def __call__(self,
input_layer,
kernel,
depth,
activation_fn=None,
stride=None,
l2loss=None,
init=None,
stddev=None,
bias=True,
bias_init=tf.zeros_initializer,
edges=PAD_SAME,
batch_normalize=False,
name=PROVIDED):
"""Adds a convolution to the stack of operations.
The current head must be a rank 4 Tensor.
Args:
input_layer: The chainable object, supplied.
kernel: The size of the patch for the pool, either an int or a length 1 or
2 sequence (if length 1 or int, it is expanded).
depth: The depth of the new Tensor.
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
stride: The strides as a length 1, 2 or 4 sequence or an integer. If an
int, length 1 or 2, the stride in the first and last dimensions are 1.
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
init: An optional initialization. If not specified, uses Xavier
initialization.
stddev: A standard deviation to use in parameter initialization.
bias: Set to False to not have a bias.
bias_init: An initializer for the bias or a Tensor.
edges: Either SAME to use 0s for the out of bounds area or VALID to shrink
the output size and only uses valid input pixels.
batch_normalize: Supply a BatchNormalizationArguments to set the
parameters for batch normalization.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
Handle to the generated layer.
Raises:
ValueError: If head is not a rank 4 tensor or the depth of the input
(4th dim) is not known.
"""
if input_layer.get_shape().ndims != 4:
raise ValueError('conv2d requires a rank 4 Tensor with a known depth %s' %
input_layer.get_shape())
if input_layer.shape[3] is None:
raise ValueError('Input depth must be known')
kernel = _kernel(kernel)
stride = _stride(stride)
size = [kernel[0], kernel[1], input_layer.shape[3], depth]
books = input_layer.bookkeeper
if init is None:
if stddev is None:
patch_size = size[0] * size[1]
init = layers.he_init(size[2] * patch_size, size[3] * patch_size,
activation_fn)
else:
tf.logging.warning(
'Passing `stddev` to initialize weight variable is deprecated and '
'will be removed in the future. Pass '
'tf.truncated_normal_initializer(stddev=stddev) or '
'tf.zeros_initializer to `init` instead.')
if stddev:
init = tf.truncated_normal_initializer(stddev=stddev)
else:
init = tf.zeros_initializer
elif stddev is not None:
raise ValueError('Do not set both init and stddev.')
dtype = input_layer.tensor.dtype
params = self.variable('weights', size, init, dt=dtype)
y = tf.nn.conv2d(input_layer, params, stride, edges)
layers.add_l2loss(books, params, l2loss)
if bias:
y += self.variable('bias', [size[-1]], bias_init, dt=dtype)
books.add_scalar_summary(
tf.reduce_mean(layers.spatial_slice_zeros(y)),
'%s/zeros_spatial' % y.op.name)
y = pretty_tensor_normalization_methods.batch_normalize_with_arguments(
y, batch_normalize)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn,)
y = layers.apply_activation(books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(self,
input_layer,
kernel,
channel_multiplier,
activation_fn=None,
stride=None,
l2loss=None,
weights=None,
bias=tf.zeros_initializer(),
edges=PAD_SAME,
batch_normalize=False,
phase=prettytensor.Phase.train,
parameter_modifier=parameters.identity,
name=PROVIDED):
"""Adds a depth-wise convolution to the stack of operations.
A depthwise convolution performs the convolutions one channel at a time and
produces an output with depth `channel_multiplier * input_depth`.
`kernel` is the patch that will be pooled and it describes the pooling
along each of the 4 dimensions. The stride is how big to take each step.
* scalar (e.g. 3): Square pooling on the image
(`[b, c, r, d] = [1, 3, 3, 1]`).
* singleton list (e.g. [3]): Square pooling on the image
(`[b, c, r, d] = [1, 3, 3, 1]`).
* list of length 2 (e.g. [3, 2]): Square pooling on the image
(`[b, c, r, d] = [1, 3, 2, 1]`).
Args:
input_layer: The chainable object, supplied.
kernel: The size of the patch for the pool, either an int or a length 1 or
2 sequence (if length 1 or int, it is expanded).
channel_multiplier: Output channels will be a multiple of input channels.
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
stride: The strides as a length 1, 2 or 4 sequence or an integer. If an
int, length 1 or 2, the stride in the first and last dimensions are 1.
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
weights: An initializer for weights or a Tensor. If not specified,
uses He's initialization.
bias: An initializer for the bias or a Tensor. No bias if set to None.
edges: Either `pt.DIM_SAME` to use 0s for the out of bounds area or
`pt.DIM_VALID` to shrink the output size and only uses valid input
pixels.
batch_normalize: Supply a BatchNormalizationArguments to set the
parameters for batch normalization.
phase: The phase of graph construction. See `pt.Phase`.
parameter_modifier: A function to modify parameters that is applied after
creation and before use.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
Handle to the generated layer.
Raises:
ValueError: If input_layer is not a rank 4 tensor or the depth of the
input (4th dim) is not known.
"""
if input_layer.get_shape().ndims != 4:
raise ValueError(
'depthwise_conv2d requires a rank 4 Tensor with a known depth %s'
% input_layer.get_shape())
if input_layer.shape[3] is None:
raise ValueError('Input depth must be known')
kernel = _kernel(kernel)
stride = _stride(stride)
size = [kernel[0], kernel[1], input_layer.shape[3], channel_multiplier]
books = input_layer.bookkeeper
if weights is None:
patch_size = size[0] * size[1]
weights = layers.he_init(size[2] * patch_size,
size[3] * patch_size, activation_fn)
dtype = input_layer.tensor.dtype
params = parameter_modifier(
'weights', self.variable('weights', size, weights, dt=dtype),
phase)
y = tf.nn.depthwise_conv2d(input_layer, params, stride, edges)
layers.add_l2loss(books, params, l2loss)
if bias is not None:
y += parameter_modifier(
'bias',
self.variable('bias',
[input_layer.shape[3] * channel_multiplier],
bias,
dt=dtype), phase)
books.add_scalar_summary(tf.reduce_mean(layers.spatial_slice_zeros(y)),
'%s/zeros_spatial' % y.op.name)
y = pretty_tensor_normalization_methods.batch_normalize_with_arguments(
y, batch_normalize)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn, )
y = layers.apply_activation(books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(self,
input_layer,
size,
activation_fn=None,
l2loss=None,
weights=None,
bias=tf.zeros_initializer(),
transpose_weights=False,
phase=prettytensor.Phase.train,
parameter_modifier=parameters.identity,
tie_groups=False,
name=PROVIDED):
"""Adds the parameters for a fully connected layer and returns a tensor.
The current PrettyTensor must have rank 2.
Args:
input_layer: The Pretty Tensor object, supplied.
size: The number of neurons
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
weights: An initializer for weights or a Tensor. If not specified,
uses He's initialization.
bias: An initializer for the bias or a Tensor. No bias if set to None.
transpose_weights: Flag indicating if weights should be transposed;
this is useful for loading models with a different shape.
phase: The phase of graph construction. See `pt.Phase`.
parameter_modifier: A function to modify parameters that is applied after
creation and before use.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
A Pretty Tensor handle to the layer.
Raises:
ValueError: if the Pretty Tensor is not rank 2 or the number of input
nodes (second dim) is not known.
"""
if input_layer.get_shape().ndims != 3:
raise ValueError(
'group_connected requires a rank 3 Tensor with known 2nd and 3rd '
'dimension: %s' % input_layer.get_shape())
group_num = input_layer.shape[1]
in_size = input_layer.shape[2]
if group_num is None:
raise ValueError('Number of groups must be known.')
if in_size is None:
raise ValueError('Number of input nodes must be known.')
books = input_layer.bookkeeper
if weights is None:
weights = layers.he_init(in_size, size, activation_fn)
dtype = input_layer.tensor.dtype
weight_shape = [group_num, size, in_size
] if transpose_weights else [group_num, in_size, size]
params_var = parameter_modifier(
'weights', self.variable('weights',
weight_shape,
weights,
dt=dtype), phase)
if tie_groups and phase == prettytensor.Phase.train:
with tf.variable_scope("weight_tying"):
params = tf.tile(
tf.reduce_mean(params_var, axis=0, keep_dims=True),
[group_num, 1, 1])
with tf.control_dependencies([tf.assign(params_var, params)]):
params = tf.identity(params)
else:
params = params_var
input_tensor = tf.expand_dims(input_layer, axis=-2)
params_tensor = tf.tile(tf.expand_dims(params, axis=0),
[tf.shape(input_tensor)[0], 1, 1, 1])
y = tf.matmul(input_tensor,
params_tensor,
transpose_b=transpose_weights,
name=name)
y = tf.squeeze(y, axis=2)
layers.add_l2loss(books, params, l2loss)
if bias is not None:
y += parameter_modifier(
'bias', self.variable('bias', [size], bias, dt=dtype), phase)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn, )
y = layers.apply_activation(books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(self,
input_layer,
size,
activation_fn=None,
l2loss=None,
init=None,
stddev=None,
bias=True,
bias_init=tf.zeros_initializer,
transpose_weights=False,
name=PROVIDED):
"""Adds the parameters for a fully connected layer and returns a tensor.
The current head must be a rank 2 Tensor.
Args:
input_layer: The Pretty Tensor object, supplied.
size: The number of neurons
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
init: An optional initialization. If not specified, uses Xavier
initialization.
stddev: A standard deviation to use in parameter initialization.
bias: Set to False to not have a bias.
bias_init: The initializer for the bias or a Tensor.
transpose_weights: Flag indicating if weights should be transposed;
this is useful for loading models with a different shape.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
A Pretty Tensor handle to the layer.
Raises:
ValueError: if the head_shape is not rank 2 or the number of input nodes
(second dim) is not known.
"""
# TODO(eiderman): bias_init shouldn't take a constant and stddev shouldn't
# exist.
if input_layer.get_shape().ndims != 2:
raise ValueError(
'fully_connected requires a rank 2 Tensor with known second '
'dimension: %s'
% input_layer.get_shape())
in_size = input_layer.shape[1]
if input_layer.shape[1] is None:
raise ValueError('Number of input nodes must be known.')
books = input_layer.bookkeeper
if init is None:
if stddev is None:
init = layers.he_init(in_size, size, activation_fn)
else:
tf.logging.warning(
'Passing `stddev` to initialize weight variable is deprecated and '
'will be removed in the future. Pass '
'tf.truncated_normal_initializer(stddev=stddev) or '
'tf.zeros_initializer to `init` instead.')
if stddev:
init = tf.truncated_normal_initializer(stddev=stddev)
else:
init = tf.zeros_initializer
elif stddev is not None:
raise ValueError('Do not set both init and stddev.')
dtype = input_layer.tensor.dtype
weight_shape = [size, in_size] if transpose_weights else [in_size, size]
params = self.variable(
'weights',
weight_shape,
init,
dt=dtype)
y = tf.matmul(input_layer, params, transpose_b=transpose_weights)
layers.add_l2loss(books, params, l2loss)
if bias:
if isinstance(bias_init, tf.compat.real_types):
bias_init = tf.constant_initializer(bias_init)
y += self.variable(
'bias',
[size],
bias_init,
dt=dtype)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn,)
y = layers.apply_activation(
books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(self,
input_layer,
kernel,
depth,
activation_fn=None,
stride=None,
l2loss=None,
init=None,
stddev=None,
bias=True,
bias_init=tf.zeros_initializer,
edges=PAD_SAME,
batch_normalize=False,
name=PROVIDED):
"""Adds a convolution to the stack of operations.
The current head must be a rank 4 Tensor.
Args:
input_layer: The chainable object, supplied.
kernel: The size of the patch for the pool, either an int or a length 1 or
2 sequence (if length 1 or int, it is expanded).
depth: The depth of the new Tensor.
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
stride: The strides as a length 1, 2 or 4 sequence or an integer. If an
int, length 1 or 2, the stride in the first and last dimensions are 1.
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
init: An optional initialization. If not specified, uses Xavier
initialization.
stddev: A standard deviation to use in parameter initialization.
bias: Set to False to not have a bias.
bias_init: An initializer for the bias or a Tensor.
edges: Either SAME to use 0s for the out of bounds area or VALID to shrink
the output size and only uses valid input pixels.
batch_normalize: Supply a BatchNormalizationArguments to set the
parameters for batch normalization.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
Handle to the generated layer.
Raises:
ValueError: If head is not a rank 4 tensor or the depth of the input
(4th dim) is not known.
"""
if input_layer.get_shape().ndims != 4:
raise ValueError(
'conv2d requires a rank 4 Tensor with a known depth %s' %
input_layer.get_shape())
if input_layer.shape[3] is None:
raise ValueError('Input depth must be known')
kernel = _kernel(kernel)
stride = _stride(stride)
size = [kernel[0], kernel[1], input_layer.shape[3], depth]
books = input_layer.bookkeeper
if init is None:
if stddev is None:
patch_size = size[0] * size[1]
init = layers.he_init(size[2] * patch_size,
size[3] * patch_size, activation_fn)
else:
tf.logging.warning(
'Passing `stddev` to initialize weight variable is deprecated and '
'will be removed in the future. Pass '
'tf.truncated_normal_initializer(stddev=stddev) or '
'tf.zeros_initializer to `init` instead.')
if stddev:
init = tf.truncated_normal_initializer(stddev=stddev)
else:
init = tf.zeros_initializer
elif stddev is not None:
raise ValueError('Do not set both init and stddev.')
dtype = input_layer.tensor.dtype
params = self.variable('weights', size, init, dt=dtype)
y = tf.nn.conv2d(input_layer, params, stride, edges)
layers.add_l2loss(books, params, l2loss)
if bias:
y += self.variable('bias', [size[-1]], bias_init, dt=dtype)
books.add_scalar_summary(tf.reduce_mean(layers.spatial_slice_zeros(y)),
'%s/zeros_spatial' % y.op.name)
y = pretty_tensor_normalization_methods.batch_normalize_with_arguments(
y, batch_normalize)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn, )
y = layers.apply_activation(books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(self,
input_layer,
size,
activation_fn=None,
l2loss=None,
init=None,
stddev=None,
bias=True,
bias_init=tf.zeros_initializer,
transpose_weights=False,
name=PROVIDED):
"""Adds the parameters for a fully connected layer and returns a tensor.
The current head must be a rank 2 Tensor.
Args:
input_layer: The Pretty Tensor object, supplied.
size: The number of neurons
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
init: An optional initialization. If not specified, uses Xavier
initialization.
stddev: A standard deviation to use in parameter initialization.
bias: Set to False to not have a bias.
bias_init: The initializer for the bias or a Tensor.
transpose_weights: Flag indicating if weights should be transposed;
this is useful for loading models with a different shape.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
A Pretty Tensor handle to the layer.
Raises:
ValueError: if the head_shape is not rank 2 or the number of input nodes
(second dim) is not known.
"""
# TODO(eiderman): bias_init shouldn't take a constant and stddev shouldn't
# exist.
if input_layer.get_shape().ndims != 2:
raise ValueError(
'fully_connected requires a rank 2 Tensor with known second '
'dimension: %s' % input_layer.get_shape())
in_size = input_layer.shape[1]
if input_layer.shape[1] is None:
raise ValueError('Number of input nodes must be known.')
books = input_layer.bookkeeper
if init is None:
if stddev is None:
init = layers.he_init(in_size, size, activation_fn)
else:
tf.logging.warning(
'Passing `stddev` to initialize weight variable is deprecated and '
'will be removed in the future. Pass '
'tf.truncated_normal_initializer(stddev=stddev) or '
'tf.zeros_initializer to `init` instead.')
if stddev:
init = tf.truncated_normal_initializer(stddev=stddev)
else:
init = tf.zeros_initializer
elif stddev is not None:
raise ValueError('Do not set both init and stddev.')
dtype = input_layer.tensor.dtype
weight_shape = [size, in_size
] if transpose_weights else [in_size, size]
params = self.variable('weights', weight_shape, init, dt=dtype)
y = tf.matmul(input_layer, params, transpose_b=transpose_weights)
layers.add_l2loss(books, params, l2loss)
if bias:
if isinstance(bias_init, tf.compat.real_types):
bias_init = tf.constant_initializer(bias_init)
y += self.variable('bias', [size], bias_init, dt=dtype)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn, )
y = layers.apply_activation(books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(
self,
input_layer,
size,
activation_fn=None,
l2loss=None,
weights=None,
bias=tf.zeros_initializer(),
transpose_weights=False,
phase=prettytensor.Phase.train,
parameter_modifier=parameters.identity,
name=PROVIDED,
):
"""Adds the parameters for a fully connected layer and returns a tensor.
The current PrettyTensor must have rank 2.
Args:
input_layer: The Pretty Tensor object, supplied.
size: The number of neurons
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
weights: An initializer for weights or a Tensor. If not specified,
uses He's initialization.
bias: An initializer for the bias or a Tensor. No bias if set to None.
transpose_weights: Flag indicating if weights should be transposed;
this is useful for loading models with a different shape.
phase: The phase of graph construction. See `pt.Phase`.
parameter_modifier: A function to modify parameters that is applied after
creation and before use.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
A Pretty Tensor handle to the layer.
Raises:
ValueError: if the Pretty Tensor is not rank 2 or the number of input
nodes (second dim) is not known.
"""
if input_layer.get_shape().ndims != 2:
raise ValueError(
"fully_connected requires a rank 2 Tensor with known second " "dimension: %s" % input_layer.get_shape()
)
in_size = input_layer.shape[1]
if input_layer.shape[1] is None:
raise ValueError("Number of input nodes must be known.")
books = input_layer.bookkeeper
if weights is None:
weights = layers.he_init(in_size, size, activation_fn)
dtype = input_layer.tensor.dtype
weight_shape = [size, in_size] if transpose_weights else [in_size, size]
params = parameter_modifier("weights", self.variable("weights", weight_shape, weights, dt=dtype), phase)
y = tf.matmul(input_layer, params, transpose_b=transpose_weights)
layers.add_l2loss(books, params, l2loss)
if bias is not None:
y += parameter_modifier("bias", self.variable("bias", [size], bias, dt=dtype), phase)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn,)
y = layers.apply_activation(books, y, activation_fn[0], activation_args=activation_fn[1:])
books.add_histogram_summary(y, "%s/activations" % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)
def __call__(self,
input_layer,
kernel,
channel_multiplier,
activation_fn=None,
stride=None,
l2loss=None,
weights=None,
bias=tf.zeros_initializer(),
edges=PAD_SAME,
batch_normalize=False,
phase=prettytensor.Phase.train,
parameter_modifier=parameters.identity,
name=PROVIDED):
"""Adds a depth-wise convolution to the stack of operations.
A depthwise convolution performs the convolutions one channel at a time and
produces an output with depth `channel_multiplier * input_depth`.
`kernel` is the patch that will be pooled and it describes the pooling
along each of the 4 dimensions. The stride is how big to take each step.
* scalar (e.g. 3): Square pooling on the image
(`[b, c, r, d] = [1, 3, 3, 1]`).
* singleton list (e.g. [3]): Square pooling on the image
(`[b, c, r, d] = [1, 3, 3, 1]`).
* list of length 2 (e.g. [3, 2]): Square pooling on the image
(`[b, c, r, d] = [1, 3, 2, 1]`).
Args:
input_layer: The chainable object, supplied.
kernel: The size of the patch for the pool, either an int or a length 1 or
2 sequence (if length 1 or int, it is expanded).
channel_multiplier: Output channels will be a multiple of input channels.
activation_fn: A tuple of (activation_function, extra_parameters). Any
function that takes a tensor as its first argument can be used. More
common functions will have summaries added (e.g. relu).
stride: The strides as a length 1, 2 or 4 sequence or an integer. If an
int, length 1 or 2, the stride in the first and last dimensions are 1.
l2loss: Set to a value greater than 0 to use L2 regularization to decay
the weights.
weights: An initializer for weights or a Tensor. If not specified,
uses He's initialization.
bias: An initializer for the bias or a Tensor. No bias if set to None.
edges: Either `pt.DIM_SAME` to use 0s for the out of bounds area or
`pt.DIM_VALID` to shrink the output size and only uses valid input
pixels.
batch_normalize: Supply a BatchNormalizationArguments to set the
parameters for batch normalization.
phase: The phase of graph construction. See `pt.Phase`.
parameter_modifier: A function to modify parameters that is applied after
creation and before use.
name: The name for this operation is also used to create/find the
parameter variables.
Returns:
Handle to the generated layer.
Raises:
ValueError: If input_layer is not a rank 4 tensor or the depth of the
input (4th dim) is not known.
"""
if input_layer.get_shape().ndims != 4:
raise ValueError(
'depthwise_conv2d requires a rank 4 Tensor with a known depth %s' %
input_layer.get_shape())
if input_layer.shape[3] is None:
raise ValueError('Input depth must be known')
kernel = _kernel(kernel)
stride = _stride(stride)
size = [kernel[0], kernel[1], input_layer.shape[3], channel_multiplier]
books = input_layer.bookkeeper
if weights is None:
patch_size = size[0] * size[1]
weights = layers.he_init(size[2] * patch_size, size[3] * patch_size,
activation_fn)
dtype = input_layer.tensor.dtype
params = parameter_modifier(
'weights',
self.variable('weights', size, weights, dt=dtype),
phase)
y = tf.nn.depthwise_conv2d(input_layer, params, stride, edges)
layers.add_l2loss(books, params, l2loss)
if bias is not None:
y += parameter_modifier(
'bias',
self.variable('bias', [input_layer.shape[3] * channel_multiplier],
bias,
dt=dtype),
phase)
books.add_scalar_summary(
tf.reduce_mean(layers.spatial_slice_zeros(y)),
'%s/zeros_spatial' % y.op.name)
y = pretty_tensor_normalization_methods.batch_normalize_with_arguments(
y, batch_normalize)
if activation_fn is not None:
if not isinstance(activation_fn, collections.Sequence):
activation_fn = (activation_fn,)
y = layers.apply_activation(books,
y,
activation_fn[0],
activation_args=activation_fn[1:])
books.add_histogram_summary(y, '%s/activations' % y.op.name)
return input_layer.with_tensor(y, parameters=self.vars)