def _residual_block(input_net,
num_outputs,
kernel_size,
stride=1,
padding_size=0,
activation_fn=tf.nn.relu,
normalizer_fn=None,
name='residual_block'):
"""Residual Block.
Input Tensor X - > Conv1 -> IN -> ReLU -> Conv2 -> IN + X
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L7
Args:
input_net: Tensor as input.
num_outputs: (int) number of output channels for Convolution.
kernel_size: (int) size of the square kernel for Convolution.
stride: (int) stride for Convolution. Default to 1.
padding_size: (int) padding size for Convolution. Default to 0.
activation_fn: Activation function.
normalizer_fn: Normalization function.
name: Name scope
Returns:
Residual Tensor with the same shape as the input tensor.
"""
with tf.variable_scope(name):
with tf.contrib.framework.arg_scope([tf.contrib.layers.conv2d],
num_outputs=num_outputs,
kernel_size=kernel_size,
stride=stride,
padding='VALID',
normalizer_fn=normalizer_fn,
activation_fn=None):
res_block = ops.pad(input_net, padding_size)
res_block = tf.contrib.layers.conv2d(inputs=res_block,
scope='conv_0')
res_block = activation_fn(res_block, name='activation_0')
res_block = ops.pad(res_block, padding_size)
res_block = tf.contrib.layers.conv2d(inputs=res_block,
scope='conv_1')
output_net = res_block + input_net
return output_net
def test_padding_with_tensor_of_invalid_shape(self):
n = 2
invalid_rank = 1
h = 128
w = 64
c = 3
pad = 3
test_input_tensor = tf.random_uniform((n, invalid_rank, h, w, c))
with self.assertRaises(ValueError):
ops.pad(test_input_tensor, padding_size=pad)
def test_padding_with_tensor_of_invalid_shape(self):
n = 2
invalid_rank = 1
h = 128
w = 64
c = 3
pad = 3
test_input_tensor = tf.random_uniform((n, invalid_rank, h, w, c))
with self.assertRaises(ValueError):
ops.pad(test_input_tensor, padding_size=pad)
def _residual_block(input_net,
num_outputs,
kernel_size,
stride=1,
padding_size=0,
activation_fn=tf.nn.relu,
normalizer_fn=None,
name='residual_block'):
"""Residual Block.
Input Tensor X - > Conv1 -> IN -> ReLU -> Conv2 -> IN + X
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L7
Args:
input_net: Tensor as input.
num_outputs: (int) number of output channels for Convolution.
kernel_size: (int) size of the square kernel for Convolution.
stride: (int) stride for Convolution. Default to 1.
padding_size: (int) padding size for Convolution. Default to 0.
activation_fn: Activation function.
normalizer_fn: Normalization function.
name: Name scope
Returns:
Residual Tensor with the same shape as the input tensor.
"""
with tf.variable_scope(name):
with tf.contrib.framework.arg_scope(
[tf.contrib.layers.conv2d],
num_outputs=num_outputs,
kernel_size=kernel_size,
stride=stride,
padding='VALID',
normalizer_fn=normalizer_fn,
activation_fn=None):
res_block = ops.pad(input_net, padding_size)
res_block = tf.contrib.layers.conv2d(inputs=res_block, scope='conv_0')
res_block = activation_fn(res_block, name='activation_0')
res_block = ops.pad(res_block, padding_size)
res_block = tf.contrib.layers.conv2d(inputs=res_block, scope='conv_1')
output_net = res_block + input_net
return output_net
def discriminator_output_source(input_net):
"""Output Layer for Source in the Discriminator.
Determine if the image is real/fake based on the feature extracted. We follow
the original paper design where the output is not a simple (batch_size) shape
Tensor but rather a (batch_size, 2, 2, 2048) shape Tensor. We will get the
correct shape later when we piece things together.
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L79
Args:
input_net: Tensor of shape (batch_size, h / 64, w / 64, 2048) as features.
Returns:
Tensor of shape (batch_size, h / 64, w / 64, 1) as the score.
"""
with tf.variable_scope('discriminator_output_source'):
output_src = ops.pad(input_net, 1)
output_src = tf.contrib.layers.conv2d(
inputs=output_src,
num_outputs=1,
kernel_size=3,
stride=1,
padding='VALID',
activation_fn=None,
normalizer_fn=None,
biases_initializer=None,
scope='conv')
return output_src
def discriminator_output_source(input_net):
"""Output Layer for Source in the Discriminator.
Determine if the image is real/fake based on the feature extracted. We follow
the original paper design where the output is not a simple (batch_size) shape
Tensor but rather a (batch_size, 2, 2, 2048) shape Tensor. We will get the
correct shape later when we piece things together.
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L79
Args:
input_net: Tensor of shape (batch_size, h / 64, w / 64, 2048) as features.
Returns:
Tensor of shape (batch_size, h / 64, w / 64, 1) as the score.
"""
with tf.variable_scope('discriminator_output_source'):
output_src = ops.pad(input_net, 1)
output_src = tf.contrib.layers.conv2d(inputs=output_src,
num_outputs=1,
kernel_size=3,
stride=1,
padding='VALID',
activation_fn=None,
normalizer_fn=None,
biases_initializer=None,
scope='conv')
return output_src
def test_padding_with_3D_tensor(self):
h = 128
w = 64
c = 3
pad = 3
test_input_tensor = tf.random_uniform((h, w, c))
test_output_tensor = ops.pad(test_input_tensor, padding_size=pad)
with self.test_session() as sess:
output = sess.run(test_output_tensor)
self.assertTupleEqual((h + pad * 2, w + pad * 2, c), output.shape)
def test_padding_with_3D_tensor(self):
h = 128
w = 64
c = 3
pad = 3
test_input_tensor = tf.random_uniform((h, w, c))
test_output_tensor = ops.pad(test_input_tensor, padding_size=pad)
with self.test_session() as sess:
output = sess.run(test_output_tensor)
self.assertTupleEqual((h + pad * 2, w + pad * 2, c), output.shape)
def generator_up_sample(input_net, num_outputs):
"""Up-sampling module in Generator.
Up sampling path for image generation in the Generator.
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L44
Args:
input_net: Tensor of shape (batch_size, h / 4, w / 4, 256).
num_outputs: (int) Number of channel for the output tensor.
Returns:
Tensor of shape (batch_size, h, w, num_outputs).
"""
with tf.variable_scope('generator_up_sample'):
with tf.contrib.framework.arg_scope(
[tf.contrib.layers.conv2d_transpose],
kernel_size=4,
stride=2,
padding='VALID',
normalizer_fn=tf.contrib.layers.instance_norm,
activation_fn=tf.nn.relu):
up_sample = tf.contrib.layers.conv2d_transpose(inputs=input_net,
num_outputs=128,
scope='deconv_0')
up_sample = up_sample[:, 1:-1, 1:-1, :]
up_sample = tf.contrib.layers.conv2d_transpose(inputs=up_sample,
num_outputs=64,
scope='deconv_1')
up_sample = up_sample[:, 1:-1, 1:-1, :]
output_net = ops.pad(up_sample, 3)
output_net = tf.contrib.layers.conv2d(inputs=output_net,
num_outputs=num_outputs,
kernel_size=7,
stride=1,
padding='VALID',
activation_fn=tf.nn.tanh,
normalizer_fn=None,
biases_initializer=None,
scope='conv_0')
return output_net
def generator_up_sample(input_net, num_outputs):
"""Up-sampling module in Generator.
Up sampling path for image generation in the Generator.
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L44
Args:
input_net: Tensor of shape (batch_size, h / 4, w / 4, 256).
num_outputs: (int) Number of channel for the output tensor.
Returns:
Tensor of shape (batch_size, h, w, num_outputs).
"""
with tf.variable_scope('generator_up_sample'):
with tf.contrib.framework.arg_scope(
[tf.contrib.layers.conv2d_transpose],
kernel_size=4,
stride=2,
padding='VALID',
normalizer_fn=tf.contrib.layers.instance_norm,
activation_fn=tf.nn.relu):
up_sample = tf.contrib.layers.conv2d_transpose(
inputs=input_net, num_outputs=128, scope='deconv_0')
up_sample = up_sample[:, 1:-1, 1:-1, :]
up_sample = tf.contrib.layers.conv2d_transpose(
inputs=up_sample, num_outputs=64, scope='deconv_1')
up_sample = up_sample[:, 1:-1, 1:-1, :]
output_net = ops.pad(up_sample, 3)
output_net = tf.contrib.layers.conv2d(
inputs=output_net,
num_outputs=num_outputs,
kernel_size=7,
stride=1,
padding='VALID',
activation_fn=tf.nn.tanh,
normalizer_fn=None,
biases_initializer=None,
scope='conv_0')
return output_net
def discriminator_input_hidden(input_net, hidden_layer=6, init_num_outputs=64):
"""Input Layer + Hidden Layer in the Discriminator.
Feature extraction pathway in the Discriminator.
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L68
Args:
input_net: Tensor of shape (batch_size, h, w, 3) as batch of images.
hidden_layer: (int) Number of hidden layers. Default to 6 per the original
implementation.
init_num_outputs: (int) Number of hidden unit in the first hidden layer. The
number of hidden unit double after each layer. Default to 64 per the
original implementation.
Returns:
Tensor of shape (batch_size, h / 64, w / 64, 2048) as features.
"""
num_outputs = init_num_outputs
with tf.variable_scope('discriminator_input_hidden'):
hidden = input_net
for i in range(hidden_layer):
hidden = ops.pad(hidden, 1)
hidden = tf.contrib.layers.conv2d(
inputs=hidden,
num_outputs=num_outputs,
kernel_size=4,
stride=2,
padding='VALID',
activation_fn=None,
normalizer_fn=None,
scope='conv_{}'.format(i))
hidden = tf.nn.leaky_relu(hidden, alpha=0.01)
num_outputs = 2 * num_outputs
return hidden
def discriminator_input_hidden(input_net, hidden_layer=6, init_num_outputs=64):
"""Input Layer + Hidden Layer in the Discriminator.
Feature extraction pathway in the Discriminator.
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L68
Args:
input_net: Tensor of shape (batch_size, h, w, 3) as batch of images.
hidden_layer: (int) Number of hidden layers. Default to 6 per the original
implementation.
init_num_outputs: (int) Number of hidden unit in the first hidden layer. The
number of hidden unit double after each layer. Default to 64 per the
original implementation.
Returns:
Tensor of shape (batch_size, h / 64, w / 64, 2048) as features.
"""
num_outputs = init_num_outputs
with tf.variable_scope('discriminator_input_hidden'):
hidden = input_net
for i in range(hidden_layer):
hidden = ops.pad(hidden, 1)
hidden = tf.contrib.layers.conv2d(inputs=hidden,
num_outputs=num_outputs,
kernel_size=4,
stride=2,
padding='VALID',
activation_fn=None,
normalizer_fn=None,
scope='conv_{}'.format(i))
hidden = tf.nn.leaky_relu(hidden, alpha=0.01)
num_outputs = 2 * num_outputs
return hidden
def generator_down_sample(input_net, final_num_outputs=256):
"""Down-sampling module in Generator.
Down sampling pathway of the Generator Architecture:
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L32
Notes:
We require dimension 1 and dimension 2 of the input_net to be fully defined
for the correct down sampling.
Args:
input_net: Tensor of shape (batch_size, h, w, c + num_class).
final_num_outputs: (int) Number of hidden unit for the final layer.
Returns:
Tensor of shape (batch_size, h / 4, w / 4, 256).
Raises:
ValueError: If final_num_outputs are not divisible by 4,
or input_net does not have a rank of 4,
or dimension 1 and dimension 2 of input_net are not defined at graph
construction time,
or dimension 1 and dimension 2 of input_net are not divisible by 4.
"""
if final_num_outputs % 4 != 0:
raise ValueError('Final number outputs need to be divisible by 4.')
# Check the rank of input_net.
input_net.shape.assert_has_rank(4)
# Check dimension 1 and dimension 2 are defined and divisible by 4.
if input_net.shape[1]:
if input_net.shape[1] % 4 != 0:
raise ValueError(
'Dimension 1 of the input should be divisible by 4, but is {} '
'instead.'.
format(input_net.shape[1]))
else:
raise ValueError('Dimension 1 of the input should be explicitly defined.')
# Check dimension 1 and dimension 2 are defined and divisible by 4.
if input_net.shape[2]:
if input_net.shape[2] % 4 != 0:
raise ValueError(
'Dimension 2 of the input should be divisible by 4, but is {} '
'instead.'.
format(input_net.shape[2]))
else:
raise ValueError('Dimension 2 of the input should be explicitly defined.')
with tf.variable_scope('generator_down_sample'):
with tf.contrib.framework.arg_scope(
[tf.contrib.layers.conv2d],
padding='VALID',
biases_initializer=None,
normalizer_fn=tf.contrib.layers.instance_norm,
activation_fn=tf.nn.relu):
down_sample = ops.pad(input_net, 3)
down_sample = tf.contrib.layers.conv2d(
inputs=down_sample,
num_outputs=final_num_outputs / 4,
kernel_size=7,
stride=1,
scope='conv_0')
down_sample = ops.pad(down_sample, 1)
down_sample = tf.contrib.layers.conv2d(
inputs=down_sample,
num_outputs=final_num_outputs / 2,
kernel_size=4,
stride=2,
scope='conv_1')
down_sample = ops.pad(down_sample, 1)
output_net = tf.contrib.layers.conv2d(
inputs=down_sample,
num_outputs=final_num_outputs,
kernel_size=4,
stride=2,
scope='conv_2')
return output_net
def generator_down_sample(input_net, final_num_outputs=256):
"""Down-sampling module in Generator.
Down sampling pathway of the Generator Architecture:
PyTorch Version:
https://github.com/yunjey/StarGAN/blob/fbdb6a6ce2a4a92e1dc034faec765e0dbe4b8164/model.py#L32
Notes:
We require dimension 1 and dimension 2 of the input_net to be fully defined
for the correct down sampling.
Args:
input_net: Tensor of shape (batch_size, h, w, c + num_class).
final_num_outputs: (int) Number of hidden unit for the final layer.
Returns:
Tensor of shape (batch_size, h / 4, w / 4, 256).
Raises:
ValueError: If final_num_outputs are not divisible by 4,
or input_net does not have a rank of 4,
or dimension 1 and dimension 2 of input_net are not defined at graph
construction time,
or dimension 1 and dimension 2 of input_net are not divisible by 4.
"""
if final_num_outputs % 4 != 0:
raise ValueError('Final number outputs need to be divisible by 4.')
# Check the rank of input_net.
input_net.shape.assert_has_rank(4)
# Check dimension 1 and dimension 2 are defined and divisible by 4.
if input_net.shape[1]:
if input_net.shape[1] % 4 != 0:
raise ValueError(
'Dimension 1 of the input should be divisible by 4, but is {} '
'instead.'.format(input_net.shape[1]))
else:
raise ValueError(
'Dimension 1 of the input should be explicitly defined.')
# Check dimension 1 and dimension 2 are defined and divisible by 4.
if input_net.shape[2]:
if input_net.shape[2] % 4 != 0:
raise ValueError(
'Dimension 2 of the input should be divisible by 4, but is {} '
'instead.'.format(input_net.shape[2]))
else:
raise ValueError(
'Dimension 2 of the input should be explicitly defined.')
with tf.variable_scope('generator_down_sample'):
with tf.contrib.framework.arg_scope(
[tf.contrib.layers.conv2d],
padding='VALID',
biases_initializer=None,
normalizer_fn=tf.contrib.layers.instance_norm,
activation_fn=tf.nn.relu):
down_sample = ops.pad(input_net, 3)
down_sample = tf.contrib.layers.conv2d(
inputs=down_sample,
num_outputs=final_num_outputs / 4,
kernel_size=7,
stride=1,
scope='conv_0')
down_sample = ops.pad(down_sample, 1)
down_sample = tf.contrib.layers.conv2d(
inputs=down_sample,
num_outputs=final_num_outputs / 2,
kernel_size=4,
stride=2,
scope='conv_1')
down_sample = ops.pad(down_sample, 1)
output_net = tf.contrib.layers.conv2d(
inputs=down_sample,
num_outputs=final_num_outputs,
kernel_size=4,
stride=2,
scope='conv_2')
return output_net
def run(self, args):
print(" [*] Running....")
if not os.path.exists(args.output_path):
os.makedirs(args.output_path)
self.sess.run(tf.global_variables_initializer())
self.sess.run(tf.local_variables_initializer())
disp_batch = self.dataloader.disp
left_batch = self.dataloader.left
right_batch = self.dataloader.right
image_name = self.dataloader.filename
num_samples = count_text_lines(self.dataset)
if self.mode == 'reprojection':
print(' [*] Generating Reprojection error')
warped = ops.generate_image_left(self.placeholders['right'],
self.placeholders['disp'])
net_output = -(tf.reduce_sum(
0.85 * ops.SSIM(warped, self.placeholders['left']) +
0.15 * tf.abs(warped - self.placeholders['left']),
-1,
keepdims=True))
elif self.mode == 'agreement':
print(' [*] Generating Disparity Agreement')
net_output = tf.py_func(ops.agreement,
[self.placeholders['disp'], 2], tf.float32)
elif self.mode == 'uniqueness':
print(' [*] Generating Uniqueness Constraint')
net_output = tf.py_func(ops.uniqueness,
[self.placeholders['disp']], tf.float32)
elif 'otb' in self.mode:
print(' [*] OTB inference')
self.saver = tf.train.Saver()
if args.checkpoint_path:
self.saver.restore(self.sess, args.checkpoint_path)
print(" [*] Load model: SUCCESS")
else:
print(" [*] Load failed...neglected")
print(" [*] End Testing...")
raise ValueError('args.checkpoint_path is None')
net_output = tf.nn.sigmoid(self.prediction)
if self.mode == 'otb-online':
print(' [*] Online Adaptation')
self.optimizer = tf.train.GradientDescentOptimizer(
self.learning_rate).minimize(
self.loss, var_list=tf.global_variables())
else:
print(" [*] Unsupported testing mode!")
raise ValueError('args.mode is not supported')
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
print(" [*] Start Testing...")
bar = progressbar.ProgressBar(max_value=num_samples)
for step in range(num_samples):
batch_left, batch_right, batch_disp, filename = self.sess.run(
[left_batch, right_batch, disp_batch, image_name])
val_disp, hpad, wpad = ops.pad(batch_disp, self.image_height,
self.image_width)
val_left, _, _ = ops.pad(batch_left, self.image_height,
self.image_width)
val_right, _, _ = ops.pad(batch_right, self.image_height,
self.image_width)
if self.mode == 'otb-online':
_, confidence = self.sess.run(
[self.optimizer, net_output],
feed_dict={
self.placeholders['disp']: val_disp,
self.placeholders['left']: val_left,
self.placeholders['right']: val_right,
self.learning_rate: self.initial_learning_rate
})
else:
confidence = self.sess.run(net_output,
feed_dict={
self.placeholders['disp']:
val_disp,
self.placeholders['left']:
val_left,
self.placeholders['right']:
val_right
})
confidence = ops.depad(confidence, hpad, wpad)
outdir = args.output_path + '/' + filename
if not os.path.exists(outdir):
os.makedirs(outdir)
confidence_file = outdir + '/' + self.mode + '.png'
c = confidence[0]
c = (c - np.min(c)) / (np.max(c) - np.min(c))
cv2.imwrite(confidence_file, (c * (2**16 - 1)).astype('uint16'))
if self.colors:
color_file = outdir + '/' + self.mode + '-color.png'
cv2.imwrite(
color_file,
cv2.applyColorMap(((1 - c) * (2**8 - 1)).astype('uint8'),
cv2.COLORMAP_WINTER))
bar.update(step + 1)
coord.request_stop()
coord.join(threads)
def test(self, args):
print("[*] Testing....")
if not os.path.exists(args.output_path):
os.makedirs(args.output_path)
self.sess.run(tf.global_variables_initializer())
self.sess.run(tf.local_variables_initializer())
if self.model == 'LGC':
self.vars = tf.all_variables()
self.vars_global = [k for k in self.vars if k.name.startswith('ConfNet')]
self.vars_local = [k for k in self.vars if (k.name.startswith('CCNN') or k.name.startswith('LFN')) ]
self.vars_lgc = [k for k in self.vars if k.name.startswith('LGC')]
self.saver_global = tf.train.Saver(self.vars_global)
self.saver_local = tf.train.Saver(self.vars_local)
self.saver_LGC = tf.train.Saver(self.vars_lgc)
if args.checkpoint_path[0] and args.checkpoint_path[1] and args.checkpoint_path[2]:
self.saver_global.restore(self.sess, args.checkpoint_path[0])
self.saver_local.restore(self.sess, args.checkpoint_path[1])
self.saver_LGC.restore(self.sess, args.checkpoint_path[2])
print(" [*] Load model: SUCCESS")
else:
print(" [*] Load failed...neglected")
print(" [*] End Testing...")
raise ValueError('args.checkpoint_path[0] or args.checkpoint_path[1] or args.checkpoint_path[2] is None')
else:
self.saver = tf.train.Saver()
if args.checkpoint_path:
self.saver.restore(self.sess, args.checkpoint_path[0])
print(" [*] Load model: SUCCESS")
else:
print(" [*] Load failed...neglected")
print(" [*] End Testing...")
raise ValueError('args.checkpoint_path is None')
disp_batch = self.dataloader.disp
left_batch = self.dataloader.left
line = self.dataloader.disp_filename
num_samples = count_text_lines(self.dataset)
if self.model == 'ConfNet':
prediction = tf.nn.sigmoid(self.prediction)
else:
prediction = tf.pad(tf.nn.sigmoid(self.prediction), tf.constant([[0, 0], [self.radius, self.radius], [self.radius, self.radius], [0, 0]]), "CONSTANT")
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
print(" [*] Start Testing...")
for step in range(num_samples):
batch_left, batch_disp, filename = self.sess.run([left_batch, disp_batch, line])
if self.model == 'ConfNet' or self.model == 'LGC':
val_disp, hpad, wpad = ops.pad(batch_disp)
val_left, _, _ = ops.pad(batch_left)
print(" [*] Test image:" + filename)
start = time.time()
if self.model == 'ConfNet' or self.model == 'LGC':
confidence = self.sess.run(prediction, feed_dict={self.left: val_left, self.disp: val_disp})
confidence = ops.depad(confidence, hpad, wpad)
else:
confidence = self.sess.run(prediction, feed_dict={self.left: batch_left, self.disp: batch_disp})
current = time.time()
output_file = args.output_path + filename.strip().split('/')[-1]
cv2.imwrite(output_file, (confidence[0] * 65535.0).astype('uint16'))
print(" [*] Confidence prediction saved in:" + output_file)
print(" [*] Running time:" + str(current - start) + "s")
coord.request_stop()
coord.join(threads)