def bottom_compress(self, inputs, name="bottom"):
"""Transform input from data space to model space.
Perform conversion of RGB pixel values to a real number in the range -1 to 1
and combine channel values for each pixel to form a representation of
size image_length x image_length dims.
Args:
inputs: A Tensor representing RGB pixel intensities as integers.
[batch, ...]
name: string, scope.
Returns:
body_input: A Tensor with shape [batch, ?, ?, body_input_depth].
"""
with tf.variable_scope(name):
inputs = tf.to_float(inputs)
hp = self._model_hparams
if hp.mode != tf.estimator.ModeKeys.PREDICT:
tf.summary.image("inputs", inputs, max_outputs=2)
inputs = common_layers.convert_rgb_to_symmetric_real(inputs)
ishape = common_layers.shape_list(inputs)
inputs = tf.reshape(inputs,
[-1, ishape[1], ishape[2] * ishape[3], 1])
inputs.set_shape([None, None, None, 1])
# We compress RGB intensities for each pixel using a conv.
x = tf.layers.conv2d(inputs,
self._body_input_depth,
(1, self.num_channels),
padding="VALID",
strides=(1, self.num_channels),
activation=tf.nn.relu,
name="conv_input")
x.set_shape([None, None, None, self._body_input_depth])
return x
def testDmlLoss(self, batch, height, width, num_mixtures, reduce_sum):
channels = 3
pred = tf.random_normal([batch, height, width, num_mixtures * 10])
labels = tf.random_uniform([batch, height, width, channels],
minval=0, maxval=256, dtype=tf.int32)
actual_loss_num, actual_loss_den = common_layers.dml_loss(
pred=pred, labels=labels, reduce_sum=reduce_sum)
actual_loss = actual_loss_num / actual_loss_den
real_labels = common_layers.convert_rgb_to_symmetric_real(labels)
expected_loss = common_layers.discretized_mix_logistic_loss(
pred=pred, labels=real_labels) / channels
if reduce_sum:
expected_loss = tf.reduce_mean(expected_loss)
actual_loss_val, expected_loss_val = self.evaluate(
[actual_loss, expected_loss])
self.assertAllClose(actual_loss_val, expected_loss_val)
def testDmlLoss(self, batch, height, width, num_mixtures, reduce_sum):
channels = 3
pred = tf.random_normal([batch, height, width, num_mixtures * 10])
labels = tf.random_uniform([batch, height, width, channels],
minval=0, maxval=256, dtype=tf.int32)
actual_loss_num, actual_loss_den = common_layers.dml_loss(
pred=pred, labels=labels, reduce_sum=reduce_sum)
actual_loss = actual_loss_num / actual_loss_den
real_labels = common_layers.convert_rgb_to_symmetric_real(labels)
expected_loss = common_layers.discretized_mix_logistic_loss(
pred=pred, labels=real_labels) / channels
if reduce_sum:
expected_loss = tf.reduce_mean(expected_loss)
actual_loss_val, expected_loss_val = self.evaluate(
[actual_loss, expected_loss])
self.assertAllClose(actual_loss_val, expected_loss_val)
def bottom_compress(self, inputs, name="bottom"):
"""Compresses channel-wise input pixels into whole pixel representions.
Perform conversion of RGB pixel values to a real number in the range -1 to
1. This combines pixel channels to form a representation of shape
[img_len, img_len].
Args:
inputs: Tensor representing RGB pixel intensities as integers, of shape
[batch, img_len, img_len, channels].
name: string, scope.
Returns:
body_input: Tensor of shape [batch, img_len, img_len, body_input_depth].
"""
with tf.variable_scope(name):
inputs = tf.to_float(inputs)
hp = self._model_hparams
if hp.mode != tf.estimator.ModeKeys.PREDICT:
tf.summary.image("inputs",
common_layers.tpu_safe_image_summary(inputs),
max_outputs=2)
inputs = common_layers.convert_rgb_to_symmetric_real(inputs)
# Reshape inputs to apply convolutions across [img_len, img_len*channels].
inputs_shape = common_layers.shape_list(inputs)
inputs = tf.reshape(
inputs,
[-1, inputs_shape[1], inputs_shape[2] * inputs_shape[3], 1])
# tf.logging.info("input shape" , inputs_shape)
# Compress RGB intensities for each pixel using a convolution.
# ValueError: Negative dimension size caused by subtracting 3 from 1 for 'imagetransformer/parallel_0_4/
# imagetransformer/imagetransformer/image_channel_bottom_identity_modality/output_bottom/conv_input/Conv2D'
# (op: 'Conv2D') with input shapes: [?,1,1,1], [1,3,1,256].
outputs = tf.layers.conv2d(inputs,
self._body_input_depth,
kernel_size=(1, self.num_channels),
padding="VALID",
strides=(1, self.num_channels),
activation=tf.nn.relu,
name="conv_input")
return outputs
def preprocess_example(self, example, unused_mode, unused_hparams):
""" Luptonize the examples, so that we can use t2t models easily
"""
p = self.get_hparams()
image = example["inputs"]
# Apply Luptonic Asinh stretch, and return uint8 rgb images
def my_func(x):
return make_lupton_rgb(x[...,2], x[...,1], x[...,0], Q=15, stretch=0.5, minimum=0)
int_image = tf.py_func(my_func, [image], tf.uint8)
int_image.set_shape(image.shape)
image = common_layers.convert_rgb_to_symmetric_real(int_image)
if hasattr(p, 'attributes'):
example["attributes"] = tf.stack([example[k] for k in p.attributes])
example["inputs"] = image
example["targets"] = image
return example
def bottom_compress(self, inputs, name="bottom"):
"""Compresses channel-wise input pixels into whole pixel representions.
Perform conversion of RGB pixel values to a real number in the range -1 to
1. This combines pixel channels to form a representation of shape
[img_len, img_len].
Args:
inputs: Tensor representing RGB pixel intensities as integers, of shape
[batch, img_len, img_len, channels].
name: string, scope.
Returns:
body_input: Tensor of shape
[batch, img_len, img_len, self._model_hparams.hidden_size].
"""
with tf.variable_scope(name):
inputs = tf.to_float(inputs)
hp = self._model_hparams
if hp.mode != tf.estimator.ModeKeys.PREDICT:
tf.summary.image(
"inputs",
common_layers.tpu_safe_image_summary(inputs),
max_outputs=2)
inputs = common_layers.convert_rgb_to_symmetric_real(inputs)
# Reshape inputs to apply convolutions across [img_len, img_len*channels].
inputs_shape = common_layers.shape_list(inputs)
inputs = tf.reshape(
inputs, [-1, inputs_shape[1], inputs_shape[2] * inputs_shape[3], 1])
# Compress RGB intensities for each pixel using a convolution.
outputs = tf.layers.conv2d(
inputs,
self._model_hparams.hidden_size,
kernel_size=(1, self.num_channels),
padding="VALID",
strides=(1, self.num_channels),
activation=tf.nn.relu,
name="conv_input")
return outputs
def bottom_compress(self, inputs, name="bottom"):
"""Transform input from data space to model space.
Perform conversion of RGB pixel values to a real number in the range -1 to 1
and combine channel values for each pixel to form a representation of
size image_length x image_length dims.
Args:
inputs: A Tensor representing RGB pixel intensities as integers.
[batch, ...]
name: string, scope.
Returns:
body_input: A Tensor with shape [batch, ?, ?, body_input_depth].
"""
with tf.variable_scope(name):
inputs = tf.to_float(inputs)
hp = self._model_hparams
if hp.mode != tf.estimator.ModeKeys.PREDICT:
tf.summary.image(
"inputs",
common_layers.tpu_safe_image_summary(inputs),
max_outputs=2)
inputs = common_layers.convert_rgb_to_symmetric_real(inputs)
ishape = common_layers.shape_list(inputs)
inputs = tf.reshape(inputs, [-1, ishape[1], ishape[2] * ishape[3], 1])
inputs.set_shape([None, None, None, 1])
# We compress RGB intensities for each pixel using a conv.
x = tf.layers.conv2d(
inputs,
self._body_input_depth, (1, self.num_channels),
padding="VALID",
strides=(1, self.num_channels),
activation=tf.nn.relu,
name="conv_input")
x.set_shape([None, None, None, self._body_input_depth])
return x