def _init_engines(self) -> Tuple[Engine, Engine]: self.train_metrics = { 'total_loss': metrics.RunningAverage(output_transform=lambda x: x['loss']), 'segment_loss': metrics.RunningAverage( output_transform=lambda x: x['segment_loss']), 'kl_div': metrics.RunningAverage(output_transform=lambda x: x['kl_div']) } self.val_metrics = { 'vae_metrics': VAEMetrics(loss_fn=self.criterion, mse_factor=0, kld_factor=self.starting_kld_factor), 'segment_metrics': SegmentationMetrics(num_classes=self.data_loaders.num_classes, threshold=self.config.binarize_threshold), } trainer = self._init_trainer_engine() evaluator = self._init_evaluator_engine() return trainer, evaluator
def _init_engines(self) -> Tuple[Engine, Engine]: if self.use_ensemble: trainer = self._init_trainer_engine_ensemble() evaluator = self._init_evaluator_engine_ensemble() else: trainer = self._init_trainer_engine() evaluator = self._init_evaluator_engine() metrics.RunningAverage(output_transform=lambda x: x).attach(trainer, 'train_loss') return trainer, evaluator
def _init_engines(self): trainer = self._init_bayesian_trainer_engine(self.model, self.optimizer, self.vi, self.device) evaluator = self._init_bayesian_evaluator_engine( self.model, self.metrics, self.device) metrics.RunningAverage(output_transform=lambda x: x).attach( trainer, 'train_loss') return trainer, evaluator
def get_engines(learning_rate=0.01): # We remove 1 because of the target column model = get_model(n_features=len(useful_columns) - 1) optimizer = optim.Adam(model.parameters(), lr=learning_rate) loss = nn.MSELoss() trainer = engine.create_supervised_trainer(model, optimizer, loss) evaluator = engine.create_supervised_evaluator(model, metrics={ 'mae': metrics.MeanAbsoluteError(), 'loss': metrics.Loss(loss), 'avg': metrics.RunningAverage(metrics.Loss(loss)) }) # workaround for using the ProgressBar with training loss Identity().attach(trainer, 'loss') Average().attach(trainer, 'avg_loss') return trainer, evaluator
def evaluator_epoch_comp_callback(engine): # save masks for each batch batch_output = engine.state.output input_filenames = batch_output['input_filename'] masks = batch_output['mask'] for i, input_filename in enumerate(input_filenames): mask = cv2.resize(masks[i], dsize=(utils.cropped_width, utils.cropped_height), interpolation=cv2.INTER_AREA) # if pad: # h_start, w_start = utils.h_start, utils.w_start # h, w = mask.shape # # recover to original shape # full_mask = np.zeros((original_height, original_width)) # full_mask[h_start:h_start + h, w_start:w_start + w] = t_mask # mask = full_mask #print("Input Filename-->", input_filename) #instrument_folder_name = input_filename.parent.parent.name instrument_folder_name = os.path.basename( os.path.dirname(os.path.dirname(input_filename))) #print("instrument_folder_name-->", instrument_folder_name) # mask_folder/instrument_dataset_x/problem_type_masks/framexxx.png mask_folder = mask_save_dir / instrument_folder_name / utils.mask_folder[ args.problem_type] mask_folder.mkdir(exist_ok=True, parents=True) mask_filename = mask_folder / os.path.basename(input_filename) #print("mask_filename-->", mask_filename) cv2.imwrite(str(mask_filename), mask) if 'TAPNet' in args.model: attmap = batch_output['attmap'][i] attmap_folder = mask_save_dir / instrument_folder_name / '_'.join( args.problem_type, 'attmaps') attmap_folder.mkdir(exist_ok=True, parents=True) attmap_filename = attmap_folder / input_filename.name cv2.imwrite(str(attmap_filename), attmap) evaluator.run(eval_loader) # validator engine validator = engine.Engine(valid_step) # monitor loss valid_ra_loss = imetrics.RunningAverage( output_transform=lambda x: x['loss'], alpha=0.98) valid_ra_loss.attach(validator, 'valid_ra_loss') # monitor validation loss over epoch valid_loss = imetrics.Loss(loss_func, output_transform=lambda x: (x['output'], x['target'])) valid_loss.attach(validator, 'valid_loss') # monitor <data> mean metrics valid_data_miou = imetrics.RunningAverage( output_transform=lambda x: x['iou'].data_mean()['mean'], alpha=0.98) valid_data_miou.attach(validator, 'mIoU') valid_data_mdice = imetrics.RunningAverage( output_transform=lambda x: x['dice'].data_mean()['mean'], alpha=0.98) valid_data_mdice.attach(validator, 'mDice') # show metrics on progress bar (after every iteration) valid_pbar = c_handlers.ProgressBar(persist=True, dynamic_ncols=True) valid_metric_names = ['valid_ra_loss', 'mIoU', 'mDice'] valid_pbar.attach(validator, metric_names=valid_metric_names) # ## monitor ignite IoU (the same as iou we are using) ### # cm = imetrics.ConfusionMatrix(num_classes, # output_transform=lambda x: (x['output'], x['target'])) # imetrics.IoU(cm, # ignore_index=0 # ).attach(validator, 'iou') # # monitor ignite mean iou (over all classes even not exist in gt) # mean_iou = imetrics.mIoU(cm, # ignore_index=0 # ).attach(validator, 'mean_iou') @validator.on(engine.Events.STARTED) def validator_start_callback(engine): pass @validator.on(engine.Events.EPOCH_STARTED) def validator_epoch_start_callback(engine): engine.state.epoch_metrics = { # directly use definition to calculate 'iou': MetricRecord(), 'dice': MetricRecord(), 'confusion_matrix': np.zeros((num_classes, num_classes), dtype=np.uint32), } # evaluate after iter finish @validator.on(engine.Events.ITERATION_COMPLETED) def validator_iter_comp_callback(engine): pass # evaluate after epoch finish @validator.on(engine.Events.EPOCH_COMPLETED) def validator_epoch_comp_callback(engine): # log ignite metrics # logging_logger.info(engine.state.metrics) # ious = engine.state.metrics['iou'] # msg = 'IoU: ' # for ins_id, iou in enumerate(ious): # msg += '{:d}: {:.3f}, '.format(ins_id + 1, iou) # logging_logger.info(msg) # logging_logger.info('nonzero mean IoU for all data: {:.3f}'.format(ious[ious > 0].mean())) # log monitored epoch metrics epoch_metrics = engine.state.epoch_metrics ######### NOTICE: Two metrics are available but different ########## ### 1. mean metrics for all data calculated by confusion matrix #### ''' compared with using confusion_matrix[1:, 1:] in original code, we use the full confusion matrix and only present non-background result ''' confusion_matrix = epoch_metrics['confusion_matrix'] # [1:, 1:] ious = calculate_iou(confusion_matrix) dices = calculate_dice(confusion_matrix) mean_ious = np.mean(list(ious.values())) mean_dices = np.mean(list(dices.values())) std_ious = np.std(list(ious.values())) std_dices = np.std(list(dices.values())) logging_logger.info('mean IoU: %.3f, std: %.3f, for each class: %s' % (mean_ious, std_ious, ious)) logging_logger.info('mean Dice: %.3f, std: %.3f, for each class: %s' % (mean_dices, std_dices, dices)) ### 2. mean metrics for all data calculated by definition ### iou_data_mean = epoch_metrics['iou'].data_mean() dice_data_mean = epoch_metrics['dice'].data_mean() logging_logger.info('data (%d) mean IoU: %.3f, std: %.3f' % (len(iou_data_mean['items']), iou_data_mean['mean'], iou_data_mean['std'])) logging_logger.info('data (%d) mean Dice: %.3f, std: %.3f' % (len(dice_data_mean['items']), dice_data_mean['mean'], dice_data_mean['std'])) # record metrics in trainer every epoch # trainer.state.metrics_records[trainer.state.epoch] = \ # {'miou': mean_ious, 'std_miou': std_ious, # 'mdice': mean_dices, 'std_mdice': std_dices} trainer.state.metrics_records[trainer.state.epoch] = \ {'miou': iou_data_mean['mean'], 'std_miou': iou_data_mean['std'], 'mdice': dice_data_mean['mean'], 'std_mdice': dice_data_mean['std']} # log interal variables(attention maps, outputs, etc.) on validation def tb_log_valid_iter_vars(engine, logger, event_name): log_tag = 'valid_iter' output = engine.state.output batch_size = output['output'].shape[0] res_grid = tvutils.make_grid( torch.cat([ output['output_argmax'].unsqueeze(1), output['target'].unsqueeze(1), ]), padding=2, normalize=False, # show origin image nrow=batch_size).cpu() logger.writer.add_image(tag='%s (outputs, targets)' % (log_tag), img_tensor=res_grid) if 'TAPNet' in args.model: # log attention maps and other internal values inter_vals_grid = tvutils.make_grid(torch.cat([ output['attmap'], ]), padding=2, normalize=True, nrow=batch_size).cpu() logger.writer.add_image(tag='%s internal vals' % (log_tag), img_tensor=inter_vals_grid) def tb_log_valid_epoch_vars(engine, logger, event_name): log_tag = 'valid_iter' # log monitored epoch metrics epoch_metrics = engine.state.epoch_metrics confusion_matrix = epoch_metrics['confusion_matrix'] # [1:, 1:] ious = calculate_iou(confusion_matrix) dices = calculate_dice(confusion_matrix) mean_ious = np.mean(list(ious.values())) mean_dices = np.mean(list(dices.values())) logger.writer.add_scalar('mIoU', mean_ious, engine.state.epoch) logger.writer.add_scalar('mIoU', mean_dices, engine.state.epoch) if args.tb_log: # log internal values tb_logger.attach(validator, log_handler=tb_log_valid_iter_vars, event_name=engine.Events.ITERATION_COMPLETED) tb_logger.attach(validator, log_handler=tb_log_valid_epoch_vars, event_name=engine.Events.EPOCH_COMPLETED) # tb_logger.attach(validator, log_handler=OutputHandler('valid_iter', valid_metric_names), # event_name=engine.Events.ITERATION_COMPLETED) tb_logger.attach(validator, log_handler=OutputHandler('valid_epoch', ['valid_loss']), event_name=engine.Events.EPOCH_COMPLETED) # score function for model saving ckpt_score_function = lambda engine: \ np.mean(list(calculate_iou(engine.state.epoch_metrics['confusion_matrix']).values())) # ckpt_score_function = lambda engine: engine.state.epoch_metrics['iou'].data_mean()['mean'] ckpt_filename_prefix = 'fold_%d' % fold # model saving handler model_ckpt_handler = handlers.ModelCheckpoint( dirname=args.model_save_dir, filename_prefix=ckpt_filename_prefix, score_function=ckpt_score_function, create_dir=True, require_empty=False, save_as_state_dict=True, atomic=True) validator.add_event_handler(event_name=engine.Events.EPOCH_COMPLETED, handler=model_ckpt_handler, to_save={ 'model': model, }) # early stop # trainer=trainer, but should be handled by validator early_stopping = handlers.EarlyStopping(patience=args.es_patience, score_function=ckpt_score_function, trainer=trainer) validator.add_event_handler(event_name=engine.Events.EPOCH_COMPLETED, handler=early_stopping) # evaluate after epoch finish @trainer.on(engine.Events.EPOCH_COMPLETED) def trainer_epoch_comp_callback(engine): validator.run(valid_loader) trainer.run(train_loader, max_epochs=args.max_epochs) if args.tb_log: # close tb_logger tb_logger.close() return trainer.state.metrics_records
def train_fold(fold, args): # loggers logging_logger = args.logging_logger if args.tb_log: tb_logger = args.tb_logger num_classes = utils.problem_class[args.problem_type] # init model model = eval(args.model)(in_channels=3, num_classes=num_classes, bn=False) model = nn.DataParallel(model, device_ids=args.device_ids).cuda() # transform for train/valid data train_transform, valid_transform = get_transform(args.model) # loss function loss_func = LossMulti(num_classes, args.jaccard_weight) if args.semi: loss_func_semi = LossMultiSemi(num_classes, args.jaccard_weight, args.semi_loss_alpha, args.semi_method) # train/valid filenames train_filenames, valid_filenames = utils.trainval_split(args.train_dir, fold) # DataLoader and Dataset args train_shuffle = True train_ds_kwargs = { 'filenames': train_filenames, 'problem_type': args.problem_type, 'transform': train_transform, 'model': args.model, 'mode': 'train', 'semi': args.semi, } valid_num_workers = args.num_workers valid_batch_size = args.batch_size if 'TAPNet' in args.model: # for TAPNet, cancel default shuffle, use self-defined shuffle in torch.Dataset instead train_shuffle = False train_ds_kwargs['batch_size'] = args.batch_size train_ds_kwargs['mf'] = args.mf if args.semi == True: train_ds_kwargs['semi_method'] = args.semi_method train_ds_kwargs['semi_percentage'] = args.semi_percentage # additional valid dataset kws valid_ds_kwargs = { 'filenames': valid_filenames, 'problem_type': args.problem_type, 'transform': valid_transform, 'model': args.model, 'mode': 'valid', } if 'TAPNet' in args.model: # in validation, num_workers should be set to 0 for sequences valid_num_workers = 0 # in validation, batch_size should be set to 1 for sequences valid_batch_size = 1 valid_ds_kwargs['mf'] = args.mf # train dataloader train_loader = DataLoader( dataset=RobotSegDataset(**train_ds_kwargs), shuffle=train_shuffle, # set to False to disable pytorch dataset shuffle num_workers=args.num_workers, batch_size=args.batch_size, pin_memory=True ) # valid dataloader valid_loader = DataLoader( dataset=RobotSegDataset(**valid_ds_kwargs), shuffle=False, # in validation, no need to shuffle num_workers=valid_num_workers, batch_size=valid_batch_size, # in valid time. have to use one image by one pin_memory=True ) # optimizer optimizer = optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay) # optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=0.9, # weight_decay=args.weight_decay, nesterov=True) # ignite trainer process function def train_step(engine, batch): # set model to train model.train() # clear gradients optimizer.zero_grad() # additional params to feed into model add_params = {} inputs = batch['input'].cuda(non_blocking=True) with torch.no_grad(): targets = batch['target'].cuda(non_blocking=True) # for TAPNet, add attention maps if 'TAPNet' in args.model: add_params['attmap'] = batch['attmap'].cuda(non_blocking=True) outputs = model(inputs, **add_params) loss_kwargs = {} if args.semi: loss_kwargs['labeled'] = batch['labeled'] if args.semi_method == 'rev_flow': loss_kwargs['optflow'] = batch['optflow'] loss = loss_func_semi(outputs, targets, **loss_kwargs) else: loss = loss_func(outputs, targets, **loss_kwargs) loss.backward() optimizer.step() return_dict = { 'output': outputs, 'target': targets, 'loss_kwargs': loss_kwargs, 'loss': loss.item(), } # for TAPNet, update attention maps after each iteration if 'TAPNet' in args.model: # output_classes and target_classes: <b, h, w> output_softmax_np = torch.softmax(outputs, dim=1).detach().cpu().numpy() # update attention maps train_loader.dataset.update_attmaps(output_softmax_np, batch['abs_idx'].numpy()) return_dict['attmap'] = add_params['attmap'] return return_dict # init trainer trainer = engine.Engine(train_step) # lr scheduler and handler # cyc_scheduler = optim.lr_scheduler.CyclicLR(optimizer, args.lr / 100, args.lr) # lr_scheduler = c_handlers.param_scheduler.LRScheduler(cyc_scheduler) # trainer.add_event_handler(engine.Events.ITERATION_COMPLETED, lr_scheduler) step_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_decay_epochs, gamma=args.lr_decay) lr_scheduler = c_handlers.param_scheduler.LRScheduler(step_scheduler) trainer.add_event_handler(engine.Events.EPOCH_STARTED, lr_scheduler) @trainer.on(engine.Events.STARTED) def trainer_start_callback(engine): logging_logger.info('training fold {}, {} train / {} valid files'. \ format(fold, len(train_filenames), len(valid_filenames))) # resume training if args.resume: # ckpt for current fold fold_<fold>_model_<epoch>.pth ckpt_dir = Path(args.ckpt_dir) ckpt_filename = ckpt_dir.glob('fold_%d_model_[0-9]*.pth' % fold)[0] res = re.match(r'fold_%d_model_(\d+).pth' % fold, ckpt_filename) # restore epoch engine.state.epoch = int(res.groups()[0]) # load model state dict model.load_state_dict(torch.load(str(ckpt_filename))) logging_logger.info('restore model [{}] from epoch {}.'.format(args.model, engine.state.epoch)) else: logging_logger.info('train model [{}] from scratch'.format(args.model)) # record metrics history every epoch engine.state.metrics_records = {} @trainer.on(engine.Events.EPOCH_STARTED) def trainer_epoch_start_callback(engine): # log learning rate on pbar train_pbar.log_message('model: %s, problem type: %s, fold: %d, lr: %.5f, batch size: %d' % \ (args.model, args.problem_type, fold, lr_scheduler.get_param(), args.batch_size)) # for TAPNet, change dataset schedule to random after the first epoch if 'TAPNet' in args.model and engine.state.epoch > 1: train_loader.dataset.set_dataset_schedule("shuffle") @trainer.on(engine.Events.ITERATION_COMPLETED) def trainer_iter_comp_callback(engine): # logging_logger.info(engine.state.metrics) pass # monitor loss # running average loss train_ra_loss = imetrics.RunningAverage(output_transform= lambda x: x['loss'], alpha=0.98) train_ra_loss.attach(trainer, 'train_ra_loss') # monitor train loss over epoch if args.semi: train_loss = imetrics.Loss(loss_func_semi, output_transform=lambda x: (x['output'], x['target'], x['loss_kwargs'])) else: train_loss = imetrics.Loss(loss_func, output_transform=lambda x: (x['output'], x['target'])) train_loss.attach(trainer, 'train_loss') # progress bar train_pbar = c_handlers.ProgressBar(persist=True, dynamic_ncols=True) train_metric_names = ['train_ra_loss'] train_pbar.attach(trainer, metric_names=train_metric_names) # tensorboardX: log train info if args.tb_log: tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer, 'lr'), event_name=engine.Events.EPOCH_STARTED) tb_logger.attach(trainer, log_handler=OutputHandler('train_iter', train_metric_names), event_name=engine.Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=OutputHandler('train_epoch', ['train_loss']), event_name=engine.Events.EPOCH_COMPLETED) tb_logger.attach(trainer, log_handler=WeightsScalarHandler(model, reduction=torch.norm), event_name=engine.Events.ITERATION_COMPLETED) # tb_logger.attach(trainer, log_handler=tb_log_train_vars, # event_name=engine.Events.ITERATION_COMPLETED) # ignite validator process function def valid_step(engine, batch): with torch.no_grad(): model.eval() inputs = batch['input'].cuda(non_blocking=True) targets = batch['target'].cuda(non_blocking=True) # additional arguments add_params = {} # for TAPNet, add attention maps if 'TAPNet' in args.model: add_params['attmap'] = batch['attmap'].cuda(non_blocking=True) # output logits outputs = model(inputs, **add_params) # loss loss = loss_func(outputs, targets) output_softmaxs = torch.softmax(outputs, dim=1) output_argmaxs = output_softmaxs.argmax(dim=1) # output_classes and target_classes: <b, h, w> output_classes = output_argmaxs.cpu().numpy() target_classes = targets.cpu().numpy() # record current batch metrics iou_mRecords = MetricRecord() dice_mRecords = MetricRecord() cm_b = np.zeros((num_classes, num_classes), dtype=np.uint32) for output_class, target_class in zip(output_classes, target_classes): # calculate metrics for each frame # calculate using confusion matrix or dirctly using definition cm = calculate_confusion_matrix_from_arrays(output_class, target_class, num_classes) iou_mRecords.update_record(calculate_iou(cm)) dice_mRecords.update_record(calculate_dice(cm)) cm_b += cm ######## calculate directly using definition ########## # iou_mRecords.update_record(iou_multi_np(target_class, output_class)) # dice_mRecords.update_record(dice_multi_np(target_class, output_class)) # accumulate batch metrics to engine state engine.state.epoch_metrics['confusion_matrix'] += cm_b engine.state.epoch_metrics['iou'].merge(iou_mRecords) engine.state.epoch_metrics['dice'].merge(dice_mRecords) return_dict = { 'loss': loss.item(), 'output': outputs, 'output_argmax': output_argmaxs, 'target': targets, # for monitoring 'iou': iou_mRecords, 'dice': dice_mRecords, } if 'TAPNet' in args.model: # for TAPNet, update attention maps after each iteration valid_loader.dataset.update_attmaps(output_softmaxs.cpu().numpy(), batch['abs_idx'].numpy()) # for TAPNet, return extra internal values return_dict['attmap'] = add_params['attmap'] # TODO: for TAPNet, return internal self-learned attention maps return return_dict # validator engine validator = engine.Engine(valid_step) # monitor loss valid_ra_loss = imetrics.RunningAverage(output_transform= lambda x: x['loss'], alpha=0.98) valid_ra_loss.attach(validator, 'valid_ra_loss') # monitor validation loss over epoch valid_loss = imetrics.Loss(loss_func, output_transform=lambda x: (x['output'], x['target'])) valid_loss.attach(validator, 'valid_loss') # monitor <data> mean metrics valid_data_miou = imetrics.RunningAverage(output_transform= lambda x: x['iou'].data_mean()['mean'], alpha=0.98) valid_data_miou.attach(validator, 'mIoU') valid_data_mdice = imetrics.RunningAverage(output_transform= lambda x: x['dice'].data_mean()['mean'], alpha=0.98) valid_data_mdice.attach(validator, 'mDice') # show metrics on progress bar (after every iteration) valid_pbar = c_handlers.ProgressBar(persist=True, dynamic_ncols=True) valid_metric_names = ['valid_ra_loss', 'mIoU', 'mDice'] valid_pbar.attach(validator, metric_names=valid_metric_names) # ## monitor ignite IoU (the same as iou we are using) ### # cm = imetrics.ConfusionMatrix(num_classes, # output_transform=lambda x: (x['output'], x['target'])) # imetrics.IoU(cm, # ignore_index=0 # ).attach(validator, 'iou') # # monitor ignite mean iou (over all classes even not exist in gt) # mean_iou = imetrics.mIoU(cm, # ignore_index=0 # ).attach(validator, 'mean_iou') @validator.on(engine.Events.STARTED) def validator_start_callback(engine): pass @validator.on(engine.Events.EPOCH_STARTED) def validator_epoch_start_callback(engine): engine.state.epoch_metrics = { # directly use definition to calculate 'iou': MetricRecord(), 'dice': MetricRecord(), 'confusion_matrix': np.zeros((num_classes, num_classes), dtype=np.uint32), } # evaluate after iter finish @validator.on(engine.Events.ITERATION_COMPLETED) def validator_iter_comp_callback(engine): pass # evaluate after epoch finish @validator.on(engine.Events.EPOCH_COMPLETED) def validator_epoch_comp_callback(engine): # log ignite metrics # logging_logger.info(engine.state.metrics) # ious = engine.state.metrics['iou'] # msg = 'IoU: ' # for ins_id, iou in enumerate(ious): # msg += '{:d}: {:.3f}, '.format(ins_id + 1, iou) # logging_logger.info(msg) # logging_logger.info('nonzero mean IoU for all data: {:.3f}'.format(ious[ious > 0].mean())) # log monitored epoch metrics epoch_metrics = engine.state.epoch_metrics ######### NOTICE: Two metrics are available but different ########## ### 1. mean metrics for all data calculated by confusion matrix #### ''' compared with using confusion_matrix[1:, 1:] in original code, we use the full confusion matrix and only present non-background result ''' confusion_matrix = epoch_metrics['confusion_matrix']# [1:, 1:] ious = calculate_iou(confusion_matrix) dices = calculate_dice(confusion_matrix) mean_ious = np.mean(list(ious.values())) mean_dices = np.mean(list(dices.values())) std_ious = np.std(list(ious.values())) std_dices = np.std(list(dices.values())) logging_logger.info('mean IoU: %.3f, std: %.3f, for each class: %s' % (mean_ious, std_ious, ious)) logging_logger.info('mean Dice: %.3f, std: %.3f, for each class: %s' % (mean_dices, std_dices, dices)) ### 2. mean metrics for all data calculated by definition ### iou_data_mean = epoch_metrics['iou'].data_mean() dice_data_mean = epoch_metrics['dice'].data_mean() logging_logger.info('data (%d) mean IoU: %.3f, std: %.3f' % (len(iou_data_mean['items']), iou_data_mean['mean'], iou_data_mean['std'])) logging_logger.info('data (%d) mean Dice: %.3f, std: %.3f' % (len(dice_data_mean['items']), dice_data_mean['mean'], dice_data_mean['std'])) # record metrics in trainer every epoch # trainer.state.metrics_records[trainer.state.epoch] = \ # {'miou': mean_ious, 'std_miou': std_ious, # 'mdice': mean_dices, 'std_mdice': std_dices} trainer.state.metrics_records[trainer.state.epoch] = \ {'miou': iou_data_mean['mean'], 'std_miou': iou_data_mean['std'], 'mdice': dice_data_mean['mean'], 'std_mdice': dice_data_mean['std']} # log interal variables(attention maps, outputs, etc.) on validation def tb_log_valid_iter_vars(engine, logger, event_name): log_tag = 'valid_iter' output = engine.state.output batch_size = output['output'].shape[0] res_grid = tvutils.make_grid(torch.cat([ output['output_argmax'].unsqueeze(1), output['target'].unsqueeze(1), ]), padding=2, normalize=False, # show origin image nrow=batch_size).cpu() logger.writer.add_image(tag='%s (outputs, targets)' % (log_tag), img_tensor=res_grid) if 'TAPNet' in args.model: # log attention maps and other internal values inter_vals_grid = tvutils.make_grid(torch.cat([ output['attmap'], ]), padding=2, normalize=True, nrow=batch_size).cpu() logger.writer.add_image(tag='%s internal vals' % (log_tag), img_tensor=inter_vals_grid) def tb_log_valid_epoch_vars(engine, logger, event_name): log_tag = 'valid_iter' # log monitored epoch metrics epoch_metrics = engine.state.epoch_metrics confusion_matrix = epoch_metrics['confusion_matrix']# [1:, 1:] ious = calculate_iou(confusion_matrix) dices = calculate_dice(confusion_matrix) mean_ious = np.mean(list(ious.values())) mean_dices = np.mean(list(dices.values())) logger.writer.add_scalar('mIoU', mean_ious, engine.state.epoch) logger.writer.add_scalar('mIoU', mean_dices, engine.state.epoch) if args.tb_log: # log internal values tb_logger.attach(validator, log_handler=tb_log_valid_iter_vars, event_name=engine.Events.ITERATION_COMPLETED) tb_logger.attach(validator, log_handler=tb_log_valid_epoch_vars, event_name=engine.Events.EPOCH_COMPLETED) # tb_logger.attach(validator, log_handler=OutputHandler('valid_iter', valid_metric_names), # event_name=engine.Events.ITERATION_COMPLETED) tb_logger.attach(validator, log_handler=OutputHandler('valid_epoch', ['valid_loss']), event_name=engine.Events.EPOCH_COMPLETED) # score function for model saving ckpt_score_function = lambda engine: \ np.mean(list(calculate_iou(engine.state.epoch_metrics['confusion_matrix']).values())) # ckpt_score_function = lambda engine: engine.state.epoch_metrics['iou'].data_mean()['mean'] ckpt_filename_prefix = 'fold_%d' % fold # model saving handler model_ckpt_handler = handlers.ModelCheckpoint( dirname=args.model_save_dir, filename_prefix=ckpt_filename_prefix, score_function=ckpt_score_function, create_dir=True, require_empty=False, save_as_state_dict=True, atomic=True) validator.add_event_handler(event_name=engine.Events.EPOCH_COMPLETED, handler=model_ckpt_handler, to_save={ 'model': model, }) # early stop # trainer=trainer, but should be handled by validator early_stopping = handlers.EarlyStopping(patience=args.es_patience, score_function=ckpt_score_function, trainer=trainer ) validator.add_event_handler(event_name=engine.Events.EPOCH_COMPLETED, handler=early_stopping) # evaluate after epoch finish @trainer.on(engine.Events.EPOCH_COMPLETED) def trainer_epoch_comp_callback(engine): validator.run(valid_loader) trainer.run(train_loader, max_epochs=args.max_epochs) if args.tb_log: # close tb_logger tb_logger.close() return trainer.state.metrics_records
class Metrics(enum.Enum): train_class_metrics: t.Dict[str, im.Metric] = { 'acc_1': im.RunningAverage(im.Accuracy(output_transform=lambda x: x[1:3])), 'acc_5': im.RunningAverage( im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: x[1:3])), 'ce_loss': im.RunningAverage(output_transform=lambda x: x[0]), 'total_loss': im.RunningAverage(output_transform=lambda x: x[0]) } train_ae_metrics: t.Dict[str, im.Metric] = { 'acc_1': im.RunningAverage( im.Accuracy(output_transform=lambda x: (x[1], x[5]))), 'acc_5': im.RunningAverage( im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: (x[1], x[5]))), 'ce_loss': train_ae_ce_loss, 'l1_loss': train_ae_l1_loss, 'total_loss': train_ae_total_loss } train_gsnn_metrics: t.Dict[str, im.Metric] = { 'acc_1': im.RunningAverage( im.Accuracy( output_transform=lambda x: (x[0].squeeze(dim=1), x[6]))), 'acc_5': im.RunningAverage( im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: (x[0].squeeze(dim=1), x[6]))), 'ce_loss': train_gsnn_ce_loss, 'kld_loss': train_gsnn_kld_loss, 'total_loss': train_gsnn_total_loss, 'kld_factor': train_gsnn_kld_factor, } train_vae_metrics: t.Dict[str, im.Metric] = { 'acc_1': im.RunningAverage( im.Accuracy( output_transform=lambda x: (x[1].squeeze(dim=1), x[7]))), 'acc_5': im.RunningAverage( im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: (x[1].squeeze(dim=1), x[7]))), 'ce_loss': train_vae_ce_loss, 'l1_loss': train_vae_l1_loss, 'kld_loss': train_vae_kld_loss, 'total_loss': train_vae_total_loss, 'kld_factor': train_vae_kld_factor, } eval_class_metrics = { 'acc_1': im.Accuracy(output_transform=lambda x: x[0:2]), 'acc_5': im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: x[0:2]), 'ce_loss': im.Loss(nn.CrossEntropyLoss(), output_transform=lambda x: x[0:2]), 'total_loss': im.Loss(nn.CrossEntropyLoss(), output_transform=lambda x: x[0:2]) } eval_ae_metrics = { 'acc_1': im.Accuracy(output_transform=lambda x: (x[1], x[5])), 'acc_5': im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: (x[1], x[5])), 'ce_loss': eval_ae_loss_metric[0], 'l1_loss': eval_ae_loss_metric[1], 'total_loss': eval_ae_total_loss, } eval_gsnn_metrics = { 'acc_1': im.Accuracy(output_transform=lambda x: (x[-1], x[-2])), 'acc_5': im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: (x[-1], x[-2])), 'ce_loss': eval_gsnn_loss_metric[0], 'kld_loss': eval_gsnn_loss_metric[1], 'total_loss': eval_gsnn_total_loss, } eval_vae_metrics = { 'acc_1': im.Accuracy(output_transform=lambda x: (x[-1], x[-2])), 'acc_5': im.TopKCategoricalAccuracy(k=5, output_transform=lambda x: (x[-1], x[-2])), 'ce_loss': eval_vae_loss_metric[0], 'l1_loss': eval_vae_loss_metric[1], 'kld_loss': eval_vae_loss_metric[2], 'total_loss': eval_vae_total_loss, }
import enum import typing as t import ignite.metrics as im import torch.nn as nn import criterion.custom as cc import metrics.custom as cm import models as mo ######################################################################################################################## # AE HELPERS ######################################################################################################################## train_ae_ce_loss = im.RunningAverage(output_transform=lambda x: x[-2]) train_ae_l1_loss = im.RunningAverage(output_transform=lambda x: x[-1]) train_ae_total_loss = im.RunningAverage( output_transform=lambda x: sum([x[-2], x[-1]])) eval_ae_loss_metric = cm.AELoss(cc.AECriterion()) eval_ae_total_loss = im.MetricsLambda(lambda x: sum(x), eval_ae_loss_metric) ######################################################################################################################## # GSNN HELPERS ######################################################################################################################## train_gsnn_ce_loss = im.RunningAverage(output_transform=lambda x: x[-3]) train_gsnn_kld_loss = im.RunningAverage(output_transform=lambda x: x[-2]) train_gsnn_kld_factor = im.RunningAverage(output_transform=lambda x: x[-1]) train_gsnn_total_loss = im.RunningAverage( output_transform=lambda x: sum([x[-3], x[-2]])) eval_gsnn_loss_metric = cm.GSNNLoss(cc.GSNNCriterion()) eval_gsnn_total_loss = im.MetricsLambda(lambda x: sum(x), eval_gsnn_loss_metric)
def run(experiment_name: str, visdom_host: str, visdom_port: int, visdom_env_path: str, model_class: str, model_args: Dict[str, Any], optimizer_class: str, optimizer_args: Dict[str, Any], dataset_class: str, dataset_args: Dict[str, Any], batch_train: int, batch_test: int, workers_train: int, workers_test: int, transforms: List[Dict[str, Union[str, Dict[str, Any]]]], epochs: int, log_interval: int, saved_models_path: str, performance_metrics: Optional = None, scheduler_class: Optional[str] = None, scheduler_args: Optional[Dict[str, Any]] = None, model_suffix: Optional[str] = None, setup_suffix: Optional[str] = None, orig_stdout: Optional[io.TextIOBase] = None): with _utils.tqdm_stdout(orig_stdout) as orig_stdout: device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') transforms_train = list() transforms_test = list() for idx, transform in enumerate(transforms): use_train = transform.get('train', True) use_test = transform.get('test', True) transform = _utils.load_class( transform['class'])(**transform['args']) if use_train: transforms_train.append(transform) if use_test: transforms_test.append(transform) transforms[idx]['train'] = use_train transforms[idx]['test'] = use_test transforms_train = tv.transforms.Compose(transforms_train) transforms_test = tv.transforms.Compose(transforms_test) Dataset: Type = _utils.load_class(dataset_class) train_loader, eval_loader = _utils.get_data_loaders( Dataset, dataset_args, batch_train, batch_test, workers_train, workers_test, transforms_train, transforms_test) Network: Type = _utils.load_class(model_class) model: _interfaces.AbstractNet = Network(**model_args) model = model.to(device) Optimizer: Type = _utils.load_class(optimizer_class) optimizer: torch.optim.Optimizer = Optimizer(model.parameters(), **optimizer_args) if scheduler_class is not None: Scheduler: Type = _utils.load_class(scheduler_class) if scheduler_args is None: scheduler_args = dict() scheduler: Optional[ torch.optim.lr_scheduler._LRScheduler] = Scheduler( optimizer, **scheduler_args) else: scheduler = None model_short_name = ''.join( [c for c in Network.__name__ if c == c.upper()]) model_name = '{}{}'.format( model_short_name, '-{}'.format(model_suffix) if model_suffix is not None else '') visdom_env_name = '{}_{}_{}{}'.format( Dataset.__name__, experiment_name, model_name, '-{}'.format(setup_suffix) if setup_suffix is not None else '') vis, vis_pid = _visdom.get_visdom_instance(visdom_host, visdom_port, visdom_env_name, visdom_env_path) prog_bar_epochs = tqdm.tqdm(total=epochs, desc='Epochs', file=orig_stdout, dynamic_ncols=True, unit='epoch') prog_bar_iters = tqdm.tqdm(desc='Batches', file=orig_stdout, dynamic_ncols=True) tqdm.tqdm.write(f'\n{repr(model)}\n') tqdm.tqdm.write('Total number of parameters: {:.2f}M'.format( sum(p.numel() for p in model.parameters()) / 1e6)) def training_step(_: ieng.Engine, batch: _interfaces.TensorPair) -> torch.Tensor: model.train() optimizer.zero_grad() x, y = batch x = x.to(device) y = y.to(device) _, loss = model(x, y) loss.backward(retain_graph=False) optimizer.step(None) return loss.item() def eval_step(_: ieng.Engine, batch: _interfaces.TensorPair) -> _interfaces.TensorPair: model.eval() with torch.no_grad(): x, y = batch x = x.to(device) y = y.to(device) y_pred = model(x) return y_pred, y trainer = ieng.Engine(training_step) validator_train = ieng.Engine(eval_step) validator_eval = ieng.Engine(eval_step) # placeholder for summary window vis.text(text='', win=experiment_name, env=visdom_env_name, opts={ 'title': 'Summary', 'width': 940, 'height': 416 }, append=vis.win_exists(experiment_name, visdom_env_name)) default_metrics = { "Loss": { "window_name": None, "x_label": "#Epochs", "y_label": model.loss_fn_name, "width": 940, "height": 416, "lines": [{ "line_label": "SMA", "object": imet.RunningAverage(output_transform=lambda x: x), "test": False, "update_rate": "iteration" }, { "line_label": "Val.", "object": imet.Loss(model.loss_fn) }] } } performance_metrics = {**default_metrics, **performance_metrics} checkpoint_metrics = list() for scope_name, scope in performance_metrics.items(): scope['window_name'] = scope.get('window_name', scope_name) or scope_name for line in scope['lines']: if 'object' not in line: line['object']: imet.Metric = _utils.load_class( line['class'])(**line['args']) line['metric_label'] = '{}: {}'.format(scope['window_name'], line['line_label']) line['update_rate'] = line.get('update_rate', 'epoch') line_suffixes = list() if line['update_rate'] == 'iteration': line['object'].attach(trainer, line['metric_label']) line['train'] = False line['test'] = False line_suffixes.append(' Train.') if line.get('train', True): line['object'].attach(validator_train, line['metric_label']) line_suffixes.append(' Train.') if line.get('test', True): line['object'].attach(validator_eval, line['metric_label']) line_suffixes.append(' Eval.') if line.get('is_checkpoint', False): checkpoint_metrics.append(line['metric_label']) for line_suffix in line_suffixes: _visdom.plot_line( vis=vis, window_name=scope['window_name'], env=visdom_env_name, line_label=line['line_label'] + line_suffix, x_label=scope['x_label'], y_label=scope['y_label'], width=scope['width'], height=scope['height'], draw_marker=(line['update_rate'] == 'epoch')) if checkpoint_metrics: score_name = 'performance' def get_score(engine: ieng.Engine) -> float: current_mode = getattr( engine.state.dataloader.iterable.dataset, dataset_args['training']['key']) val_mode = dataset_args['training']['no'] score = 0.0 if current_mode == val_mode: for metric_name in checkpoint_metrics: try: score += engine.state.metrics[metric_name] except KeyError: pass return score model_saver = ihan.ModelCheckpoint(os.path.join( saved_models_path, visdom_env_name), filename_prefix=visdom_env_name, score_name=score_name, score_function=get_score, n_saved=3, save_as_state_dict=True, require_empty=False, create_dir=True) validator_eval.add_event_handler(ieng.Events.EPOCH_COMPLETED, model_saver, {model_name: model}) @trainer.on(ieng.Events.EPOCH_STARTED) def reset_progress_iterations(engine: ieng.Engine): prog_bar_iters.clear() prog_bar_iters.n = 0 prog_bar_iters.last_print_n = 0 prog_bar_iters.start_t = time.time() prog_bar_iters.last_print_t = time.time() prog_bar_iters.total = len(engine.state.dataloader) @trainer.on(ieng.Events.ITERATION_COMPLETED) def log_training(engine: ieng.Engine): prog_bar_iters.update(1) num_iter = (engine.state.iteration - 1) % len(train_loader) + 1 early_stop = np.isnan(engine.state.output) or np.isinf( engine.state.output) if num_iter % log_interval == 0 or num_iter == len( train_loader) or early_stop: tqdm.tqdm.write( 'Epoch[{}] Iteration[{}/{}] Loss: {:.4f}'.format( engine.state.epoch, num_iter, len(train_loader), engine.state.output)) x_pos = engine.state.epoch + num_iter / len(train_loader) - 1 for scope_name, scope in performance_metrics.items(): for line in scope['lines']: if line['update_rate'] == 'iteration': line_label = '{} Train.'.format(line['line_label']) line_value = engine.state.metrics[ line['metric_label']] if engine.state.epoch > 1: _visdom.plot_line( vis=vis, window_name=scope['window_name'], env=visdom_env_name, line_label=line_label, x_label=scope['x_label'], y_label=scope['y_label'], x=np.full(1, x_pos), y=np.full(1, line_value)) if early_stop: tqdm.tqdm.write( colored('Early stopping due to invalid loss value.', 'red')) trainer.terminate() def log_validation(engine: ieng.Engine, train: bool = True): if train: run_type = 'Train.' data_loader = train_loader validator = validator_train else: run_type = 'Eval.' data_loader = eval_loader validator = validator_eval prog_bar_validation = tqdm.tqdm(data_loader, desc=f'Validation {run_type}', file=orig_stdout, dynamic_ncols=True, leave=False) validator.run(prog_bar_validation) prog_bar_validation.clear() prog_bar_validation.close() tqdm_info = ['Epoch: {}'.format(engine.state.epoch)] for scope_name, scope in performance_metrics.items(): for line in scope['lines']: if line['update_rate'] == 'epoch': try: line_label = '{} {}'.format( line['line_label'], run_type) line_value = validator.state.metrics[ line['metric_label']] _visdom.plot_line(vis=vis, window_name=scope['window_name'], env=visdom_env_name, line_label=line_label, x_label=scope['x_label'], y_label=scope['y_label'], x=np.full(1, engine.state.epoch), y=np.full(1, line_value), draw_marker=True) tqdm_info.append('{}: {:.4f}'.format( line_label, line_value)) except KeyError: pass tqdm.tqdm.write('{} results - {}'.format(run_type, '; '.join(tqdm_info))) @trainer.on(ieng.Events.EPOCH_COMPLETED) def log_validation_train(engine: ieng.Engine): log_validation(engine, True) @trainer.on(ieng.Events.EPOCH_COMPLETED) def log_validation_eval(engine: ieng.Engine): log_validation(engine, False) if engine.state.epoch == 1: summary = _utils.build_summary_str( experiment_name=experiment_name, model_short_name=model_name, model_class=model_class, model_args=model_args, optimizer_class=optimizer_class, optimizer_args=optimizer_args, dataset_class=dataset_class, dataset_args=dataset_args, transforms=transforms, epochs=epochs, batch_train=batch_train, log_interval=log_interval, saved_models_path=saved_models_path, scheduler_class=scheduler_class, scheduler_args=scheduler_args) _visdom.create_summary_window(vis=vis, visdom_env_name=visdom_env_name, experiment_name=experiment_name, summary=summary) vis.save([visdom_env_name]) prog_bar_epochs.update(1) if scheduler is not None: scheduler.step(engine.state.epoch) trainer.run(train_loader, max_epochs=epochs) if vis_pid is not None: tqdm.tqdm.write('Stopping visdom') os.kill(vis_pid, signal.SIGTERM) del vis del train_loader del eval_loader prog_bar_iters.clear() prog_bar_iters.close() prog_bar_epochs.clear() prog_bar_epochs.close() tqdm.tqdm.write('\n')
def _stacked_ae_fit(stacked_ae, train_data, val_data, batch_size, optimizer, max_epochs, early_stopping=False, check_every=100, patience=10, writer=None): """ Function used to fit a stacked autoencoder on data. MSELoss used as lossfunction. ### Parameters - *ae*: Autoencoder, required + The autoencoder to fit. - *train_data*: torch.Tensor, required + Dataset with the data to train on where samples are in rows - *val_data*: torch.Tensor, required + Dataset with the data to validate on where samples are in rows - *optimizer*: torch.optim.Optimizer, required - *batch_size*: int, required - *max_epochs*: int, required + Training is at most run for this number of epochs - *early_stopping*: bool, required + Whether to use early stopping or not - *check_every*: int, optional, default=1500 + Scores are calculated every *check_every* iteration. - *patience*: int, optional, default=10 + If the validation score has not increased after *patience* checks and early stopping is True, training is stopped. ### Returns - *time*: Time taken for fitting """ start_time = time() num_nets = stacked_ae.num_nets train_loss = StackedMSELoss(size_average=None, reduce=None, reduction='elementwise_mean') val_loss = StackedMSELoss(size_average=None, reduce=None, reduction='stacked_elementwise_mean') train_dataset = StackedTensorDataset(train_data, train_data, shuffle=True) train_loader = StackedDataLoader(train_dataset, batch_size=batch_size, shuffle=True) val_dataset = StackedTensorDataset(val_data, val_data, shuffle=False) val_loader = StackedDataLoader(val_dataset, batch_size=4 * batch_size, shuffle=False) logger = logging.getLogger('stacked_ae_fit') trainer = create_supervised_trainer(stacked_ae, optimizer, train_loss) run_av = metrics.RunningAverage( output_transform=lambda x: x / len(stacked_ae.net_list), alpha=0.9) run_av.attach(trainer, 'av_loss') pbar_train = ProgressBar() pbar_train.attach(trainer, ['av_loss']) val_loss_metric = StackedLossMetric(val_loss) evaluator = create_supervised_evaluator( stacked_ae, metrics={'stacked_loss': val_loss_metric}) def score_function(evaluator): return -evaluator.state.metrics['stacked_loss'] evaluator.register_events(*EarlyStoppingEvents) evaluator.add_event_handler( Events.COMPLETED, StackedEarlyStopping(patience, score_function, stacked_ae.num_nets)) evaluator.num_converged = 0 @evaluator.on(EarlyStoppingEvents.NET_CONVERGED) def update_count(evaluator): evaluator.num_converged += 1 if early_stopping: evaluator.best_state = {} for name, tensor in stacked_ae.state_dict().items(): # Create a new instance of the tensor on the cpu evaluator.best_state[name] = torch.tensor(tensor, device='cpu') @evaluator.on(EarlyStoppingEvents.HAS_IMPROVED) def store_states(evaluator): for name, tensor in stacked_ae.state_dict().items(): # Create a new instance of the tensor on the cpu evaluator.best_state[name] = torch.tensor(tensor, device='cpu') @evaluator.on(EarlyStoppingEvents.EARLY_STOPPING_DONE) def terminate(evaluator): trainer.terminate() class CustomEvents(Enum): N_ITERATIONS_COMPLETED = "n_iterations_completed" trainer.register_events(CustomEvents['N_ITERATIONS_COMPLETED']) @trainer.on(Events.ITERATION_COMPLETED, check_every) def every_n(trainer, check_every): if trainer.state.iteration % check_every == 0: trainer.fire_event(CustomEvents.N_ITERATIONS_COMPLETED) @trainer.on(CustomEvents.N_ITERATIONS_COMPLETED) def run_validation(trainer): iteration = trainer.state.iteration epoch = trainer.state.epoch evaluator.run(val_loader, max_epochs=1) msg = (("Iteration {} (Epoch {})" + "training error: {:.5f}, " + "mean validation error: {:.5f}, " + "{}/{} nets convered.").format( iteration, epoch, trainer.state.metrics['av_loss'], evaluator.state.metrics['stacked_loss'].mean(), evaluator.num_converged, num_nets)) logger.info(msg) # clear some memory evaluator.state.output = None if writer is not None: writer.add_scalar('train_error_estimate', trainer.state.output, iteration) # Write to tensorboard for i, loss in enumerate(evaluator.state.metrics['stacked_loss']): if writer is not None: writer.add_scalar('val_error_{}'.format(i), loss, iteration) trainer.run(train_loader, max_epochs=max_epochs) if early_stopping: # Load the best states stacked_ae.load_state_dict(evaluator.best_state) stacked_ae.update_list() train_time = time() - start_time if early_stopping and evaluator.num_converged < num_nets: msg = ("Only {}/{} nets converged".format(evaluator.num_converged, num_nets)) logger.warning(msg) msg = ('Trained for {} epochs, {}/{} nets converged.' '\n Training took {}s'.format(trainer.state.epoch, evaluator.num_converged, num_nets, train_time)) logger.info(msg) return (train_time)
def fit(self, train_loader: _data.DataLoader, val_loader: _data.DataLoader, epochs: int = 1, batches: int = None, learning_rate: float = 1e-3) -> None: if batches is None: batches = VocalExtractor.get_number_of_batches(train_loader) loss_fn = nn.BCELoss() optimizer = _optim.Adam(self.model.parameters(), lr=learning_rate) trainer = _engine.create_supervised_trainer(self.model, optimizer, loss_fn, device=self.device) _metrics.RunningAverage(output_transform=lambda x: x, device=self.device).attach(trainer, 'loss') progressbar = _chandlers.ProgressBar( bar_format= "{desc}[{n_fmt}/{total_fmt}] {percentage:3.0f}%|{bar:20}| " "[{elapsed}<{remaining}]{postfix}", persist=True, ascii=" #") progressbar.attach(trainer, ['loss']) def get_metrics_fn() -> Dict[str, _metrics.Metric]: def rounded_transform(output): y_pred, y = output return torch.round(y_pred), y transform = rounded_transform accuracy = _metrics.Accuracy(transform, device=self.device) precision = _metrics.Precision(transform, device=self.device) recall = _metrics.Recall(transform, device=self.device) f1 = precision * recall * 2 / (precision + recall + 1e-20) return { 'loss': _metrics.Loss(loss_fn), 'accuracy': accuracy, 'precision': precision, 'recall': recall, 'f1': f1 } evaluator = _engine.create_supervised_evaluator( self.model, metrics=get_metrics_fn(), device=self.device) score_fn_name = "f1" def score_function(engine: _engine.Engine): return engine.state.metrics[score_fn_name] best_model_saver = _handlers.ModelCheckpoint( dirname="best_models", filename_prefix="vocal_extractor", score_name=score_fn_name, score_function=score_function, n_saved=5, create_dir=True) evaluator.add_event_handler(_engine.Events.COMPLETED, best_model_saver, {"model": self.model}) each_model_saver = _handlers.ModelCheckpoint( dirname="all_models", filename_prefix="vocal_extractor", score_name=score_fn_name, score_function=score_function, n_saved=None, create_dir=True) evaluator.add_event_handler(_engine.Events.COMPLETED, each_model_saver, {"model": self.model}) @trainer.on(_engine.Events.EPOCH_COMPLETED) def on_epoch_completed(engine: _engine.Engine) -> None: metrics = VocalExtractor.compute_metrics(val_loader, evaluator) string = ", ".join(f"val_{k}: {v:.4f}" for k, v in metrics.items()) progressbar.log_message(string + "\n") with _tb_logger.TensorboardLogger(log_dir="tb_logs") as tb_logger: global_step = _tb_logger.global_step_from_engine(trainer) train_running_loss_log_handler = _tb_logger.OutputHandler( tag="training", output_transform=lambda x: {'running_loss': x}) tb_logger.attach(trainer, log_handler=train_running_loss_log_handler, event_name=_engine.Events.ITERATION_COMPLETED) val_metrics_log_handler = _tb_logger.OutputHandler( tag="validation", metric_names=[name for name, _ in get_metrics_fn().items()], global_step_transform=global_step) tb_logger.attach(evaluator, log_handler=val_metrics_log_handler, event_name=_engine.Events.EPOCH_COMPLETED) tb_logger.attach( trainer, log_handler=_tb_logger.OptimizerParamsHandler(optimizer), event_name=_engine.Events.ITERATION_STARTED) tb_logger.attach(trainer, log_handler=_tb_logger.WeightsScalarHandler( self.model), event_name=_engine.Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=_tb_logger.WeightsHistHandler( self.model), event_name=_engine.Events.EPOCH_COMPLETED) tb_logger.attach(trainer, log_handler=_tb_logger.GradsScalarHandler( self.model), event_name=_engine.Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=_tb_logger.GradsHistHandler( self.model), event_name=_engine.Events.EPOCH_COMPLETED) torchsummary.summary(self.model, input_size=(1, self.freq_bins, self.time_bins), batch_size=train_loader.batch_size, device=self.device) trainer.run(data=train_loader, epoch_length=batches, max_epochs=epochs)