def train_robust_cae(x_train, model, criterion, optimizer, lmbda, inner_epochs, outer_epochs, reinit=True): batch_size = 128 S = np.zeros_like(x_train) # reside on numpy as x_train def get_reconstruction(loader): model.eval() rc = [] for batch, _ in loader: with torch.no_grad(): rc.append(model(batch.cuda()).cpu().numpy()) out = np.concatenate(rc, axis=0) # NOTE: transform_train swaps the channel axis, swap back to yield the same shape out = out.transpose((0, 2, 3, 1)) return out for oe in range(outer_epochs): # update AE if reinit: # Since our CAE_pytorch does not implement reset_parameters, regenerate a new model if reinit. del model model = CAE_pytorch(in_channels=x_train.shape[get_channels_axis()]).cuda() model.train() trainset = trainset_pytorch(x_train-S, train_labels=np.ones((x_train.shape[0], )), transform=transform_train) trainloader = data.DataLoader(trainset, batch_size=batch_size, shuffle=True) for ie in range(inner_epochs): for batch_idx, (inputs, _) in enumerate(trainloader): inputs = inputs.cuda() outputs = model(inputs) loss = criterion(inputs, outputs) # compute gradient and do SGD step optimizer.zero_grad() loss.backward() optimizer.step() if (batch_idx + 1) % 10 == 0: print('Epoch: [{} | {} ({} | {})], batch: {}, loss: {}'.format( ie+1, inner_epochs, oe+1, outer_epochs, batch_idx+1, loss.item()) ) # update S via l21 proximal operator testloader = data.DataLoader(trainset, batch_size=1024, shuffle=False) recon = get_reconstruction(testloader) S = prox_l21(x_train - recon, lmbda) # get final reconstruction finalset = trainset_pytorch(x_train - S, train_labels=np.ones((x_train.shape[0],)), transform=transform_train) finalloader = data.DataLoader(finalset, batch_size=1024, shuffle=False) reconstruction = get_reconstruction(finalloader) losses = ((x_train-S-reconstruction) ** 2).sum(axis=(1, 2, 3), keepdims=False) return losses
def _DRAE_experiment(x_train, y_train, dataset_name, single_class_ind, gpu_q, p): gpu_to_use = gpu_q.get() n_channels = x_train.shape[get_channels_axis()] model = CAE_pytorch(in_channels=n_channels) batch_size = 128 model = model.cuda() trainset = trainset_pytorch(train_data=x_train, train_labels=y_train, transform=transform_train) trainloader = data.DataLoader(trainset, batch_size=batch_size, shuffle=True) cudnn.benchmark = True criterion = DRAELossAutograd(lamb=0.1) optimizer = optim.Adam(model.parameters(), eps=1e-7, weight_decay=0.0005) epochs = 250 # #########################Training######################## train_cae(trainloader, model, criterion, optimizer, epochs) # #######################Testin############################ testloader = data.DataLoader(trainset, batch_size=1024, shuffle=False) losses, reps = test_cae_pytorch(testloader, model) losses = losses - losses.min() losses = losses / (1e-8+losses.max()) scores = 1 - losses res_file_name = '{}_drae-{}_{}_{}.npz'.format(dataset_name, p, get_class_name_from_index(single_class_ind, dataset_name), datetime.now().strftime('%Y-%m-%d-%H%M')) res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name) os.makedirs(os.path.join(RESULTS_DIR, dataset_name), exist_ok=True) save_roc_pr_curve_data(scores, y_train, res_file_path) gpu_q.put(gpu_to_use)
def _E3Outlier_experiment(x_train, y_train, dataset_name, single_class_ind, gpu_q, p): """Surrogate Supervision Discriminative Network training.""" gpu_to_use = gpu_q.get() n_channels = x_train.shape[get_channels_axis()] if OP_TYPE == 'RA': transformer = RA(8, 8) elif OP_TYPE == 'RA+IA': transformer = RA_IA(8, 8, 12) elif OP_TYPE == 'RA+IA+PR': transformer = RA_IA_PR(8, 8, 12, 23, 2) else: raise NotImplementedError print(transformer.n_transforms) if BACKEND == 'wrn': n, k = (10, 4) model = WideResNet(num_classes=transformer.n_transforms, depth=n, widen_factor=k, in_channel=n_channels) elif BACKEND == 'resnet20': n = 20 model = ResNet(num_classes=transformer.n_transforms, depth=n, in_channels=n_channels) elif BACKEND == 'resnet50': n = 50 model = ResNet(num_classes=transformer.n_transforms, depth=n, in_channels=n_channels) elif BACKEND == 'densenet22': n = 22 model = DenseNet(num_classes=transformer.n_transforms, depth=n, in_channels=n_channels) elif BACKEND == 'densenet40': n = 40 model = DenseNet(num_classes=transformer.n_transforms, depth=n, in_channels=n_channels) else: raise NotImplementedError('Unimplemented backend: {}'.format(BACKEND)) print('Using backend: {} ({})'.format(type(model).__name__, BACKEND)) x_train_task = x_train transformations_inds = np.tile(np.arange(transformer.n_transforms), len(x_train_task)) x_train_task_transformed = transformer.transform_batch( np.repeat(x_train_task, transformer.n_transforms, axis=0), transformations_inds) # parameters for training trainset = trainset_pytorch(train_data=x_train_task_transformed, train_labels=transformations_inds, transform=transform_train) batch_size = 128 trainloader = data.DataLoader(trainset, batch_size=batch_size, shuffle=True) cudnn.benchmark = True criterion = nn.CrossEntropyLoss() model = torch.nn.DataParallel(model).cuda() if dataset_name in ['mnist', 'fashion-mnist']: optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9, weight_decay=0.0005) else: optimizer = optim.Adam(model.parameters(), eps=1e-7, weight_decay=0.0005) epochs = int(np.ceil(250 / transformer.n_transforms)) train_pytorch(trainloader, model, criterion, optimizer, epochs) # SSD-IF test_set = testset_pytorch(test_data=x_train_task, transform=transform_test) x_train_task_rep = get_features_pytorch(testloader=data.DataLoader( test_set, batch_size=batch_size, shuffle=False), model=model).numpy() clf = IsolationForest(contamination=p, n_jobs=4).fit(x_train_task_rep) if_scores = clf.decision_function(x_train_task_rep) res_file_name = '{}_ssd-iforest-{}_{}_{}.npz'.format( dataset_name, p, get_class_name_from_index(single_class_ind, dataset_name), datetime.now().strftime('%Y-%m-%d-%H%M')) res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name) os.makedirs(os.path.join(RESULTS_DIR, dataset_name), exist_ok=True) save_roc_pr_curve_data(if_scores, y_train, res_file_path) # E3Outlier if SCORE_MODE == 'pl_mean': preds = np.zeros((len(x_train_task), transformer.n_transforms)) original_preds = np.zeros((transformer.n_transforms, len(x_train_task), transformer.n_transforms)) for t in range(transformer.n_transforms): idx = np.squeeze( np.array([range(x_train_task.shape[0])]) * transformer.n_transforms + t) test_set = testset_pytorch( test_data=x_train_task_transformed[idx, :], transform=transform_test) original_preds[t, :, :] = softmax( test_pytorch(testloader=data.DataLoader(test_set, batch_size=batch_size, shuffle=False), model=model)) preds[:, t] = original_preds[t, :, :][:, t] scores = preds.mean(axis=-1) elif SCORE_MODE == 'max_mean': preds = np.zeros((len(x_train_task), transformer.n_transforms)) original_preds = np.zeros((transformer.n_transforms, len(x_train_task), transformer.n_transforms)) for t in range(transformer.n_transforms): idx = np.squeeze( np.array([range(x_train_task.shape[0])]) * transformer.n_transforms + t) test_set = testset_pytorch( test_data=x_train_task_transformed[idx, :], transform=transform_test) original_preds[t, :, :] = softmax( test_pytorch(testloader=data.DataLoader(test_set, batch_size=batch_size, shuffle=False), model=model)) preds[:, t] = np.max(original_preds[t, :, :], axis=1) scores = preds.mean(axis=-1) elif SCORE_MODE == 'neg_entropy': preds = np.zeros((len(x_train_task), transformer.n_transforms)) original_preds = np.zeros((transformer.n_transforms, len(x_train_task), transformer.n_transforms)) for t in range(transformer.n_transforms): idx = np.squeeze( np.array([range(x_train_task.shape[0])]) * transformer.n_transforms + t) test_set = testset_pytorch( test_data=x_train_task_transformed[idx, :], transform=transform_test) original_preds[t, :, :] = softmax( test_pytorch(testloader=data.DataLoader(test_set, batch_size=batch_size, shuffle=False), model=model)) for s in range(len(x_train_task)): preds[s, t] = neg_entropy(original_preds[t, s, :]) scores = preds.mean(axis=-1) else: raise NotImplementedError res_file_name = '{}_e3outlier-{}_{}_{}.npz'.format( dataset_name, p, get_class_name_from_index(single_class_ind, dataset_name), datetime.now().strftime('%Y-%m-%d-%H%M')) res_file_path = os.path.join(RESULTS_DIR, dataset_name, res_file_name) os.makedirs(os.path.join(RESULTS_DIR, dataset_name), exist_ok=True) save_roc_pr_curve_data(scores, y_train, res_file_path) gpu_q.put(gpu_to_use)