def weight_init(name='undefined',
method='msra',
height=None,
width=None,
input_channel=None,
output_channel=None,
collection='TRAINABLED'):
assert method in ['msra'], print('mothod not support')
if method == 'msra':
initer = tf.contrib.layer.variance_scaling_initializer(factor=2.0,
model='FAN_IN',
uniform=False)
# elif:
# sys.exit()
else:
sys.exit()
w = tf.get_variable(name=name + '_weight',
shape=[height, width, input_channel, output_channel],
dtype=tf.float32,
initializer=initer)
Util.AddToCollectionInfo(tf.GraphKeys.GLOBAL_VARIABLES, w)
tf.add_to_collection(collection, w)
Util.AddToCollectionInfo(collection, w)
return w
def Conv_Layer(
name='undefined',
input=None,
height=None,
width=None,
output_channel=None
):
conv = tf.nn.conv2d(
input=input,
filter=init.weight_init(
name=name,
height=height,
width=width,
input_channel=input.get_shape().as_list()[-1],
output_channel=output_channel,
collection='ConvWeight'
),
strides=[1, 1, 1, 1],
padding='SAME',
name=name
)
tf.add_to_collection(name='ConvOut', value=conv)
Util.AddToCollectionInfo('ConvOut', conv)
Util.CLayerInfo(name, input, conv)
return conv
pass
def Activate_Layer(
name='undefined',
input=None,
method='LeakReLU'
):
"""
默认格式'NHWC'
:param name:
:param input:
:param method:
:return:
"""
assert method in ['LeakReLU', 'ReLU'], Util.CError('method is not supported')
if method == 'LeakReLU':
activate = keras.layers.LeakyReLU(
alpha=0.1,
name=name + 'LeakReLU'
)(input)
elif method == 'ReLU':
activate = keras.layers.ReL
else:
Util.CError('method is not supported!')
sys.exit()
tf.add_to_collection(name='ActiOut', value=activate)
Util.CLayerInfo(name, input, activate)
Util.AddToCollectionInfo('ActiOut', activate)
return activate
pass
def Bias_Layer(
name='undefined',
input=None,
):
"""
默认格式'NHWC'
:param name:
:param input:
:return:
"""
bias = tf.nn.bias_add(
value=input,
bias=init.bias_init(
name=name,
output_channel=input.get_shape().as_list()[-1],
collection='BiasBias'
),
data_format='NHWC',
name=name
)
tf.add_to_collection(name='BiasOut', value=bias)
Util.AddToCollectionInfo('BiasOut', bias)
Util.CLayerInfo(name, input, bias)
return bias
pass
def bias_init(name='undefined',
method='zero',
output_channel=None,
collection='TRAINABLED'):
assert method in ['zero', 'one'], print('method not support')
if method == 'zero':
b = tf.get_variable(name=name + '_bias',
shape=[output_channel],
dtype=tf.float32,
initializer=tf.zeros_initializer())
elif method == 'one':
b = tf.get_variable(name=name + '_bias',
shape=[output_channel],
dtype=tf.float32,
initializer=tf.ones_initializer())
else:
sys.exit()
Util.AddToCollectionInfo(tf.GraphKeys.GLOBAL_VARIABLES, b)
tf.add_to_collection(collection, b)
Util.AddToCollectionInfo(collection, b)
return b
pass
def BatchNormal_Layer(
name='undefined',
input=None,
train=tf.bool(True),
):
"""
默认格式'NHWC'
:param input:
:return:
"""
global MOVING_DECAY
global BNEPS
assert Util.CGlobalExit('MOVING_DECAY')
assert Util.CGlobalExit('BNEPS')
train_mean = tf.reduce_mean(
input_tensor=input,
axis=3,
name=name + '_t_mean',
)
train_var = tf.reduce_mean(
tf.square(
x=tf.subtract(
x=input,
y=train_mean
)
),
axis=[0, 1, 2]
)
beta = init.bias_init(
name=name + '_beta',
method='zero',
output_channel=input.get_shape().as_list()[-1],
collection='BnBeta'
)
gama= init.bias_init(
name=name + '_gama',
method='one',
output_channel=input.get_shape().as_list()[-1],
collection='BnGama'
)
ema = tf.train.ExponentialMovingAverage(MOVING_DECAY)
predict_mean, predict_var = ema.apply([train_mean, train_var])
def depend_in_train():
with tf.control_dependencies([predict_mean, predict_var]):
return tf.identity(train_mean), tf.identity(train_var)
pass
mean, var = tf.cond(
train,
lambda: depend_in_train(),
lambda: (predict_mean, predict_var)
)
bn = tf.nn.batch_normalization(
x=input,
mean=mean,
variance=var,
offset=beta,
scale=gama,
variance_epsilon=BNEPS
)
tf.add_to_collection(name='BnOut', value=bn)
Util.AddToCollectionInfo('BnOut', bn)
Util.CLayerInfo(name, input, bn)
# bn = tf.add(
# x=beta,
# y=tf.div(
# tf.multiply(
# x=gama,
# y=tf.subtract(
# x=input,
# y=mean
# )
# ),
# y=var
# ),
# name=name + 'bn_output'
# )
return bn
pass
