def quant_fc(self,
input_tensor,
shape,
init_func=msra_init,
enable_bias=True,
quant_bits=8,
quant_input=True,
enable_prune=False):
W = init_func(shape, "fc_weights")
if enable_prune:
W = pruning.apply_mask(W)
self.prune_weights.append(W)
W = fake_quant(W,
min=tf.reduce_min(W),
max=tf.reduce_max(W),
num_bits=quant_bits)
if quant_input == True:
input_tensor = fake_quant(input_tensor,
min=tf.reduce_min(input_tensor),
max=tf.reduce_max(input_tensor),
num_bits=quant_bits)
current_tensor = tf.matmul(input_tensor, W)
if enable_bias:
bias = zeros_init([shape[1]], "fc_bias")
current_tensor = tf.add(current_tensor, bias)
return current_tensor
def quant_conv2d_bn(self,
name,
input_tensor,
kernel,
stride,
init_func=msra_init,
quant_bits=8,
enable_scale=True,
enable_bias=True,
enable_prune=False):
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
current_tensor = self.quant_conv2d(input_tensor,
kernel,
stride,
init_func,
quant_bits=quant_bits,
enable_prune=enable_prune)
scale = None
bias = None
if enable_scale:
scale = ones_init([kernel[3]], "bn_scale")
if enable_bias:
bias = zeros_init([kernel[3]], "bn_bias")
current_tensor = self.bn(current_tensor, scale, bias)
return current_tensor
def bnn_fc(self,
input_tensor,
shape,
init_func=msra_init,
enable_bn=False,
enable_bias=True):
W = init_func(shape, "fc_weights")
W = binary(W)
input_tensor = binary(input_tensor)
current_tensor = tf.matmul(input_tensor, W)
if enable_bias:
bias = zeros_init([shape[1]], "fc_bias")
current_tensor = tf.add(current_tensor, bias)
if enable_bn:
scale = ones_init([shape[1]], "bn_scale")
bias = zeros_init([shape[1]], "bn_bias")
current_tensor = self.bn(current_tensor,
scale,
bias,
is_conv=False)
return current_tensor
def fc(self,
input_tensor,
shape,
init_func=msra_init,
enable_bn=False,
enable_bias=True,
enable_prune=False):
W = init_func(shape, "fc_weights")
if enable_prune:
W = pruning.apply_mask(W)
self.prune_weights.append(W)
current_tensor = tf.matmul(input_tensor, W)
if enable_bias:
bias = zeros_init([shape[1]], "fc_bias")
current_tensor = tf.add(current_tensor, bias)
if enable_bn:
scale = ones_init([shape[1]], "bn_scale")
bias = zeros_init([shape[1]], "bn_bias")
current_tensor = self.bn(current_tensor,
scale,
bias,
is_conv=False)
return current_tensor
def bnn_conv2d_bn(self,
name,
input_tensor,
kernel,
stride,
init_func=msra_init,
bn_input=True,
enable_scale=True,
enable_bias=True,
padding="SAME"):
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
current_tensor = self.bnn_conv2d(input_tensor, kernel, stride,
init_func, bn_input, padding)
scale = None
bias = None
if enable_scale:
scale = ones_init([kernel[3]], "bn_scale")
if enable_bias:
bias = zeros_init([kernel[3]], "bn_bias")
current_tensor = self.bn(current_tensor, scale, bias)
return current_tensor