def initial_block(self, x, scope):
# convolution branch:
W_conv = self.get_variable_weight_decay(
scope + "/W",
shape=[3, 3, 3, 13],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
b_conv = self.get_variable_weight_decay(
scope + "/b",
shape=[13], # ([out_depth])
initializer=tf.constant_initializer(0),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(
x, W_conv, strides=[1, 2, 2, 1], padding="SAME") + b_conv
# max pooling branch:
pool_branch = tf.nn.max_pool(x,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding="VALID")
# concatenate the branches:
concat = tf.concat([conv_branch, pool_branch],
axis=3) # (3: the depth axis)
# apply batch normalization and PReLU:
output = tf.contrib.slim.batch_norm(concat)
output = PReLU(output, scope=scope)
return output
def encoder_bottleneck_asymmetric(self,
x,
output_depth,
keep_prob,
scope,
proj_ratio=4):
input_shape = x.get_shape().as_list()
input_depth = input_shape[3]
internal_depth = int(output_depth / proj_ratio)
# convolution branch:
conv_branch = x
# # 1x1 projection:
W_proj = self.get_variable_weight_decay(
scope + "/W_proj",
shape=[1, 1, input_depth, internal_depth],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(conv_branch,
W_proj,
strides=[1, 1, 1, 1],
padding="VALID") # NOTE! no bias terms
# # # batch norm and PReLU:
conv_branch = tf.contrib.slim.batch_norm(conv_branch)
conv_branch = PReLU(conv_branch, scope=scope + "/proj")
# # asymmetric conv:
# # # asymmetric conv 1:
W_conv1 = self.get_variable_weight_decay(
scope + "/W_conv1",
shape=[5, 1, internal_depth, internal_depth],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(conv_branch,
W_conv1,
strides=[1, 1, 1, 1],
padding="SAME") # NOTE! no bias terms
# # # asymmetric conv 2:
W_conv2 = self.get_variable_weight_decay(
scope + "/W_conv2",
shape=[1, 5, internal_depth, internal_depth],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
b_conv2 = self.get_variable_weight_decay(
scope + "/b_conv2",
shape=[internal_depth], # ([out_depth])
initializer=tf.constant_initializer(0),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(
conv_branch, W_conv2, strides=[1, 1, 1, 1
], padding="SAME") + b_conv2
# # # batch norm and PReLU:
conv_branch = tf.contrib.slim.batch_norm(conv_branch)
conv_branch = PReLU(conv_branch, scope=scope + "/conv")
# # 1x1 expansion:
shape = [1, 1, internal_depth, output_depth]
W_exp = self.get_variable_weight_decay(
scope + "/W_exp",
shape=shape,
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
W_exp = tf.reshape(drop_connect(W_exp, keep_prob), shape=shape)
conv_branch = tf.nn.conv2d(conv_branch,
W_exp,
strides=[1, 1, 1, 1],
padding="VALID") # NOTE! no bias terms
# # # batch norm:
conv_branch = tf.contrib.slim.batch_norm(conv_branch)
# NOTE! no PReLU here
# # regularizer:
# conv_branch = dropout(conv_branch, self.keep_prob_pl)
# main branch:
main_branch = x
# add the branches:
merged = conv_branch + main_branch
# apply PReLU:
output = PReLU(merged, scope=scope + "/output")
return output
def encoder_bottleneck_regular(self,
x,
output_depth,
scope,
keep_prob,
proj_ratio=4,
downsampling=False):
input_shape = x.get_shape().as_list()
input_depth = input_shape[3]
internal_depth = int(output_depth / proj_ratio)
# convolution branch:
conv_branch = x
# # 1x1 projection:
if downsampling:
W_conv = self.get_variable_weight_decay(
scope + "/W_proj",
shape=[2, 2, input_depth, internal_depth],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(conv_branch,
W_conv,
strides=[1, 2, 2, 1],
padding="VALID") # NOTE! no bias terms
else:
W_proj = self.get_variable_weight_decay(
scope + "/W_proj",
shape=[1, 1, input_depth, internal_depth],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(conv_branch,
W_proj,
strides=[1, 1, 1, 1],
padding="VALID") # NOTE! no bias terms
# # # batch norm and PReLU:
conv_branch = tf.contrib.slim.batch_norm(conv_branch)
conv_branch = PReLU(conv_branch, scope=scope + "/proj")
# # conv:
W_conv = self.get_variable_weight_decay(
scope + "/W_conv",
shape=[3, 3, internal_depth, internal_depth],
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
b_conv = self.get_variable_weight_decay(
scope + "/b_conv",
shape=[internal_depth], # ([out_depth])
initializer=tf.constant_initializer(0),
loss_category="encoder_wd_losses")
conv_branch = tf.nn.conv2d(
conv_branch, W_conv, strides=[1, 1, 1, 1], padding="SAME") + b_conv
# # # batch norm and PReLU:
conv_branch = tf.contrib.slim.batch_norm(conv_branch)
conv_branch = PReLU(conv_branch, scope=scope + "/conv")
# # 1x1 expansion:
shape = [1, 1, internal_depth, output_depth]
W_exp = self.get_variable_weight_decay(
scope + "/W_exp",
shape=shape,
# ([filter_height, filter_width, in_depth, out_depth])
initializer=tf.contrib.layers.xavier_initializer(),
loss_category="encoder_wd_losses")
W_exp = tf.reshape(drop_connect(W_exp, self.keep_prob_pl), shape=shape)
conv_branch = tf.nn.conv2d(conv_branch,
W_exp,
strides=[1, 1, 1, 1],
padding="VALID") # NOTE! no bias terms
# # # batch norm:
conv_branch = tf.contrib.slim.batch_norm(conv_branch)
# NOTE! no PReLU here
# # regularizer:
# conv_branch = dropout(conv_branch, self.keep_prob_pl)
# main branch:
main_branch = x
if downsampling:
# max pooling with argmax (for use in max_unpool in the decoder):
main_branch, pooling_indices = tf.nn.max_pool_with_argmax(
main_branch,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding="SAME")
# (everytime we downsample, we also increase the feature block depth)
# pad with zeros so that the feature block depth matches:
depth_to_pad = output_depth - input_depth
paddings = tf.convert_to_tensor([[0, 0], [0, 0], [0, 0],
[0, depth_to_pad]])
# (paddings is an integer tensor of shape [4, 2] where 4 is the rank
# of main_branch. For each dimension D (D = 0, 1, 2, 3) of main_branch,
# paddings[D, 0] is the no of values to add before the contents of
# main_branch in that dimension, and paddings[D, 0] is the no of
# values to add after the contents of main_branch in that dimension)
main_branch = tf.pad(main_branch,
paddings=paddings,
mode="CONSTANT")
# add the branches:
merged = conv_branch + main_branch
# apply PReLU:
output = PReLU(merged, scope=scope + "/output")
if downsampling:
return output, pooling_indices
else:
return output