def _get_output(self, loader, net: BaseNet):
logger = logging.getLogger()
epoch_loss = 0.0
n_batches = 0
# Set device for network
net = net.to(self.device)
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_output = []
net.eval()
with torch.no_grad():
for data in loader:
inputs, labels, idx, _ = data
inputs = inputs.to(self.device)
outputs = net(inputs)
# Save triples of (idx, label, score) in a list
idx_label_output += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
outputs.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
_, labels, outputs = zip(*idx_label_output)
labels = np.array(labels)
outputs = np.array(outputs)
return labels, outputs
def test(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Set device for network
net = net.to(self.device)
# Testing
logger.info('Starting testing...')
epoch_loss = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, semi_targets, idx = data
inputs = inputs.to(self.device)
labels = labels.to(self.device)
semi_targets = semi_targets.to(self.device)
print('Unique Semi Targets: ',
np.unique(semi_targets.data.cpu().numpy()))
idx = idx.to(self.device)
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
losses = torch.where(
semi_targets == 0, dist,
self.eta * ((dist + self.eps)**semi_targets.float()))
loss = torch.mean(losses)
scores = dist
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
epoch_loss += loss.item()
n_batches += 1
self.test_time = time.time() - start_time
self.test_scores = idx_label_score
# Compute AUC
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
self.test_auc = roc_auc_score(labels, scores)
# Log results
logger.info('Test Loss: {:.6f}'.format(epoch_loss / n_batches))
logger.info('Test AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Test Time: {:.3f}s'.format(self.test_time))
logger.info('Finished testing.')
def pretrain(self, deepSVDD, cfg, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
# Get train data loader
train_loader, _ = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Set optimizer (Adam optimizer for now)
optimizer = optim.Adam(net.parameters(),
lr=self.lr,
weight_decay=self.weight_decay,
amsgrad=self.optimizer_name == 'amsgrad')
# Set learning rate scheduler
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=self.lr_milestones, gamma=0.1)
# Training
logger.info('Starting pretraining...')
start_time = time.time()
net.train()
best_score = 0
for epoch in range(self.pre_training_epochs):
loss_epoch = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in tqdm(train_loader):
inputs, _, _ = data
inputs = inputs.to(self.device)
# Zero the network parameter gradients
optimizer.zero_grad()
# Update network parameters via backpropagation: forward + backward + optimize
_, rec_images = net(inputs)
loss = torch.mean(
torch.sum(torch.abs(rec_images - inputs),
dim=tuple(range(1, rec_images.dim()))))
loss.backward()
optimizer.step()
loss_epoch += loss.item()
n_batches += 1
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}'.format(
epoch + 1, self.pre_training_epochs, epoch_train_time,
loss_epoch / n_batches))
self.train_time = time.time() - start_time
logger.info('Training time: %.3f' % self.train_time)
logger.info('Finished training.')
return net
def init_center_c(self, train_loader: DataLoader, net: BaseNet, eps=0.1):
"""Initialize hypersphere center c as the mean from an initial forward pass on the data."""
n_samples = 0
c = torch.zeros(net.rep_dim, device=self.device)
net.eval()
with torch.no_grad():
for data in train_loader:
# get the inputs of the batch
if self.dataset_name == 'object' or self.dataset_name == 'texture':
inputs, _, _ = data
else:
inputs, _, _ = data
inputs = inputs.to(self.device)
outputs, _, _ = net(inputs)
n_samples += outputs.shape[0]
c += torch.sum(outputs, dim=0)
c /= n_samples
# If c_i is too close to 0, set to +-eps. Reason: a zero unit can be trivially matched with zero weights.
c[(abs(c) < eps) & (c < 0)] = -eps
c[(abs(c) < eps) & (c > 0)] = eps
return c
def train_one_step(self, net: BaseNet, epoch: int):
logger = logging.getLogger()
# Set device for network
# net = net.to(self.device)
# Training
logger.info('Starting training...')
start_time = time.time()
net.train()
if (True):
self.scheduler.step()
if epoch in self.lr_milestones:
logger.info(' LR scheduler: new learning rate is %g' %
float(self.scheduler.get_lr()[0]))
epoch_loss = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in self.train_loader:
inputs, _, _, _ = data
inputs = inputs.to(self.device)
# Zero the network parameter gradients
self.optimizer.zero_grad()
# Update network parameters via backpropagation: forward + backward + optimize
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
if self.objective == 'soft-boundary':
scores = dist - self.R**2
loss = self.R**2 + (1 / self.nu) * torch.mean(
torch.max(torch.zeros_like(scores), scores))
else:
loss = torch.mean(dist)
loss.backward()
self.optimizer.step()
# Update hypersphere radius R on mini-batch distances
if (self.objective == 'soft-boundary') and (
epoch >= self.warm_up_n_epochs):
self.R.data = torch.tensor(get_radius(dist, self.nu),
device=self.device)
epoch_loss += loss.item()
n_batches += 1
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}'.format(
epoch + 1, self.n_epochs, epoch_train_time,
epoch_loss / n_batches))
self.train_time = time.time() - start_time
logger.info('Training time: %.3f' % self.train_time)
logger.info('Finished training.')
return net
def train(self, dataset: BaseADDataset, svm_net: BaseNet):
""" 训练 svm 模型 """
logger = logging.getLogger()
# Set device for networks
svm_net = svm_net.to(self.device)
train_loader, _ = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
optimizer = optim.SGD(svm_net.parameters(),
lr=self.lr,
momentum=self.momentum)
scheduler = optim.lr_scheduler.StepLR(optimizer,
step_size=self.step_size,
gamma=self.gamma)
# Training
logger.info('Starting train svm_trainer ...')
start_time = time.time()
svm_net.train()
for epoch in range(self.n_epochs):
scheduler.step()
if epoch in self.lr_milestones:
logger.info(' LR scheduler: new learning rate is %g' %
float(scheduler.get_lr()[0]))
loss_epoch = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in train_loader:
inputs, labels, _ = data
inputs = inputs.to(self.device)
# Zero the networks parameter gradients
optimizer.zero_grad()
# Update networks parameters via back propagation: forward + backward + optimize
outputs = svm_net(inputs)
# get loss
loss = self.hinge_loss(outputs, labels)
loss.backward()
optimizer.step()
loss_epoch += loss.item()
n_batches += 1
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}'.format(
epoch + 1, self.n_epochs, epoch_train_time,
loss_epoch / n_batches))
pretrain_time = time.time() - start_time
logger.info('svm_trainer train time: %.3f' % pretrain_time)
logger.info('Finished train svm_trainer.')
return svm_net
def train(self, dataset: BaseADDataset, ae_net: BaseNet):
logger = logging.getLogger()
# Get train data loader
train_loader, _, _ = dataset.loaders(batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
# Set loss
criterion = nn.MSELoss(reduction='none')
# Set device
#ae_net = ae_net.to(self.device)
#criterion = criterion.to(self.device)
# Set optimizer (Adam optimizer for now)
optimizer = optim.Adam(ae_net.parameters(), lr=self.lr, weight_decay=self.weight_decay)
# Set learning rate scheduler
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=self.lr_milestones, gamma=0.1)
# Training
logger.info('Starting pretraining...')
start_time = time.time()
ae_net.train()
for epoch in range(self.n_epochs):
scheduler.step()
if epoch in self.lr_milestones:
logger.info(' LR scheduler: new learning rate is %g' % float(scheduler.get_lr()[0]))
epoch_loss = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in train_loader:
inputs, _, _, _ = data
inputs = inputs.to(self.device)
# Zero the network parameter gradients
optimizer.zero_grad()
# Update network parameters via backpropagation: forward + backward + optimize
rec = ae_net(inputs)
rec_loss = criterion(rec, inputs)
loss = torch.mean(rec_loss)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
n_batches += 1
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(f'| Epoch: {epoch + 1:03}/{self.n_epochs:03} | Train Time: {epoch_train_time:.3f}s '
f'| Train Loss: {epoch_loss / n_batches:.6f} |')
self.train_time = time.time() - start_time
logger.info('Pretraining Time: {:.3f}s'.format(self.train_time))
logger.info('Finished pretraining.')
return ae_net
def test(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Set device for network
net = net.to(self.device)
# Testing
logger.info('Starting testing...')
epoch_loss = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, semi_targets, idx = data
inputs = inputs.to(self.device)
labels = labels.to(self.device)
semi_targets = semi_targets.to(self.device)
idx = idx.to(self.device)
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
losses = torch.where(
semi_targets == 0, dist,
self.eta * ((dist + self.eps)**semi_targets.float()))
loss = torch.mean(losses)
scores = dist
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
epoch_loss += loss.item()
n_batches += 1
self.test_time = time.time() - start_time
self.test_scores = idx_label_score
# Compute metrics
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
# AUC
self.auc_roc = roc_auc_score(labels, scores)
# PR-curve
self.pr_curve = precision_recall_curve(labels, scores)
precision, recall, thresholds = self.pr_curve
self.auc_pr = auc(recall, precision)
self.test_loss = epoch_loss / n_batches
def init_center_c_w(self, train_loader: DataLoader, net: BaseNet, eps=0.1):
"""Initialize hypersphere center c as the mean from an initial forward pass on the data."""
n_samples = 0
c = torch.zeros(net.rep_dim, device=self.device)
net.eval()
grad_max = torch.tensor([-np.inf], device=self.device)
for data in train_loader:
# get the inputs of the batch
inputs, _, _ = data
inputs = inputs.to(self.device)
inputs.requires_grad_(True)
outputs = net(inputs)
n_samples += outputs.shape[0]
if self.mode == 'weight':
pass
# grads = torch.autograd.grad(outputs=outputs.sum(), inputs=layer, create_graph=True, retain_graph=True)[0]
# grads = grads / (torch.sum(grads**2) + 1e-5)
elif self.mode == 'input':
grads = \
torch.autograd.grad(outputs=outputs.sum(), inputs=inputs, create_graph=False, retain_graph=False)[0]
b = grads.shape[0]
grads_norm = (torch.sum(grads.view(b, -1) ** 2) + 1e-5)
grad_max = torch.maximum(grad_max, grads_norm.max())
inputs.requires_grad_(False)
# with torch.no_grad():
for data in train_loader:
# get the inputs of the batch
inputs, _, _ = data
inputs = inputs.to(self.device)
inputs.requires_grad_(True)
outputs = net(inputs)
if self.mode == 'weight':
pass
# grads = torch.autograd.grad(outputs=outputs.sum(), inputs=layer, create_graph=True, retain_graph=True)[0]
# grads = grads / (torch.sum(grads**2) + 1e-5)
elif self.mode == 'input':
grads = \
torch.autograd.grad(outputs=outputs.sum(), inputs=inputs, create_graph=False, retain_graph=False)[0]
b = grads.shape[0]
grads_norm = (torch.sum(grads.view(b, -1) ** 2) + 1e-5)
outputs = (1 - grads_norm / grad_max) * outputs
inputs.requires_grad_(False)
n_samples += outputs.shape[0]
c += torch.sum(outputs.detach(), dim=0).squeeze()
c /= n_samples
# If c_i is too close to 0, set to +-eps. Reason: a zero unit can be trivially matched with zero weights.
c[(abs(c) < eps) & (c < 0)] = -eps
c[(abs(c) < eps) & (c > 0)] = eps
return c
def train(self, dataset: BaseADDataset, ae_net: BaseNet):
logger = logging.getLogger()
# Set device for network
ae_net = ae_net.to(self.device)
# Get train data loader
train_loader, _ = dataset.loaders(batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
# Set optimizer (Adam optimizer for now)
optimizer = optim.Adam(ae_net.parameters(), lr=self.lr, weight_decay=self.weight_decay,
amsgrad=self.optimizer_name == 'amsgrad')
# Set learning rate scheduler
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=self.lr_milestones, gamma=0.1)
# Training
logger.info('Starting pretraining...')
start_time = time.time()
ae_net.train()
for epoch in range(self.n_epochs):
scheduler.step()
if epoch in self.lr_milestones:
logger.info(' LR scheduler: new learning rate is %g' % float(scheduler.get_lr()[0]))
loss_epoch = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in train_loader:
inputs, _, _ = data
inputs = inputs.to(self.device)
# Zero the network parameter gradients
optimizer.zero_grad()
# Update network parameters via backpropagation: forward + backward + optimize
outputs = ae_net(inputs)
scores = torch.sum((outputs.float() - inputs.float()) ** 2, dim=tuple(range(1, outputs.dim())))
loss = torch.mean(scores)
loss.backward()
optimizer.step()
loss_epoch += loss.item()
n_batches += 1
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}'
.format(epoch + 1, self.n_epochs, epoch_train_time, loss_epoch / n_batches))
pretrain_time = time.time() - start_time
logger.info('Pretraining time: %.3f' % pretrain_time)
logger.info('Finished pretraining.')
return ae_net
def test(self, dataset: BaseADDataset, vae: BaseNet):
logger = logging.getLogger()
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Set device
vae = vae.to(self.device)
# Testing
logger.info('Starting testing...')
epoch_loss = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
vae.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, _, idx = data
inputs, labels, idx = inputs.to(self.device), labels.to(
self.device), idx.to(self.device)
inputs = inputs.view(inputs.size(0), -1)
rec = vae(inputs)
likelihood = -binary_cross_entropy(rec, inputs)
scores = -likelihood # negative likelihood as anomaly score
# Save triple of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
# Overall loss
elbo = likelihood - vae.kl_divergence
loss = -torch.mean(elbo)
epoch_loss += loss.item()
n_batches += 1
self.test_time = time.time() - start_time
# Compute AUC
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
self.test_auc = roc_auc_score(labels, scores)
# Log results
logger.info('Test Loss: {:.6f}'.format(epoch_loss / n_batches))
logger.info('Test AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Test Time: {:.3f}s'.format(self.test_time))
logger.info('Finished testing variational autoencoder.')
def train(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Set device for networks
net = net.to(self.device)
train_loader, _ = dataset.loaders(batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
optimizer = optim.RMSprop(net.parameters(), lr=self.lr, weight_decay=self.weight_decay, eps=self.epsilon,
momentum=self.momentum)
# Set learning rate scheduler
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=self.lr_milestones, gamma=0.1)
# Training
logger.info('Starting train lstm_autoencoder ...')
start_time = time.time()
net.train()
for epoch in range(self.n_epochs):
scheduler.step()
if epoch in self.lr_milestones:
logger.info(' LR scheduler: new learning rate is %g' % float(scheduler.get_lr()[0]))
loss_epoch = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in train_loader:
inputs, _, _ = data
inputs = inputs.to(self.device)
# Zero the networks parameter gradients
optimizer.zero_grad()
# Update networks parameters via back propagation: forward + backward + optimize
_, outputs = net(inputs.view(-1, 1, self.n_features))
scores = torch.sum((outputs - inputs) ** 2, dim=tuple(range(1, outputs.dim())))
loss = torch.mean(scores)
loss.backward()
optimizer.step()
loss_epoch += loss.item()
n_batches += 1
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(' Epoch {}/{}\t Time: {:.3f}\t Loss: {:.8f}'
.format(epoch + 1, self.n_epochs, epoch_train_time, loss_epoch / n_batches))
self.train_time = time.time() - start_time
logger.info('lstm_autoencoder train time: %.3f' % self.train_time)
logger.info('Finished train lstm_autoencoder.')
return net
def test(self, dataset: BaseADDataset, ae_net: BaseNet):
logger = logging.getLogger()
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Set loss
criterion = nn.MSELoss(reduction='none')
# Set device for network
ae_net = ae_net.to(self.device)
criterion = criterion.to(self.device)
# Testing
logger.info('Testing autoencoder...')
epoch_loss = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
ae_net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, _, idx = data
inputs, labels, idx = inputs.to(self.device), labels.to(
self.device), idx.to(self.device)
rec = ae_net(inputs)
rec_loss = criterion(rec, inputs)
scores = torch.mean(rec_loss, dim=tuple(range(1, rec.dim())))
# Save triple of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
loss = torch.mean(rec_loss)
epoch_loss += loss.item()
n_batches += 1
self.test_time = time.time() - start_time
# Compute AUC
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
self.test_auc = roc_auc_score(labels, scores)
# Log results
logger.info('Test Loss: {:.6f}'.format(epoch_loss / n_batches))
logger.info('Test AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Test Time: {:.3f}s'.format(self.test_time))
logger.info('Finished testing autoencoder.')
def test(self, dataset: BaseADDataset, net: BaseNet, corner_cracks=True):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
# Get test data loader
if not corner_cracks:
_, test_loader, _ = dataset.loaders(
batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
else:
_, _, test_loader = dataset.loaders(
batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
if self.objective == 'soft-boundary':
scores = dist - self.R**2
else:
scores = dist
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
self.test_scores = idx_label_score
# Compute AUC
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
self.test_auc = roc_auc_score(labels, scores)
if not corner_cracks:
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
else:
logger.info('Test set AUC (corner): {:.2f}%'.format(100. *
self.test_auc))
logger.info('Finished testing.')
def test(self, dataset: BaseADDataset, ae_net: BaseNet, test_image):
logger = logging.getLogger()
# Set device for network
ae_net = ae_net.to(self.device)
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Testing
logger.info('Testing autoencoder...')
loss_epoch = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
ae_net.eval()
with torch.no_grad():
for i, data in enumerate(test_loader):
inputs, labels, idx = data
inputs = inputs.to(self.device)
outputs = ae_net(inputs)
# import pdb;pdb.set_trace()
if labels == 0:
check_autoencoder_quality(inputs, test_image[i], outputs)
scores = torch.sum((outputs - inputs)**2,
dim=tuple(range(1, outputs.dim())))
loss = torch.mean(scores)
# Save triple of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
loss_epoch += loss.item()
n_batches += 1
logger.info('Test set Loss: {:.8f}'.format(loss_epoch / n_batches))
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
auc = roc_auc_score(labels, scores)
logger.info('Test set AUC: {:.2f}%'.format(100. * auc))
test_time = time.time() - start_time
logger.info('Autoencoder testing time: %.3f' % test_time)
logger.info('Finished testing autoencoder.')
def test(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
list_output = []
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
print('num of test_loader : {}'.format(len(test_loader)))
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_score = []
net.eval()
print('deepSVDD test()---------------')
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
print(dist)
print(labels)
dist_ = dist.cpu().numpy().tolist()
if self.objective == 'soft-boundary':
scores = dist - self.R**2
else:
scores = dist
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
self.test_scores = idx_label_score
# Compute AUC
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
# self.test_auc = roc_auc_score(labels, scores)
test_acc = accuracy_score(labels, list_output)
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
#logger.info('Test set AUCCURAY : {:.2f}%'.format(100. * test_acc))
logger.info('Finished testing.')
return str(test_acc)
def _test(self, loader, net: BaseNet):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
criterion = BCEWithLogitsLoss()
# Testing
logger.info('Starting testing...')
epoch_loss = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in loader:
inputs, labels, semi_targets, idx = data
inputs = inputs.to(self.device)
labels = labels.to(self.device)
semi_targets = semi_targets.to(self.device)
idx = idx.to(self.device)
outputs = net(inputs)
labels = labels.type_as(outputs)
loss = criterion(outputs, labels.unsqueeze(1))
scores = outputs.sigmoid()
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
epoch_loss += loss.item()
n_batches += 1
self.test_time = time.time() - start_time
self.test_scores = idx_label_score
# Compute metrics
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
test_loss = epoch_loss / n_batches
return labels, scores, test_loss
def test(self, dataset: BaseADDataset, ae_net: BaseNet):
logger = logging.getLogger()
# Set device for network
ae_net = ae_net.to(self.device)
# Get test data loader
test_loader = dataset.test_set
# Testing
logger.info('Testing autoencoder...')
loss_epoch = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
ae_net.eval()
with torch.no_grad():
for inputs, labels in test_loader:
if len(inputs) == 32:
inputs = inputs.to(self.device)
inputs = inputs.unsqueeze(1)
outputs = ae_net(inputs.float())
scores = torch.sum((outputs.float() - inputs.float())**2,
dim=tuple(range(1, outputs.dim())))
loss = torch.mean(scores)
# Save triple of (idx, label, score) in a list
idx_label_score += list(
zip(labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
loss_epoch += loss.item()
n_batches += 1
logger.info('Test set Loss: {:.8f}'.format(loss_epoch / n_batches))
labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
fpr, tpr, thresholds = roc_curve(labels, scores, pos_label=1)
test_auc = auc(fpr, tpr)
logger.info('Test set AUC: {:.2f}%'.format(100. * test_auc))
test_time = time.time() - start_time
logger.info('Autoencoder testing time: %.3f' % test_time)
logger.info('Finished testing autoencoder.')
def train_one_step(self, net: BaseNet, epoch: int):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
# Training
logger.info('Starting training...')
start_time = time.time()
net.train()
epoch_loss = 0.0
n_batches = 0
epoch_start_time = time.time()
for data in self.train_loader:
inputs, targets, _, _ = data
inputs, targets = inputs.to(self.device), targets.to(self.device)
# Zero the network parameter gradients
self.optimizer.zero_grad()
# Update network parameters via backpropagation: forward + backward + optimize
outputs = net(inputs)
targets = targets.type_as(outputs)
loss = self.criterion(outputs, targets.unsqueeze(1))
loss.backward()
self.optimizer.step()
epoch_loss += loss.item()
n_batches += 1
self.scheduler.step()
if epoch in self.lr_milestones:
logger.info(' LR scheduler: new learning rate is %g' %
float(scheduler.get_lr()[0]))
# log epoch statistics
epoch_train_time = time.time() - epoch_start_time
logger.info(
f'| Epoch: {epoch + 1:03}/{self.n_epochs:03} | Train Time: {epoch_train_time:.3f}s '
f'| Train Loss: {epoch_loss / n_batches:.6f} |')
return {'train_loss': epoch_loss / n_batches}
def test(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
# Get test data loader
test_loader = dataset.test_set
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for inputs, labels in test_loader:
inputs = inputs.to(self.device)
inputs = inputs.unsqueeze(1)
outputs = net(inputs.float())
dist = torch.sum((outputs.float() - self.c)**2, dim=1)
if self.objective == 'soft-boundary':
scores = dist - self.R**2
else:
scores = dist
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
self.test_scores = idx_label_score
# Compute AUC
labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
fpr, tpr, thresholds = roc_curve(labels, scores, pos_label=1)
self.test_auc = auc(fpr, tpr)
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Finished testing.')
def t_sne(self, dataset: BaseADDataset, net: BaseNet, data_path, xp_path):
logger = logging.getLogger()
center = np.array(self.c.cpu()).reshape(1, 100)
save_path = xp_path
with open(os.path.join(data_path, 'test_label.pickle'), 'rb') as f:
test_class = pickle.load(f)
test_class = np.array(test_class)
test_class = np.append(test_class, 2) # 2: center
# Set device for network
net = net.to(self.device)
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# t_sne
logger.info('Start plot t_sne')
t_sne_array = np.empty((0, 100))
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
outputs = net(inputs)
t_sne_array = np.append(t_sne_array,
outputs.cpu().numpy(),
axis=0)
t_sne_array = np.append(t_sne_array, center, axis=0)
tsne = TSNE(n_components=2, random_state=32)
tsne_results = tsne.fit_transform(t_sne_array)
plt.figure(figsize=(16, 10))
normal_index = (test_class == 0)
abnormal_index = (test_class == 1)
plt.scatter(tsne_results[normal_index, 0],
tsne_results[normal_index, 1],
c='b',
label='normal',
s=1,
marker=',')
plt.scatter(tsne_results[abnormal_index, 0],
tsne_results[abnormal_index, 1],
c='r',
label='abnormal',
s=1,
marker=',')
plt.scatter(tsne_results[-1, 0],
tsne_results[-1, 1],
c='k',
label='center',
s=20,
marker='D')
plt.legend()
plt.savefig(os.path.join(save_path, 't_sne.png'))
def _test(self, loader, net: BaseNet):
logger = logging.getLogger()
epoch_loss = 0.0
n_batches = 0
# Set device for network
net = net.to(self.device)
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in loader:
inputs, labels, idx, _ = data
inputs = inputs.to(self.device)
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
if self.objective == 'soft-boundary':
scores = dist - self.R**2
loss = self.R**2 + (1 / self.nu) * torch.mean(
torch.max(torch.zeros_like(scores), scores))
else:
loss = torch.mean(dist)
scores = dist
epoch_loss += loss.item()
n_batches += 1
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
test_loss = epoch_loss / n_batches
return labels, scores, test_loss
def test(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
outputs = net(inputs)
scores = self.lastlay(outputs)
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
self.test_scores = idx_label_score
# Compute AUC
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
self.test_auc = roc_auc_score(labels, scores)
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Finished testing.')
def apply_model(self, dataset: BaseADDataset, net: BaseNet):
logger = logging.getLogger()
# Set device for network
net = net.to(self.device)
# Get apply_model data loader
_, _, apply_loader = dataset.loaders(
batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
# Applying model
logger.info('Starting Deep SVDD application.')
start_time = time.time()
idx_score = []
net.eval()
with torch.no_grad():
for data in apply_loader:
inputs, nolabels, idx = data # nolables are NaN
inputs = inputs.to(self.device)
outputs = net(inputs)
dist = torch.sum((outputs - self.c)**2, dim=1)
if self.objective == 'soft-boundary':
scores = dist - self.R**2
else:
scores = dist
# Save triples of (idx, label, score) in a list
idx_score += list(
zip(idx.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.apply_time = time.time() - start_time
logger.info('Deep SVDD application time: %.3f' % self.apply_time)
ind, scores = zip(*idx_score)
self.ind = np.array(ind)
self.scores = np.array(scores)
logger.info('Finished Deep SVDD application.')
def init_center_c_grad(self, train_loader: DataLoader, net: BaseNet, layer: torch.nn.Module, eps=0.1):
"""Initialize hypersphere center c as the mean from an initial forward pass on the data."""
n_samples = 0
c = None
net.eval()
# with torch.no_grad():
for data in train_loader:
# get the inputs of the batch
inputs, _, _ = data
inputs = inputs.to(self.device)
inputs.requires_grad_(True)
outputs = net(inputs)
n_samples += outputs.shape[0]
if self.mode == 'weight':
grads = torch.autograd.grad(outputs=outputs.sum(), inputs=layer, create_graph=True, retain_graph=True)[
0]
grads = grads / (torch.sum(grads ** 2) + 1e-5)
elif self.mode == 'input':
grads = torch.autograd.grad(outputs=outputs.sum(), inputs=inputs, create_graph=True, retain_graph=True)[
0]
if 'grad_norm' in self.add_params:
grads = grads / (torch.sqrt(
torch.sum(grads ** 2, dim=tuple(range(1, len(grads.shape))), keepdim=True)) + 1e-5)
grads = torch.sum(grads, dim=0)
inputs.requires_grad_(False)
if c is None:
c = torch.zeros_like(grads)
c += grads.detach()
c /= n_samples
# If c_i is too close to 0, set to +-eps. Reason: a zero unit can be trivially matched with zero weights.
c[(abs(c) < eps) & (c < 0)] = -eps
c[(abs(c) < eps) & (c > 0)] = eps
return c
def test(self, dataset: BaseADDataset, net1: BaseNet, net2: BaseNet):
logger = logging.getLogger()
print('R', self.R)
print('c', self.c)
# Set device for networks
net1 = net1.to(self.device)
net2 = net2.to(self.device)
# Get test data loader
_, test_loader = dataset.loaders(batch_size=self.batch_size,
num_workers=self.n_jobs_dataloader)
# Testing
logger.info('Starting testing...')
start_time = time.time()
idx_label_score = []
net1.eval()
net2.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
code, _ = net1(inputs.view(-1, 1, 9))
outputs = net2(code)
dist = torch.sum((outputs - self.c)**2, dim=1)
if self.objective == 'soft-boundary':
scores = dist - self.R**2
else:
scores = dist
# Save triples of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
self.test_time = time.time() - start_time
logger.info('Testing time: %.3f' % self.test_time)
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
self.test_ftr, self.test_tpr, _ = roc_curve(labels, scores)
self.test_score = scores
self.test_auc = roc_auc_score(labels, scores)
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Finished testing.')
def test(dataset: BaseADDataset, ae_net: BaseNet):
# Set device for network
ae_net = ae_net.to(device)
# Get test data loader
letter, labels = dataset.loaders(batch_size=batch_size,
num_workers=0,
shuffle_test=False,
shuffle_train=False)
loss_epoch = 0.0
n_batches = 0
start_time = time.time()
with torch.no_grad():
i = 0
for data, label in zip(letter, labels):
i += 1
inputs, _ = data
lab, _ = label
inputs = inputs.to(device)
lab = lab.to(device)
# Zero the network parameter gradients
outputs = ae_net(inputs)
plot_images_grid(inputs[0:16],
export_img='./log/test/input' + str(i),
title='Input ',
nrow=4,
padding=4)
plot_images_grid(lab[0:16],
export_img='./log/test/label' + str(i),
title='Label ',
nrow=4,
padding=4)
plot_images_grid(outputs[0:16],
export_img='./log/test/output' + str(i),
title='Output ',
nrow=4,
padding=4)
def __init__(self, dataset: BaseADDataset, network: BaseNet, k: int,
lr: float, n_epochs: int, batch_size: int, rep_dim: int,
K: int, weight_decay: float, device: str,
n_jobs_dataloader: int, w_rec: float, w_feat: float, cfg):
super().__init__(lr, n_epochs, batch_size, rep_dim, K, weight_decay,
device, n_jobs_dataloader, w_rec, w_feat)
self.ae_net = network.to(self.device)
self.train_loader, self.test_loader = dataset.loaders(
batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
self.optimizer = optim.Adam(self.ae_net.parameters(),
lr=self.lr,
weight_decay=self.weight_decay)
self.rec_loss = torch.nn.L1Loss()
self.feat_loss = torch.nn.MSELoss()
self.best_score = 0
self.min_loss = 1000
self.k = k
self.cfg = cfg
self.logger = logging.getLogger()
self.memory = torch.randn(size=(len(self.train_loader.dataset),
self.rep_dim)).to(self.device)
def test(self, dataset: BaseADDataset, net: BaseNet, is_test=0):
"""
dt_type：数据集的类型， 测试集 0 / 训练集 1
"""
logger = logging.getLogger()
# Set device for networks
net = net.to(self.device)
# Get test data loader
if is_test == 0: # 测试集加载器
_, test_loader = dataset.loaders(batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
if is_test == 1: # 训练集加载器
test_loader, _ = dataset.loaders(batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
# Testing
logger.info('Testing lstm_autoencoder...')
loss_epoch = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
# get lstm test label，label.shape = (128,)
label = labels.numpy()
if is_test == 0:
for i in range(len(label)):
self.test_label.append(label[i])
if is_test == 1:
for i in range(len(label)):
self.train_label.append(label[i])
code, outputs = net(inputs.view(-1, 1, self.n_features))
code = code.detach().numpy()
if is_test == 0:
for i in range(len(code)):
self.test_code.append(code[i])
if is_test == 1:
for i in range(len(code)):
self.train_code.append(code[i])
scores = torch.sum((outputs - inputs) ** 2, dim=tuple(range(1, outputs.dim())))
loss = torch.mean(scores)
# Save triple of (idx, label, score) in a list
idx_label_score += list(zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist()))
loss_epoch += loss.item()
n_batches += 1
logger.info('Test set Loss: {:.8f}'.format(loss_epoch / n_batches))
self.test_time = time.time() - start_time
logger.info('lstm_autoencoder testing time: %.3f' % self.test_time)
self.test_scores = idx_label_score
_, labels, scores = zip(*idx_label_score)
labels = np.array(labels)
scores = np.array(scores)
print(len(labels))
print(len(scores))
""" 测试集 """
if is_test == 0:
self.test_auc = roc_auc_score(labels, scores)
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('Finished test lstm_autoencoder.')
def test(self, dataset: BaseADDataset, ae_net: BaseNet, flg=0):
"""
训练集 获取正常数据簇 -- 中心点，半径
测试集 Kmeans 对数据进行预测，超过簇半径为异常数据，否则正常数据
"""
logger = logging.getLogger()
# Set device for networks
ae_net = ae_net.to(self.device)
# 训练集 flg==1 测试集 flg==0
if flg == 1:
test_loader, _ = dataset.loaders(
batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
else:
_, test_loader = dataset.loaders(
batch_size=self.batch_size, num_workers=self.n_jobs_dataloader)
# Testing
logger.info('Testing ae...')
loss_epoch = 0.0
n_batches = 0
start_time = time.time()
idx_label_score = []
ae_net.eval()
with torch.no_grad():
for data in test_loader:
inputs, labels, idx = data
inputs = inputs.to(self.device)
outputs = ae_net(inputs)
scores = torch.sum((outputs - inputs)**2,
dim=tuple(range(1, outputs.dim())))
error = (outputs - inputs)**2
loss = torch.mean(scores)
# Save triple of (idx, label, score) in a list
idx_label_score += list(
zip(idx.cpu().data.numpy().tolist(),
labels.cpu().data.numpy().tolist(),
scores.cpu().data.numpy().tolist(),
error.cpu().data.numpy().tolist()), )
loss_epoch += loss.item()
n_batches += 1
logger.info('Test set Loss: {:.8f}'.format(loss_epoch / n_batches))
_, labels, scores, error = zip(*idx_label_score)
labels = np.array(labels) # labels.shape(97278, )
scores = np.array(scores) # scores.shape(97278, )
error = np.array(error) # scores.shape(97278, )
if flg == 1: # 训练集
X = error
self.kmeans = KMeans(n_clusters=self.clusters).fit(X)
self.center = self.kmeans.cluster_centers_.tolist()
self.radius = self.get_radius(X)
print("roc_self.center", self.center)
print("roc_self.radius", self.radius)
else: # 测试集
Y = error
pred_labels = [] # 实际标签
pred_km = self.kmeans.predict(Y)
print(pred_km.shape)
print(pred_km)
for i in range(len(pred_km)):
dis = self.manhattan_distance(self.center[pred_km[i]],
Y[i]) # dis：簇中心到点的距离，作为分类依据
if dis > self.radius[pred_km[i]]:
pred_labels.append(1)
else:
pred_labels.append(0)
pred_labels = np.array(pred_labels)
self.test_ftr, self.test_tpr, _ = roc_curve(labels, pred_labels)
# roc_self.test_auc = roc_auc_score(pred_labels, labels)
fpr, tpr, thresholds = roc_curve(labels, pred_labels) # 面积作为准确率
print(fpr, tpr)
self.test_auc = auc(fpr, tpr)
self.test_mcc = matthews_corrcoef(labels, pred_labels)
_, _, f_score, _ = precision_recall_fscore_support(labels,
pred_labels,
labels=[0, 1])
self.test_f_score = f_score[1]
print(len(labels))
print(len(scores))
self.test_time = time.time() - start_time
if flg == 0:
logger.info('Test set AUC: {:.2f}%'.format(100. * self.test_auc))
logger.info('ae testing time: %.3f' % self.test_time)
logger.info('Finished testing ae.')
