def train_encoder_classifier_epoch(dataloader,
encoder,
classifier,
encoder_opt,
classifier_opt,
criterion,
device,
batch_size=64):
encoder.train()
classifier.train()
losses = AverageMeter()
acc = AverageMeter()
batch_time = AverageMeter()
end = time.time()
batches = len(dataloader)
for batch_idx, (imgs, labels) in enumerate(dataloader):
y_batch = labels.to(device).float()
imgs = imgs.to(device).float()
classifier_opt.zero_grad()
encoder_opt.zero_grad()
z = encoder(imgs)
# (1) Forward
y_hat = classifier(z)
# (2) Compute diff
loss = criterion(y_hat, y_batch)
# (3) Compute gradients
loss.backward()
# (4) update weights
classifier_opt.step()
encoder_opt.step()
batch_time.update(time.time() - end)
end = time.time()
# (3) Compute gradients
losses.update(loss.item(), imgs.size(0))
preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
accuracy = sum(preds == y_batch).cpu().numpy() / len(y_batch)
# print(y_batch)
# print(preds)
# print(accuracy)
acc.update(accuracy, imgs.size(0))
if batch_idx % print_freq == 0:
print('Batch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Accuracy {acc.val:.4f} ({acc.avg:.4f})'.format(
batch_idx,
batches,
batch_time=batch_time,
loss=losses,
acc=acc))
return losses.avg, acc.avg
def alfr_train_eo(dataloader,
encoder,
classifier,
adversary,
opt_en,
opt_cls,
opt_adv,
cls_criterion,
adv_criterion,
device,
adv_update=0.6,
cls_update=0.5,
batch_size=64):
cls_losses = AverageMeter()
adv_losses = AverageMeter()
cls_en_combinedLosses = AverageMeter()
cls_en_accs = AverageMeter()
adv_combinedLosses = AverageMeter()
adv_accs = AverageMeter()
# combined_loss = AverageMeter()
batch_time = AverageMeter()
end = time.time()
batches = len(dataloader)
updateEncoder = False
for batch_idx, (imgs, shape, color) in enumerate(dataloader):
y_cls_batch = shape.to(device).float()
y_adv_batch = color.to(device).float()
imgs = imgs.to(device).float()
# zero out accumulated gradients
opt_cls.zero_grad()
opt_en.zero_grad()
opt_adv.zero_grad()
# fix adversary take gradient step with classifier and encoder
z = encoder(imgs)
y_hat = classifier(z)
a_hat = adversary(z)
cls_loss = cls_criterion(y_hat.squeeze(), y_cls_batch)
adv_loss = adv_criterion(a_hat.squeeze(), y_adv_batch, y_cls_batch)
# adv_loss = adv_criterion(a_hat, y_adv_batch)
cls_preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
cls_acc = sum(
cls_preds == y_cls_batch).cpu().numpy() / len(y_cls_batch)
cls_en_accs.update(cls_acc, imgs.shape[0])
adv_preds = torch.round(a_hat.data).squeeze(1).cpu().numpy()
adv_acc = sum(
adv_preds == y_adv_batch).cpu().numpy() / len(y_adv_batch)
adv_accs.update(adv_acc, imgs.shape[0])
if (updateEncoder
and adv_acc > adv_update) or (updateEncoder
and cls_acc < cls_update):
print('*', end='')
# if adv_acc < 0.6:
# # print('Skipping encoder update')
# continue
combinedLoss = cls_loss + adv_loss
combinedLoss.backward()
opt_en.step()
opt_cls.step()
cls_en_combinedLosses.update(combinedLoss.item(), imgs.shape[0])
elif not updateEncoder and adv_acc < 0.9:
print('$', end='')
# if adv_acc > 0.9:
# # print('Skipping adv update')
# continue
combinedLoss = -(adv_loss + cls_loss)
combinedLoss.backward()
opt_adv.step()
adv_combinedLosses.update(combinedLoss.item(), imgs.shape[0])
cls_losses.update(cls_loss.item(), imgs.shape[0])
adv_losses.update(adv_loss.item(), imgs.shape[0])
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % print_freq == 0 and batch_idx > 0:
print(
'\nBatch: [{0}/{1}]\t'
# 'Step - Encoder: {2}\n'
#'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\n'
'Cls step loss:{cls_loss.val:.4f} ({cls_loss.avg:.4f})\t'
'Adv step loss:{adv_loss.val:.4f} ({adv_loss.avg:.4f})\n'
# 'Combined loss {combined_loss.val:.4f} ({combined_loss.avg:.4f})\n'
'Cls Acc:{cls_acc.val:.4f} ({cls_acc.avg:.4f})\t'
'Adv Acc:{adv_acc.val:.4f} ({adv_acc.avg:.4f})'.format(
batch_idx,
batches, #updateEncoder,
#batch_time=batch_time,
# cls_loss=cls_losses, adv_loss=adv_losses,
cls_loss=cls_en_combinedLosses,
adv_loss=adv_combinedLosses,
# combined_loss=combined_loss,
cls_acc=cls_en_accs,
adv_acc=adv_accs))
updateEncoder = not updateEncoder
return cls_losses.avg, cls_en_accs.avg, adv_losses.avg, adv_accs.avg, cls_en_combinedLosses.avg, adv_combinedLosses.avg
def laftr_validate_dp(dataloader,
encoder,
classifier,
adversary,
loss_cls,
loss_adv,
device,
batch_size=64):
combined_losses = AverageMeter()
cls_en_losses = AverageMeter()
cls_en_accs = AverageMeter()
adv_losses = AverageMeter()
adv_accs = AverageMeter()
encoder.eval()
classifier.eval()
adversary.eval()
batch_time = AverageMeter()
batches = len(dataloader)
with torch.no_grad():
end = time.time()
for batch_idx, (imgs, shape, color) in enumerate(dataloader):
y_cls_batch = shape.to(device).float()
y_adv_batch = color.to(device).float()
imgs = imgs.to(device).float()
# fix adversary take gradient step with classifier and encoder
z = encoder(imgs)
y_hat = classifier(z)
a_hat = adversary(z)
cls_en_loss = loss_cls(y_hat, y_cls_batch)
adv_loss = loss_adv(a_hat.squeeze(1), y_adv_batch)
combined_loss = cls_en_loss + adv_loss
cls_en_losses.update(cls_en_loss.item(), imgs.shape[0])
adv_losses.update(adv_loss.item(), imgs.shape[0])
combined_losses.update(combined_loss, imgs.shape[0])
cls_preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
cls_acc = sum(
cls_preds == y_cls_batch).cpu().numpy() / len(y_cls_batch)
cls_en_accs.update(cls_acc, imgs.shape[0])
adv_preds = torch.round(a_hat.data).squeeze(1).cpu().numpy()
adv_acc = sum(
adv_preds == y_adv_batch).cpu().numpy() / len(y_adv_batch)
adv_accs.update(adv_acc, imgs.shape[0])
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % print_freq == 0 and batch_idx > 0:
print(
'Test batch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\n'
'Classifier loss {cls_loss.val:.4f} ({cls_loss.avg:.4f})\t'
'Adversary loss {adv_loss.val:.4f} ({adv_loss.avg:.4f})\n'
'Combined Loss {combloss.val:.4f} ({combloss.avg:.4f})\t'
'Classifier Accuracy {cls_acc.val:.4f} ({cls_acc.avg:.4f})\t'
'Adversary Accuracy {adv_acc.val:.4f} ({adv_acc.avg:.4f})'.
format(batch_idx,
batches,
batch_time=batch_time,
cls_loss=cls_en_losses,
adv_loss=adv_losses,
combloss=combined_losses,
cls_acc=cls_en_accs,
adv_acc=adv_accs))
return combined_losses.avg, cls_en_losses.avg, cls_en_accs.avg, adv_losses.avg, adv_accs.avg
def laftr_epoch_dp(dataloader,
encoder,
classifier,
adversary,
opt_en,
opt_cls,
opt_adv,
cls_criterion,
adv_criterion,
device,
batch_size=64):
'''
one training laftr epoch
'''
cls_losses = AverageMeter()
adv_losses = AverageMeter()
cls_en_combinedLosses = AverageMeter()
cls_en_accs = AverageMeter()
adv_combinedLosses = AverageMeter()
adv_accs = AverageMeter()
batch_time = AverageMeter()
end = time.time()
batches = len(dataloader)
for batch_idx, (imgs, shape, color) in enumerate(dataloader):
y_cls_batch = shape.to(device).float()
y_adv_batch = color.to(device).float()
imgs = imgs.to(device).float()
# fix adversary take gradient step with classifier and encoder
unfreeze_model(encoder)
unfreeze_model(classifier)
z = encoder(imgs)
y_hat = classifier(z)
freeze_model(adversary)
a_fixed = adversary(z)
opt_cls.zero_grad()
opt_en.zero_grad()
cls_loss = cls_criterion(y_hat, y_cls_batch)
adv_loss_fixed = adv_criterion(a_fixed, y_adv_batch)
cls_en_combinedLoss = cls_loss + adv_loss_fixed
cls_en_combinedLoss.backward()
opt_cls.step()
opt_en.step()
cls_losses.update(cls_loss.item(), imgs.shape[0])
# fix encoder and classifier and take gradient step with adversary
freeze_model(encoder)
freeze_model(classifier)
z_fixed = encoder(imgs)
y_hat_fixed = classifier(z_fixed)
unfreeze_model(adversary)
a_hat = adversary(z_fixed)
opt_adv.zero_grad()
cls_loss_fixed = cls_criterion(y_hat_fixed, y_cls_batch)
adv_loss = adv_criterion(a_hat, y_adv_batch)
adv_combinedLoss = -(cls_loss_fixed + adv_loss)
adv_combinedLoss.backward()
opt_adv.step()
adv_losses.update(adv_loss.item(), imgs.shape[0])
cls_en_combinedLosses.update(cls_en_combinedLoss.item(), imgs.shape[0])
adv_combinedLosses.update(adv_combinedLoss.item(), imgs.shape[0])
cls_preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
cls_acc = sum(
cls_preds == y_cls_batch).cpu().numpy() / len(y_cls_batch)
cls_en_accs.update(cls_acc, imgs.shape[0])
adv_preds = torch.round(a_hat.data).squeeze(1).cpu().numpy()
adv_acc = sum(
adv_preds == y_adv_batch).cpu().numpy() / len(y_adv_batch)
adv_accs.update(adv_acc, imgs.shape[0])
batch_time.update(time.time() - end)
end = time.time()
if batch_idx % print_freq == 0:
print(
'Batch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\n'
'Classifier loss {cls_loss.val:.4f} ({cls_loss.avg:.4f})\t'
'Adversary loss {adv_loss.val:.4f} ({adv_loss.avg:.4f})\n'
'Combined Loss during classifier step {cls_combloss.val:.4f} ({cls_combloss.avg:.4f})\t'
'Combined Loss during adversary step {adv_combloss.val:.4f} ({adv_combloss.avg:.4f})\n'
'Classifier Accuracy {cls_acc.val:.4f} ({cls_acc.avg:.4f})\t'
'Adversary Accuracy {adv_acc.val:.4f} ({adv_acc.avg:.4f})'.
format(batch_idx,
batches,
batch_time=batch_time,
cls_loss=cls_losses,
adv_loss=adv_losses,
cls_combloss=cls_en_combinedLosses,
adv_combloss=adv_combinedLosses,
cls_acc=cls_en_accs,
adv_acc=adv_accs))
return cls_losses.avg, cls_en_combinedLosses.avg, cls_en_accs.avg, adv_losses.avg, adv_combinedLosses.avg, adv_accs.avg
def train_encoder_classifier_epoch(encoder,
classifier,
X,
Y,
encoder_opt,
classifier_opt,
criterion,
device,
batch_size=64):
encoder.train()
classifier.train()
losses = AverageMeter()
acc = AverageMeter()
batch_time = AverageMeter()
end = time.time()
batches = int(X.shape[0] / batch_size)
for beg_i in range(0, X.shape[0], batch_size):
i = int(beg_i / batch_size)
x_batch = X[beg_i:beg_i + batch_size]
y_batch = Y[beg_i:beg_i + batch_size]
x_batch = torch.from_numpy(x_batch).to(device).float()
y_batch = torch.from_numpy(y_batch).to(device).float()
x_batch = x_batch.permute((0, 3, 1, 2))
x_batch /= 255.
classifier_opt.zero_grad()
encoder_opt.zero_grad()
z = encoder(x_batch)
# (1) Forward
y_hat = classifier(z)
# (2) Compute diff
loss = criterion(y_hat, y_batch)
# (3) Compute gradients
loss.backward()
# (4) update weights
classifier_opt.step()
encoder_opt.step()
batch_time.update(time.time() - end)
end = time.time()
# (3) Compute gradients
losses.update(loss.item(), x_batch.size(0))
preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
accuracy = sum(preds == y_batch).cpu().numpy() / len(y_batch)
# print(y_batch)
# print(preds)
# print(accuracy)
acc.update(accuracy, x_batch.size(0))
if i % print_freq == 0:
print('Batch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Accuracy {acc.val:.4f} ({acc.avg:.4f})'.format(
i, batches, batch_time=batch_time, loss=losses, acc=acc))
return losses.avg, acc.avg
def laftr_validate_dp(encoder,
classifier,
adversary,
X,
y_cls,
y_adv,
loss_cls,
loss_adv,
device,
batch_size=64):
combined_losses = AverageMeter()
cls_en_losses = AverageMeter()
cls_en_accs = AverageMeter()
adv_losses = AverageMeter()
adv_accs = AverageMeter()
encoder.eval()
classifier.eval()
adversary.eval()
batch_time = AverageMeter()
batches = int(X.shape[0] / batch_size)
with torch.no_grad():
end = time.time()
for beg_i in range(0, X.shape[0], batch_size):
i = int(beg_i / batch_size)
x_batch = X[beg_i:beg_i + batch_size]
y_cls_batch = y_cls[beg_i:beg_i + batch_size]
y_adv_batch = y_adv[beg_i:beg_i + batch_size]
x_batch = torch.from_numpy(x_batch).to(device).float()
y_cls_batch = torch.from_numpy(y_cls_batch).to(device).float()
y_adv_batch = torch.from_numpy(y_adv_batch).to(device).float()
x_batch = x_batch.permute((0, 3, 1, 2))
x_batch /= 255.
# fix adversary take gradient step with classifier and encoder
z = encoder(x_batch)
y_hat = classifier(z)
a_hat = adversary(z)
cls_en_loss = loss_cls(y_hat, y_cls_batch)
adv_loss = loss_adv(a_hat, y_adv_batch)
combined_loss = cls_en_loss + adv_loss
cls_en_losses.update(cls_en_loss.item(), x_batch.shape[0])
adv_losses.update(adv_loss.item(), x_batch.shape[0])
combined_losses.update(combined_loss, x_batch.shape[0])
cls_preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
cls_acc = sum(
cls_preds == y_cls_batch).cpu().numpy() / len(y_cls_batch)
cls_en_accs.update(cls_acc, x_batch.shape[0])
adv_preds = torch.round(a_hat.data).squeeze(1).cpu().numpy()
adv_acc = sum(
adv_preds == y_adv_batch).cpu().numpy() / len(y_adv_batch)
adv_accs.update(adv_acc, x_batch.shape[0])
batch_time.update(time.time() - end)
end = time.time()
if i % print_freq == 0:
print(
'Test batch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\n'
'Classifier loss {cls_loss.val:.4f} ({cls_loss.avg:.4f})\t'
'Adversary loss {adv_loss.val:.4f} ({adv_loss.avg:.4f})\n'
'Combined Loss {combloss.val:.4f} ({combloss.avg:.4f})\t'
'Classifier Accuracy {cls_acc.val:.4f} ({cls_acc.avg:.4f})\t'
'Adversary Accuracy {adv_acc.val:.4f} ({adv_acc.avg:.4f})'.
format(i,
batches,
batch_time=batch_time,
cls_loss=cls_en_losses,
adv_loss=adv_losses,
combloss=combined_losses,
cls_acc=cls_en_accs,
adv_acc=adv_accs))
return combined_losses.avg, cls_en_losses.avg, cls_en_accs.avg, adv_losses.avg, adv_accs.avg
def laftr_epoch_dp(encoder,
classifier,
adversary,
X,
y_cls,
y_adv,
opt_en,
opt_cls,
opt_adv,
cls_criterion,
adv_criterion,
device,
batch_size=64):
'''
one training laftr epoch
'''
cls_losses = AverageMeter()
adv_losses = AverageMeter()
cls_en_combinedLosses = AverageMeter()
cls_en_accs = AverageMeter()
adv_combinedLosses = AverageMeter()
adv_accs = AverageMeter()
batch_time = AverageMeter()
end = time.time()
batches = int(X.shape[0] / batch_size)
for beg_i in range(0, X.shape[0], batch_size):
i = int(beg_i / batch_size)
x_batch = X[beg_i:beg_i + batch_size]
y_cls_batch = y_cls[beg_i:beg_i + batch_size]
y_adv_batch = y_adv[beg_i:beg_i + batch_size]
x_batch = torch.from_numpy(x_batch).to(device).float()
y_cls_batch = torch.from_numpy(y_cls_batch).to(device).float()
y_adv_batch = torch.from_numpy(y_adv_batch).to(device).float()
x_batch = x_batch.permute((0, 3, 1, 2))
x_batch /= 255.
# fix adversary take gradient step with classifier and encoder
unfreeze_model(encoder)
unfreeze_model(classifier)
z = encoder(x_batch)
y_hat = classifier(z)
freeze_model(adversary)
a_fixed = adversary(z)
opt_cls.zero_grad()
opt_en.zero_grad()
cls_loss = cls_criterion(y_hat, y_cls_batch)
adv_loss_fixed = adv_criterion(a_fixed, y_adv_batch)
cls_en_combinedLoss = cls_loss + adv_loss_fixed
cls_en_combinedLoss.backward()
opt_cls.step()
opt_en.step()
cls_losses.update(cls_loss.item(), x_batch.shape[0])
# fix encoder and classifier and take gradient step with adversary
freeze_model(encoder)
freeze_model(classifier)
z_fixed = encoder(x_batch)
y_hat_fixed = classifier(z_fixed)
unfreeze_model(adversary)
a_hat = adversary(z_fixed)
opt_adv.zero_grad()
cls_loss_fixed = cls_criterion(y_hat_fixed, y_cls_batch)
adv_loss = adv_criterion(a_hat, y_adv_batch)
adv_combinedLoss = -(cls_loss_fixed + adv_loss)
adv_combinedLoss.backward()
opt_adv.step()
adv_losses.update(adv_loss.item(), x_batch.shape[0])
cls_en_combinedLosses.update(cls_en_combinedLoss.item(),
x_batch.shape[0])
adv_combinedLosses.update(adv_combinedLoss.item(), x_batch.shape[0])
cls_preds = torch.round(y_hat.data).squeeze(1).cpu().numpy()
cls_acc = sum(
cls_preds == y_cls_batch).cpu().numpy() / len(y_cls_batch)
cls_en_accs.update(cls_acc, x_batch.shape[0])
adv_preds = torch.round(a_hat.data).squeeze(1).cpu().numpy()
adv_acc = sum(
adv_preds == y_adv_batch).cpu().numpy() / len(y_adv_batch)
adv_accs.update(adv_acc, x_batch.shape[0])
batch_time.update(time.time() - end)
end = time.time()
if i % print_freq == 0:
print(
'Batch: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\n'
'Classifier loss {cls_loss.val:.4f} ({cls_loss.avg:.4f})\t'
'Adversary loss {adv_loss.val:.4f} ({adv_loss.avg:.4f})\n'
'Combined Loss during classifier step {cls_combloss.val:.4f} ({cls_combloss.avg:.4f})\t'
'Combined Loss during adversary step {adv_combloss.val:.4f} ({adv_combloss.avg:.4f})\n'
'Classifier Accuracy {cls_acc.val:.4f} ({cls_acc.avg:.4f})\t'
'Adversary Accuracy {adv_acc.val:.4f} ({adv_acc.avg:.4f})'.
format(i,
batches,
batch_time=batch_time,
cls_loss=cls_losses,
adv_loss=adv_losses,
cls_combloss=cls_en_combinedLosses,
adv_combloss=adv_combinedLosses,
cls_acc=cls_en_accs,
adv_acc=adv_accs))
return cls_losses.avg, cls_en_combinedLosses.avg, cls_en_accs.avg, adv_losses.avg, adv_combinedLosses.avg, adv_accs.avg