def main():
best_loss = 100
eps = 1e-4
epoch_counter = 0
for epoch in count(1):
batch_losses = []
samp2_batches = enumerate(sample2_loader)
for batch_idx, batch1 in enumerate(sample1_loader):
try:
_, batch2 = next(samp2_batches)
except:
samp2_batches = enumerate(sample2_loader)
_, batch2 = next(samp2_batches)
batch1 = batch1[0].to(device=device)
batch2 = batch2[0].to(device=device)
batch_loss = training_step(batch1, batch2)
batch_losses.append(batch_loss)
epoch_loss = np.mean(batch_losses)
if epoch_loss < best_loss - eps:
best_loss = epoch_loss
epoch_counter = 0
else:
epoch_counter += 1
print('Epoch {}, loss: {:.3f}, counter: {}'.format(epoch,
epoch_loss,
epoch_counter)
)
update_lr()
if epoch_counter == args.epochs_wo_im:
break
print('Finished training')
# calibrate sample2 -> batch 1
mmd_resnet.train(False)
calibrated_sample2 = []
for batch_idx, batch2 in enumerate(sample2_loader):
batch2 = batch2[0].to(device=device)
calibrated_batch = mmd_resnet(batch2)
calibrated_sample2 += [calibrated_batch.detach().cpu().numpy()]
calibrated_sample2 = np.concatenate(calibrated_sample2)
# ==============================================================================
# = visualize calibration =
# ==============================================================================
# PCA
pca = decomposition.PCA()
pca.fit(sample1)
pc1 = 0
pc2 = 1
axis1 = 'PC'+str(pc1)
axis2 = 'PC'+str(pc2)
# plot data before calibration
sample1_pca = pca.transform(sample1)
sample2_pca = pca.transform(sample2)
sh.scatterHist(sample1_pca[:,pc1],
sample1_pca[:,pc2],
sample2_pca[:,pc1],
sample2_pca[:,pc2],
axis1,
axis2,
title="Data before calibration",
name1='sample1',
name2='sample2')
# plot data after calibration
calibrated_sample2_pca = pca.transform(calibrated_sample2)
sh.scatterHist(sample1_pca[:,pc1],
sample1_pca[:,pc2],
calibrated_sample2_pca[:,pc1],
calibrated_sample2_pca[:,pc2],
axis1,
axis2,
title="Data after calibration",
name1='sample1',
name2='sample2')
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
np.save(args.save_dir + '/sample1.csv', sample1)
np.save(args.save_dir + '/calibrated_sample2.csv', calibrated_sample2)
t_rec_train, t_c_train = sess.run([rec_a1, c_a1], feed_dict={input_a: target_train_data})
s_cal_train = sess.run(rec_a1, feed_dict={input_a: source_train_data})
s_rec_train, s_c_train = sess.run([rec_b1, c_b1], feed_dict={input_b: source_train_data})
if use_test:
t_rec_test, t_c_test = sess.run([rec_a1, c_a1], feed_dict={input_a: target_test_data})
s_cal_test = sess.run(rec_a1, feed_dict={input_a: source_test_data})
s_rec_test, s_c_test = sess.run([rec_b1, c_b1], feed_dict={input_b: source_test_data})
sess.close()
target_pca = pca.transform(target_train_data)
source_pca = pca.transform(source_train_data)
sh.scatterHist(target_pca[:,pc1], target_pca[:,pc2], source_pca[:,pc1],
source_pca[:,pc2], axis1, axis2, title="train data before calibration",
name1='target', name2='source')
target_rec_pca = pca.transform(t_rec_train)
source_cal_pca = pca.transform(s_cal_train)
sh.scatterHist(target_rec_pca[:,pc1], target_rec_pca[:,pc2],
source_cal_pca[:,pc1], source_cal_pca[:,pc2], axis1, axis2,
title="train data after calibration", name1='target', name2='source')
# ==============================================================================
# = save data =
# ==============================================================================
# save data for visualization
save_dir = './output/%s/calibrated_data' % experiment_name
before_calib = np.concatenate([source_train_data, target_train_data], axis=0)
after_calib = np.concatenate(
[calibrated_source_train_data, reconstructed_target_train_data], axis=0)
# embed data
embedding_before = TSNE(n_components=2,
n_iter=1200).fit_transform(before_calib)
embedding_after = TSNE(n_components=2, n_iter=1200).fit_transform(after_calib)
# visualize
sh.scatterHist(embedding_before[:n_s, 0],
embedding_before[:n_s, 1],
embedding_before[n_s:, 0],
embedding_before[n_s:, 1],
axis1='',
axis2='',
title='TSNE embedding before calibration',
name1='batch 1',
name2='batch 2',
plots_dir=plots_dir)
sh.scatterHist(embedding_after[:n_s, 0],
embedding_after[:n_s, 1],
embedding_after[n_s:, 0],
embedding_after[n_s:, 1],
axis1='',
axis2='',
title='TSNE embedding after calibration',
name1='batch 1',
name2='batch 2',
plots_dir=plots_dir)
if use_test:
source_test_data_pca = pca.transform(source_test_data)
target_test_data_pca = pca.transform(target_test_data)
reconstructed_source_test_data_pca = pca.transform(
reconstructed_source_test_data)
calibrated_source_test_data_pca = pca.transform(
calibrated_source_test_data)
reconstructed_target_test_data_pca = pca.transform(
reconstructed_target_test_data)
# plot reconstructions
sh.scatterHist(target_train_data_pca[:, pc1],
target_train_data_pca[:, pc2],
reconstructed_target_train_data_pca[:, pc1],
reconstructed_target_train_data_pca[:, pc2],
axis1,
axis2,
title="target train data reconstruction",
name1='true',
name2='recon',
plots_dir=plots_dir)
sh.scatterHist(source_train_data_pca[:, pc1],
source_train_data_pca[:, pc2],
reconstructed_source_train_data_pca[:, pc1],
reconstructed_source_train_data_pca[:, pc2],
axis1,
axis2,
title="source train data reconstruction",
name1='true',
name2='recon',
plots_dir=plots_dir)