def get_train_transform(resizing='default', random_horizontal_flip=True, random_color_jitter=True, random_gray_scale=True): """ resizing mode: - default: random resized crop with scale factor(0.7, 1.0) and size 224; - cen.crop: take the center crop of 224; - res.|cen.crop: resize the image to 256 and take the center crop of size 224; - res: resize the image to 224; - res2x: resize the image to 448; - res.|crop: resize the image to 256 and take a random crop of size 224; - res.sma|crop: resize the image keeping its aspect ratio such that the smaller side is 256, then take a random crop of size 224; – inc.crop: “inception crop” from (Szegedy et al., 2015); – cif.crop: resize the image to 224, zero-pad it by 28 on each side, then take a random crop of size 224. """ if resizing == 'default': transform = T.RandomResizedCrop(224, scale=(0.7, 1.0)) elif resizing == 'cen.crop': transform = T.CenterCrop(224) elif resizing == 'res.|cen.crop': transform = T.Compose([ResizeImage(256), T.CenterCrop(224)]) elif resizing == 'res': transform = ResizeImage(224) elif resizing == 'res2x': transform = ResizeImage(448) elif resizing == 'res.|crop': transform = T.Compose([T.Resize((256, 256)), T.RandomCrop(224)]) elif resizing == "res.sma|crop": transform = T.Compose([T.Resize(256), T.RandomCrop(224)]) elif resizing == 'inc.crop': transform = T.RandomResizedCrop(224) elif resizing == 'cif.crop': transform = T.Compose([ T.Resize((224, 224)), T.Pad(28), T.RandomCrop(224), ]) else: raise NotImplementedError(resizing) transforms = [transform] if random_horizontal_flip: transforms.append(T.RandomHorizontalFlip()) if random_color_jitter: transforms.append( T.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.3)) if random_gray_scale: transforms.append(T.RandomGrayscale()) transforms.extend([ T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) return T.Compose(transforms)
def get_val_transform(resizing='default'): """ resizing mode: - default: resize the image to 256 and take the center crop of size 224; – res.: resize the image to 224 – res.|crop: resize the image such that the smaller side is of size 256 and then take a central crop of size 224. """ if resizing == 'default': transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), ]) elif resizing == 'res.': transform = T.Resize((224, 224)) elif resizing == 'res.|crop': transform = T.Compose([ T.Resize(256), T.CenterCrop(224), ]) else: raise NotImplementedError(resizing) return T.Compose([ transform, T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ])
def get_val_transform(resizing='default', resize_size=224, norm_mean=(0.485, 0.456, 0.406), norm_std=(0.229, 0.224, 0.225)): """ resizing mode: - default: resize the image to 256 and take the center crop of size 224; – res.: resize the image to 224 """ if resizing == 'default': transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), ]) elif resizing == 'res.': transform = ResizeImage(resize_size) else: raise NotImplementedError(resizing) return T.Compose([ transform, T.ToTensor(), T.Normalize(mean=norm_mean, std=norm_std) ])
def get_val_transform(resizing='default'): """ resizing mode: - default: resize the image to 224; - res2x: resize the image to 448; - res.|cen.crop: resize the image to 256 and take the center crop of size 224; """ if resizing == 'default': transform = ResizeImage(224) elif resizing == 'res2x': transform = ResizeImage(448) elif resizing == 'res.|cen.crop': transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), ]) else: raise NotImplementedError(resizing) return T.Compose([ transform, T.ToTensor(), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ])
def get_train_transform(resizing='default', random_horizontal_flip=True, random_color_jitter=False, resize_size=224, norm_mean=(0.485, 0.456, 0.406), norm_std=(0.229, 0.224, 0.225)): """ resizing mode: - default: resize the image to 256 and take a random resized crop of size 224; - cen.crop: resize the image to 256 and take the center crop of size 224; - res: resize the image to 224; """ if resizing == 'default': transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224) ]) elif resizing == 'cen.crop': transform = T.Compose([ ResizeImage(256), T.CenterCrop(224) ]) elif resizing == 'ran.crop': transform = T.Compose([ ResizeImage(256), T.RandomCrop(224) ]) elif resizing == 'res.': transform = ResizeImage(resize_size) else: raise NotImplementedError(resizing) transforms = [transform] if random_horizontal_flip: transforms.append(T.RandomHorizontalFlip()) if random_color_jitter: transforms.append(T.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5)) transforms.extend([ T.ToTensor(), T.Normalize(mean=norm_mean, std=norm_std) ]) return T.Compose(transforms)
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) else: train_transform = T.Compose([ ResizeImage(256), T.RandomCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] train_source_dataset = dataset(root=args.root, task=args.source, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_target_dataset = dataset(root=args.root, task=args.target, download=True, transform=train_transform) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = dataset(root=args.root, task=args.target, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) if args.data == 'DomainNet': test_dataset = dataset(root=args.root, task=args.target, split='test', download=True, transform=val_transform) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) else: test_loader = val_loader train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) classifier = ImageClassifier(backbone, train_source_dataset.num_classes, args.num_blocks, bottleneck_dim=args.bottleneck_dim, dropout_p=args.dropout_p).to(device) adaptive_feature_norm = AdaptiveFeatureNorm(args.delta).to(device) # define optimizer # the learning rate is fixed according to origin paper optimizer = SGD(classifier.get_parameters(), args.lr, weight_decay=args.weight_decay) # resume from the best checkpoint if args.phase != 'train': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) # analysis the model if args.phase == 'analysis': # extract features from both domains feature_extractor = nn.Sequential(classifier.backbone, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device) target_feature = collect_feature(train_target_loader, feature_extractor, device) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(test_loader, classifier, args) print(acc1) return # start training best_acc1 = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, train_target_iter, classifier, adaptive_feature_norm, optimizer, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) acc1 = validate(test_loader, classifier, args) print("test_acc1 = {:3.1f}".format(acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] train_dataset = dataset(root=args.root, split='train', sample_rate=args.sample_rate, download=True, transform=train_transform) train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_iter = ForeverDataIterator(train_loader) val_dataset = dataset(root=args.root, split='test', sample_rate=100, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) num_classes = train_dataset.num_classes classifier = Classifier(backbone, num_classes).to(device) classifier = convert_model(classifier, p=args.prob) # define optimizer and lr scheduler optimizer = SGD(classifier.get_parameters(args.lr), lr=args.lr, momentum=args.momentum, weight_decay=args.wd, nesterov=True) lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, args.lr_decay_epochs, gamma=args.lr_gamma) # resume from the best checkpoint if args.phase == 'test': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) acc1 = validate(val_loader, classifier, args) print(acc1) return # start training best_acc1 = 0.0 for epoch in range(args.epochs): print(lr_scheduler.get_lr()) # train for one epoch train(train_iter, classifier, optimizer, epoch, args) lr_scheduler.step() # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code if args.num_channels == 3: mode = 'RGB' mean = std = [0.5, 0.5, 0.5] else: mode = 'L' mean = std = [ 0.5, ] normalize = T.Normalize(mean=mean, std=std) train_transform = T.Compose([ ResizeImage(args.image_size), # T.RandomRotation(10), # TODO need results T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(args.image_size), T.ToTensor(), normalize]) source_dataset = datasets.__dict__[args.source] train_source_dataset = source_dataset(root=args.source_root, mode=mode, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) target_dataset = datasets.__dict__[args.target] train_target_dataset = target_dataset(root=args.target_root, mode=mode, download=True, transform=train_transform) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = target_dataset(root=args.target_root, mode=mode, split='test', download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) arch = models.__dict__[args.arch]() bottleneck = nn.Sequential( nn.Flatten(), nn.Linear(arch.bottleneck_dim, arch.bottleneck_dim), nn.BatchNorm1d(arch.bottleneck_dim), nn.ReLU(), nn.Dropout(0.5)) head = arch.head() adv_head = arch.head() classifier = GeneralModule(arch.backbone(), arch.num_classes, bottleneck, head, adv_head, finetune=False) mdd = MarginDisparityDiscrepancy(args.margin).to(device) # define optimizer and lr scheduler optimizer = Adam(classifier.get_parameters(), args.lr, betas=args.betas, weight_decay=args.wd) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # resume from the best checkpoint if args.phase != 'train': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) # analysis the model if args.phase == 'analysis': # extract features from both domains feature_extractor = torch.nn.Sequential( classifier.backbone, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device, 10) target_feature = collect_feature(val_loader, feature_extractor, device, 10) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(val_loader, classifier, args) print(acc1) return # start training best_acc1 = 0. for epoch in range(args.epochs): print(lr_scheduler.get_lr()) # train for one epoch train(train_source_iter, train_target_iter, classifier, mdd, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) else: train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] # If sources is not set, # then use all domains except the target domain. if args.sources is None: args.sources = dataset.domains() args.sources.remove(args.target) print("Source: {} Target: {}".format(args.sources, args.target)) train_source_dataset = ConcatDataset([ dataset(root=args.root, task=source, download=True, transform=train_transform) for source in args.sources ]) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = dataset(root=args.root, task=args.target, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) if args.data == 'DomainNet': test_dataset = dataset(root=args.root, task=args.target, split='test', download=True, transform=val_transform) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) else: test_loader = val_loader train_source_iter = ForeverDataIterator(train_source_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) num_classes = val_dataset.num_classes classifier = Classifier(backbone, num_classes).to(device) # define optimizer and lr scheduler optimizer = SGD(classifier.get_parameters(), args.lr, momentum=args.momentum, weight_decay=args.wd, nesterov=True) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) if args.phase == 'test': # resume from the best checkpoint checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) acc1 = validate(test_loader, classifier, args) print(acc1) return # start training best_acc1 = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, classifier, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) acc1 = validate(test_loader, classifier, args) print("test_acc1 = {:3.1f}".format(acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) print(args) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] train_dataset = dataset(root=args.root, split='train', sample_rate=args.sample_rate, download=True, transform=train_transform) train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = dataset(root=args.root, split='test', sample_rate=100, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) train_iter = ForeverDataIterator(train_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) backbone_source = models.__dict__[args.arch](pretrained=True) num_classes = train_dataset.num_classes classifier = Classifier(backbone, num_classes).to(device) source_classifier = Classifier( backbone_source, head=backbone_source.copy_head(), num_classes=backbone_source.fc.out_features).to(device) for param in source_classifier.parameters(): param.requires_grad = False source_classifier.eval() # define optimizer and lr scheduler optimizer = SGD(classifier.get_parameters(args.lr), momentum=args.momentum, weight_decay=args.wd, nesterov=True) lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, args.lr_decay_epochs, gamma=args.lr_gamma) # resume from the best checkpoint if args.phase == 'test': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) acc1 = validate(val_loader, classifier, args) print(acc1) return # create intermediate layer getter if args.arch == 'resnet50': return_layers = [ 'backbone.layer1.2.conv3', 'backbone.layer2.3.conv3', 'backbone.layer3.5.conv3', 'backbone.layer4.2.conv3' ] elif args.arch == 'resnet101': return_layers = [ 'backbone.layer1.2.conv3', 'backbone.layer2.3.conv3', 'backbone.layer3.5.conv3', 'backbone.layer4.2.conv3' ] else: raise NotImplementedError(args.arch) source_getter = IntermediateLayerGetter(source_classifier, return_layers=return_layers) target_getter = IntermediateLayerGetter(classifier, return_layers=return_layers) # get regularization if args.regularization_type == 'l2_sp': backbone_regularization = SPRegularization(source_classifier.backbone, classifier.backbone) elif args.regularization_type == 'feature_map': backbone_regularization = BehavioralRegularization() elif args.regularization_type == 'attention_feature_map': attention_file = os.path.join(logger.root, args.attention_file) if not os.path.exists(attention_file): attention = calculate_channel_attention(train_dataset, return_layers, args) torch.save(attention, attention_file) else: print("Loading channel attention from", attention_file) attention = torch.load(attention_file) attention = [a.to(device) for a in attention] backbone_regularization = AttentionBehavioralRegularization(attention) else: raise NotImplementedError(args.regularization_type) head_regularization = L2Regularization( nn.ModuleList([classifier.head, classifier.bottleneck])) # start training best_acc1 = 0.0 for epoch in range(args.epochs): print(lr_scheduler.get_lr()) # train for one epoch train(train_iter, classifier, backbone_regularization, head_regularization, target_getter, source_getter, optimizer, epoch, args) lr_scheduler.step() # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) logger.close()
def main(args: argparse.Namespace): if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = transforms.Compose([ ResizeImage(256), transforms.CenterCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize ]) else: train_transform = transforms.Compose([ ResizeImage(256), transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize ]) val_transform = transforms.Compose([ ResizeImage(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize ]) dataset = datasets.__dict__[args.data] partial_dataset = partial(dataset) train_source_dataset = dataset(root=args.root, task=args.source, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_target_dataset = partial_dataset(root=args.root, task=args.target, download=True, transform=train_transform) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = partial_dataset(root=args.root, task=args.target, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) if args.data == 'DomainNet': test_dataset = partial_dataset(root=args.root, task=args.target, split='test', download=True, transform=val_transform) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) else: test_loader = val_loader train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True).to(device) num_classes = train_source_dataset.num_classes classifier = ImageClassifier(backbone, num_classes, bottleneck_dim=args.bottleneck_dim, width=args.bottleneck_dim).to(device) mdd = MarginDisparityDiscrepancy(args.margin).to(device) class_weight_module = AutomaticUpdateClassWeightModule( args.class_weight_update_steps, train_target_loader, classifier, num_classes, device, args.temperature, train_target_dataset.partial_classes_idx) # define optimizer and lr_scheduler # The learning rate of the classifiers are set 10 times to that of the feature extractor by default. optimizer = SGD(classifier.get_parameters(), args.lr, momentum=args.momentum, weight_decay=args.wd, nesterov=True) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # start training best_acc1 = 0. best_model = classifier.state_dict() for epoch in range(args.epochs): # train for one epoch train(train_source_iter, train_target_iter, classifier, mdd, class_weight_module, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint if acc1 > best_acc1: best_model = copy.deepcopy(classifier.state_dict()) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) # evaluate on test set classifier.load_state_dict(best_model) acc1 = validate(test_loader, classifier, args) print("test_acc1 = {:3.1f}".format(acc1))
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) print(args) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) else: train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] train_source_dataset = dataset(root=args.root, task=args.source, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_target_dataset = dataset(root=args.root, task=args.target, download=True, transform=train_transform) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = dataset(root=args.root, task=args.target, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) if args.data == 'DomainNet': test_dataset = dataset(root=args.root, task=args.target, split='test', download=True, transform=val_transform) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) else: test_loader = val_loader train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) num_classes = train_source_dataset.num_classes classifier = ImageClassifier(backbone, num_classes, bottleneck_dim=args.bottleneck_dim).to(device) # define loss function if args.adversarial: thetas = [ Theta(dim).to(device) for dim in (classifier.features_dim, num_classes) ] else: thetas = None jmmd_loss = JointMultipleKernelMaximumMeanDiscrepancy( kernels=([GaussianKernel(alpha=2**k) for k in range(-3, 2)], (GaussianKernel(sigma=0.92, track_running_stats=False), )), linear=args.linear, thetas=thetas).to(device) parameters = classifier.get_parameters() if thetas is not None: parameters += [{ "params": theta.parameters(), 'lr': 0.1 } for theta in thetas] # define optimizer optimizer = SGD(parameters, args.lr, momentum=args.momentum, weight_decay=args.wd, nesterov=True) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # resume from the best checkpoint if args.phase != 'train': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) # analysis the model if args.phase == 'analysis': # extract features from both domains feature_extractor = nn.Sequential(classifier.backbone, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device) target_feature = collect_feature(train_target_loader, feature_extractor, device) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(test_loader, classifier, args) print(acc1) return # start training best_acc1 = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, train_target_iter, classifier, jmmd_loss, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) acc1 = validate(test_loader, classifier, args) print("test_acc1 = {:3.1f}".format(acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code if args.num_channels == 3: mode = 'RGB' mean = std = [0.5, 0.5, 0.5] else: mode = 'L' mean = std = [ 0.5, ] normalize = T.Normalize(mean=mean, std=std) if args.resume_cyclegan is not None: print("Use CycleGAN to translate source images into target style") checkpoint = torch.load(args.resume_cyclegan, map_location='cpu') nc = args.num_channels netG_S2T = cyclegan.generator.__dict__[args.netG]( ngf=args.ngf, norm=args.norm, use_dropout=False, input_nc=nc, output_nc=nc).to(device) print("Loading CycleGAN model from", args.resume_cyclegan) netG_S2T.load_state_dict(checkpoint['netG_S2T']) train_transform = T.Compose([ ResizeImage(args.image_size), cyclegan.transform.Translation(netG_S2T, device, mean=mean, std=std), T.ToTensor(), normalize ]) else: train_transform = T.Compose( [ResizeImage(args.image_size), T.ToTensor(), normalize]) val_transform = T.Compose( [ResizeImage(args.image_size), T.ToTensor(), normalize]) source_dataset = datasets.__dict__[args.source] train_source_dataset = source_dataset(root=args.source_root, mode=mode, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) target_dataset = datasets.__dict__[args.target] val_dataset = target_dataset(root=args.target_root, mode=mode, split='test', download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) train_source_iter = ForeverDataIterator(train_source_loader) print(len(train_source_dataset)) # create model print("=> using pre-trained model '{}'".format(args.arch)) arch = models.__dict__[args.arch]() classifier = Classifier(arch.backbone(), arch.num_classes, arch.bottleneck(), arch.bottleneck_dim, arch.head(), False).to(device) # define optimizer and lr scheduler optimizer = Adam(classifier.get_parameters(), args.lr, betas=args.betas, weight_decay=args.wd) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # resume from the best checkpoint if args.phase != 'train': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) # analysis the model if args.phase == 'analysis': # using shuffled val loader val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers) # extract features from both domains feature_extractor = classifier.backbone.to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device, 10) target_feature = collect_feature(val_loader, feature_extractor, device, 10) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(val_loader, classifier, args) print(acc1) return # start training best_acc1 = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, classifier, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) else: train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) # __dict__这里可以查一下百度,实际上就是类里面的各种静态变量之类的 # 官方示例--data office31,实际上这里是common/vision/datasets/office31.py 里面定义类的实例化 # 可以在这里直接实例化common/vision/datasets/imagelist.py 里面定义的类ImageList dataset = datasets.__dict__[args.data] train_source_dataset = dataset(root=args.root, task=args.source, download=False, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_target_dataset = dataset(root=args.root, task=args.target, download=False, transform=train_transform) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = dataset(root=args.root, task=args.validation, download=False, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) if args.data == 'DomainNet': test_dataset = dataset(root=args.root, task=args.target, split='test', download=True, transform=val_transform) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) else: test_loader = val_loader train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) classifier = ImageClassifier(backbone, train_source_dataset.num_classes, bottleneck_dim=args.bottleneck_dim).to(device) domain_discri = DomainDiscriminator(in_feature=classifier.features_dim, hidden_size=1024).to(device) # define optimizer and lr scheduler optimizer = SGD(classifier.get_parameters() + domain_discri.get_parameters(), args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # define loss function domain_adv = DomainAdversarialLoss(domain_discri).to(device) # resume from the best checkpoint if args.phase != 'train': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) # analysis the model if args.phase == 'analysis': # extract features from both domains feature_extractor = nn.Sequential(classifier.backbone, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device) target_feature = collect_feature(train_target_loader, feature_extractor, device) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(test_loader, classifier, args) print(acc1) return # start training best_acc1 = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, train_target_iter, classifier, domain_adv, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) acc1 = validate(test_loader, classifier, args) print("test_acc1 = {:3.1f}".format(acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) print(args) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_transform = T.Compose([ ResizeImage(256), T.RandomCrop(224), T.RandomHorizontalFlip(), T.ColorJitter(brightness=0.7, contrast=0.7, saturation=0.7, hue=0.5), T.RandomGrayscale(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) train_source_dataset, train_target_dataset, val_dataset, test_dataset, num_classes, args.class_names = \ utils.get_dataset(args.data, args.root, args.source, args.target, train_transform, val_transform, MultipleApply([train_transform, val_transform])) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) # create model print("=> using model '{}'".format(args.arch)) backbone = utils.get_model(args.arch, pretrain=not args.scratch) pool_layer = nn.Identity() if args.no_pool else None classifier = ImageClassifier(backbone, num_classes, bottleneck_dim=args.bottleneck_dim, pool_layer=pool_layer, finetune=not args.scratch).to(device) # define optimizer and lr scheduler optimizer = Adam(classifier.get_parameters(), args.lr) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # resume from the best checkpoint if args.phase != 'train': checkpoint = torch.load(logger.get_checkpoint_path('best'), map_location='cpu') classifier.load_state_dict(checkpoint) # analysis the model if args.phase == 'analysis': # extract features from both domains feature_extractor = nn.Sequential(classifier.backbone, classifier.pool_layer, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device) target_feature = collect_feature(train_target_loader, feature_extractor, device) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.pdf') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = utils.validate(test_loader, classifier, args, device) print(acc1) return if args.pretrain is None: # first pretrain the classifier wish source data print("Pretraining the model on source domain.") args.pretrain = logger.get_checkpoint_path('pretrain') pretrain_model = ImageClassifier(backbone, num_classes, bottleneck_dim=args.bottleneck_dim, pool_layer=pool_layer, finetune=not args.scratch).to(device) pretrain_optimizer = Adam(pretrain_model.get_parameters(), args.pretrain_lr) pretrain_lr_scheduler = LambdaLR( pretrain_optimizer, lambda x: args.pretrain_lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # start pretraining for epoch in range(args.pretrain_epochs): # pretrain for one epoch utils.pretrain(train_source_iter, pretrain_model, pretrain_optimizer, pretrain_lr_scheduler, epoch, args, device) # validate to show pretrain process utils.validate(val_loader, pretrain_model, args, device) torch.save(pretrain_model.state_dict(), args.pretrain) print("Pretraining process is done.") checkpoint = torch.load(args.pretrain, map_location='cpu') classifier.load_state_dict(checkpoint) teacher = EmaTeacher(classifier, alpha=args.alpha) consistent_loss = L2ConsistencyLoss().to(device) class_balance_loss = ClassBalanceLoss(num_classes).to(device) # start training best_acc1 = 0. for epoch in range(args.epochs): print(lr_scheduler.get_lr()) # train for one epoch train(train_source_iter, train_target_iter, classifier, teacher, consistent_loss, class_balance_loss, optimizer, lr_scheduler, epoch, args) # evaluate on validation set acc1 = utils.validate(val_loader, classifier, args, device) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if acc1 > best_acc1: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_acc1 = max(acc1, best_acc1) print("best_acc1 = {:3.1f}".format(best_acc1)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) acc1 = utils.validate(test_loader, classifier, args, device) print("test_acc1 = {:3.1f}".format(acc1)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) print(args) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) else: train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] source_dataset = open_set(dataset, source=True) target_dataset = open_set(dataset, source=False) train_source_dataset = source_dataset(root=args.root, task=args.source, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = target_dataset(root=args.root, task=args.target, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) if args.data == 'DomainNet': test_dataset = target_dataset(root=args.root, task=args.target, split='test', download=True, transform=val_transform) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) else: test_loader = val_loader train_source_iter = ForeverDataIterator(train_source_loader) # create model print("=> using pre-trained model '{}'".format(args.arch)) backbone = models.__dict__[args.arch](pretrained=True) num_classes = train_source_dataset.num_classes classifier = Classifier(backbone, num_classes).to(device) # define optimizer and lr scheduler optimizer = SGD(classifier.get_parameters(), args.lr, momentum=args.momentum, weight_decay=args.wd, nesterov=True) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # analysis the model if args.phase == 'analysis': # using shuffled val loader val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers) # extract features from both domains feature_extractor = nn.Sequential(classifier.backbone, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device) target_feature = collect_feature(val_loader, feature_extractor, device) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(test_loader, classifier, args) print(acc1) return # start training best_h_score = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, classifier, optimizer, lr_scheduler, epoch, args) # evaluate on validation set h_score = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if h_score > best_h_score: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_h_score = max(h_score, best_h_score) print("best_h_score = {:3.1f}".format(best_h_score)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) h_score = validate(test_loader, classifier, args) print("test_h_score = {:3.1f}".format(h_score)) logger.close()
def main(args: argparse.Namespace): logger = CompleteLogger(args.log, args.phase) print(args) if args.seed is not None: random.seed(args.seed) torch.manual_seed(args.seed) cudnn.deterministic = True warnings.warn('You have chosen to seed training. ' 'This will turn on the CUDNN deterministic setting, ' 'which can slow down your training considerably! ' 'You may see unexpected behavior when restarting ' 'from checkpoints.') cudnn.benchmark = True # Data loading code normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if args.center_crop: train_transform = T.Compose([ ResizeImage(256), T.CenterCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) else: train_transform = T.Compose([ ResizeImage(256), T.RandomResizedCrop(224), T.RandomHorizontalFlip(), T.ToTensor(), normalize ]) val_transform = T.Compose( [ResizeImage(256), T.CenterCrop(224), T.ToTensor(), normalize]) dataset = datasets.__dict__[args.data] """ dataset settings for SECC """ from common.vision.datasets.office31 import Office31 public_classes = Office31.CLASSES[:10] source_private = Office31.CLASSES[10:20] target_private = Office31.CLASSES[20:] source_dataset = open_set(dataset, public_classes, source_private) target_dataset = open_set(dataset, public_classes, target_private) """""" train_source_dataset = source_dataset(root=args.root, task=args.source, download=True, transform=train_transform) train_source_loader = DataLoader(train_source_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) train_target_dataset = target_dataset(root=args.root, task=args.target, download=True, transform=train_transform) train_target_loader = DataLoader(train_target_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, drop_last=True) val_dataset = target_dataset(root=args.root, task=args.target, download=True, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers) train_source_iter = ForeverDataIterator(train_source_loader) train_target_iter = ForeverDataIterator(train_target_loader) test_loader = val_loader # create model print("=> using pre-trained model '{}'".format(args.arch)) num_classes = train_source_dataset.num_classes backbone = models.__dict__[args.arch](pretrained=True) """ mean teacher model for SECC """ classifier = ImageClassifier(backbone, num_classes, bottleneck_dim=args.bottleneck_dim).to(device) teacher = EmaTeacher(classifier, 0.9) k = 25 """ distribution cluster for SECC """ print("=> initiating k-means clusters") feature_extractor = FeatureExtractor(backbone) cluster_distribution = ClusterDistribution(train_target_loader, feature_extractor, k=k) """ cluster assignment for SECC """ cluster_assignment = ASoftmax( feature_extractor, num_clusters=k, num_features=cluster_distribution.num_features).to(device) """ loss functions for SECC """ kl_loss = nn.KLDivLoss().to(device) conditional_loss = ConditionalEntropyLoss().to(device) consistent_loss = L2ConsistencyLoss().to(device) class_balance_loss = ClassBalanceLoss(num_classes).to(device) # define optimizer and lr scheduler optimizer = SGD(classifier.get_parameters() + cluster_assignment.get_parameters(), args.lr, momentum=args.momentum, weight_decay=args.weight_decay, nesterov=True) lr_scheduler = LambdaLR( optimizer, lambda x: args.lr * (1. + args.lr_gamma * float(x))**(-args.lr_decay)) # analysis the model if args.phase == 'analysis': # extract features from both domains feature_extractor = nn.Sequential(classifier.backbone, classifier.bottleneck).to(device) source_feature = collect_feature(train_source_loader, feature_extractor, device) target_feature = collect_feature(train_target_loader, feature_extractor, device) # plot t-SNE tSNE_filename = osp.join(logger.visualize_directory, 'TSNE.png') tsne.visualize(source_feature, target_feature, tSNE_filename) print("Saving t-SNE to", tSNE_filename) # calculate A-distance, which is a measure for distribution discrepancy A_distance = a_distance.calculate(source_feature, target_feature, device) print("A-distance =", A_distance) return if args.phase == 'test': acc1 = validate(test_loader, classifier, args) print(acc1) return # start training best_h_score = 0. for epoch in range(args.epochs): # train for one epoch train(train_source_iter, train_target_iter, classifier, teacher, cluster_assignment, cluster_distribution, consistent_loss, class_balance_loss, kl_loss, conditional_loss, optimizer, lr_scheduler, epoch, args) # evaluate on validation set h_score = validate(val_loader, classifier, args) # remember best acc@1 and save checkpoint torch.save(classifier.state_dict(), logger.get_checkpoint_path('latest')) if h_score > best_h_score: shutil.copy(logger.get_checkpoint_path('latest'), logger.get_checkpoint_path('best')) best_h_score = max(h_score, best_h_score) print("best_h_score = {:3.1f}".format(best_h_score)) # evaluate on test set classifier.load_state_dict(torch.load(logger.get_checkpoint_path('best'))) h_score = validate(test_loader, classifier, args) print("test_h_score = {:3.1f}".format(h_score)) logger.close()