def save_model(args, epoch, model, model_without_ddp, optimizer, loss_scaler, model_ema=None): output_dir = Path(args.output_dir) epoch_name = str(epoch) if loss_scaler is not None: checkpoint_paths = [output_dir / ('checkpoint-%s.pth' % epoch_name)] for checkpoint_path in checkpoint_paths: to_save = { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'epoch': epoch, 'scaler': loss_scaler.state_dict(), 'args': args, } if model_ema is not None: to_save['model_ema'] = get_state_dict(model_ema) save_on_master(to_save, checkpoint_path) else: client_state = {'epoch': epoch} if model_ema is not None: client_state['model_ema'] = get_state_dict(model_ema) model.save_checkpoint(save_dir=args.output_dir, tag="checkpoint-%s" % epoch_name, client_state=client_state)
def __getstate__(self): d = self.__dict__.copy() try: import torch net = d.pop('net', None) model_ema = d.pop('_model_ema', None) optimizer = d.pop('_optimizer', None) loss_scaler = d.pop('_loss_scaler', None) save_state = {} if net is not None: if not self._custom_net: if isinstance(net, torch.nn.DataParallel): save_state['state_dict'] = get_state_dict(net.module, unwrap_model) else: save_state['state_dict'] = get_state_dict(net, unwrap_model) else: net_pickle = pickle.dumps(net) save_state['net_pickle'] = net_pickle if optimizer is not None: save_state['optimizer'] = optimizer.state_dict() if loss_scaler is not None: save_state[loss_scaler.state_dict_key] = loss_scaler.state_dict() if model_ema is not None: save_state['state_dict_ema'] = get_state_dict(model_ema, unwrap_model) except ImportError: pass d['save_state'] = save_state d['_logger'] = None d['_reporter'] = None return d
def save(self, filename): d = dict() current_states = self.__dict__.copy() if self.net: if not self._custom_net: if isinstance(self.net, torch.nn.DataParallel): d['model_state_dict'] = get_state_dict(self.net.module, unwrap_model) else: d['model_state_dict'] = get_state_dict(self.net, unwrap_model) else: net_pickle = pickle.dumps(self.net) d['net_pickle'] = net_pickle self.net = None if self._optimizer: d['optimizer_state_dict'] = self._optimizer.state_dict() self._optimizer = None if hasattr(self, '_loss_scaler') and self._loss_scaler: d[self._loss_scaler.state_dict_key] = self._loss_scaler.state_dict() d['_loss_scaler_state_dict_key'] = self._loss_scaler.state_dict_key if self._model_ema: d['ema_state_dict'] = get_state_dict(self._model_ema, unwrap_model) self._model_ema = None self._logger = None self._reporter = None d['estimator'] = self torch.save(d, filename) self.__dict__.update(current_states)
def save_checkpoint( epoch: int, model: nn.Module, optimizer: optim.Optimizer, models_path: Path, exp_name: str, epoch_metrics: Dict[str, float], model_ema: Optional[Any], amp_scaler: Optional[Any], scheduler: Optional[Any] = None, ) -> None: save_state = { "epoch": epoch + 1, # increment epoch (to not repeat then resume) "state_dict": get_state_dict(model, unwrap_model), "optimizer": optimizer.state_dict(), "val_loss": epoch_metrics["val_loss"], "val_score": epoch_metrics["val_score"], "threshold": epoch_metrics["threshold"], } if model_ema is not None: save_state["state_dict_ema"] = get_state_dict(model_ema, unwrap_model) if amp_scaler is not None: save_state[amp_scaler.state_dict_key] = amp_scaler.state_dict() if scheduler is not None: save_state["lr_scheduler"] = scheduler.state_dict() torch.save( save_state, f"{models_path}/{exp_name}.pth", )
def save_checkpoint( epoch: int, model: torch.nn.Module, optimizer: torch.optim.Optimizer, models_path: Path, exp_name: str, ) -> None: save_state = { "epoch": epoch + 1, # increment epoch (to not repeat then resume) "state_dict": get_state_dict(model, unwrap_model), "optimizer": optimizer.state_dict(), } torch.save( save_state, f"{models_path}/{exp_name}.pth", )
def main(): """Load the pretrained moco model from args.moco_path and save both moco encoders to the Moco project models folder""" # !WIP args = parser.parse_args() checkpoint = Path(args.moco_path) assert checkpoint.exists() model = ModelMoCo( dim=Config["moco_dim"], K=Config["moco_K"], m=Config["moco_m"], T=Config["moco_T"], arch=Config["moco_arch"], ) _ = resume_checkpoint(model, checkpoint) encoder_q = model.encoder_q.net encoder_q.reset_classifier(11) # hard code save_state = get_state_dict(encoder_q, unwrap_model) torch.save( save_state, f"{MODELS_PATH}/{checkpoint.stem}_q.pth", )
def main(args): utils.init_distributed_mode(args) print(args) if args.distillation_type != 'none' and args.finetune and not args.eval: raise NotImplementedError( "Finetuning with distillation not yet supported") device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) if args.dist_eval: if len(dataset_val) % num_tasks != 0: print( 'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' 'This will slightly alter validation results as extra duplicate entries are added to achieve ' 'equal num of samples per-process.') sampler_val = torch.utils.data.DistributedSampler( dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=int( 1.5 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, drop_block_rate=None, ) if args.finetune: if args.finetune.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.finetune, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.finetune, map_location='cpu') checkpoint_model = checkpoint['model'] state_dict = model.state_dict() for k in [ 'head.weight', 'head.bias', 'head_dist.weight', 'head_dist.bias' ]: if k in checkpoint_model and checkpoint_model[ k].shape != state_dict[k].shape: print(f"Removing key {k} from pretrained checkpoint") del checkpoint_model[k] # interpolate position embedding pos_embed_checkpoint = checkpoint_model['pos_embed'] embedding_size = pos_embed_checkpoint.shape[-1] num_patches = model.patch_embed.num_patches num_extra_tokens = model.pos_embed.shape[-2] - num_patches # height (== width) for the checkpoint position embedding orig_size = int( (pos_embed_checkpoint.shape[-2] - num_extra_tokens)**0.5) # height (== width) for the new position embedding new_size = int(num_patches**0.5) # class_token and dist_token are kept unchanged extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate(pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model['pos_embed'] = new_pos_embed model.load_state_dict(checkpoint_model, strict=False) model.to(device) model_ema = None if args.model_ema: # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper model_ema = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() teacher_model = None if args.distillation_type != 'none': assert args.teacher_path, 'need to specify teacher-path when using distillation' print(f"Creating teacher model: {args.teacher_model}") teacher_model = create_model( args.teacher_model, pretrained=False, num_classes=args.nb_classes, global_pool='avg', ) if args.teacher_path.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.teacher_path, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.teacher_path, map_location='cpu') teacher_model.load_state_dict(checkpoint['model']) teacher_model.to(device) teacher_model.eval() # wrap the criterion in our custom DistillationLoss, which # just dispatches to the original criterion if args.distillation_type is 'none' criterion = DistillationLoss(criterion, teacher_model, args.distillation_type, args.distillation_alpha, args.distillation_tau) output_dir = Path(args.output_dir) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn, set_training_mode=args.finetune == '' # keep in eval mode during finetuning ) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / ('checkpoint_%04d.pth' % (epoch))] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) if not args.train_without_eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } else: log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(args): utils.init_distributed_mode(args) print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader(dataset_val, batch_size=int( 1.5 * args.batch_size), shuffle=False, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, drop_block_rate=args.drop_block, ) # TODO: finetuning model.to(device) model_ema = None if args.model_ema: # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper model_ema = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() output_dir = Path(args.output_dir) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print("Start training") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) train_stats = train_one_epoch(model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'args': args, }, checkpoint_path) test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(args): utils.init_distributed_mode(args) # disable any harsh augmentation in case of Self-supervise training if args.training_mode == 'SSL': print("NOTE: Smoothing, Mixup, CutMix, and AutoAugment will be disabled in case of Self-supervise training") args.smoothing = args.reprob = args.reprob = args.recount = args.mixup = args.cutmix = 0.0 args.aa = '' if args.SiT_LinearEvaluation == 1: print("Warning: Linear Evaluation should be set to 0 during SSL training - changing SiT_LinearEvaluation to 0") args.SiT_LinearEvaluation = 0 utils.print_args(args) device = torch.device(args.device) seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) cudnn.benchmark = True print("Loading dataset ....") dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) sampler_val = torch.utils.data.SequentialSampler(dataset_val) data_loader_train = torch.utils.data.DataLoader(dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, collate_fn=collate_fn) data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=int(1.5 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False, collate_fn=collate_fn) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup( mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, representation_size=args.representation_size, drop_block_rate=None, training_mode=args.training_mode) if args.finetune: checkpoint = torch.load(args.finetune, map_location='cpu') checkpoint_model = checkpoint['model'] state_dict = model.state_dict() for k in ['rot_head.weight', 'rot_head.bias', 'contrastive_head.weight', 'contrastive_head.bias']: if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape: print(f"Removing key {k} from pretrained checkpoint") del checkpoint_model[k] # interpolate position embedding pos_embed_checkpoint = checkpoint_model['pos_embed'] embedding_size = pos_embed_checkpoint.shape[-1] num_patches = model.patch_embed.num_patches num_extra_tokens = model.pos_embed.shape[-2] - num_patches orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5) new_size = int(num_patches ** 0.5) extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model['pos_embed'] = new_pos_embed model.load_state_dict(checkpoint_model, strict=False) model.to(device) # Freeze the backbone in case of linear evaluation if args.SiT_LinearEvaluation == 1: requires_grad(model, False) model.rot_head.weight.requires_grad = True model.rot_head.bias.requires_grad = True model.contrastive_head.weight.requires_grad = True model.contrastive_head.bias.requires_grad = True if args.representation_size is not None: model.pre_logits_rot.fc.weight.requires_grad = True model.pre_logits_rot.fc.bias.requires_grad = True model.pre_logits_contrastive.fc.weight.requires_grad = True model.pre_logits_contrastive.fc.bias.requires_grad = True model_ema = None if args.model_ema: model_ema = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size() / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) if args.training_mode == 'SSL': criterion = MTL_loss(args.device, args.batch_size) elif args.training_mode == 'finetune' and args.mixup > 0.: criterion = SoftTargetCrossEntropy() else: criterion = torch.nn.CrossEntropyLoss() output_dir = Path(args.output_dir) if args.resume: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.eval: test_stats = evaluate_SSL(data_loader_val, model, device) print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) if args.training_mode == 'SSL': train_stats = train_SSL( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn) else: train_stats = train_finetune( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn) lr_scheduler.step(epoch) if epoch%args.validate_every == 0: if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master({ 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) if args.training_mode == 'SSL': test_stats = evaluate_SSL(data_loader_val, model, device, epoch, args.output_dir) else: test_stats = evaluate_finetune(data_loader_val, model, device) print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters} if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))