def conv_pool_bn(pre_layer,
kernel_num,
kernel_size,
padding,
pool_size,
activation,
training,
epsilon=1e-4,
alpha=.1,
binary=True,
stochastic=False,
H=1.,
W_LR_scale="Glorot"):
conv = binary_layer.conv2d_binary(pre_layer,
kernel_num,
kernel_size,
padding=padding,
binary=binary,
stochastic=stochastic,
H=H,
W_LR_scale=W_LR_scale)
pool = tf.layers.max_pooling2d(conv,
pool_size=pool_size,
strides=pool_size)
bn = binary_layer.batch_normalization(pool,
epsilon=epsilon,
momentum=1 - alpha,
training=training)
output = activation(bn)
return output
def conv_latent(pre_layer,
kernel_num,
kernel_size,
padding,
activation,
training,
epsilon=1e-4,
alpha=.1,
binary=True,
stochastic=False,
H=1.,
W_LR_scale="Glorot"):
conv = binary_layer.conv2d_latent(pre_layer,
kernel_num,
kernel_size,
padding=padding,
binary=binary,
stochastic=stochastic,
H=H,
W_LR_scale=W_LR_scale)
bn = binary_layer.batch_normalization(conv,
epsilon=epsilon,
momentum=1 - alpha,
training=training)
output = activation(bn)
return output
def fully_connect_latent(pre_layer,
output_dim,
act,
use_bias,
training,
epsilon=1e-4,
alpha=.1,
binary=True,
stochastic=False,
H=1.,
W_LR_scale="Glorot"):
pre_act = binary_layer.dense_latent(
pre_layer,
output_dim,
use_bias=use_bias,
kernel_constraint=lambda w: tf.clip_by_value(w, -1.0, 1.0))
bn = binary_layer.batch_normalization(pre_act,
momentum=1 - alpha,
epsilon=epsilon,
training=training)
if act == None:
output = bn
else:
output = act(bn)
return output
def fully_connect_bn(pre_layer, output_dim, act, use_bias, training):
pre_act = binary_layer.dense_binary(
pre_layer,
output_dim,
use_bias=use_bias,
activation=None,
kernel_constraint=lambda w: tf.clip_by_value(w, -1.0, 1.0))
bn = binary_layer.batch_normalization(pre_act,
momentum=0.9,
epsilon=1e-4,
training=training)
if act == None:
output = bn
else:
output = act(bn)
return output
