def top(self, body_output, _):
"""Transform inputs from model space to target space.
Perform the Xception "Exit flow", consisting of a single residual block and
two separable convolutional upscalings followed by global spatial average
pooling.
Args:
body_output: A Tensor with shape [batch, ?, ?, body_output_size].
Returns:
a Tensors, each with shape [batch_size, ?, ?, vocab_size]
"""
with tf.variable_scope(self.name):
x = body_output
# Assume input is a square with self._body_input_depth channels.
if self._is_2d:
length_float = tf.to_float(tf.shape(x)[1])
spatial_dim_float = tf.sqrt(length_float)
spatial_dim = tf.to_int32(spatial_dim_float)
x = tf.reshape(
x, [-1, spatial_dim, spatial_dim, self._body_input_depth])
x = common_layers.conv_block_downsample(x, self._kernel,
self._strides,
self._padding)
x = tf.nn.relu(x)
x = tf.reduce_mean(x, axis=[1, 2], keep_dims=True)
res = common_layers.conv(x, self._vocab_size, (1, 1))
return tf.expand_dims(res, 3)
def testConv(self):
x = np.random.rand(5, 7, 1, 11)
with self.test_session() as session:
y = common_layers.conv(tf.constant(x, dtype=tf.float32), 13, (3, 3))
session.run(tf.global_variables_initializer())
res = session.run(y)
self.assertEqual(res.shape, (5, 5, 1, 13))
def top(self, body_output, _):
"""Transform inputs from model space to target space.
Perform the Xception "Exit flow", consisting of a single residual block and
two separable convolutional upscalings followed by global spatial average
pooling.
Args:
body_output: A Tensor with shape [batch, ?, ?, body_output_size].
Returns:
a Tensors, each with shape [batch_size, ?, ?, vocab_size]
"""
with tf.variable_scope(self.name):
x = body_output
# Assume input is a square with self._body_input_depth channels.
if self._is_2d:
length_float = tf.to_float(tf.shape(x)[1])
spatial_dim_float = tf.sqrt(length_float)
spatial_dim = tf.to_int32(spatial_dim_float)
x = tf.reshape(x,
[-1, spatial_dim, spatial_dim, self._body_input_depth])
x = common_layers.conv_block_downsample(x, self._kernel, self._strides,
self._padding)
x = tf.nn.relu(x)
x = tf.reduce_mean(x, axis=[1, 2], keep_dims=True)
res = common_layers.conv(x, self._vocab_size, (1, 1))
return tf.expand_dims(res, 3)
def testConv(self):
x = np.random.rand(5, 7, 1, 11)
with self.test_session() as session:
y = common_layers.conv(tf.constant(x, dtype=tf.float32), 13,
(3, 1))
session.run(tf.global_variables_initializer())
res = session.run(y)
self.assertEqual(res.shape, (5, 5, 1, 13))
def targets_top_simple(self, body_output, _):
# TODO(lukaszkaiser): work on a better way to generate large images.
with tf.variable_scope(self.name):
decompressed_inputs = common_layers.deconv_stride2_multistep(
body_output,
self._model_hparams.compress_steps,
body_output.get_shape()[-1],
name="deconv")
return common_layers.conv(
decompressed_inputs, self._vocab_size, (1, 1), padding="SAME")
def top(self, body_output, _):
# TODO(lukaszkaiser): work on a better way to generate large images.
with tf.variable_scope(self.name):
decompressed_inputs = common_layers.deconv_stride2_multistep(
body_output,
self._model_hparams.compress_steps,
body_output.get_shape()[-1],
name="deconv")
return common_layers.conv(
decompressed_inputs, self._vocab_size, (1, 1), padding="SAME")
