=( batch_size, 3种scale框, size1, size2, no) ''' pred = model(imgs) # forward # Loss # 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(lbox, lobj, lcls, langle, loss) loss, loss_items = compute_loss( pred, targets.to(device), model, csl_label_flag=True) # 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 # 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: ema.update(model) # Print if rank in [-1, 0]: # mloss (lbox, lobj, lcls, langle, loss) # 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' * 7) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) # 进度条显示以上信息 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 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) # 存储的格式为[H, W, C] # 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: # 更新EMA的属性 # 添加include的属性 ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) final_epoch = epoch + 1 == epochs # # 判断该epoch是否为最后一轮 # if not opt.notest or final_epoch: # Calculate mAP # # 对测试集进行测试,计算mAP等指标 # # 测试时使用的是EMA模型 # 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 # 将测试指标写入result.txt with open(results_file, 'a') as f: f.write( s + '%10.4g' * 8 % 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 # 添加指标,损失等信息到tensorboard显示 if tb_writer: tags = [ 'train/box_loss', 'train/obj_loss', 'train/cls_loss', 'train/angle_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/angle_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 # 更新best_fitness 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信息在最后一轮完成后不会进行保存 未完成训练则保存该信息 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() } # 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 """ 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 # 上传结果到谷歌云盘 os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish # 可视化results.txt文件 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)) # 释放显存 dist.destroy_process_group() if rank not in [-1, 0] else None torch.cuda.empty_cache() return results
def train(opt, device): epochs, batch_size, noval, nosave, workers, freeze, = \ opt.epochs, opt.batch_size, opt.noval, opt.nosave, opt.workers, opt.freeze d = datetime.datetime.now() run_id = '{:%Y-%m-%d__%H-%M-%S}'.format(d) save_dir = Path(opt.save_dir) / run_id # Directories w = save_dir / 'weights' # weights dir w.mkdir(parents=True, exist_ok=True) # make dir last, best = w / 'last.pt', w / 'best.pt' # Get hyperparameter dict hyp, hyp_loss = get_hyperparameter_dict(opt.dataset_name, opt.hp_config) # 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) tb_writer = SummaryWriter(save_dir) opt.img_dir = Path(opt.img_dir) # Config cuda = device.type != 'cpu' init_seeds(1 + RANK) # Dataloaders dataset_kwargs = {} if opt.train_img_res: dataset_kwargs = {'img_size': opt.train_img_res} dataset_splits = get_data_splits_by_name( data_root=opt.img_dir, dataset_name=opt.dataset_name, model_name=opt.model_name, batch_size=batch_size, num_workers=workers, distributed=(cuda and RANK != -1), **dataset_kwargs ) test_img_size = dataset_splits["test"].dataset._img_size train_img_size = dataset_splits["train"].dataset._img_size if opt.test_img_res: test_img_size = opt.test_img_res train_loader = dataset_splits["train"] dataset = train_loader.dataset nc = dataset.num_classes nb = len(train_loader) # number of batches # Model model = create_model( model_name=opt.model_name, pretraining_dataset=opt.pretraining_source_dataset, pretrained=opt.pretrained, num_classes=nc, progress=True, device=device, ) # Freeze freeze = [f'model.{x}.' for x in range(freeze)] # 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): print(f'freezing {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']}") 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.adam: optimizer = Adam(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, {len(g1)} weight (no decay), {len(g2)} bias") del g0, g1, g2 # Scheduler 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 start_epoch, best_fitness = 0, 0.0 # 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']) # 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: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) LOGGER.info('Using SyncBatchNorm()') # Process 0 if RANK in [-1, 0]: # Anchors 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['giou'] *= 3. / nl # scale to layers hyp['box'] = hyp['giou'] hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers hyp['obj'] *= (train_img_size / 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 eval_function = get_eval_function(dataset_name=opt.dataset_name, model_name=opt.model_name) criterion = YoloV5Loss( model=model, num_classes=nc, device=device, hyp_cfg=hyp_loss, ) if opt.eval_before_train: ap_dict = evaluate(model, eval_function, opt.dataset_name, opt.img_dir, nc, test_img_size, device) LOGGER.info(f'Eval metrics: {ap_dict}') # Start training t0 = time.time() nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) last_opt_step = -1 scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) stopper = EarlyStopping(patience=opt.patience) loss_giou_mean = AverageMeter() loss_conf_mean = AverageMeter() loss_cls_mean = AverageMeter() loss_mean = AverageMeter() LOGGER.info(f'Image sizes {train_img_size} train, {test_img_size} val\n' f'Using {train_loader.num_workers} 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() 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) # progress bar optimizer.zero_grad() for i, (imgs, targets, labels_length, _) in pbar: # batch ni = i + nb * epoch # number integrated batches (since train start) imgs = imgs.to(device, non_blocking=True).float() # Warmup if ni <= nw: xi = [0, nw] # x interp 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(train_img_size * 0.5, train_img_size * 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_giou, loss_conf, loss_cls = criterion( pred, targets, labels_length, imgs.shape[-1] ) # Update running mean of tracked metrics loss_items = torch.tensor([loss_giou, loss_conf, loss_cls]).to(device) if RANK in (-1, 0): loss_giou_mean.update(loss_giou, imgs.size(0)) loss_conf_mean.update(loss_conf, imgs.size(0)) loss_cls_mean.update(loss_cls, imgs.size(0)) loss_mean.update(loss, imgs.size(0)) if RANK != -1: loss *= WORLD_SIZE # gradient averaged between devices in DDP mode # 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])) # end batch # Scheduler scheduler.step() if RANK in [-1, 0]: for idx, param_group in enumerate(optimizer.param_groups): tb_writer.add_scalar(f'learning_rate/gr{idx}', param_group['lr'], epoch) tb_writer.add_scalar('train/giou_loss', loss_giou_mean.avg, epoch) tb_writer.add_scalar('train/conf_loss', loss_conf_mean.avg, epoch) tb_writer.add_scalar('train/cls_loss', loss_cls_mean.avg, epoch) tb_writer.add_scalar('train/loss', loss_mean.avg, epoch) # mAP 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) and epoch % opt.eval_freq == 0: # Calculate mAP ap_dict = evaluate(ema.ema, eval_function, opt.dataset_name, opt.img_dir, nc, test_img_size, device) LOGGER.info(f'Eval metrics: {ap_dict}') tb_writer.add_scalar('eval/mAP', ap_dict['mAP'], epoch) for eval_key, eval_value in ap_dict.items(): if eval_key != 'mAP': tb_writer.add_scalar(f'ap_per_class/{eval_key}', eval_value, epoch) # Update best mAP fi = ap_dict['mAP'] if fi > best_fitness: best_fitness = fi # Save model if (not nosave) or final_epoch: # 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()} # 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 # Stop Single-GPU if RANK == -1 and stopper(epoch=epoch, fitness=fi): break # 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}...') ckpt = torch.load(f, map_location=device) model = ckpt['ema' if ckpt.get('ema') else 'model'] model.float().eval() ap_dict = evaluate(model, eval_function, opt.dataset_name, opt.img_dir, nc, test_img_size, device) LOGGER.info(f'Eval metrics: {ap_dict}') LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}") torch.cuda.empty_cache()
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(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(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) # 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 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 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 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) # Resume start_epoch, best_fitness = 0, 0.0 # 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()') # 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) # 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 plot_labels(labels, save_dir=log_dir) 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 # 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 # 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', '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\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() 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) 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 accumulate = max( 1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) for j, x in enumerate(optimizer.param_groups): 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 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 = 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) # 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 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, loss_name = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank, opt.loss if loss_name == 'ciou': loss_fn = compute_loss_ciou elif loss_name == 'giou': loss_fn = compute_loss_giou elif loss_name == 'gioupp': loss_fn = compute_loss_gioupp # 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]: if loss_items: 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 if not opt.notest or final_epoch: # Calculate mAP 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, compute_loss=loss_fn) # 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: if loss_items: 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) else: tags = [ "train/" + loss_name + "_loss", 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', "val/" + loss_name + "_loss" ] for x, tag in zip(list(mloss) + 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.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): 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): logger.info(f'Hyperparameters {hyp}') # 获取记录训练日志的路径 # 如果设置进化算法则不会传入tb_writer(则为None),设置一个evolve文件夹作为日志目录 log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory # 设置保存权重的路径 wdir = log_dir / 'weights' # weights directory os.makedirs(wdir, exist_ok=True) last = wdir / 'last.pt' best = wdir / 'best.pt' # 设置保存results的路径 results_file = str(log_dir / 'results.txt') # 获取轮次、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) epochs, batch_size, total_batch_size, weights, rank = \ opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank # rank = -1 # Save run settings # 保存hyp和opt 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') # 设置随机种子 init_seeds(2 + rank) with open(opt.data) as f: # 加载数据配置信息 data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict with torch_distributed_zero_first( rank): # torch_distributed_zero_first同步所有进程 check_dataset( data_dict ) # 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) # 获取训练集、测试集图片路径 train_path = data_dict['train'] test_path = data_dict['val'] # 获取类别数量和类别名字, 如果设置了opt.single_cls则为一类 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: # 如果采用预训练 # 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: # 创建模型, ch为输入图片通道 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) if any(freeze): for k, v in model.named_parameters(): # print(k,v) 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 accumulate = 4 # 根据accumulate设置权重衰减系数 hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale 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_parameters(): # print(k) 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 # 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 scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) # Resume # 初始化开始训练的epoch和最好的结果 # 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'] # 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 # print(ckpt['epoch']) start_epoch = ckpt['epoch'] + 1 # 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) shutil.copytree(wdir, wdir.parent / f'weights_backup_epoch{start_epoch - 1}' ) # save previous weights """ 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 # Image sizes # 获取模型总步长和模型输入图片分辨率 gs = int(max(model.stride)) # grid size (max stride) # 检查输入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 # imgsz, imgsz_test 都是640 # 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) # 执行了 # 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()') # Exponential moving average 指数滑动平均,或指数加权平均 # 为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建 ema = ModelEMA(model) if rank in [-1, 0] else None # 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 # 创建训练集dataloader 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) """ 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 # 创建测试集dataloader 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: # 将所有样本的标签拼接到一起shape为(total, 5),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)) # 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=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 """ 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['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset # 设置类别数,超参数 model.nc = nc # attach number of classes to model model.hyp = hyp # attach hyperparameters to model """ 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 """ model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou) # 根据labels初始化图片采样权重 model.class_weights = labels_to_class_weights(dataset.labels, nc).to( device) # attach class weights # 获取类别的名字 model.names = names # Start training t0 = time.time() # 获取热身训练的迭代次数 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 # 初始化mAP和results 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' """ 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( '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 """ 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 # 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 # 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() # 广播索引到其他group 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', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size')) if rank in [-1, 0]: # tqdm 创建进度条,方便训练时 信息的展示 pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, ( imgs, targets, paths, _ ) in pbar: # batch ------------------------------------------------------------- # 计算迭代的次数iteration 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]) # 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 """ 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) ]) # 动量momentum也从0.9慢慢变到hyp['momentum'](default=0.937) 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) # Forward # 混合精度 with amp.autocast(enabled=cuda): pred = model(imgs) # forward 前向传播 # Loss # 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 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 # Backward # 反向传播 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: 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]) # 进度条显示以上信息 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 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 # 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 if ema: # 更新EMA的属性 # 添加include的属性 ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) # 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 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 = 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) # Write # 将指标写入result.txt with open(results_file, 'a') as f: f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls) # 如果设置opt.bucket, 上传results.txt到谷歌云盘 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 # 添加指标,损失等信息到tensorboard显示 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 # 更新best_fitness 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 """ 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() } # 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 """ 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 # 上传结果到谷歌云盘 os.system( 'gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket else None # upload # Finish # 可视化results.txt文件 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)) # 释放显存 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, ): 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(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, 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.dump(hyp, f, sort_keys=False) with open(save_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.SafeLoader) # 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, 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 # Model pretrained = weights.endswith('.pt') model = SR_Model(opt.cfg, ch=3).to(device) # Freeze freeze = [ ] # parameter names to freeze (full or partial) 'model.%s.' % x for x in range(8) 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']}") optimizer = optim.Adam(model.parameters(), lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) scheduler = lr_scheduler.StepLR(optimizer, step_size=hyp['lr_decay'], gamma=hyp['gamma']) # EMA ema = ModelEMA(model) if rank in [-1, 0] else None # Resume start_epoch, best_fitness = 0, 0.0 # 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 # DIV2k dataset dataloader, dataset = create_SRdataloader(opt, train=True, batch_size=opt.batch_size, rank=rank, world_size=opt.world_size, workers=opt.workers) nb = len(dataloader) scaler = amp.GradScaler(enabled=cuda) scheduler.last_epoch = start_epoch - 1 # do not move sr_loss = SR_Loss(opt, device) testloader, _ = create_SRdataloader(opt, train=False, batch_size=1, rank=rank, world_size=opt.world_size, workers=opt.workers) for epoch in range( start_epoch, epochs ): # epoch ------------------------------------------------------------------ model.train() mloss = torch.zeros(1, device=device) if rank != -1: dataloader.sampler.set_epoch(epoch) pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 4) % ('Epoch', 'gpu_mem', 'loss', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() for i, (lr, hr, _) in pbar: ni = i + nb * epoch # number integrated batches (since train start) idx_scale = opt.scale lr = lr.to(device).float() hr = hr.to(device).float() # Forward with amp.autocast(enabled=cuda): pred = model(lr) # forward loss = sr_loss(pred, hr) # 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) / (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' * 2) % ('%g/%g' % (epoch, epochs - 1), mem, mloss, lr.shape[-1]) pbar.set_description(s) # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler learning_rate = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # DDP process 0 or single-GPU if rank in [-1, 0]: # PSNR ema.update_attr( model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'class_weights']) final_epoch = epoch + 1 == epochs model.eval() with torch.no_grad(): for idx_scale, scale in enumerate(opt.scale): eval_acc = 0 #testloader.dataset.set_scale(idx_scale) pbar = enumerate(testloader) pbar = tqdm(pbar, total=len(testloader)) for idx_img, (lr, hr, filename) in pbar: lr = lr.to(device).float() hr = hr.to(device).float() filename = filename[0] pred = model(lr, idx_scale) pred = quantize(pred, opt.rgb_range) save_list = [pred] eval_acc += calc_psnr(pred, hr, scale, opt.rgb_range) save_list.extend([lr, hr]) # PSNR 로그로 표시 results = eval_acc / len(testloader) logger.info(f'[DIV2K x{opt.scale}]\tPSNR: {results}') # Write with open(results_file, 'a') as f: f.write(s + '%10.4g' * 1 % (results) + '\n') # append metrics, val_loss # Update best PSNR 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 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
def train(hyp, opt, device): 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 is_coco = opt.data.endswith('coco.yaml') # Logging- Doing this before checking the dataset. Might update data_dict opt.hyp = hyp # add hyperparameters nc = int(data_dict['nc']) # number of classes names = 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: 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 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 loss before optimizing accumulate = max(round(nbs / total_batch_size), 1) 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) # add pg1 with weight_decay optimizer.add_param_group({ 'params': pg1, 'weight_decay': hyp['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: def lf(x): return (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 == -1 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) # number of detection layers (used for scaling hyp['obj']) nl = model.model[-1].nl # verify imgsz are gs-multiples imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size] # Trainloader dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, 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 == -1: 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, 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 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 # Model parameters hyp['box'] *= 3. / nl # scale to layers hyp['cls'] *= nc / 80. * 3. / nl # scale to classes and layers # scale to image size and layers hyp['obj'] *= (imgsz / 640)**2 * 3. / nl 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() # number of warmup iterations, max(3 epochs, 1k iterations) nw = max(round(hyp['warmup_epochs'] * nb), 1000) # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training maps = np.zeros(nc) # mAP per class # P, R, [email protected], [email protected], val_loss(box, obj, cls) results = (0, 0, 0, 0, 0, 0, 0) 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...') # epoch ------------------------------------------------------------------ for epoch in range(start_epoch, epochs): model.train() # Update image weights (optional) if opt.image_weights: # Generate indices 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 mloss = torch.zeros(4, device=device) # mean losses pbar = enumerate(dataloader) logger.info( ('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'labels', 'img_size')) if rank == -1: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() # batch ------------------------------------------------------------- for i, (imgs, targets, paths, _) in pbar: # number integrated batches (since train start) ni = i + nb * epoch 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: # new shape (stretched to gs-multiple) ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] 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 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 == -1: 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() # end batch ------------------------------------------------------------------------------------------------ # end epoch ---------------------------------------------------------------------------------------------------- # Scheduler lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard scheduler.step() # single-GPU if rank == -1: # 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, dataloader=testloader, save_dir=save_dir, verbose=nc < 50 and final_epoch, plots=plots and final_epoch, compute_loss=compute_loss, is_coco=is_coco) # Write with open(results_file, 'a') as f: # append metrics, val_loss f.write(s + '%10.4g' * 7 % results + '\n') # 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 # Update best mAP # weighted combination of [P, R, [email protected], [email protected]] fi = fitness(np.array(results).reshape(1, -1)) if fi > best_fitness: 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).half(), 'ema': deepcopy(ema.ema).half(), 'updates': ema.updates, '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 if rank == -1: # Plots if plots: plot_results(save_dir=save_dir) # save as results.png # 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(), 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 else: dist.destroy_process_group() torch.cuda.empty_cache() return results
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 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,超参数,训练para 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): # torch_distributed_zero_first同步所有进程 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 # 所以这里主要是设定一个,如果加载预训练权重进行训练的话,就去除掉权重中的anchor,采用用户自定义的; # 如果是resume的话,就是不去除anchor,就权重和anchor一起加载, 接着训练; 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 resume时将opt.cfg设为空 exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [ ] # exclude keys 如果resume,则加载权重中保存的anchor来继续训练; 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)) # 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) 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 #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 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 DataParallel模式,仅支持单机多卡 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()') # Exponential moving average 为模型创建EMA指数滑动平均,如果GPU进程数大于1,则不创建 ema = ModelEMA(model) if rank in [-1, 0] else None # 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) 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) """ 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 """ # 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, 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)) # 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 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) scheduler.last_epoch = start_epoch - 1 # do not move scaler = amp.GradScaler(enabled=cuda) # 通过torch1.6自带的api设置混合精度训练 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) # 广播索引到其他group 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) # 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 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 """ 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 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) # 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 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() # 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: 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) if best_fitness == fi: torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training # 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, 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 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 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 = 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 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 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 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 # 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 # TODO 将cfg添加到配置变量中 cfg_model = Darknet('cfg/yolov5s_v4.cfg', (opt.img_size[0], opt.img_size[0])).to(device) # cfg_model = Darknet('cfg/yolov5s_v3.cfg', (416, 416)).to(device) copy_weight_v4(model, cfg_model) # 剪枝操作 sr开启稀疏训练 prune 不同的剪枝策略 # 剪枝操作 if opt.prune == 1: CBL_idx, _, prune_idx, shortcut_idx, _ = parse_module_defs2( cfg_model.module_defs) if opt.sr: print('shortcut sparse training') elif opt.prune == 0: CBL_idx, _, prune_idx = parse_module_defs(cfg_model.module_defs) if opt.sr: print('normal sparse training ') # 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, quad=opt.quad) 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: plot_labels(labels, 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 parameters hyp['cls'] *= nc / 80. # scale hyp['cls'] to class count hyp['obj'] *= imgsz**2 / 640.**2 * 3. / nl # scale hyp['obj'] to image size and output 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) model.class_weights = labels_to_class_weights( dataset.labels, nc).to(device) * nc # attach class weights model.names = names for idx in prune_idx: bn_weights = gather_bn_weights(cfg_model.module_list, [idx]) tb_writer.add_histogram('before_train_perlayer_bn_weights/hist', bn_weights.numpy(), idx, bins='doane') # 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 / 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', 'targets', 'img_size')) if rank in [-1, 0]: pbar = tqdm(pbar, total=nb) # progress bar optimizer.zero_grad() sr_flag = get_sr_flag(epoch, opt.sr) 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 # if opt.quad: # loss *= 4. # 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() loss.backward() idx2mask = None # if opt.sr and opt.prune==1 and epoch > opt.epochs * 0.5: # idx2mask = get_mask2(model, prune_idx, 0.85) # copy_weight(model,cfg_model) BNOptimizer.updateBN(sr_flag, cfg_model.module_list, opt.s, prune_idx, epoch, idx2mask, opt) # 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 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', 'class_weights' ]) 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 #剪枝后bn层权重 bn_weights = gather_bn_weights(cfg_model.module_list, prune_idx) tb_writer.add_histogram('bn_weights/hist', bn_weights.numpy(), epoch, bins='doane') # 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 for idx in prune_idx: bn_weights = gather_bn_weights(cfg_model.module_list, [idx]) tb_writer.add_histogram('after_train_perlayer_bn_weights/hist', bn_weights.numpy(), idx, bins='doane') 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', '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() wandb.run.finish() if wandb and wandb.run else None torch.cuda.empty_cache() return results
class YoloTrainer: def __init__(self, model, hyp, opt, nb, nc): self.nb = nb self.nc = nc self.n_burn = max(3 * nb, 500) self.optimizer = None self.model = model self.batch_size = opt.batch_size self.accumulate = max(round(64 / opt.batch_size), 1) self.gs = 32 # (pixels) grid size imgsz_min, imgsz_max, imgsz_test = opt.img_size # img sizes (min, max, test) opt.multi_scale |= imgsz_min != imgsz_max if opt.multi_scale: if imgsz_min == imgsz_max: imgsz_min //= 1.5 imgsz_max //= 0.667 self.grid_min, self.grid_max = imgsz_min // self.gs, imgsz_max // self.gs imgsz_min, imgsz_max = int(self.grid_min *self.gs), int(self.grid_max * self.gs) self.img_size = imgsz_max hyp['cls'] *= nc / 80 # update coco-tuned hyp['cls'] to current dataset self.ema = ModelEMA(model) self.hyp = hyp self.opt = opt # Scheduler https://arxiv.org/pdf/1812.01187.pdf self.lf = lambda x: (((1 + math.cos(x * math.pi / opt.epochs)) / 2) ** 1.0) * 0.95 + 0.05 # cosine assert math.fmod(imgsz_min, self.gs) == 0, '--img-size %g must be a %g-multiple' % (imgsz_min, self.gs) # validation stuff self.stats = [] self.seen = 0 iouv = torch.linspace(0.5, 0.95, 10).to("cuda" if torch.cuda.is_available() else "cpu") # iou vector for [email protected]:0.95 self.iouv = iouv[0].view(1) # comment for [email protected]:0.95 self.niou = iouv.numel() def pre_train_step(self, batch, batch_idx, epoch): if self.optimizer is None: raise Exception("Set optimizer for the yolo trainer.") imgs, _, _, _, _ = batch ni = self.calc_ni(batch_idx, epoch) # number integrated batches (since train start) if imgs.dtype == torch.uint8: imgs = imgs.float() / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0 # Burn-in if ni <= self.n_burn: xi = [0, self.n_burn] # x interp self.opt.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou) self.accumulate = max(1, np.interp(ni, xi, [1, 64 / self.batch_size]).round()) for j, x in enumerate(self.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'] * self.lf(epoch)]) x['weight_decay'] = np.interp(ni, xi, [0.0, self.hyp['weight_decay'] if j == 1 else 0.0]) if 'momentum' in x: x['momentum'] = np.interp(ni, xi, [0.9, self.hyp['momentum']]) # Multi-Scale if self.opt.multi_scale: if ni / self.accumulate % 1 == 0: # adjust img_size (67% - 150%) every 1 batch self.img_size = random.randrange(self.grid_min, self.grid_max + 1) * self.gs sf = self.img_size / max(imgs.shape[2:]) # scale factor if sf != 1: ns = [math.ceil(x * sf / self.gs) * self.gs for x in imgs.shape[2:]] # new shape (stretched to 32-multiple) imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) return imgs def post_train_step(self, outputs, batch, batch_idx, epoch): _, targets, _, _, _ = batch # Loss loss, loss_items = compute_loss(outputs, targets, self.model) if not torch.isfinite(loss): print('WARNING: non-finite loss, ending training ', loss_items) exit(-1) loss *= self.batch_size / 64 # scale loss if self.calc_ni(batch_idx, epoch) % self.accumulate == 0: self.ema.update(self.model) return loss, loss_items def train_epoch_end(self): self.ema.update_attr(self.model) def validation_epoch_start(self): self.seen = 0 self.stats = [] def validation_step(self, opt, outputs, batch, batch_idx, epoch): imgs, targets, paths, shapes, pad = batch _, _, height, width = imgs.shape inf_out, train_out = outputs whwh = torch.Tensor([width, height, width, height]).to(imgs.device) losses = compute_loss(train_out, targets, self.model)[1][:3] # GIoU, obj, cls output = non_max_suppression(inf_out, conf_thres=opt.conf_thres, iou_thres=opt.iou_thres, multi_label=self.calc_ni(batch_idx, epoch) > self.n_burn) # 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 self.seen += 1 if pred is None: if nl: self.stats.append((torch.zeros(0, self.niou, dtype=torch.bool), 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] # Clip boxes to image bounds clip_coords(pred, (height, width)) # Assign all predictions as incorrect correct = torch.zeros(pred.shape[0], self.niou, dtype=torch.bool, device=imgs.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().view(-1) # target indices pi = (cls == pred[:, 5]).nonzero().view(-1) # prediction 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 > self.iouv[0].to(ious.device)).nonzero(): d = ti[i[j]] # detected target if d not in detected: detected.append(d) correct[pi[j]] = ious[j] > self.iouv # iou_thres is 1xn if len(detected) == nl: # all targets already located in image break # Append statistics (correct, conf, pcls, tcls) self.stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) return losses def validation_epoch_end(self): stats = [np.concatenate(x, 0) for x in zip(*self.stats)] # to numpy p, r, ap, f1, ap_class = ap_per_class(*stats) if self.niou > 1: p, r, ap, f1 = p[:, 0], r[:, 0], ap.mean(1), ap[:, 0] # [P, R, [email protected]:0.95, [email protected]] mp, mr, map, mf1 = p.mean(), r.mean(), ap.mean(), f1.mean() nt = np.bincount(stats[3].astype(np.int64), minlength=self.nc) # number of targets per class maps = np.zeros(self.nc) + map for i, c in enumerate(ap_class): maps[c] = ap[i] # Return results return (mp, mr, map, mf1), maps def calc_ni(self, batch_idx, epoch): return batch_idx + self.nb * epoch def set_optimizer(self, optimizer): self.optimizer = optimizer
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), 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) 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, 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, wandb=None): logger.info(f'Hyperparameters {hyp}') log_dir = Path(tb_writer.log_dir) if tb_writer else Path( opt.logdir) / 'evolve' # logging directory log_dir = Path("/project/train/log") wdir = Path("/project/train/models") # weights directory os.makedirs(wdir / 'last', exist_ok=True) os.makedirs(wdir / 'final', exist_ok=True) last = wdir / 'last' / 'last.pt' best = wdir / 'final' / 'best.pt' results_file = str(log_dir / 'log.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) # data dict with torch_distributed_zero_first(rank): print(data_dict) 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: 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 # 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()') # 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) # 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]: if i % 500 == 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 / f'train_batch{ni}.jpg') # 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 log_imgs=opt.log_imgs) # 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/log%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 # logger.info(results) 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 } ckpt = {'model': ema.ema} # 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'log{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 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 log_dir = Path('/home/data/yolov5_pt', log_dir) 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) # data dict print(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 # load MNN Model model_file = './weights/20201231_exp25_599_800_forT_bs128_320x320.mnn' net = nn.load_module_from_file(model_file, for_training=True) nn.compress.train_quant(net, quant_bits=8) mnn_opt = MNN.optim.SGD(1e-9, 0.9, 0) mnn_opt.append(net.parameters) net.train(True) # 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 - 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) 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()') # 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 t0 = time.time() 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) t1 = time.time() print('dataloader is created in {}s'.format(t1 - t0)) 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, 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 / 104. # 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 # 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\nUsing %g dataloader workers\nLogging results to %s\n' 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, log_dir, epochs)) # add by zlf on 20201113 # write down all the results # save_path = '/home/data/yolov5_pt/new_val_work/val_result/20201124_finetune_800/' # if not os.path.exists(save_path): # os.makedirs(save_path) # count_f = open(save_path+'result_count.txt','w') # add by zlf on 20201113 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) 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 with amp.autocast(enabled=cuda): # data_image1 = imgs[0].cpu() # images_t3 = tf.ToPILImage()(data_image1) # images_t3.show() # for n,p in model.named_parameters(): # print(p.dtype) # pred = model(imgs) # 20201231 by zlf # MNN forward data = MNNF.const(imgs.flatten().tolist(), [128, 3, 320, 320], MNNF.data_format.NCHW) predict = net.forward(data) predict.read() p1 = MNNF.Var.read(predict) print(predict.shape) print(predict.size) print(predict.dtype) p1 = torch.tensor(p1).cuda() x1, x2, x3 = torch.split(p1, [4800, 1200, 300], 1) x1 = x1.view(-1, 3, 109, 40, 40).permute(0, 1, 3, 4, 2).contiguous() x2 = x2.view(-1, 3, 109, 20, 20).permute(0, 1, 3, 4, 2).contiguous() x3 = x3.view(-1, 3, 109, 10, 10).permute(0, 1, 3, 4, 2).contiguous() x = [x1, x2, x3] # by zlf 20201016 now,pred is (bs,327,6,10), # but complute loss needs (bs,3,6,10,109) # for idx in range(3): # bs, _, ny, nx = pred[idx].shape # pred[idx] = pred[idx].view(bs, 3, 109, ny, nx).permute(0, 1, 3, 4, 2).contiguous() # 20201016 by zlf # loss, loss_items = compute_loss(pred, targets.to(device), model) loss1, loss_items1 = compute_loss(x, targets.to(device), model) loss1 = np.array(loss1.cpu()) loss1 = MNNF.const(loss1.flatten().tolist(), [1], MNNF.data_format.NCHW) # print('loss:',loss)# loss scaled by batch_size # if rank != -1: # loss *= opt.world_size # gradient averaged between devices in DDP mode # Backward # scaler.scale(loss).backward() mnn_opt.step(loss1) # 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/%g' % (epoch, epochs - 1,i), 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 # 20201016 by zlf # save model # with open(results_file, 'a') as f: # f.write(s + '%10.4g' * 7 % results + '\n') # # with open(results_file, 'r') as f: # create checkpoint # ckpt = {'epoch': epoch, # 'best_fitness': best_fitness, # 'training_results': None, # 'model': ema.ema, # 'optimizer': None} # if epoch>120 and (i % 450 == 0) and (i != 0): # torch.save(ckpt, wdir / 'ckpt_model_{}_{}_{}.pt'.format(epoch,i,round((mloss.sum()/2.).item(),5))) # add by zlf on 20201113 New Val Project # wdir_s = str(wdir)+'/' # modelv = ModelV(wdir_s+'ckpt_model_{}_{}_{}.pt'.format(epoch,i,round((mloss.sum()/2.).item(),5))) # imgs_dir = '%s_%s/'%(epoch,i) # if not os.path.exists(save_path+imgs_dir): # os.mkdir(save_path+imgs_dir) # l = val(modelv,'new_val_work/val.txt', # issave=False, # save_path=save_path+imgs_dir, # result_txt=save_path+imgs_dir+'result.txt') # result_txt = save_path+imgs_dir+'result.txt' # d1, d2, miss_detect, wrong_detect, extra_detect, total_error = check_result(result_txt) # line1 = 'model_epoch%s_%s: '%(epoch,i) # line2 = '{"total":%d,"miss":%d,"wrong":%d,"extra":%d}'%(total_error,miss_detect,wrong_detect,extra_detect) # print(line1+line2) # line = line1+line2+'\n'+str(d2) # count_f.write(line+'\n') # del ckpt # end batch ------------------------------------------------------------------------------------------------ # save model file_name = './weights/%d_20201231test.mnn' % epoch net.train(False) predict = net.forward(MNNF.placeholder([1, 3, 192, 320], MNNF.NC4HW4)) print("Save to " + file_name) MNNF.save([predict], file_name) # 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 = 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) # 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': None, '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: # best = wdir / 'epoch%d_best.pt'%epoch torch.save(ckpt, best) del ckpt # end epoch ---------------------------------------------------------------------------------------------------- # end training # count_f.close() 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 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 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