def __call__(self, example_string):
"""Processes a single example string.
Extracts and processes the image, and ignores the label. We assume that the
image has three channels.
Args:
example_string: str, an Example protocol buffer.
Returns:
image_rescaled: the image, resized to `image_size x image_size` and
rescaled to [-1, 1]. Note that Gaussian data augmentation may cause values
to go beyond this range.
"""
image_decoded = read_example_and_parse_image(example_string)['image']
image_resized = tf.image.resize_images(
image_decoded, [self.image_size, self.image_size],
method=tf.image.ResizeMethod.BILINEAR,
align_corners=True)
image_resized = tf.cast(image_resized, tf.float32)
image = 2 * (image_resized / 255.0 - 0.5) # Rescale to [-1, 1].
if self.data_augmentation is not None:
if self.data_augmentation.enable_gaussian_noise:
image = image + tf.random_normal(tf.shape(
image)) * self.data_augmentation.gaussian_noise_std
if self.data_augmentation.enable_jitter:
j = self.data_augmentation.jitter_amount
paddings = tf.constant([[j, j], [j, j], [0, 0]])
image = tf.pad(image, paddings, 'REFLECT')
image = tf.image.random_crop(
image, [self.image_size, self.image_size, 3])
return image
def process_example(example_string, image_size, data_augmentation=None):
"""Processes a single example string.
Extracts and processes the image, and ignores the label. We assume that the
image has three channels.
Args:
example_string: str, an Example protocol buffer.
image_size: int, desired image size. The extracted image will be resized to
`[image_size, image_size]`.
data_augmentation: A DataAugmentation object with parameters for perturbing
the images.
Returns:
image_rescaled: the image, resized to `image_size x image_size` and rescaled
to [-1, 1]. Note that Gaussian data augmentation may cause values to
go beyond this range.
"""
image_string = tf.parse_single_example(example_string,
features={
'image':
tf.FixedLenFeature(
[], dtype=tf.string),
'label':
tf.FixedLenFeature([], tf.int64)
})['image']
image_decoded = tf.image.decode_jpeg(image_string, channels=3)
image_resized = tf.image.resize_images(
image_decoded, [image_size, image_size],
method=tf.image.ResizeMethod.BILINEAR,
align_corners=True)
image = 2 * (image_resized / 255.0 - 0.5) # Rescale to [-1, 1].
if data_augmentation is not None:
if data_augmentation.enable_gaussian_noise:
image = image + tf.random_normal(
tf.shape(image)) * data_augmentation.gaussian_noise_std
if data_augmentation.enable_jitter:
j = data_augmentation.jitter_amount
paddings = tf.constant([[j, j], [j, j], [0, 0]])
image = tf.pad(image, paddings, 'REFLECT')
image = tf.image.random_crop(image, [image_size, image_size, 3])
return image
def build_graph(self):
"""Builds the neural network graph."""
# define graph
self.g = tf.Graph()
with self.g.as_default():
# create and store a new session for the graph
self.sess = tf.Session()
# define placeholders
self.x = tf.placeholder(shape=[None, self.dim_input],
dtype=tf.float32)
self.y = tf.placeholder(shape=[None, self.num_classes],
dtype=tf.float32)
# linear layer(WX + b)
with tf.variable_scope('last_layer/dense') as scope:
weights = tf.get_variable('kernel',
[self.dim_input, self.num_classes],
dtype=tf.float32)
biases = tf.get_variable('bias', [self.num_classes],
dtype=tf.float32)
wb = tf.concat([weights, tf.expand_dims(biases, axis=0)], 0)
wb_renorm = tf.matmul(self.sigma_half_inv, wb)
weights_renorm = wb_renorm[:self.dim_input, :]
biases_renorm = wb_renorm[-1, :]
self.z = tf.add(tf.matmul(self.x, weights_renorm),
biases_renorm,
name=scope.name)
# Gaussian prior
# prior = tf.nn.l2_loss(weights) + tf.nn.l2_loss(biases)
# Non normalized loss, because of the preconditioning
self.loss = self.n * tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.y,
logits=self.z))
# Bayesian loss
self.bayesian_loss = self.loss # + prior
self.output_probs = tf.nn.softmax(self.z)
# Variables of the last layer
self.ll_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
self.ll_vars_concat = tf.concat(
[self.ll_vars[0],
tf.expand_dims(self.ll_vars[1], axis=0)], 0)
# Summary
_variable_summaries(self.ll_vars_concat)
# saving the weights of last layer when running SGLD/SGD/MCMC algorithm
self.saver = tf.train.Saver(var_list=self.ll_vars,
max_to_keep=self.num_samples)
self.gd_opt = tf.train.GradientDescentOptimizer(self.step_size)
# SGLD optimizer for the last layer
if self.sampler in ['sgld', 'lmc']:
grads_vars = self.gd_opt.compute_gradients(self.bayesian_loss)
grads_vars_sgld = []
for g, v in grads_vars:
if g is not None:
s = list(v.name)
s[v.name.rindex(':')] = '_'
# Adding Gaussian noise to the gradient
gaussian_noise = (np.sqrt(2. / self.step_size) *
tf.random_normal(tf.shape(g)))
g_sgld = g + gaussian_noise
tf.summary.histogram(''.join(s) + '/grad_hist_mcmc', g)
tf.summary.histogram(
''.join(s) + '/gaussian_noise_hist_mcmc',
gaussian_noise)
tf.summary.histogram(
''.join(s) + '/grad_total_hist_mcmc', g_sgld)
grads_vars_sgld.append((g_sgld, v))
self.train_op = self.gd_opt.apply_gradients(grads_vars_sgld)
# SGD optimizer for the last layer
if self.sampler == 'sgd':
grads_vars_sgd = self.gd_opt.compute_gradients(self.loss)
self.train_op = self.gd_opt.apply_gradients(grads_vars_sgd)
for g, v in grads_vars_sgd:
if g is not None:
s = list(v.name)
s[v.name.rindex(':')] = '_'
tf.summary.histogram(''.join(s) + '/grad_hist_sgd', g)
# Merge all the summaries and write them out
self.all_summaries = tf.summary.merge_all()
location = os.path.join(self.working_dir, 'logs')
self.writer = tf.summary.FileWriter(location, graph=self.g)
saver_network = tf.train.Saver(var_list=self.ll_vars)
print('loading the network ...')
# Restores from checkpoint
saver_network.restore(self.sess, self.model_dir)
print('Graph successfully loaded.')
def build_graph(self):
"""Builds the neural network graph."""
# define graph
self.g = tf.Graph()
with self.g.as_default():
# create and store a new session for the graph
self.sess = tf.Session()
# define placeholders
self.x = tf.placeholder(shape=[None, self.dim_input],
dtype=tf.float32)
self.y = tf.placeholder(shape=[None, self.num_classes],
dtype=tf.float32)
# define simple model
with tf.variable_scope('last_layer'):
self.z = tf.layers.dense(inputs=self.x, units=self.num_classes)
self.loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.y,
logits=self.z))
self.output_probs = tf.nn.softmax(self.z)
# Variables of the last layer
self.ll_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
self.ll_vars_concat = tf.concat(
[self.ll_vars[0],
tf.expand_dims(self.ll_vars[1], axis=0)], 0)
# Summary
_variable_summaries(self.ll_vars_concat)
# add regularization that acts as a unit Gaussian prior on the last layer
regularizer = tf.contrib.layers.l2_regularizer(1.0)
# regularization
prior = tf.contrib.layers.apply_regularization(
regularizer, self.ll_vars)
self.bayesian_loss = self.n * self.loss + prior
# saving the weights of last layer when running SGLD/SGD/MCMC algorithm
self.saver = tf.train.Saver(var_list=self.ll_vars,
max_to_keep=self.num_samples)
# SGLD optimizer for the last layer
if self.sampler in ['sgld', 'lmc']:
step = self.step_size / self.n
gd_opt = tf.train.GradientDescentOptimizer(step)
grads_vars = gd_opt.compute_gradients(self.bayesian_loss)
grads_vars_sgld = []
for g, v in grads_vars:
if g is not None:
s = list(v.name)
s[v.name.rindex(':')] = '_'
# Adding Gaussian noise to the gradient
gaussian_noise = (np.sqrt(2. / step) *
tf.random_normal(tf.shape(g)))
g_sgld = g + gaussian_noise
tf.summary.histogram(''.join(s) + '/grad_hist_mcmc',
g / self.n)
tf.summary.histogram(
''.join(s) + '/gaussian_noise_hist_mcmc',
gaussian_noise / self.n)
tf.summary.histogram(
''.join(s) + '/grad_total_hist_mcmc',
g_sgld / self.n)
grads_vars_sgld.append((g_sgld, v))
self.train_op = gd_opt.apply_gradients(grads_vars_sgld)
# SGD optimizer for the last layer
if self.sampler == 'sgd':
gd_opt = tf.train.GradientDescentOptimizer(self.step_size)
grads_vars_sgd = gd_opt.compute_gradients(self.loss)
self.train_op = gd_opt.apply_gradients(grads_vars_sgd)
for g, v in grads_vars_sgd:
if g is not None:
s = list(v.name)
s[v.name.rindex(':')] = '_'
tf.summary.histogram(''.join(s) + '/grad_hist_sgd', g)
# Merge all the summaries and write them out
self.all_summaries = tf.summary.merge_all()
location = os.path.join(self.working_dir, 'logs')
self.writer = tf.summary.FileWriter(location, graph=self.g)
saver_network = tf.train.Saver(var_list=self.ll_vars)
print('loading the network ...')
# Restores from checkpoint
# self.sess.run(tf.global_variables_initializer())
saver_network.restore(self.sess, self.model_dir)
print('Graph successfully loaded.')
def __call__(self, example_string):
"""Processes a single example string.
Extracts and processes the image, and ignores the label. We assume that the
image has three channels.
Args:
example_string: str, an Example protocol buffer.
Returns:
image_rescaled: the image, resized to `image_size x image_size` and
rescaled to [-1, 1]. Note that Gaussian data augmentation may cause values
to go beyond this range.
"""
image_string = tf.parse_single_example(
example_string,
features={
'image': tf.FixedLenFeature([], dtype=tf.string),
'label': tf.FixedLenFeature([], tf.int64)
})['image']
image_decoded = tf.image.decode_image(image_string, channels=3)
image_decoded.set_shape([None, None, 3])
image_resized = tf.image.resize_images(
image_decoded, [self.image_size, self.image_size],
method=tf.image.ResizeMethod.BILINEAR,
align_corners=True)
image = tf.cast(image_resized, tf.float32)
if self.data_augmentation is not None:
if self.data_augmentation.enable_random_brightness:
delta = self.data_augmentation.random_brightness_delta
image = tf.image.random_brightness(image, delta)
if self.data_augmentation.enable_random_saturation:
delta = self.data_augmentation.random_saturation_delta
image = tf.image.random_saturation(image, 1 - delta, 1 + delta)
if self.data_augmentation.enable_random_contrast:
delta = self.data_augmentation.random_contrast_delta
image = tf.image.random_contrast(image, 1 - delta, 1 + delta)
if self.data_augmentation.enable_random_hue:
delta = self.data_augmentation.random_hue_delta
image = tf.image.random_hue(image, delta)
if self.data_augmentation.enable_random_flip:
image = tf.image.random_flip_left_right(image)
image = 2 * (image / 255.0 - 0.5) # Rescale to [-1, 1].
if self.data_augmentation is not None:
if self.data_augmentation.enable_gaussian_noise:
image = image + tf.random_normal(tf.shape(
image)) * self.data_augmentation.gaussian_noise_std
if self.data_augmentation.enable_jitter:
j = self.data_augmentation.jitter_amount
paddings = tf.constant([[j, j], [j, j], [0, 0]])
image = tf.pad(image, paddings, 'REFLECT')
image = tf.image.random_crop(
image, [self.image_size, self.image_size, 3])
return image
