def train(args, epoch, model, data_loader, optimizer): """ TRAINING PROCEDURE Parameters: ----------- - args: various arguments - epoch: number of epochs - model: specified model to test - data_loader: specified train data_loader - optimizer: specified optimizer to use Returns: -------- - average_loss: average loss per batch """ statistics = [] total_loss = 0 model.train() title = 'Training Epoch {}'.format(epoch) progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=len(data_loader), smoothing=.9, miniters=1, leave=True, desc=title) sys.stdout.flush() for batch_idx, (data, target) in enumerate(progress): #data, target = data.to(args.device), target.to(args.device) optimizer.zero_grad() d = model(data[0].to(args.device), im_2=data[1].to(args.device)) loss = _apply_loss(d, target).mean() loss.backward() optimizer.step() total_loss += loss.item() assert not np.isnan(total_loss) # Print out statistics statistics.append(loss.item()) title = '{} Epoch {}'.format('Training', epoch) progress.set_description(title + '\tLoss:\t' + str(statistics[-1])) sys.stdout.flush() progress.close() return total_loss / float(batch_idx + 1)
def test(args, epoch, model, data_loader): """ TESTING PROCEDURE Parameters: ----------- - args: various arguments - epoch: number of epochs - model: specified model to test - data_loader: specified test data_loader Returns: -------- - average_loss: average loss per batch - pck: Percentage of Correct Keypoints metric """ statistics = [] total_loss = 0 model.eval() title = 'Validating Epoch {}'.format(epoch) progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=len(data_loader), smoothing=.9, miniters=1, leave=True, desc=title) predictions = [] gt = [] sys.stdout.flush() with torch.no_grad(): for batch_idx, (data, target) in enumerate(progress): d = model(data[0].to(args.device), im_2=data[1].to(args.device)) loss = _apply_loss(d, target).mean() total_loss += loss.item() predictions.extend(d.numpy()) gt.extend(target.numpy()) # Print out statistics statistics.append(loss.item()) title = '{} Epoch {}'.format('Validating', epoch) progress.set_description(title + '\tLoss:\t' + str(statistics[-1])) sys.stdout.flush() progress.close() pck = tools.calc_pck(np.asarray(predictions), np.asarray(gt)) print('PCK for epoch %d is %f' % (epoch, pck)) return total_loss / float(batch_idx + 1), pck
def train(args, epoch, start_iteration, data_loader, model, optimizer, loss, logger, is_validate=False, offset=0): statistics = [] total_loss = 0 gpu_mem = tools.gpumemusage() if is_validate: model.eval() title = 'Validating {} Epoch {}'.format(gpu_mem, epoch) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training {} Epoch {}'.format(tools.gpumemusage(), epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) last_log_time = progress._time() for batch_idx, (data, target) in enumerate(progress): data, target = [Variable(d, volatile=is_validate) for d in data], [ Variable(t, volatile=is_validate) for t in target ] if args.cuda: data, target = [d.cuda(async=True) for d in data ], [t.cuda(async=True) for t in target] optimizer.zero_grad() if not is_validate else None output = model(data[0]) loss_labels, loss_values = loss(output, target[0]) loss_val = loss_values[0] total_loss += loss_val.data[0] loss_values = [v.data[0] for v in loss_values] assert not np.isnan(total_loss) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) optimizer.step() # Update hyperparameters if needed global_iteration = start_iteration + batch_idx if not is_validate: tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer) loss_labels.append('lr') loss_values.append(optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(progress.iterable.last_duration) # Print out statistics statistics.append(loss_values) title = '{} {} Epoch {}'.format( 'Validating' if is_validate else 'Training', tools.gpumemusage(), epoch) progress.set_description( title + ' ' + tools.format_dictionary_of_losses(loss_labels, statistics[-1])) if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == args.validation_n_batches - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( 'batch logs per second', len(statistics) / (progress._time() - last_log_time), global_iteration) last_log_time = progress._time() all_losses = np.array(statistics) for i, key in enumerate(loss_labels): logger.add_scalar('average batch ' + key, all_losses[:, i].mean(), global_iteration) logger.add_histogram(key, all_losses[:, i], global_iteration) # Reset Summary statistics = [] if (is_validate and (batch_idx == args.validation_n_batches)): break if ((not is_validate) and (batch_idx == (args.train_n_batches))): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def main(args): dataset = args.dataset bsize = args.batch_size root = args.data_root cache_root = args.cache prediction_root = args.pre train_root = root + dataset + '/train' val_root = root + dataset + '/val' # validation dataset # mkdir( path [,mode] ):创建一个目录,可以是相对或者绝对路径,mode的默认模式是0777。 # 如果目录有多级,则创建最后一级。如果最后一级目录的上级目录有不存在的,则会抛出一个OSError。 # makedirs( path [,mode] ):创建递归的目录树,可以是相对或者绝对路径,mode的默认模式是 # 0777。如果子目录创建失败或者已经存在,会抛出一个OSError的异常,Windows上Error 183即为 # 目录已经存在的异常错误。如果path只有一级,与mkdir相同。 check_root_opti = cache_root + '/opti' # save checkpoint parameters if not os.path.exists(check_root_opti): os.makedirs(check_root_opti) check_root_feature = cache_root + '/feature' # save checkpoint parameters if not os.path.exists(check_root_feature): os.makedirs(check_root_feature) # 获取调整后的数据集 train_loader = torch.utils.data.DataLoader( MyData(train_root, transform=True), batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True ) val_loader = torch.utils.data.DataLoader( MyTestData(val_root, transform=True), batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True ) model = densenet169(pretrained=True, new_block=RCL_Module).cuda() criterion = nn.BCELoss() optimizer_feature = torch.optim.Adam(model.parameters(), lr=args.lr) # http://www.spytensor.com/index.php/archives/32/ scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer_feature, 'max', verbose=1, patience=10 ) progress = tqdm( range(args.start_epoch, args.total_epochs + 1), miniters=1, ncols=100, desc='Overall Progress', leave=True, position=0 ) offset = 1 best = 0 result = {'epoch': [], 'F_measure': [], 'MAE': []} for epoch in progress: # ===============================TRAIN================================= title = 'Training Epoch {}'.format(epoch) progress_epoch = tqdm( tools.IteratorTimer(train_loader), ncols=120, total=len(train_loader), smoothing=0.9, miniters=1, leave=True, position=offset, desc=title ) train(model, progress_epoch, criterion, optimizer_feature, epoch, args) # ==============================TEST=================================== if epoch % args.val_rate == 0: epoch, F_measure, mae = test( model, val_loader, epoch, prediction_root, check_root_feature, check_root_opti, val_root ) result['epoch'].append(int(epoch)) result['F_measure'].append(round(float(F_measure), 3)) result['MAE'].append(round(float(mae), 3)) df = pd.DataFrame(result).set_index('epoch') df.to_csv('./lart/result.csv') if epoch == 0: best = F_measure - mae elif (F_measure - mae) > best: best = F_measure - mae # 存储最好的权重和偏置 filename = ('%s/feature-best.pth' % check_root_feature) torch.save(model.state_dict(), filename) # 存储最好的优化器状态 filename_opti = ('%s/opti-best.pth' % check_root_opti) torch.save(optimizer_feature.state_dict(), filename_opti) # 只在验证期间考虑更改学习率 scheduler.step(best)
def train(args, epoch, start_iteration, data_loader, model, optimizer, logger, is_validate=False, offset=0): statistics = [] total_loss = 0 if is_validate: model.eval() title = 'Validating Epoch {}'.format(epoch) args.validation_n_batches = len( data_loader ) - 1 if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(epoch) args.train_n_batches = len( data_loader ) - 1 if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) last_log_time = progress._time() for batch_idx, (data, target) in enumerate(progress): data, target = [Variable(d, volatile=is_validate) for d in data], [ Variable(t, volatile=is_validate) for t in target ] if args.cuda and args.number_gpus == 1: data, target = [d.cuda(async=True) for d in data ], [t.cuda(async=True) for t in target] optimizer.zero_grad() if not is_validate else None losses = model(data[0], target[0]) losses = [torch.mean(loss_value) for loss_value in losses] loss_val = losses[1] # Collect first loss for weight update total_loss += loss_val.data[0] loss_values = [v.data[0] for v in losses] # gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather' loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) optimizer.step() # Update hyperparameters if needed global_iteration = start_iteration + batch_idx if not is_validate: tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer) loss_labels.append('lr') loss_values.append(optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(progress.iterable.last_duration) # Print out statistics statistics.append(loss_values) title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', epoch) if (type(loss_labels[0]) is list) or (type(loss_labels[0]) is tuple): progress.set_description(title + ' ' + tools.format_dictionary_of_losses( loss_labels[0], statistics[-1])) else: progress.set_description(title + ' ' + tools.format_dictionary_of_losses( loss_labels, statistics[-1])) if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == args.validation_n_batches - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( 'batch logs per second', len(statistics) / (progress._time() - last_log_time), global_iteration) last_log_time = progress._time() all_losses = np.array(statistics) for i, key in enumerate(loss_labels[0] if ( type(loss_labels[0]) is list) or ( type(loss_labels[0]) is tuple) else loss_labels): logger.add_scalar('average batch ' + str(key), all_losses[:, i].mean(), global_iteration) #logger.add_histogram(str(key), all_losses[:, i], global_iteration) if is_validate: _, output = model(data[0], target[0], inference=True) render_flow = output[0].data.cpu().numpy().transpose( 1, 2, 0) ground_truth = target[0][0].data.cpu().numpy().transpose( 1, 2, 0) render_img = tools.flow_to_image(render_flow).transpose( 2, 0, 1) true_img = tools.flow_to_image(ground_truth).transpose( 2, 0, 1) render_img = torch.Tensor(render_img) / 255.0 true_img = torch.Tensor(true_img) / 255.0 input_img = data[0][0, :, 0, :, :].data.cpu() / 255.0 logger.add_image('renderimg', torchvision.utils.make_grid(render_img), global_iteration) logger.add_image('ground_truth', torchvision.utils.make_grid(true_img), global_iteration) logger.add_image('input_img', torchvision.utils.make_grid(input_img), global_iteration) # Reset Summary statistics = [] if (is_validate and (batch_idx == args.validation_n_batches)): break if ((not is_validate) and (batch_idx == (args.train_n_batches))): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(input_args, train_epoch, start_iteration, files_loader, model, model_optimizer, logger, is_validate=False, offset=0): statistics = [] total_loss = 0 if is_validate: model.eval() title = 'Validating Epoch {}'.format(train_epoch) input_args.validation_n_batches = np.inf if input_args.validation_n_batches < 0 else input_args.validation_n_batches file_progress = tqdm(tools.IteratorTimer(files_loader), ncols=100, total=np.minimum( len(files_loader), input_args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(train_epoch) input_args.train_n_batches = np.inf if input_args.train_n_batches < 0 else input_args.train_n_batches file_progress = tqdm(tools.IteratorTimer(files_loader), ncols=120, total=np.minimum(len(files_loader), input_args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) last_log_time = file_progress._time() for batch_idx, (data_file) in enumerate(file_progress): video_dataset = datasets_video.VideoFileDataJIT( input_args, data_file[0]) video_loader = DataLoader(video_dataset, batch_size=args.effective_batch_size, shuffle=True, **gpuargs) global_iteration = start_iteration + batch_idx # note~ for debugging purposes # video_frame_progress = tqdm(tools.IteratorTimer(video_loader), ncols=120, # total=len(video_loader), smoothing=0.9, miniters=1, # leave=True, desc=data_file[0]) for i_batch, (data, target) in enumerate(video_loader): data, target = [Variable(d) for d in data], [Variable(t) for t in target] if input_args.cuda and input_args.number_gpus == 1: data, target = [d.cuda(async=True) for d in data ], [t.cuda(async=True) for t in target] model_optimizer.zero_grad() if not is_validate else None losses = model(data[0], target[0]) losses = [torch.mean(loss_value) for loss_value in losses] loss_val = losses[0] # Collect first loss for weight update total_loss += loss_val.data loss_values = [v.data for v in losses] # gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather' loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss.cpu().numpy()) if not is_validate and input_args.fp16: loss_val.backward() if input_args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), input_args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / input_args.loss_scale) model_optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if input_args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), input_args.gradient_clip) model_optimizer.step() # Update hyperparameters if needed if not is_validate: tools.update_hyperparameter_schedule( input_args, train_epoch, global_iteration, model_optimizer) loss_labels.append('lr') loss_values.append(model_optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(file_progress.iterable.last_duration) # Print out statistics statistics.append(loss_values) title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', train_epoch) file_progress.set_description( title + ' ' + tools.format_dictionary_of_losses( tools.flatten_list(loss_labels), statistics[-1])) if ((((global_iteration + 1) % input_args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == input_args.validation_n_batches - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( 'batch logs per second', len(statistics) / (file_progress._time() - last_log_time), global_iteration) last_log_time = file_progress._time() all_losses = np.array(statistics) for i, key in enumerate(tools.flatten_list(loss_labels)): if isinstance(all_losses[:, i].item(), torch.Tensor): average_batch = all_losses[:, i].item().mean() else: average_batch = all_losses[:, i].item() logger.add_scalar('average batch ' + str(key), average_batch, global_iteration) logger.add_histogram(str(key), all_losses[:, i], global_iteration) # Reset Summary statistics = [] if is_validate and (batch_idx == input_args.validation_n_batches): break if (not is_validate) and (batch_idx == (input_args.train_n_batches)): break file_progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args, epoch, start_iteration, data_loader, model, optimizer, logger, is_validate=False, offset=0): statistics = [] total_loss = 0 if is_validate: model.eval() title = "Validating Epoch {}".format(epoch) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm( tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title, ) else: model.train() title = "Training Epoch {}".format(epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm( tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=0.9, miniters=1, leave=True, position=offset, desc=title, ) last_log_time = progress._time() for batch_idx, (data, target) in enumerate(progress): data, target = [Variable(d) for d in data], [Variable(t) for t in target] if args.cuda and args.number_gpus == 1: data, target = [d.cuda(non_blocking=True) for d in data], [t.cuda(non_blocking=True) for t in target] optimizer.zero_grad() if not is_validate else None losses = model(data[0], target[0]) losses = [torch.mean(loss_value) for loss_value in losses] loss_val = losses[0] # Collect first loss for weight update total_loss += loss_val.item() loss_values = [v.item() for v in losses] # gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather' loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as(params[i]).detach() param_copy[i].grad.mul_(1.0 / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) optimizer.step() # Update hyperparameters if needed global_iteration = start_iteration + batch_idx if not is_validate: tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer) loss_labels.append("lr") loss_values.append(optimizer.param_groups[0]["lr"]) loss_labels.append("load") loss_values.append(progress.iterable.last_duration) # Print out statistics statistics.append(loss_values) title = "{} Epoch {}".format("Validating" if is_validate else "Training", epoch) progress.set_description(title + " " + tools.format_dictionary_of_losses(loss_labels, statistics[-1])) if (((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or ( is_validate and batch_idx == args.validation_n_batches - 1 ): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( "batch logs per second", len(statistics) / (progress._time() - last_log_time), global_iteration ) last_log_time = progress._time() all_losses = np.array(statistics) for i, key in enumerate(loss_labels): logger.add_scalar("average batch " + str(key), all_losses[:, i].mean(), global_iteration) logger.add_histogram(str(key), all_losses[:, i], global_iteration) # Reset Summary statistics = [] if is_validate and (batch_idx == args.validation_n_batches): break if (not is_validate) and (batch_idx == (args.train_n_batches)): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def main(args): dataset = args.dataset bsize = args.batch_size root = args.data_root cache_root = args.cache prediction_root = args.pre train_root = root + dataset + '/train' val_root = root + dataset + '/val' # validation dataset check_root_opti = cache_root + '/opti' # save checkpoint parameters if not os.path.exists(check_root_opti): os.mkdir(check_root_opti) check_root_feature = cache_root + '/feature' # save checkpoint parameters if not os.path.exists(check_root_feature): os.mkdir(check_root_feature) train_loader = torch.utils.data.DataLoader(MyData(train_root, transform=True), batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True) val_loader = torch.utils.data.DataLoader(MyTestData(val_root, transform=True), batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True) model = Feature(RCL_Module) model.cuda() criterion = nn.BCELoss() optimizer_feature = torch.optim.Adam(model.parameters(), lr=args.lr) train_losses = [] progress = tqdm(range(args.start_epoch, args.total_epochs + 1), miniters=1, ncols=100, desc='Overall Progress', leave=True, position=0) offset = 1 best = 0 evaluation = [] result = {'epoch': [], 'F_measure': [], 'MAE': []} for epoch in progress: if (epoch != 0): print("\nloading parameters") model.load_state_dict( torch.load(check_root_feature + '/feature-current.pth')) optimizer_feature.load_state_dict( torch.load(check_root_opti + '/opti-current.pth')) # title = 'Training Epoch {}'.format(epoch) progress_epoch = tqdm(tools.IteratorTimer(train_loader), ncols=120, total=len(train_loader), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) for ib, (input, gt) in enumerate(progress_epoch): inputs = Variable(input).cuda() gt = Variable(gt.unsqueeze(1)).cuda() gt_28 = functional.interpolate(gt, size=28, mode='bilinear') gt_56 = functional.interpolate(gt, size=56, mode='bilinear') gt_112 = functional.interpolate(gt, size=112, mode='bilinear') msk1, msk2, msk3, msk4, msk5 = model.forward(inputs) loss = criterion(msk1, gt_28) + criterion(msk2, gt_28) + criterion( msk3, gt_56) + criterion(msk4, gt_112) + criterion(msk5, gt) model.zero_grad() loss.backward() optimizer_feature.step() train_losses.append(round(float(loss.data.cpu()), 3)) title = '{} Epoch {}/{}'.format('Training', epoch, args.total_epochs) progress_epoch.set_description(title + ' ' + 'loss:' + str(loss.data.cpu().numpy())) filename = ('%s/feature-current.pth' % (check_root_feature)) filename_opti = ('%s/opti-current.pth' % (check_root_opti)) torch.save(model.state_dict(), filename) torch.save(optimizer_feature.state_dict(), filename_opti) #--------------------------validation on the test set every n epoch-------------- if (epoch % args.val_rate == 0): fileroot = ('%s/feature-current.pth' % (check_root_feature)) model.load_state_dict(torch.load(fileroot)) val_output_root = (prediction_root + '/epoch_current') if not os.path.exists(val_output_root): os.mkdir(val_output_root) print("\ngenerating output images") for ib, (input, img_name, _) in enumerate(val_loader): inputs = Variable(input).cuda() _, _, _, _, output = model.forward(inputs) output = functional.sigmoid(output) out = output.data.cpu().numpy() for i in range(len(img_name)): imsave(os.path.join(val_output_root, img_name[i] + '.png'), out[i, 0], cmap='gray') print("\nevaluating mae....") F_measure, mae = get_FM(salpath=val_output_root + '/', gtpath=val_root + '/gt/') evaluation.append([int(epoch), float(F_measure), float(mae)]) result['epoch'].append(int(epoch)) result['F_measure'].append(round(float(F_measure), 3)) result['MAE'].append(round(float(mae), 3)) df = pd.DataFrame(result).set_index('epoch') df.to_csv('./result.csv') if (epoch == 0): best = F_measure - mae elif ((F_measure - mae) > best): best = F_measure - mae filename = ('%s/feature-best.pth' % (check_root_feature)) filename_opti = ('%s/opti-best.pth' % (check_root_opti)) torch.save(model.state_dict(), filename) torch.save(optimizer_feature.state_dict(), filename_opti)
def train(args, epoch, start_iteration, data_loader, model, optimizer, logger, is_validate=False, offset=0): statistics = [] all_gradient_norms = [] total_loss = 0 if is_validate: model.eval() title = 'Validating Epoch {}'.format(epoch) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=200, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=200, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) last_log_time = progress._time() for batch_idx, (data, target) in enumerate(progress): data, target = [Variable(d) for d in data], [Variable(t) for t in target] if args.cuda and args.number_gpus == 1: data, target = [d.cuda(non_blocking=True) for d in data ], [t.cuda(non_blocking=True) for t in target] optimizer.zero_grad() if not is_validate else None losses, flow = model(data[0], target[0]) #print('Losses shape {} {}'.format(losses[0].shape, losses[1].shape)) losses = [torch.mean(loss_value) for loss_value in losses] loss_val = losses[0] # Collect first loss for weight update total_loss += loss_val.item() loss_values = [v.item() for v in losses] loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: gradient_norm = torch.nn.utils.clip_grad_norm( model.parameters(), args.gradient_clip) all_gradient_norms.append(gradient_norm) optimizer.step() # Update hyperparameters if needed global_iteration = start_iteration + batch_idx if not is_validate: tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer) loss_labels.append('lr') loss_values.append(optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(progress.iterable.last_duration) # Print out statistics statistics.append(loss_values) title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', epoch) progress.set_description( title + ' ' + tools.format_dictionary_of_losses(loss_labels, statistics[-1])) if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == args.validation_n_batches - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( 'batch logs per second', len(statistics) / (progress._time() - last_log_time), global_iteration) last_log_time = progress._time() all_losses = np.array(statistics) for i, key in enumerate(loss_labels): logger.add_scalar('average batch ' + str(key), all_losses[:, i].mean(), global_iteration) logger.add_histogram(str(key), all_losses[:, i], global_iteration) if args.gradient_clip: logger.add_scalar('average batch gradient_norm', np.array(all_gradient_norms).mean(), global_iteration) all_gradient_norms = [] # Returns multiscale flow, get largest scale and first element in batch if args.multiframe or args.multiframe_two_output: flow = flow_utils.flow_postprocess(flow)[0][0] num_flows = len(args.frame_weights) flows_scaled = [ cv2.resize(flow[:, :, i:i + 2], None, fx=4.0, fy=4.0) for i in range(0, 2 * num_flows, 2) ] target = target[0].detach().cpu().numpy() target_flow = np.transpose(target[0], (1, 2, 3, 0)) results_images = [ visualize_results(flows_scaled[i], target_flow[i], data[0][0] if i == 0 else None) for i in range(0, num_flows) ] for i in range(0, num_flows): logger.add_image('flow{} and target'.format(i), ToTensor()(results_images[i]), global_iteration) else: flow = flow_utils.flow_postprocess(flow)[0][0] flow_scaled = cv2.resize(flow, None, fx=4.0, fy=4.0) target_flow = flow_utils.flow_postprocess(target)[0][0] results_image = visualize_results(flow_scaled, target_flow, data[0][0]) logger.add_image('flow and target', ToTensor()(results_image), global_iteration) # logger.add_histogram('flow_values', flow[0], global_iteration) # Reset Summary statistics = [] if (is_validate and (batch_idx == args.validation_n_batches)): break if ((not is_validate) and (batch_idx == (args.train_n_batches))): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args, epoch, start_iteration, data_loader, model, optimizer, logger, is_validate=False, offset=0): statistics = [] total_loss = 0 if is_validate: model.eval() title = 'Validating Epoch {}'.format(epoch) #print("validation_n_batches", args.validation_n_batches) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches #print("validation_n_batches", args.validation_n_batches) progress = tqdm(tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) last_log_time = progress._time() for batch_idx, (data, target) in enumerate(progress): data, target = [Variable(d) for d in data], [Variable(t) for t in target] if args.cuda and args.number_gpus == 1: data, target = [d.cuda(async=True) for d in data ], [t.cuda(async=True) for t in target] optimizer.zero_grad() if not is_validate else None #print("this is data type",data[0].type()) #print("\n") #print("this is target type",target[0].type()) #print("\n") losses = model(data[0], target[0]) losses = [torch.mean(loss_value) for loss_value in losses] # taking mean of batches loss_val = losses[ 0] # Collect first loss for weight update #take first loss, second is EPE total_loss += loss_val.data.cpu() loss_values = [v.data.cpu() for v in losses] #collect loss values # gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather' #loss_labels = [y for x in model.module.loss.loss_labels for y in x] #list(model.module.loss.loss_labels) loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss.cpu()) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) optimizer.step() # Update hyperparameters if needed global_iteration = start_iteration + batch_idx if not is_validate: tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer) loss_labels.append('lr') loss_values.append(optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(progress.iterable.last_duration) #add load #if is_validate: # print("this is EPE length", len(loss_values[:,1])) # Print out statistics #if is_validate: # print(statistics) statistics.append(loss_values) #if is_validate: # print(statistics) title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', epoch) progress.set_description( title + ' ' + tools.format_dictionary_of_losses(loss_labels, statistics[-1])) #if is_validate: # print(batch_idx) # args.log_frequency == 1 by default if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or is_validate and batch_idx == min(args.validation_n_batches, len(data_loader) - 1)): #if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == args.validation_n_batches - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( 'batch logs per second', len(statistics) / (progress._time() - last_log_time), global_iteration) last_log_time = progress._time() all_losses = np.array(statistics) #if is_validate: # print(all_losses) for i, key in enumerate(loss_labels): logger.add_scalar('average batch ' + str(key), all_losses[:, i].mean(), global_iteration) logger.add_histogram(str(key), all_losses[:, i], global_iteration) # Reset Summary statistics = [] if (is_validate and (batch_idx == args.validation_n_batches)): break if ((not is_validate) and (batch_idx == (args.train_n_batches))): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args, epoch, data_loader, model, optimizer, is_validate=False, offset=0): total_loss = 0 if is_validate: model.eval() title = 'Validating Epoch {}'.format(epoch) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) def torch2numpy(i): return i[0].numpy() for batch_idx, datas in enumerate(progress): data = np.array([list(map(down_scailing, d)) for d in datas]) target = np.array([list(map(torch2numpy, d))[1] for d in datas]) high_frames = np.array([list(map(torch2numpy, d)) for d in datas]) # if args.cuda and args.number_gpus >= 1: # data, target, high_frames = [d.cuda for d in data], [t.cuda for t in target], [hf.cuda for hf in high_frames] estimated_image = None for x, y in zip(data, target): optimizer.zero_grad() if not is_validate else None output, losses = model(x, y, high_frames, estimated_image) estimated_image = output loss_val = torch.mean(losses) total_loss += loss_val.item() if not is_validate: loss_val.backward() optimizer.step() title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', epoch) progress.set_description(title) if (is_validate and (batch_idx == args.validation_n_batches)) or \ ((not is_validate) and (batch_idx == (args.train_n_batches))): progress.close() break return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args, epoch, start_iteration, data_loader, model, optimizer, logger, is_validate=False, offset=0): #print(str(model)) statistics = [] total_loss = 0 debug = False if is_validate: model.eval() title = 'Validating Epoch {}'.format(epoch) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) last_log_time = progress._time() for batch_idx, (data, target, cdm) in enumerate(progress): data, target, cdm = [ Variable(d, volatile=is_validate) for d in data ], [Variable(t, volatile=is_validate) for t in target ], [Variable(q, volatile=is_validate) for q in cdm] if args.cuda and args.number_gpus == 1: data, target, cdm = [d.cuda(async=True) for d in data ], [t.cuda(async=True) for t in target ], [q.cuda(async=True) for q in cdm] if debug: print( '****************************************************************' ) print('data_0') print(data[0]) print('target_0') print(target[0]) print('cdm') print(type(cdm)) temp1 = cdm[0].data.cpu().numpy() print(np.max(temp1)) print(temp1.shape) print( '****************************************************************' ) optimizer.zero_grad() if not is_validate else None losses = model(data[0], target[0]) losses = [torch.mean(loss_value) for loss_value in losses] loss_val = losses[0] # Collect first loss for weight update #A[batch_idx] = loss_val.data[0] #np.savetxt('test_loss.out', np.array(A) , delimiter=',' , newline='\r\n' ) total_loss += loss_val.data[0] loss_values = [v.data[0] for v in losses] # gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather' loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) optimizer.step() # Update hyperparameters if needed global_iteration = start_iteration + batch_idx if not is_validate: tools.update_hyperparameter_schedule(args, epoch, global_iteration, optimizer) loss_labels.append('lr') loss_values.append(optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(progress.iterable.last_duration) # Print out statistics statistics.append(loss_values) title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', epoch) progress.set_description( title + ' ' + tools.format_dictionary_of_losses(loss_labels, statistics[-1])) if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (is_validate and batch_idx == args.validation_n_batches - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar( 'batch logs per second', len(statistics) / (progress._time() - last_log_time), global_iteration) last_log_time = progress._time() all_losses = np.array(statistics) for i, key in enumerate(loss_labels): logger.add_scalar('average batch ' + str(key), all_losses[:, i].mean(), global_iteration) logger.add_histogram(str(key), all_losses[:, i], global_iteration) # Reset Summary statistics = [] if (is_validate and (batch_idx == args.validation_n_batches)): break if ((not is_validate) and (batch_idx == (args.train_n_batches))): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def train(args, epoch, start_iteration, data_loader, model, optimizer, scheduler, logger, is_validate=False, offset=0, max_flows_to_show=8): running_statistics = None # Initialize below when the first losses are collected all_losses = None # Initialize below when the first losses are collected total_loss = 0 if is_validate: model.eval() title = 'Validating Epoch {}'.format(epoch) args.validation_n_batches = np.inf if args.validation_n_batches < 0 else args.validation_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=100, total=np.minimum(len(data_loader), args.validation_n_batches), leave=True, position=offset, desc=title) else: model.train() title = 'Training Epoch {}'.format(epoch) args.train_n_batches = np.inf if args.train_n_batches < 0 else args.train_n_batches progress = tqdm(tools.IteratorTimer(data_loader), ncols=120, total=np.minimum(len(data_loader), args.train_n_batches), smoothing=.9, miniters=1, leave=True, position=offset, desc=title) def convert_flow_to_image(flow_converter, flows_viz): imgs = [] for flow_pair in flows_viz: for flow in flow_pair: flow = flow.numpy().transpose((1, 2, 0)) img = flow_converter._flowToColor(flow) imgs.append(torch.from_numpy(img.transpose((2, 0, 1)))) epe_img = torch.sqrt( torch.sum(torch.pow(flow_pair[0] - flow_pair[1], 2), dim=0)) max_epe = torch.max(epe_img) if max_epe == 0: max_epe = torch.ones(1) normalized_epe_img = epe_img / max_epe normalized_epe_img = (255 * normalized_epe_img).type( torch.uint8) normalized_epe_img = torch.stack( (normalized_epe_img, normalized_epe_img, normalized_epe_img), dim=0) imgs.append(normalized_epe_img) saturated_epe_img = torch.min(epe_img, 5.0 * torch.ones_like(epe_img)) saturated_epe_img = (51 * saturated_epe_img).type(torch.uint8) saturated_epe_img = torch.stack( (saturated_epe_img, saturated_epe_img, saturated_epe_img), dim=0) imgs.append(saturated_epe_img) return imgs max_iters = min(len(data_loader), (args.validation_n_batches if (is_validate and args.validation_n_batches > 0) else len(data_loader)), (args.train_n_batches if (not is_validate and args.train_n_batches > 0) else len(data_loader))) if is_validate: flow_converter = f2i.Flow() collect_flow_interval = int( np.ceil(float(max_iters) / max_flows_to_show)) flows_viz = [] last_log_batch_idx = 0 last_log_time = progress._time() for batch_idx, (data, target) in enumerate(progress): global_iteration = start_iteration + batch_idx data, target = [Variable(d) for d in data], [Variable(t) for t in target] if args.cuda and args.number_gpus == 1: data, target = [d.cuda() for d in data], [t.cuda() for t in target] optimizer.zero_grad() if not is_validate else None losses, output = model(data[0], target[0], inference=True) losses = [torch.mean(loss_value) for loss_value in losses] loss_val = losses[0] # Collect first loss for weight update total_loss += loss_val.item() loss_values = [v.item() for v in losses] if is_validate and batch_idx % collect_flow_interval == 0: flows_viz.append( (target[0][0].detach().cpu(), output[0].detach().cpu())) if is_validate and args.validation_log_images and batch_idx == ( max_iters - 1): imgs = convert_flow_to_image(flow_converter, flows_viz) imgs = torchvision_utils.make_grid(imgs, nrow=4, normalize=False, scale_each=False) logger.add_image('target/predicted flows', imgs, global_iteration) # gather loss_labels, direct return leads to recursion limit error as it looks for variables to gather' loss_labels = list(model.module.loss.loss_labels) assert not np.isnan(total_loss) if not is_validate and args.fp16: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) params = list(model.parameters()) for i in range(len(params)): param_copy[i].grad = params[i].grad.clone().type_as( params[i]).detach() param_copy[i].grad.mul_(1. / args.loss_scale) optimizer.step() for i in range(len(params)): params[i].data.copy_(param_copy[i].data) elif not is_validate: loss_val.backward() if args.gradient_clip: torch.nn.utils.clip_grad_norm(model.parameters(), args.gradient_clip) optimizer.step() # Update hyperparameters if needed if not is_validate: scheduler.step() loss_labels.append('lr') loss_values.append(optimizer.param_groups[0]['lr']) loss_labels.append('load') loss_values.append(progress.iterable.last_duration) if running_statistics is None: running_statistics = np.array(loss_values) all_losses = np.zeros((len(data_loader), len(loss_values)), np.float32) else: running_statistics += np.array(loss_values) all_losses[batch_idx] = loss_values.copy() title = '{} Epoch {}'.format( 'Validating' if is_validate else 'Training', epoch) progress.set_description(title + ' ' + tools.format_dictionary_of_losses( loss_labels, running_statistics / (batch_idx + 1))) if ((((global_iteration + 1) % args.log_frequency) == 0 and not is_validate) or (batch_idx == max_iters - 1)): global_iteration = global_iteration if not is_validate else start_iteration logger.add_scalar('batch logs per second', (batch_idx - last_log_batch_idx) / (progress._time() - last_log_time), global_iteration) last_log_time = progress._time() last_log_batch_idx = batch_idx for i, key in enumerate(loss_labels): logger.add_scalar('average batch ' + str(key), all_losses[:batch_idx + 1, i].mean(), global_iteration) logger.add_histogram(str(key), all_losses[:batch_idx + 1, i], global_iteration) if (is_validate and (batch_idx == args.validation_n_batches)): break if ((not is_validate) and (batch_idx == (args.train_n_batches))): break progress.close() return total_loss / float(batch_idx + 1), (batch_idx + 1)
def main(args): dataset = args.dataset bsize = args.batch_size root = args.data_root cache_root = args.cache prediction_root = args.pre train_root = root + dataset + '/Train' val_root = root + dataset + '/Test' # validation dataset # mkdir( path [,mode] ):创建一个目录,可以是相对或者绝对路径,mode的默认模式是0777。 # 如果目录有多级,则创建最后一级。如果最后一级目录的上级目录有不存在的,则会抛出一个OSError。 # makedirs( path [,mode] ):创建递归的目录树,可以是相对或者绝对路径,mode的默认模式是 # 0777。如果子目录创建失败或者已经存在,会抛出一个OSError的异常,Windows上Error 183即为 # 目录已经存在的异常错误。如果path只有一级,与mkdir相同。 check_root_opti = cache_root + '/opti' # save checkpoint parameters if not os.path.exists(check_root_opti): os.makedirs(check_root_opti) check_root_feature = cache_root + '/feature' # save checkpoint parameters if not os.path.exists(check_root_feature): os.makedirs(check_root_feature) check_root_model = cache_root + '/model' # save checkpoint parameters if not os.path.exists(check_root_model): os.makedirs(check_root_model) # 获取调整后的数据集 train_loader = torch.utils.data.DataLoader( MyData(train_root, transform=True), batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True ) val_loader = torch.utils.data.DataLoader( MyTestData(val_root, transform=True), batch_size=bsize, shuffle=True, num_workers=4, pin_memory=True ) model = Vgg(RCL_Module) model.cuda() criterion = nn.BCELoss() optimizer_feature = torch.optim.Adam(model.parameters(), lr=args.lr) # http://www.spytensor.com/index.php/archives/32/ scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer_feature, 'max', verbose=1, patience=10 ) progress = tqdm( range(args.start_epoch, args.total_epochs + 1), miniters=1, ncols=100, desc='Overall Progress', leave=True, position=0 ) offset = 1 best = 0 evaluation = [] result = {'epoch': [], 'F_measure': [], 'MAE': []} for epoch in progress: # ===============================TRAIN================================= # if epoch != 0: # print("\nloading parameters") # # 载入上一次的训练结果(权重和偏置项), 进一步的训练 # model.load_state_dict( # torch.load(check_root_feature + '/feature-current.pth') # ) # # 载入优化器状态 # optimizer_feature.load_state_dict( # torch.load(check_root_opti + '/opti-current.pth') # ) title = 'Training Epoch {}'.format(epoch) progress_epoch = tqdm( tools.IteratorTimer(train_loader), ncols=120, total=len(train_loader), smoothing=0.9, miniters=1, leave=True, position=offset, desc=title ) # 一个周期内部进行迭代计算 for ib, (input_, gt) in enumerate(progress_epoch): # 获取对应的5个掩膜预测结果 inputs = Variable(input_).cuda() msk1, msk2, msk3, msk4, msk5 = model.forward(inputs) gt = Variable(gt.unsqueeze(1)).cuda() gt_28 = functional.interpolate(gt, size=28, mode='bilinear') gt_56 = functional.interpolate(gt, size=56, mode='bilinear') gt_112 = functional.interpolate(gt, size=112, mode='bilinear') loss = criterion(msk1, gt_28) + criterion(msk2, gt_28) \ + criterion(msk3, gt_56) + criterion(msk4, gt_112) \ + criterion(msk5, gt) model.zero_grad() loss.backward() optimizer_feature.step() title = '{} Epoch {}/{}'.format( 'Training', epoch, args.total_epochs ) progress_epoch.set_description( title + ' ' + 'loss:' + str(loss.data.cpu().numpy()) ) # 存储一个epoch后的模型(权重和偏置项), 以便后期使用 filename = ('%s/feature-current.pth' % check_root_feature) torch.save(model.state_dict(), filename) # 存储优化器状态 filename_opti = ('%s/opti-current.pth' % check_root_opti) torch.save(optimizer_feature.state_dict(), filename_opti) # ==============================TEST=================================== if epoch % args.val_rate == 0: fileroot = ('%s/feature-current.pth' % check_root_feature) # 基于torch.save(model.state_dict(), filename)存储方法的对应的恢复方法 model.load_state_dict(torch.load(fileroot)) val_output_root = (prediction_root + '/epoch_current') if not os.path.exists(val_output_root): os.makedirs(val_output_root) print("\ngenerating output images") for ib, (input_, img_name, _) in enumerate(val_loader): inputs = Variable(input_).cuda() _, _, _, _, output = model.forward(inputs) out = output.data.cpu().numpy() for i in range(len(img_name)): print(out[i]) imsave(os.path.join(val_output_root, img_name[i] + '.png'), out[i, 0], cmap='gray') print("\nevaluating mae....") # mean = np.array([0.485, 0.456, 0.406]) # std = np.array([0.229, 0.224, 0.225]) # img = Image.open("./data/ILSVRC2012_test_00000004_224x224.jpg") # img = np.array(img) # img = img.astype(np.float64) / 255 # img -= mean # img /= std # img = img.transpose(2, 0, 1) # img = np.array(img)[np.newaxis, :, :, :].astype(np.float32) # img = torch.from_numpy(img).float() # inputs = Variable(img).cuda() # _, _, _, _, output = model.forward(inputs) # out = output.data.cpu().numpy() # print(out) # imsave(os.path.join(val_output_root, 'caffe2_test' + '.png'), # out[0, 0], cmap='gray') # 计算F测度和平均绝对误差 F_measure, mae = get_FM( salpath=val_output_root + '/', gtpath=val_root + '/masks/' ) evaluation.append([int(epoch), float(F_measure), float(mae)]) result['epoch'].append(int(epoch)) result['F_measure'].append(round(float(F_measure), 3)) result['MAE'].append(round(float(mae), 3)) df = pd.DataFrame(result).set_index('epoch') df.to_csv('./result.csv') if epoch == 0: best = F_measure - mae elif (F_measure - mae) > best: best = F_measure - mae # 存储最好的权重和偏置 filename = ('%s/feature-best.pth' % check_root_feature) torch.save(model.state_dict(), filename) # 存储最好的优化器状态 filename_opti = ('%s/opti-best.pth' % check_root_opti) torch.save(optimizer_feature.state_dict(), filename_opti) # # 存储最好的完整网络 # filename_opti = ('%s/model-best.pth' % check_root_model) # torch.save(model, filename_opti) print("完成一次保存") # 只在验证期间考虑更改学习率 scheduler.step(best) print("完成一次测试")