test_nat_acc = 0
fgsm_acc = 0
test_pgd20_acc = 0
cw_acc = 0
best_epoch = 0
for epoch in range(start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch + 1)
train_time, train_loss, bp_count_avg = train(
model, train_loader, optimizer, adjust_tau(epoch + 1, args.dynamictau))
## Evalutions the same as DAT.
loss, test_nat_acc = attack.eval_clean(model, test_loader)
loss, fgsm_acc = attack.eval_robust(model,
test_loader,
perturb_steps=1,
epsilon=0.031,
step_size=0.031,
loss_fn="cent",
category="Madry",
rand_init=True)
loss, test_pgd20_acc = attack.eval_robust(model,
test_loader,
perturb_steps=20,
epsilon=0.031,
step_size=0.031 / 4,
loss_fn="cent",
category="Madry",
rand_init=True)
loss, cw_acc = attack.eval_robust(model,
test_loader,
perturb_steps=30,
epsilon=0.031,
for epoch in range(start_epoch, args.epochs):
# Get lambda
Lambda = adjust_Lambda(epoch + 1)
# Adversarial training
train_robust_loss, lr = train(epoch, model, train_loader, optimizer,
Lambda)
# Evalutions similar to DAT.
_, test_nat_acc = attack.eval_clean(model, test_loader)
_, test_pgd20_acc = attack.eval_robust(model,
test_loader,
perturb_steps=20,
epsilon=0.031,
step_size=0.031 / 4,
loss_fn="cent",
category="Madry",
random=True)
print(
'Epoch: [%d | %d] | Learning Rate: %f | Natural Test Acc %.2f | PGD20 Test Acc %.2f |\n'
% (epoch, args.epochs, lr, test_nat_acc, test_pgd20_acc))
logger_test.append([epoch + 1, test_nat_acc, test_pgd20_acc])
# Save the best checkpoint
if test_pgd20_acc > best_acc:
best_acc = test_pgd20_acc
save_checkpoint(
{
