def bottom(self, inputs):
"""Transform input from data space to model space.
Perform the Xception "Entry flow", which consists of two convolutional
filter upscalings followed by three residually connected separable
convolution blocks.
Args:
inputs: A Tensor with shape [batch, ...]
Returns:
body_input: A Tensor with shape [batch, ?, ?, body_input_depth].
"""
with tf.variable_scope(self.name):
def xnet_resblock(x, filters, res_relu, name):
with tf.variable_scope(name):
y = common_layers.separable_conv_block(
x,
filters, [((1, 1), (3, 3)), ((1, 1), (3, 3))],
first_relu=True,
padding="SAME",
force2d=True,
name="sep_conv_block")
y = common_layers.pool(y, (3, 3),
"MAX",
"SAME",
strides=(2, 2))
return y + common_layers.conv_block(x,
filters, [((1, 1),
(1, 1))],
padding="SAME",
strides=(2, 2),
first_relu=res_relu,
force2d=True,
name="res_conv0")
inputs = common_layers.standardize_images(inputs)
# TODO(lukaszkaiser): summaries here don't work in multi-problem case yet.
# tf.summary.image("inputs", inputs, max_outputs=2)
x = common_layers.conv_block(inputs,
32, [((1, 1), (3, 3))],
first_relu=False,
padding="SAME",
strides=(2, 2),
force2d=True,
name="conv0")
x = common_layers.conv_block(x,
64, [((1, 1), (3, 3))],
padding="SAME",
force2d=True,
name="conv1")
x = xnet_resblock(x, min(128, self._body_input_depth), True,
"block0")
x = xnet_resblock(x, min(256, self._body_input_depth), False,
"block1")
return xnet_resblock(x, self._body_input_depth, False, "block2")
def targets_bottom_simple(self, inputs):
with tf.variable_scope(self.name):
inputs = common_layers.standardize_images(inputs)
if self._model_hparams.compress_steps > 0:
kernel, strides = (2, 2), (2, 2) # Crucial to not leak!
else:
kernel, strides = (1, 1), (1, 1)
return common_layers.conv_block(inputs,
self._body_input_depth,
[((1, 1), kernel)],
first_relu=False,
strides=strides,
force2d=True,
name="small_image_conv")
def bottom(self, inputs):
"""Transform input from data space to model space.
Perform the Xception "Entry flow", which consists of two convolutional
filter upscalings followed by three residually connected separable
convolution blocks.
Args:
inputs: A Tensor with shape [batch, ...]
Returns:
body_input: A Tensor with shape [batch, ?, ?, body_input_depth].
"""
with tf.variable_scope(self.name):
def xnet_resblock(x, filters, res_relu, name):
with tf.variable_scope(name):
y = common_layers.separable_conv_block(
x,
filters, [((1, 1), (3, 3)), ((1, 1), (3, 3))],
first_relu=True,
padding="SAME",
force2d=True,
name="sep_conv_block")
y = common_layers.pool(y, (3, 3), "MAX", "SAME", strides=(2, 2))
return y + common_layers.conv_block(
x,
filters, [((1, 1), (1, 1))],
padding="SAME",
strides=(2, 2),
first_relu=res_relu,
force2d=True,
name="res_conv0")
inputs = common_layers.standardize_images(inputs)
# TODO(lukaszkaiser): summaries here don't work in multi-problem case yet.
# tf.summary.image("inputs", inputs, max_outputs=2)
x = common_layers.conv_block(
inputs,
32, [((1, 1), (3, 3))],
first_relu=False,
padding="SAME",
strides=(2, 2),
force2d=True,
name="conv0")
x = common_layers.conv_block(
x, 64, [((1, 1), (3, 3))], padding="SAME", force2d=True, name="conv1")
x = xnet_resblock(x, min(128, self._body_input_depth), True, "block0")
x = xnet_resblock(x, min(256, self._body_input_depth), False, "block1")
return xnet_resblock(x, self._body_input_depth, False, "block2")
def bottom(self, inputs):
with tf.variable_scope(self.name):
inputs = common_layers.standardize_images(inputs)
# TODO(lukaszkaiser): summaries here don't work in multi-problem case yet.
# tf.summary.image("inputs", inputs, max_outputs=2)
if self._model_hparams.compress_steps > 0:
strides = (2, 2)
else:
strides = (1, 1)
return common_layers.conv_block(inputs,
self._body_input_depth,
[((1, 1), (3, 3))],
first_relu=False,
strides=strides,
padding="SAME",
force2d=True,
name="small_image_conv")
def bottom(self, inputs):
with tf.variable_scope(self.name):
inputs = common_layers.standardize_images(inputs)
# TODO(lukaszkaiser): summaries here don't work in multi-problem case yet.
# tf.summary.image("inputs", inputs, max_outputs=2)
if self._model_hparams.compress_steps > 0:
strides = (2, 2)
else:
strides = (1, 1)
return common_layers.conv_block(
inputs,
self._body_input_depth, [((1, 1), (3, 3))],
first_relu=False,
strides=strides,
padding="SAME",
force2d=True,
name="small_image_conv")
def testStandardizeImages(self):
x = np.random.rand(5, 7, 7, 3)
with self.test_session() as session:
y = common_layers.standardize_images(tf.constant(x))
res = session.run(y)
self.assertEqual(res.shape, (5, 7, 7, 3))
def testStandardizeImages(self):
x = np.random.rand(5, 7, 7, 3)
with self.test_session() as session:
y = common_layers.standardize_images(tf.constant(x))
res = session.run(y)
self.assertEqual(res.shape, (5, 7, 7, 3))
