def residual_block(x, in_filter, out_filter, stride, hparams=None, layer=None):
"""Adds residual connection to `x` in addition to applying BN ->Act->3x3 Conv.
Args:
x: Tensor that is the output of the previous layer in the model.
in_filter: Number of filters `x` has.
out_filter: Number of filters that the output of this layer will have.
stride: Integer that specified what stride should be applied to `x`.
hparams: hparams
layer: layer
Returns:
A Tensor that is the result of applying two sequences of BN->Act->3x3 Conv
and then adding that Tensor to `x`.
"""
if layer is not None:
tf.logging.info('layer: {}'.format(layer))
if hparams is None:
hparams = {'use_swish': 0, 'beta_swish': 0, 'use_BN': 0}
with tf.variable_scope('residual_activation'):
block_x = ops.maybe_normalize(x, scope='init_bn')
block_x = activation_fn(block_x, hparams=hparams, layer=layer)
if in_filter != out_filter:
x = ops.conv2d(block_x,
out_filter,
3,
stride=stride,
scope='conva')
with tf.variable_scope('sub1'):
block_x = ops.conv2d(block_x,
out_filter,
3,
stride=stride,
scope='conv1',
position='residual')
with tf.variable_scope('sub2'):
block_x = ops.maybe_normalize(block_x, scope='bn2')
block_x = activation_fn(block_x, hparams=hparams, layer=layer)
block_x = ops.conv2d(block_x,
out_filter,
3,
stride=1,
scope='conv2',
position='residual_last')
return x + block_x
def _shake_shake_branch(x, output_filters, stride, rand_forward, rand_backward,
is_training):
"""Building a 2 branching convnet."""
x = tf.nn.relu(x)
x = ops.conv2d(x, output_filters, 3, stride=stride, scope='conv1')
x = ops.maybe_normalize(x, scope='bn1')
x = tf.nn.relu(x)
x = ops.conv2d(x, output_filters, 3, scope='conv2')
x = ops.maybe_normalize(x, scope='bn2')
if is_training:
x = x * rand_backward + tf.stop_gradient(x * rand_forward -
x * rand_backward)
else:
x *= 1.0 / 2
return x
def _shake_shake_skip_connection(x, output_filters, stride):
"""Adds a residual connection to the filter x for the shake-shake model."""
curr_filters = int(x.shape[3])
if curr_filters == output_filters:
return x
stride_spec = ops.stride_arr(stride, stride)
# Skip path 1
path1 = tf.nn.avg_pool(x, [1, 1, 1, 1],
stride_spec,
'VALID',
data_format='NHWC')
path1 = ops.conv2d(path1, int(output_filters / 2), 1, scope='path1_conv')
# Skip path 2
# First pad with 0's then crop
pad_arr = [[0, 0], [0, 1], [0, 1], [0, 0]]
path2 = tf.pad(x, pad_arr)[:, 1:, 1:, :]
concat_axis = 3
path2 = tf.nn.avg_pool(path2, [1, 1, 1, 1],
stride_spec,
'VALID',
data_format='NHWC')
path2 = ops.conv2d(path2, int(output_filters / 2), 1, scope='path2_conv')
# Concat and apply BN
final_path = tf.concat(values=[path1, path2], axis=concat_axis)
final_path = ops.maybe_normalize(final_path, scope='final_path_bn')
return final_path
def build_shake_shake_model(images, num_classes, hparams, is_training):
"""Builds the Shake-Shake model.
Build the Shake-Shake model from https://arxiv.org/abs/1705.07485.
Args:
images: Tensor of images that will be fed into the Wide ResNet Model.
num_classes: Number of classed that the model needs to predict.
hparams: tf.HParams object that contains additional hparams needed to
construct the model. In this case it is the `shake_shake_widen_factor`
that is used to determine how many filters the model has.
is_training: Is the model training or not.
Returns:
The logits of the Shake-Shake model.
"""
depth = 26
k = hparams.shake_shake_widen_factor # The widen factor
n = int((depth - 2) / 6)
x = images
x = ops.conv2d(x, 16, 3, scope='init_conv')
x = ops.maybe_normalize(x, scope='init_bn')
with tf.variable_scope('L1'):
x = _shake_shake_layer(x, 16 * k, n, 1, is_training)
with tf.variable_scope('L2'):
x = _shake_shake_layer(x, 32 * k, n, 2, is_training)
with tf.variable_scope('L3'):
x = _shake_shake_layer(x, 64 * k, n, 2, is_training)
x = tf.nn.relu(x)
x = ops.global_avg_pool(x)
# Fully connected
logits = ops.fc(x, num_classes)
return logits, x
def bottleneck_layer(x, n, stride, prob, is_training, alpha, beta):
"""Bottleneck layer for shake drop model."""
assert alpha[1] > alpha[0]
assert beta[1] > beta[0]
with tf.variable_scope('bottleneck_{}'.format(prob)):
input_layer = x
x = ops.maybe_normalize(x, scope='bn_1_pre')
x = ops.conv2d(x, n, 1, scope='1x1_conv_contract')
x = ops.maybe_normalize(x, scope='bn_1_post')
x = tf.nn.relu(x)
x = ops.conv2d(x, n, 3, stride=stride, scope='3x3')
x = ops.maybe_normalize(x, scope='bn_2')
x = tf.nn.relu(x)
x = ops.conv2d(x, n * 4, 1, scope='1x1_conv_expand')
x = ops.maybe_normalize(x, scope='bn_3')
# Apply regularization here
# Sample bernoulli with prob
if is_training:
batch_size = tf.shape(x)[0]
bern_shape = [batch_size, 1, 1, 1]
random_tensor = prob
random_tensor += tf.random_uniform(bern_shape, dtype=tf.float32)
binary_tensor = tf.floor(random_tensor)
alpha_values = tf.random_uniform([batch_size, 1, 1, 1],
minval=alpha[0],
maxval=alpha[1],
dtype=tf.float32)
beta_values = tf.random_uniform([batch_size, 1, 1, 1],
minval=beta[0],
maxval=beta[1],
dtype=tf.float32)
rand_forward = (binary_tensor + alpha_values -
binary_tensor * alpha_values)
rand_backward = (binary_tensor + beta_values -
binary_tensor * beta_values)
x = x * rand_backward + tf.stop_gradient(x * rand_forward -
x * rand_backward)
else:
expected_alpha = (alpha[1] + alpha[0]) / 2
# prob is the expectation of the bernoulli variable
x = (prob + expected_alpha - prob * expected_alpha) * x
res = shortcut(input_layer, n * 4, stride)
return x + res
def build_wrn_model(images, num_classes, hparams):
"""Builds the WRN model.
Build the Wide ResNet model from https://arxiv.org/abs/1605.07146.
Args:
images: Tensor of images that will be fed into the Wide ResNet Model.
num_classes: Number of classed that the model needs to predict.
hparams: hparams.
Returns:
The logits of the Wide ResNet model.
"""
kernel_size = hparams.wrn_size
filter_size = 3
num_blocks_per_resnet = 4
filters = [
min(kernel_size, 16), kernel_size, kernel_size * 2, kernel_size * 4
]
strides = [1, 2, 2] # stride for each resblock
# Run the first conv
with tf.variable_scope('init'):
x = images
output_filters = filters[0]
x = ops.conv2d(x,
output_filters,
filter_size,
scope='init_conv',
position='input')
layer = 0
for block_num in range(1, 4):
with tf.variable_scope('unit_{}_0'.format(block_num)):
x = residual_block(x,
filters[block_num - 1],
filters[block_num],
strides[block_num - 1],
hparams=hparams,
layer=layer)
layer += 1
for i in range(1, num_blocks_per_resnet):
with tf.variable_scope('unit_{}_{}'.format(block_num, i)):
x = residual_block(x,
filters[block_num],
filters[block_num],
1,
hparams=hparams,
layer=layer)
layer += 1
with tf.variable_scope('unit_last'):
x = ops.maybe_normalize(x, scope='final_bn')
x = activation_fn(x, hparams=hparams, layer=layer)
hiddens = ops.global_avg_pool(x)
logits = ops.fc(hiddens, num_classes)
return logits, hiddens
def build_shake_drop_model(images, num_classes, is_training):
"""Builds the PyramidNet Shake-Drop model.
Build the PyramidNet Shake-Drop model from https://arxiv.org/abs/1802.02375.
Args:
images: Tensor of images that will be fed into the Wide ResNet Model.
num_classes: Number of classed that the model needs to predict.
is_training: Is the model training or not.
Returns:
The logits of the PyramidNet Shake-Drop model.
"""
# ShakeDrop Hparams
p_l = 0.5
alpha_shake = [-1, 1]
beta_shake = [0, 1]
# PyramidNet Hparams
alpha = 200
depth = 272
# This is for the bottleneck architecture specifically
n = int((depth - 2) / 9)
start_channel = 16
add_channel = alpha / (3 * n)
# Building the models
x = images
x = ops.conv2d(x, 16, 3, scope='init_conv')
x = ops.maybe_normalize(x, scope='init_bn')
layer_num = 1
total_layers = n * 3
start_channel += add_channel
prob = calc_prob(layer_num, total_layers, p_l)
x = bottleneck_layer(x, round_int(start_channel), 1, prob, is_training,
alpha_shake, beta_shake)
layer_num += 1
for _ in range(1, n):
start_channel += add_channel
prob = calc_prob(layer_num, total_layers, p_l)
x = bottleneck_layer(x, round_int(start_channel), 1, prob, is_training,
alpha_shake, beta_shake)
layer_num += 1
start_channel += add_channel
prob = calc_prob(layer_num, total_layers, p_l)
x = bottleneck_layer(x, round_int(start_channel), 2, prob, is_training,
alpha_shake, beta_shake)
layer_num += 1
for _ in range(1, n):
start_channel += add_channel
prob = calc_prob(layer_num, total_layers, p_l)
x = bottleneck_layer(x, round_int(start_channel), 1, prob, is_training,
alpha_shake, beta_shake)
layer_num += 1
start_channel += add_channel
prob = calc_prob(layer_num, total_layers, p_l)
x = bottleneck_layer(x, round_int(start_channel), 2, prob, is_training,
alpha_shake, beta_shake)
layer_num += 1
for _ in range(1, n):
start_channel += add_channel
prob = calc_prob(layer_num, total_layers, p_l)
x = bottleneck_layer(x, round_int(start_channel), 1, prob, is_training,
alpha_shake, beta_shake)
layer_num += 1
assert layer_num - 1 == total_layers
x = ops.maybe_normalize(x, scope='final_bn')
x = tf.nn.relu(x)
x = ops.global_avg_pool(x)
# Fully connected
logits = ops.fc(x, num_classes)
return logits, x
