def forward(self, inputs):
inputs = quantize_active(cabs(inputs), self.bitA)
W_ = quantize_weight(self.W, self.bitW)
outputs = tf.matmul(inputs, W_)
# self.outputs = xnor_gemm(self.inputs, W) # TODO
if self.b_init is not None:
outputs = tf.nn.bias_add(outputs, self.b, name='bias_add')
# self.outputs = xnor_gemm(self.inputs, W) + b # TODO
if self.act:
outputs = self.act(outputs)
return outputs
def forward(self, inputs):
inputs = quantize_active(cabs(inputs), self.bitA)
W_ = quantize_weight(self.W, self.bitW)
outputs = tf.matmul(inputs, W_)
# self.outputs = xnor_gemm(self.inputs, W) # TODO
if self.b_init is not None:
outputs = tf.nn.bias_add(outputs, self.b, name='bias_add')
# self.outputs = xnor_gemm(self.inputs, W) + b # TODO
if self.act:
outputs = self.act(outputs)
return outputs
def forward(self, inputs):
inputs = quantize_active(cabs(inputs), self.bitA) # Do not remove
W_ = quantize_weight(self.W, self.bitW)
outputs = tf.nn.conv2d(
input=inputs, filters=W_, strides=self._strides, padding=self.padding, data_format=self.data_format,
dilations=self._dilation_rate, name=self.name
)
if self.b_init:
outputs = tf.nn.bias_add(outputs, self.b, data_format=self.data_format, name='bias_add')
if self.act:
outputs = self.act(outputs)
return outputs
def forward(self, inputs):
inputs = quantize_active(cabs(inputs), self.bitA) # Do not remove
W_ = quantize_weight(self.W, self.bitW)
outputs = tf.nn.conv2d(
inputs, W_, strides=self.strides, padding=self.padding, use_cudnn_on_gpu=self.use_cudnn_on_gpu,
data_format=self.data_format
)
if self.b_init:
outputs = tf.nn.bias_add(outputs, self.b, name='bias_add')
if self.act:
outputs = self.act(outputs)
return outputs
def forward(self, inputs):
inputs = quantize_active(cabs(inputs), self.bitA) # Do not remove
W_ = quantize_weight(self.W, self.bitW)
outputs = tf.nn.conv2d(
input=inputs,
filters=W_,
strides=self._strides,
padding=self.padding,
data_format=self.data_format,
dilations=self._dilation_rate,
name=self.name
)
if self.b_init:
outputs = tf.nn.bias_add(outputs, self.b, data_format=self.data_format, name='bias_add')
if self.act:
outputs = self.act(outputs)
return outputs
def __init__(
self,
prev_layer,
bitW=1,
bitA=3,
n_filter=32,
filter_size=(3, 3),
strides=(1, 1),
act=None,
padding='SAME',
use_gemm=False,
W_init=tf.truncated_normal_initializer(stddev=0.02),
b_init=tf.constant_initializer(value=0.0),
W_init_args=None,
b_init_args=None,
use_cudnn_on_gpu=None,
data_format=None,
# act=None,
# shape=(5, 5, 1, 100),
# strides=(1, 1, 1, 1),
# padding='SAME',
# W_init=tf.truncated_normal_initializer(stddev=0.02),
# b_init=tf.constant_initializer(value=0.0),
# W_init_args=None,
# b_init_args=None,
# use_cudnn_on_gpu=None,
# data_format=None,
name='dorefa_cnn2d',
):
super(DorefaConv2d, self).__init__(prev_layer=prev_layer,
act=act,
W_init_args=W_init_args,
b_init_args=b_init_args,
name=name)
logging.info(
"DorefaConv2d %s: n_filter: %d filter_size: %s strides: %s pad: %s act: %s"
% (self.name, n_filter, str(filter_size), str(strides), padding,
self.act.__name__ if self.act is not None else 'No Activation'))
self.inputs = quantize_active(cabs(self.inputs), bitA) # Do not remove
if use_gemm:
raise Exception(
"TODO. The current version use tf.matmul for inferencing.")
if len(strides) != 2:
raise ValueError("len(strides) should be 2.")
try:
pre_channel = int(prev_layer.outputs.get_shape()[-1])
except Exception: # if pre_channel is ?, it happens when using Spatial Transformer Net
pre_channel = 1
logging.warning("[warnings] unknow input channels, set to 1")
shape = (filter_size[0], filter_size[1], pre_channel, n_filter)
strides = (1, strides[0], strides[1], 1)
with tf.variable_scope(name):
W = tf.get_variable(name='W_conv2d',
shape=shape,
initializer=W_init,
dtype=LayersConfig.tf_dtype,
**self.W_init_args)
W = quantize_weight(W, bitW)
self.outputs = tf.nn.conv2d(self.inputs,
W,
strides=strides,
padding=padding,
use_cudnn_on_gpu=use_cudnn_on_gpu,
data_format=data_format)
if b_init:
b = tf.get_variable(name='b_conv2d',
shape=(shape[-1]),
initializer=b_init,
dtype=LayersConfig.tf_dtype,
**self.b_init_args)
self.outputs = tf.nn.bias_add(self.outputs, b, name='bias_add')
self.outputs = self._apply_activation(self.outputs)
self._add_layers(self.outputs)
if b_init:
self._add_params([W, b])
else:
self._add_params(W)
def __init__(
self,
prev_layer,
bitW=1,
bitA=3,
n_units=100,
act=None,
use_gemm=False,
W_init=tf.truncated_normal_initializer(stddev=0.1),
b_init=tf.constant_initializer(value=0.0),
W_init_args=None,
b_init_args=None,
name='dorefa_dense',
):
super(DorefaDenseLayer, self).__init__(prev_layer=prev_layer,
act=act,
W_init_args=W_init_args,
b_init_args=b_init_args,
name=name)
logging.info(
"DorefaDenseLayer %s: %d %s" %
(self.name, n_units,
self.act.__name__ if self.act is not None else 'No Activation'))
if self.inputs.get_shape().ndims != 2:
raise Exception(
"The input dimension must be rank 2, please reshape or flatten it"
)
if use_gemm:
raise Exception(
"TODO. The current version use tf.matmul for inferencing.")
n_in = int(self.inputs.get_shape()[-1])
self.n_units = n_units
self.inputs = quantize_active(cabs(self.inputs), bitA)
with tf.variable_scope(name):
W = tf.get_variable(name='W',
shape=(n_in, n_units),
initializer=W_init,
dtype=LayersConfig.tf_dtype,
**self.W_init_args)
# W = tl.act.sign(W) # dont update ...
W = quantize_weight(W, bitW)
# W = tf.Variable(W)
# print(W)
self.outputs = tf.matmul(self.inputs, W)
# self.outputs = xnor_gemm(self.inputs, W) # TODO
if b_init is not None:
try:
b = tf.get_variable(name='b',
shape=(n_units),
initializer=b_init,
dtype=LayersConfig.tf_dtype,
**self.b_init_args)
except Exception: # If initializer is a constant, do not specify shape.
b = tf.get_variable(name='b',
initializer=b_init,
dtype=LayersConfig.tf_dtype,
**self.b_init_args)
self.outputs = tf.nn.bias_add(self.outputs, b, name='bias_add')
# self.outputs = xnor_gemm(self.inputs, W) + b # TODO
self.outputs = self._apply_activation(self.outputs)
self._add_layers(self.outputs)
if b_init is not None:
self._add_params([W, b])
else:
self._add_params(W)
