E, R, source_training_loader_d, length_source_training)
# 3. Target
target_loss, target_error = evaluate(E, R, target_loader_d,
length_target)
print('Source training loss:', source_training_loss,
'Source training error:', source_training_error)
print('Source validation loss:', source_validation_loss,
'Source validation error:', source_validation_error)
print('Target loss:', target_loss, 'Target error', target_error)
# Save Step
if epoch == 0:
target_loss_minimum = target_loss
target_error_minimum = target_error
else:
if target_loss < target_loss_minimum:
torch.save(E.state_dict(), 'E_l1.pkl')
torch.save(D.state_dict(), 'D_l1.pkl')
torch.save(R.state_dict(), 'R_l1.pkl')
target_loss_minimum = target_loss
if target_error < target_error_minimum:
torch.save(E.state_dict(), 'E_l2.pkl')
torch.save(D.state_dict(), 'D_l2.pkl')
torch.save(R.state_dict(), 'R_l2.pkl')
target_error_minimum = target_error
# Optim Step
for num, ((source_batch_x_d, source_batch_y_d),
(target_batch_x_d, target_batch_y_d), (source_batch_x_e,
source_batch_y_e),
(target_batch_x_e, target_batch_y_e)) in enumerate(
zip(source_training_loader_d, target_loader_d,
source_training_loader_e, target_loader_e)):
s3_cls = s3_cls.data.cpu().numpy()
res = s1_cls * s1_weight + s2_cls * s2_weight + s3_cls * s3_weight
#pred = res.argmax(axis=1)
pred = np.argmax(res, axis=1)
labels = labels.numpy()
correct += np.equal(labels, pred).sum()
current_accuracy = correct * 1.0 / len(t_set_test)
current_accuracy = current_accuracy.item()
logger.info("Current accuracy is:{}".format(current_accuracy))
if current_accuracy >= max_correct:
max_correct = current_accuracy
max_step = step
max_epoch = cls_epoch
torch.save(
extractor.state_dict(),
os.path.join(
snapshot, "p2_extractor_" + str(step) + "_" +
str(cls_epoch) + ".pth"))
torch.save(
s1_classifier.state_dict(),
os.path.join(
snapshot,
os.path.abspath(__file__).split('/')[-1].split('.')[0] +
"p2_s1_cls_" + str(step) + "_" + str(cls_epoch) + ".pth"))
torch.save(
s2_classifier.state_dict(),
os.path.join(
snapshot,
os.path.abspath(__file__).split('/')[-1].split('.')[0] +
"p2_s2_cls_" + str(step) + "_" + str(cls_epoch) + ".pth"))
s2_cls = F.softmax(s2_cls, dim=1)
s1_cls = s1_cls.data.cpu().numpy()
s2_cls = s2_cls.data.cpu().numpy()
res = s1_cls * s1_weight + s2_cls * s2_weight
pred = np.argmax(res, axis=1)
labels = labels.numpy()
correct += np.equal(labels, pred).sum()
current_accuracy = correct * 1.0 / len(t_set_test)
current_accuracy = current_accuracy
logger.info("Current accuracy is:{}".format(current_accuracy))
if current_accuracy >= max_correct:
max_correct = current_accuracy
max_step = step
max_epoch = cls_epoch
torch.save(extractor.state_dict(),
os.path.join(snapshot, "_extractor" + ".pth"))
torch.save(
s1_classifier.state_dict(),
os.path.join(
snapshot,
os.path.abspath(__file__).split('/')[-1].split('.')[0] +
"_s1_cls" + ".pth"))
torch.save(
s2_classifier.state_dict(),
os.path.join(
snapshot,
os.path.abspath(__file__).split('/')[-1].split('.')[0] +
"_s2_cls" + ".pth"))
# Part 3: train discriminator and generate mix feature