def DeConv3D(x, out_channel, kernel_shape,
stride, padding='SAME',
W_init=None, b_init=None,
nl=tf.identity, use_bias=True,
data_format='NDHWC'):
in_shape = x.get_shape().as_list()
channel_axis = 4
in_channel = in_shape[channel_axis]
assert in_channel is not None, "[DeConv3D] Input cannot have unknown channel!"
assert isinstance(out_channel, int), out_channel
if W_init is None:
W_init = tf.variance_scaling_initializer(scale=2.0)
if b_init is None:
b_init = tf.constant_initializer()
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv3DTranspose(
out_channel, kernel_shape,
strides=stride, padding=padding,
data_format='channels_last' if data_format == 'NDHWC' else 'channels_first',
activation=lambda x: nl(x, name='output'),
use_bias=use_bias,
kernel_initializer=W_init,
bias_initializer=b_init,
trainable=True)
ret = layer.apply(x, scope=tf.get_variable_scope())
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
return ret
def mpusim_fully_connected(inputs,
units,
activation=None,
use_bias=True,
kernel_initializer=None,
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
activations_datatype_size_byte=1,
weights_datatype_size_byte=1,
results_datatype_size_byte=4,
systolic_array_height=256,
systolic_array_width=256,
activation_fifo_depth=8,
accumulator_array_height=4096,
log_file_output_dir='.',
model_name='unnamed'):
"""
A wrapper around `mpusim_fc`.
One difference to maintain backward-compatibility:
Default weight initializer is variance_scaling_initializer(2.0).
Variable Names:
* ``W``: weights of shape [in_dim, out_dim]
* ``b``: bias
"""
if kernel_initializer is None:
if get_tf_version_tuple() <= (1, 12):
kernel_initializer = tf.contrib.layers.variance_scaling_initializer(2.0) # deprecated
else:
kernel_initializer = tf.keras.initializers.VarianceScaling(2.0, distribution='untruncated_normal')
inputs = batch_flatten(inputs)
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = mpusim_fc(units=units,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
activations_datatype_size_byte=activations_datatype_size_byte,
weights_datatype_size_byte=weights_datatype_size_byte,
results_datatype_size_byte=results_datatype_size_byte,
systolic_array_height=systolic_array_height,
systolic_array_width=systolic_array_width,
activation_fifo_depth=activation_fifo_depth,
accumulator_array_height=accumulator_array_height,
log_file_output_dir=log_file_output_dir,
model_name=model_name,
_reuse=tf.get_variable_scope().reuse)
ret = layer.apply(inputs, scope=tf.get_variable_scope())
ret = tf.identity(ret, name='output')
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
return ret
def Conv3D(
inputs,
filters,
kernel_size,
strides=(1, 1, 1),
padding='same',
data_format='channels_last',
dilation_rate=(1, 1, 1),
activation=None,
use_bias=True,
kernel_initializer=tf.contrib.layers.variance_scaling_initializer(2.0),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
split=1):
"""
A wrapper around `tf.layers.Conv3D`.
Some differences to maintain backward-compatibility:
1. Default kernel initializer is variance_scaling_initializer(2.0).
2. Default padding is 'same'.
3. Support 'split' argument to do group conv.
Variable Names:
* ``W``: weights
* ``b``: bias
"""
if split == 1:
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv3D(
filters,
kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
_reuse=tf.get_variable_scope().reuse)
ret = layer.apply(inputs, scope=tf.get_variable_scope())
ret = tf.identity(ret, name='output')
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
else:
# group conv implementation
pass
return ret
def GrConv2D(x,
out_channel,
kernel_shape,
padding='SAME',
stride=1,
dilation_rate=1,
W_init=None,
b_init=None,
nl=tf.identity,
split=1,
use_bias=True,
data_format='channels_last'):
if data_format == 'NHWC' or data_format == 'channels_last':
data_format = 'channels_last'
elif data_format == 'NCHW' or data_format == 'channels_first':
data_format = 'channels_first'
else:
print "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa unknown data format"
in_shape = x.get_shape().as_list()
channel_axis = 3 if data_format == 'NHWC' else 1
in_channel = in_shape[channel_axis]
assert in_channel is not None, "[GrConv2D] Input cannot have unknown channel!"
assert in_channel % split == 0
assert out_channel % split == 0
kernel_shape = shape2d(kernel_shape)
padding = padding.upper()
filter_shape = kernel_shape + [in_channel / split, out_channel]
stride = shape2d(stride)
if W_init is None:
W_init = tf.contrib.layers.variance_scaling_initializer()
if b_init is None:
b_init = tf.constant_initializer()
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv2D(filters=out_channel,
kernel_size=kernel_shape,
strides=stride,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=lambda x: nl(x, name='output'),
use_bias=use_bias,
kernel_initializer=W_init,
bias_initializer=b_init,
trainable=True)
ret = layer.apply(x, scope=tf.get_variable_scope())
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
return ret
def Conv3DTranspose(
inputs,
filters,
kernel_size,
strides=(1, 1, 1),
padding='same',
data_format='channels_last',
activation=None,
use_bias=False,
kernel_initializer=tf.contrib.layers.
variance_scaling_initializer(
2.0
), #tf.contrib.layers.xavier_initializer(), #tf.initializers.variance_scaling(distribution='uniform'),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None):
"""
A wrapper around `tf.layers.Conv2DTranspose`.
Some differences to maintain backward-compatibility:
1. Default kernel initializer is variance_scaling_initializer(2.0),.
2. Default padding is 'same'
Variable Names:
* ``W``: weights
* ``b``: bias
"""
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv3DTranspose(
filters,
kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer)
ret = layer.apply(inputs, scope=tf.get_variable_scope())
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
return tf.identity(ret, name='output')
def Conv3D(
inputs,
filters,
kernel_size,
strides=(1, 1, 1),
padding='same',
data_format='channels_last',
dilation_rate=(1, 1, 1),
activation=None,
use_bias=True,
kernel_initializer=tf.contrib.layers.variance_scaling_initializer(2.0),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
split=1):
"""
A wrapper around `tf.layers.Conv2D`.
Some differences to maintain backward-compatibility:
1. Default kernel initializer is variance_scaling_initializer(2.0).
2. Default padding is 'same'.
3. Support 'split' argument to do group conv.
Variable Names:
* ``W``: weights
* ``b``: bias
"""
if split == 1:
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv3D(filters,
kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer)
ret = layer.apply(inputs, scope=tf.get_variable_scope())
ret = tf.identity(ret, name='output')
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
else:
# group conv implementation
data_format = get_data_format(data_format, tfmode=False)
in_shape = inputs.get_shape().as_list()
channel_axis = 3 if data_format == 'NHWC' else 1
in_channel = in_shape[channel_axis]
assert in_channel is not None, "[Conv3D] Input cannot have unknown channel!"
assert in_channel % split == 0
assert kernel_regularizer is None and bias_regularizer is None and activity_regularizer is None, \
"Not supported by group conv now!"
out_channel = filters
assert out_channel % split == 0
assert dilation_rate == (1, 1) or get_tf_version_number(
) >= 1.5, 'TF>=1.5 required for group dilated conv'
kernel_shape = shape2d(kernel_size)
filter_shape = kernel_shape + [in_channel / split, out_channel]
stride = shape4d(strides, data_format=data_format)
kwargs = dict(data_format=data_format)
if get_tf_version_number() >= 1.5:
kwargs['dilations'] = shape4d(dilation_rate,
data_format=data_format)
W = tf.get_variable('W', filter_shape, initializer=kernel_initializer)
if use_bias:
b = tf.get_variable('b', [out_channel],
initializer=bias_initializer)
inputs = tf.split(inputs, split, channel_axis)
kernels = tf.split(W, split, 3)
outputs = [
tf.nn.conv2d(i, k, stride, padding.upper(), **kwargs)
for i, k in zip(inputs, kernels)
]
conv = tf.concat(outputs, channel_axis)
if activation is None:
activation = tf.identity
ret = activation(tf.nn.bias_add(conv, b, data_format=data_format)
if use_bias else conv,
name='output')
ret.variables = VariableHolder(W=W)
if use_bias:
ret.variables.b = b
return ret
def MaskedConv2D(
inputs,
filters,
kernel_size,
strides=(1, 1),
padding='same',
data_format='channels_last',
dilation_rate=(1, 1),
activation=None,
use_bias=True,
kernel_initializer=None,
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
split=1,
masking=False):
"""
A wrapper around `tf.layers.Conv2D`.
Some differences to maintain backward-compatibility:
1. Default kernel initializer is variance_scaling_initializer(2.0).
2. Default padding is 'same'.
3. Support 'split' argument to do group conv.
Variable Names:
* ``W``: weights
* ``b``: bias
"""
if kernel_initializer is None:
if get_tf_version_tuple() <= (1, 12):
kernel_initializer = tf.contrib.layers.variance_scaling_initializer(2.0)
else:
kernel_initializer = tf.keras.initializers.VarianceScaling(2.0, distribution='untruncated_normal')
dilation_rate = shape2d(dilation_rate)
if (masking == False) and (split == 1) and (dilation_rate == [1, 1]):
# tf.layers.Conv2D has bugs with dilations (https://github.com/tensorflow/tensorflow/issues/26797)
with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
layer = tf.layers.Conv2D(
filters,
kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
_reuse=tf.get_variable_scope().reuse)
ret = layer.apply(inputs, scope=tf.get_variable_scope())
ret = tf.identity(ret, name='output')
ret.variables = VariableHolder(W=layer.kernel)
if use_bias:
ret.variables.b = layer.bias
else:
if masking == True:
assert split == 1, "Pruining group conv is not supported yet"
# group conv implementation
data_format = get_data_format(data_format, keras_mode=False)
in_shape = inputs.get_shape().as_list()
channel_axis = 3 if data_format == 'NHWC' else 1
in_channel = in_shape[channel_axis]
assert in_channel is not None, "[Conv2D] Input cannot have unknown channel!"
assert in_channel % split == 0
assert kernel_regularizer is None and bias_regularizer is None and activity_regularizer is None, \
"Not supported by group conv or dilated conv!"
out_channel = filters
assert out_channel % split == 0
assert dilation_rate == [1, 1] or get_tf_version_tuple() >= (1, 5), 'TF>=1.5 required for dilated conv.'
kernel_shape = shape2d(kernel_size)
filter_shape = kernel_shape + [in_channel / split, out_channel]
stride = shape4d(strides, data_format=data_format)
kwargs = dict(data_format=data_format)
if get_tf_version_tuple() >= (1, 5):
kwargs['dilations'] = shape4d(dilation_rate, data_format=data_format)
W = tf.get_variable(
'W', filter_shape, initializer=kernel_initializer)
if use_bias:
b = tf.get_variable('b', [out_channel], initializer=bias_initializer)
if split == 1:
if masking:
W = pruning.apply_mask(W)
conv = tf.nn.conv2d(inputs, W, stride, padding.upper(), **kwargs)
else:
conv = None
if get_tf_version_tuple() >= (1, 13):
try:
conv = tf.nn.conv2d(inputs, W, stride, padding.upper(), **kwargs)
except ValueError:
log_once("CUDNN group convolution support is only available with "
"https://github.com/tensorflow/tensorflow/pull/25818 . "
"Will fall back to a loop-based slow implementation instead!", 'warn')
if conv is None:
inputs = tf.split(inputs, split, channel_axis)
kernels = tf.split(W, split, 3)
outputs = [tf.nn.conv2d(i, k, stride, padding.upper(), **kwargs)
for i, k in zip(inputs, kernels)]
conv = tf.concat(outputs, channel_axis)
ret = tf.nn.bias_add(conv, b, data_format=data_format) if use_bias else conv
if activation is not None:
ret = activation(ret)
ret = tf.identity(ret, name='output')
ret.variables = VariableHolder(W=W)
if use_bias:
ret.variables.b = b
return ret
def BatchNorm(inputs,
axis=None,
training=None,
momentum=0.9,
epsilon=1e-5,
center=True,
scale=True,
beta_initializer=tf.zeros_initializer(),
gamma_initializer=tf.ones_initializer(),
virtual_batch_size=None,
internal_update=False):
"""
Mostly equivalent to `tf.layers.batch_normalization`, but different in
the following:
1. Accepts `data_format` when `axis` is None. For 2D input, this argument will be ignored.
2. Default value for `momentum` and `epsilon` is different.
3. Default value for `training` is automatically obtained from `TowerContext`.
4. Support the `internal_update` option.
Args:
internal_update (bool): if False, add EMA update ops to
`tf.GraphKeys.UPDATE_OPS`. If True, update EMA inside the layer
by control dependencies.
Variable Names:
* ``beta``: the bias term. Will be zero-inited by default.
* ``gamma``: the scale term. Will be one-inited by default. Input will be transformed by ``x * gamma + beta``.
* ``mean/EMA``: the moving average of mean.
* ``variance/EMA``: the moving average of variance.
Note:
1. About multi-GPU training: moving averages across GPUs are not aggregated.
Batch statistics are computed independently. This is consistent with most frameworks.
2. Combinations of ``training`` and ``ctx.is_training``:
* ``training == ctx.is_training``: standard BN, EMA are
maintained during training and used during inference. This is
the default.
* ``training and not ctx.is_training``: still use batch statistics in inference.
* ``not training and ctx.is_training``: use EMA to normalize in
training. This is useful when you load a pre-trained BN and
don't want to fine tune the EMA. EMA will not be updated in
this case.
"""
# parse shapes
shape = inputs.get_shape().as_list()
ndims = len(shape)
assert axis is not None
# parse training/ctx
ctx = get_current_tower_context()
if training is None:
training = ctx.is_training
training = bool(training)
TF_version = get_tf_version_number()
if not training and ctx.is_training:
assert TF_version >= 1.4, \
"Fine tuning a BatchNorm model with fixed statistics is only " \
"supported after https://github.com/tensorflow/tensorflow/pull/12580 "
if ctx.is_main_training_tower: # only warn in first tower
logger.warn(
"[BatchNorm] Using moving_mean/moving_variance in training.")
# Using moving_mean/moving_variance in training, which means we
# loaded a pre-trained BN and only fine-tuning the affine part.
coll_bk = backup_collection([tf.GraphKeys.UPDATE_OPS])
with rename_get_variable({
'moving_mean': 'mean/EMA',
'moving_variance': 'variance/EMA'
}):
if TF_version >= 1.5:
layer = tf.layers.BatchNormalization(
axis=axis,
momentum=momentum,
epsilon=epsilon,
center=center,
scale=scale,
beta_initializer=beta_initializer,
gamma_initializer=gamma_initializer,
virtual_batch_size=virtual_batch_size,
fused=True,
_reuse=tf.get_variable_scope().reuse)
else:
assert virtual_batch_size is None, "Feature not supported in this version of TF!"
layer = tf.layers.BatchNormalization(
axis=axis,
momentum=momentum,
epsilon=epsilon,
center=center,
scale=scale,
beta_initializer=beta_initializer,
gamma_initializer=gamma_initializer,
fused=True,
_reuse=tf.get_variable_scope().reuse)
xn = layer.apply(inputs,
training=training,
scope=tf.get_variable_scope())
# maintain EMA only on one GPU is OK, even in replicated mode.
# because training time doesn't use EMA
if ctx.is_main_training_tower:
for v in layer.non_trainable_variables:
add_model_variable(v)
if not ctx.is_main_training_tower or internal_update:
restore_collection(coll_bk)
if training and internal_update:
assert layer.updates
with tf.control_dependencies(layer.updates):
ret = tf.identity(xn, name='output')
else:
ret = tf.identity(xn, name='output')
vh = ret.variables = VariableHolder(
moving_mean=layer.moving_mean,
mean=layer.moving_mean, # for backward-compatibility
moving_variance=layer.moving_variance,
variance=layer.moving_variance) # for backward-compatibility
if scale:
vh.gamma = layer.gamma
if center:
vh.beta = layer.beta
return ret
