def _condition_conv(self, inputs, inputs_extra, layer_sizes):
channels = inputs.get_shape().as_list()[-1]
inputs_extra_flat = tf.keras.layers.Flatten()(inputs_extra)
scales = utils.mlp(inputs_extra_flat, layer_sizes + [channels])
shifts = utils.mlp(inputs_extra_flat, layer_sizes + [channels])
scales = tf.expand_dims(tf.expand_dims(scales, -2), -2)
shifts = tf.expand_dims(tf.expand_dims(shifts, -2), -2)
if self._conditioning_postprocessing is not None:
scales, shifts = self._conditioning_postprocessing(scales, shifts)
return scales * inputs + shifts
def _call_conv(self, inputs, inputs_extra):
endpoints = {}
x = inputs
for nblock in range(self._num_blocks):
if (inputs_extra is not None and
(self._conditioning_type == "concat" or
(self._conditioning_type == "input" and nblock == 0))):
x = broadcast_and_concat(x, inputs_extra)
num_channels = self._base_num_channels * (2**nblock)
with tf.variable_scope("conv_block_{}".format(nblock)):
x = self._conv_block(x, self._layers_per_block, num_channels)
if (inputs_extra is not None
and self._conditioning_type == "mult_and_add"):
endpoints["conv_block_before_cond_{}".format(nblock)] = x
with tf.variable_scope("conditioning"):
x = self._condition_conv(
x, inputs_extra, self._conditioning_layer_sizes)
endpoints["conv_block_before_nonlin_{}".format(nblock)] = x
if nblock < self._num_blocks - 1:
x = self._nonlinearity(x)
endpoints["conv_block_{}".format(nblock)] = x
if self._fc_layer_sizes is not None:
x = tf.keras.layers.Flatten()(x)
x = utils.mlp(x, self._fc_layer_sizes)
return x, endpoints
def _call_upconv(self, inputs, inputs_extra):
endpoints = {}
x = inputs
if self._fc_layer_sizes is not None:
x = tf.keras.layers.Flatten()(x)
x = utils.mlp(x, self._fc_layer_sizes)
x = tf.reshape(x, [-1] + self._upconv_reshape_size)
for nblock in range(self._num_blocks - 1, -1, -1):
num_channels = self._base_num_channels * (2**nblock)
if (inputs_extra is not None and
(self._conditioning_type == "concat" or
(self._conditioning_type == "input" and nblock == 0))):
x = broadcast_and_concat(x, inputs_extra)
with tf.variable_scope("upconv_block_{}".format(nblock)):
x = self._upconv_block(x, self._layers_per_block, num_channels)
if (inputs_extra is not None
and self._conditioning_type == "mult_and_add"):
endpoints["upconv_block_before_cond_{}".format(nblock)] = x
with tf.variable_scope("conditioning"):
x = self._condition_conv(
x, inputs_extra, self._conditioning_layer_sizes)
endpoints["upconv_block_before_nonlin_{}".format(nblock)] = x
x = self._nonlinearity(x)
endpoints["upconv_block_{}".format(nblock)] = x
# final layer that outputs the required number of channels
x = tf.layers.conv2d(inputs=x,
filters=self._channels_out,
kernel_size=3,
strides=(1, 1),
padding="same",
kernel_initializer=self._kernel_initializer)
return x, endpoints