def _deconv(self, name, x, out_channels, kernel_size=(3, 3), stride=2):
with tf.variable_scope(name):
h, w = x.shape.as_list()[1:3]
h, w = h * stride, w * stride
output_shape = [self.args.batch_size, h, w, out_channels]
stride = [1, stride, stride, 1]
kernel_shape = [
kernel_size[0], kernel_size[1], out_channels,
x.shape.as_list()[-1]
]
w = get_deconv_filter(kernel_shape, self.args.weight_decay)
variable_summaries(w)
out = tf.nn.conv2d_transpose(x,
w,
tf.stack(output_shape),
strides=stride,
padding="SAME")
if self.args.use_bias:
bias = tf.get_variable('biases', [output_shape[-1]],
initializer=tf.constant_initializer(
self.args.bias))
variable_summaries(bias)
out = tf.nn.bias_add(out, bias)
tf.add_to_collection('debug_layers', out)
tf.add_to_collection('decoding_trainable_vars', w)
tf.add_to_collection('decoding_trainable_vars', bias)
return out
def _conv(name,
x,
num_filters=16,
kernel_size=(3, 3),
padding='SAME',
stride=(1, 1),
initializer=tf.contrib.layers.xavier_initializer(),
l2_strength=0.0,
dilation=1.0,
bias=-1):
with tf.variable_scope(name):
stride = [1, stride[0], stride[1], 1]
kernel_shape = [
kernel_size[0], kernel_size[1], x.shape[-1], num_filters
]
w = variable_with_weight_decay(kernel_shape, initializer,
l2_strength)
variable_summaries(w)
if dilation > 1:
conv = tf.nn.atrous_conv2d(x, w, dilation, padding)
else:
if type(padding) == type(''):
conv = tf.nn.conv2d(x, w, stride, padding)
else:
conv = tf.pad(x, padding, "CONSTANT")
conv = tf.nn.conv2d(conv, w, stride, padding='VALID')
if bias != -1:
bias = tf.get_variable(
'biases', [num_filters],
initializer=tf.constant_initializer(bias))
variable_summaries(bias)
conv = tf.nn.bias_add(conv, bias)
tf.add_to_collection('debug_layers', conv)
return conv
def _fc(name,
x,
output_dim=128,
initializer=tf.contrib.layers.xavier_initializer(),
l2_strength=0.0,
bias=0.0):
with tf.variable_scope(name):
n_in = x.get_shape()[-1].value
w = variable_with_weight_decay([n_in, output_dim], initializer,
l2_strength)
variable_summaries(w)
if isinstance(bias, float):
bias = tf.get_variable("biases", [output_dim], tf.float32,
tf.constant_initializer(bias))
variable_summaries(bias)
output = tf.nn.bias_add(tf.matmul(x, w), bias)
return output