def test_test(self):
"""
Test on testing set.
"""
num_batches = int(
math.ceil(self.test_images.shape[0] / self.args.batch_size))
for b in range(num_batches):
b_start = b * self.args.batch_size
b_end = min((b + 1) * self.args.batch_size,
self.test_images.shape[0])
batch_images = common.torch.as_variable(
self.test_images[b_start:b_end], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
# Important to get the correct codes!
output_codes, output_logvar = self.encoder(batch_images)
output_images = self.decoder(output_codes)
e = self.reconstruction_loss(batch_images, output_images)
self.reconstruction_error += e.data
self.code_mean += torch.mean(output_codes).item()
self.code_var += torch.var(output_codes).item()
output_images = numpy.squeeze(
numpy.transpose(output_images.cpu().detach().numpy(),
(0, 2, 3, 1)))
self.pred_images = common.numpy.concatenate(
self.pred_images, output_images)
output_codes = output_codes.cpu().detach().numpy()
self.pred_codes = common.numpy.concatenate(self.pred_codes,
output_codes)
if b % 100 == 50:
log('[Testing] %d' % b)
assert self.pred_images.shape[0] == self.test_images.shape[
0], 'computed invalid number of test images'
if self.args.reconstruction_file:
utils.write_hdf5(self.args.reconstruction_file, self.pred_images)
log('[Testing] wrote %s' % self.args.reconstruction_file)
if self.args.test_theta_file:
assert self.pred_codes.shape[0] == self.test_images.shape[
0], 'computed invalid number of test codes'
utils.write_hdf5(self.args.test_theta_file, self.pred_codes)
log('[Testing] wrote %s' % self.args.test_theta_file)
threshold = 0.9
percentage = 0
# values = numpy.linalg.norm(pred_codes, ord=2, axis=1)
values = numpy.max(numpy.abs(self.pred_codes), axis=1)
while percentage < 0.9:
threshold += 0.1
percentage = numpy.sum(values <= threshold) / float(
values.shape[0])
log('[Testing] threshold %g percentage %g' %
(threshold, percentage))
log('[Testing] taking threshold %g with percentage %g' %
(threshold, percentage))
if self.args.output_directory and utils.display():
# fit = 10
# plot_file = os.path.join(self.args.output_directory, 'test_codes')
# plot.manifold(plot_file, pred_codes[::fit], None, None, 'tsne', None, title='t-SNE of Test Codes')
# log('[Testing] wrote %s' % plot_file)
for d in range(1, self.pred_codes.shape[1]):
plot_file = os.path.join(self.args.output_directory,
'test_codes_%s' % d)
plot.scatter(
plot_file,
self.pred_codes[:, 0],
self.pred_codes[:, d], (values <= threshold).astype(int),
['greater %g' % threshold,
'smaller %g' % threshold],
title='Dimensions 0 and %d of Test Codes' % d)
log('[Testing] wrote %s' % plot_file)
self.reconstruction_error /= num_batches
log('[Testing] reconstruction error %g' % self.reconstruction_error)
def visualize_perturbations(self):
"""
Visualize perturbations.
"""
num_attempts = self.perturbations.shape[1]
num_attempts = min(num_attempts, 6)
utils.makedir(self.args.output_directory)
count = 0
for i in range(min(1000, self.perturbations.shape[0])):
log('[Visualization] sample %d, iterations %s and correctly classified: %s'
% (i + 1, ' '.join(list(map(
str, self.success[i]))), self.accuracy[i]))
if not numpy.any(self.success[i] >= 0) or not self.accuracy[i]:
continue
elif count > 200:
break
#fig, axes = pyplot.subplots(num_attempts, 8)
#if num_attempts == 1:
# axes = [axes] # dirty hack for axis indexing
for j in range(num_attempts):
theta = self.test_theta[i]
theta_attack = self.perturbations[i][j]
theta_perturbation = theta_attack - theta
image = self.test_images[i]
image_attack = self.perturbation_images[i][j]
image_perturbation = image_attack - image
max_theta_perturbation = numpy.max(
numpy.abs(theta_perturbation))
theta_perturbation /= max_theta_perturbation
max_image_perturbation = numpy.max(
numpy.abs(image_perturbation))
image_perturbation /= max_image_perturbation
image_representation = self.theta_representations[i]
attack_representation = self.perturbation_representations[i][j]
image_label = numpy.argmax(image_representation)
attack_label = numpy.argmax(attack_representation)
#vmin = min(numpy.min(theta), numpy.min(theta_attack))
#vmax = max(numpy.max(theta), numpy.max(theta_attack))
#axes[j][0].imshow(theta.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][1].imshow(numpy.squeeze(image), interpolation='nearest', cmap='gray', vmin=0, vmax=1)
#axes[j][2].imshow(theta_perturbation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][2].text(0, -1, 'x' + str(max_theta_perturbation))
#axes[j][3].imshow(numpy.squeeze(image_perturbation), interpolation='nearest', cmap='seismic', vmin=-1, vmax=1)
#axes[j][3].text(0, -image.shape[1]//8, 'x' + str(max_image_perturbation))
#axes[j][4].imshow(theta_attack.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][5].imshow(numpy.squeeze(image_attack), interpolation='nearest', cmap='gray', vmin=0, vmax=1)
#axes[j][6].imshow(image_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][6].text(0, -1, 'Label:' + str(image_label))
#axes[j][7].imshow(attack_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][7].text(0, -1, 'Label:' + str(attack_label))
image_file = os.path.join(
self.args.output_directory,
'%d_%d_image_%d.png' % (i, j, image_label))
attack_file = os.path.join(
self.args.output_directory,
'%d_%d_attack_%d.png' % (i, j, attack_label))
perturbation_file = os.path.join(
self.args.output_directory, '%d_%d_perturbation_%g.png' %
(i, j, max_image_perturbation))
vis.image(image_file, image, scale=10)
vis.image(attack_file, image_attack, scale=10)
vis.perturbation(perturbation_file,
image_perturbation,
scale=10)
#plot_file = os.path.join(self.args.output_directory, str(i) + '.png')
#pyplot.savefig(plot_file)
#pyplot.close(fig)
count += 1
def visualize_perturbations(self):
"""
Visualize perturbations.
"""
num_attempts = self.perturbations.shape[1]
num_attempts = min(num_attempts, 6)
utils.makedir(self.args.output_directory)
count = 0
for i in range(min(1000, self.perturbations.shape[0])):
if not numpy.any(self.success[i]) or not self.accuracy[i]:
continue
elif count > 200:
break
#fig, axes = pyplot.subplots(num_attempts, 5)
#if num_attempts == 1:
# axes = [axes] # dirty hack for axis indexing
for j in range(num_attempts):
image = self.test_images[i]
attack = self.perturbations[i][j]
perturbation = attack - image
max_perturbation = numpy.max(numpy.abs(perturbation))
perturbation /= max_perturbation
image_representation = self.image_representations[i]
attack_representation = self.perturbation_representations[i][j]
image_label = numpy.argmax(image_representation)
attack_label = numpy.argmax(attack_representation)
#axes[j][0].imshow(numpy.squeeze(image), interpolation='nearest', cmap='gray', vmin=0, vmax=1)
#axes[j][1].imshow(numpy.squeeze(perturbation), interpolation='nearest', cmap='seismic', vmin=-1, vmax=1)
#axes[j][1].text(0, -image.shape[1]//8, 'x' + str(max_perturbation))
#axes[j][2].imshow(numpy.squeeze(attack), interpolation='nearest', cmap='gray', vmin=0, vmax=1)
#vmin = min(numpy.min(image_representation), numpy.min(attack_representation))
#vmax = max(numpy.max(image_representation), numpy.max(attack_representation))
#axes[j][3].imshow(image_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][3].text(0, -1, 'Label:' + str(image_label))
#axes[j][4].imshow(attack_representation.reshape(1, -1), interpolation='nearest', vmin=vmin, vmax=vmax)
#axes[j][4].text(0, -1, 'Label:' + str(attack_label))
image_file = os.path.join(
self.args.output_directory,
'%d_%d_image_%d.png' % (i, j, image_label))
attack_file = os.path.join(
self.args.output_directory,
'%d_%d_attack_%d.png' % (i, j, attack_label))
perturbation_file = os.path.join(
self.args.output_directory,
'%d_%d_perturbation_%g.png' % (i, j, max_perturbation))
vis.image(image_file, image, scale=10)
vis.image(attack_file, attack, scale=10)
vis.perturbation(perturbation_file, perturbation, scale=10)
if len(perturbation.shape) > 2:
perturbation_magnitude = numpy.linalg.norm(perturbation,
ord=2,
axis=2)
max_perturbation_magnitude = numpy.max(
numpy.abs(perturbation_magnitude))
perturbation_magnitude /= max_perturbation_magnitude
perturbation_file = os.path.join(
self.args.output_directory,
'%d_%d_perturbation_magnitude_%g.png' %
(i, j, max_perturbation_magnitude))
vis.perturbation(perturbation_file,
perturbation_magnitude,
scale=10)
#plot_file = os.path.join(self.args.output_directory, str(i) + '.png')
#pyplot.savefig(plot_file)
#pyplot.close(fig)
count += 1
def load_data(self):
"""
Load data.
"""
assert self.args.batch_size % 4 == 0
self.train_images = utils.read_hdf5(
self.args.train_images_file).astype(numpy.float32)
log('[Training] read %s' % self.args.train_images_file)
self.test_images = utils.read_hdf5(self.args.test_images_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.test_images_file)
# For handling both color and gray images.
if len(self.train_images.shape) < 4:
self.train_images = numpy.expand_dims(self.train_images, axis=3)
self.test_images = numpy.expand_dims(self.test_images, axis=3)
log('[Training] no color images, adjusted size')
self.resolution = self.test_images.shape[2]
log('[Training] resolution %d' % self.resolution)
self.train_codes = utils.read_hdf5(self.args.train_codes_file).astype(
numpy.int)
assert self.train_codes.shape[1] >= self.args.label_index + 1
self.train_codes = self.train_codes[:, self.args.label_index]
log('[Training] read %s' % self.args.train_codes_file)
self.N_class = numpy.max(self.train_codes) + 1
self.test_codes = utils.read_hdf5(self.args.test_codes_file).astype(
numpy.int)
assert self.test_codes.shape[1] >= self.args.label_index + 1
self.test_codes = self.test_codes[:, self.args.label_index]
log('[Training] read %s' % self.args.test_codes_file)
self.train_theta = utils.read_hdf5(self.args.train_theta_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.train_theta_file)
assert self.test_images.shape[0] == self.test_codes.shape[0]
self.min_bound = numpy.min(self.train_theta, axis=0)
self.max_bound = numpy.max(self.train_theta, axis=0)
log('[Training] min bound: %s' % ' '.join(
['%g' % self.min_bound[i]
for i in range(self.min_bound.shape[0])]))
log('[Training] max bound: %s' % ' '.join(
['%g' % self.max_bound[i]
for i in range(self.max_bound.shape[0])]))
self.test_theta = utils.read_hdf5(self.args.test_theta_file).astype(
numpy.float32)
log('[Training] read %s' % self.args.test_theta_file)
assert self.train_codes.shape[0] == self.train_images.shape[0]
assert self.test_codes.shape[0] == self.test_images.shape[0]
assert self.train_theta.shape[
0] == self.train_images.shape[0], '%s != %s' % ('x'.join(
list(map(str, self.train_theta.shape))), 'x'.join(
list(map(str, self.train_images.shape))))
assert self.test_theta.shape[0] == self.test_images.shape[0]
# Select subset of samples
if self.args.training_samples < 0:
self.args.training_samples = self.train_images.shape[0]
else:
self.args.training_samples = min(self.args.training_samples,
self.train_images.shape[0])
log('[Training] using %d training samples' %
self.args.training_samples)
if self.args.test_samples < 0:
self.args.test_samples = self.test_images.shape[0]
else:
self.args.test_samples = min(self.args.test_samples,
self.test_images.shape[0])
if self.args.early_stopping:
assert self.args.validation_samples > 0
assert self.args.training_samples + self.args.validation_samples <= self.train_images.shape[
0]
self.val_images = self.train_images[self.train_images.shape[0] -
self.args.validation_samples:]
self.val_codes = self.train_codes[self.train_codes.shape[0] -
self.args.validation_samples:]
self.train_images = self.train_images[:self.train_images.shape[0] -
self.args.validation_samples]
self.train_codes = self.train_codeſ[:self.train_codes.shape[0] -
self.args.validation_samples]
assert self.val_images.shape[
0] == self.args.validation_samples and self.val_codes.shape[
0] == self.args.validation_samples
if self.args.random_samples:
perm = numpy.random.permutation(self.train_images.shape[0] // 10)
perm = perm[:self.args.training_samples // 10]
perm = numpy.repeat(perm, self.N_class, axis=0) * 10 + numpy.tile(
numpy.array(range(self.N_class)), (perm.shape[0]))
self.train_images = self.train_images[perm]
self.train_codes = self.train_codes[perm]
self.train_theta = self.train_theta[perm]
else:
self.train_images = self.train_images[:self.args.training_samples]
self.train_codes = self.train_codes[:self.args.training_samples]
self.train_theta = self.train_theta[:self.args.training_samples]
self.train_valid = (numpy.max(numpy.abs(self.train_theta), axis=1) <=
self.args.bound).astype(int)
self.test_valid = (numpy.max(numpy.abs(self.test_theta), axis=1) <=
self.args.bound).astype(int)
# Check that the dataset is balanced.
number_samples = self.train_codes.shape[0] // self.N_class
for c in range(self.N_class):
number_samples_ = numpy.sum(self.train_codes == c)
if number_samples_ != number_samples:
log(
'[Training] dataset not balanced, class %d should have %d samples but has %d'
% (c, number_samples, number_samples_), LogLevel.WARNING)
def test(self):
"""
Test the model.
"""
assert self.model is not None
assert self.model.training is False
assert self.test_images.shape[0] == self.test_codes.shape[
0], 'number of samples have to match'
self.loss = 0.
self.error = 0.
num_batches = int(
math.ceil(self.test_images.shape[0] / self.args.batch_size))
for b in range(num_batches):
b_start = b * self.args.batch_size
b_end = min((b + 1) * self.args.batch_size,
self.test_images.shape[0])
batch_images = common.torch.as_variable(
self.test_images[b_start:b_end], self.args.use_gpu)
batch_classes = common.torch.as_variable(
self.test_codes[b_start:b_end], self.args.use_gpu)
batch_images = batch_images.permute(0, 3, 1, 2)
output_classes = self.model(batch_images)
e = torch.nn.functional.cross_entropy(output_classes,
batch_classes,
size_average=True)
self.loss += e.item()
values, indices = torch.max(torch.nn.functional.softmax(
output_classes, dim=1),
dim=1)
errors = torch.abs(indices - batch_classes)
e = torch.sum(errors > 0).float() / batch_classes.size()[0]
self.error += e.item()
self.accuracy = common.numpy.concatenate(self.accuracy,
errors.data.cpu().numpy())
self.loss /= num_batches
self.error /= num_batches
log('[Testing] test loss %g; test error %g' % (self.loss, self.error))
self.accuracy = self.accuracy == 0
if self.args.accuracy_file:
utils.write_hdf5(self.args.accuracy_file, self.accuracy)
log('[Testing] wrote %s' % self.args.accuracy_file)
accuracy = numpy.sum(self.accuracy) / self.accuracy.shape[0]
if numpy.abs(1 - accuracy - self.error) < 1e-4:
log('[Testing] accuracy file is with %g accuracy correct' %
accuracy)
self.results = {
'loss': self.loss,
'error': self.error,
}
if self.args.results_file:
utils.write_pickle(self.args.results_file, self.results)
log('[Testing] wrote %s' % self.args.results_file)