def apply(self, x, index, model):
with tf.name_scope(self.name):
input_shape = x.get_shape()
fan_in = input_shape[-1].value
stddev = math.sqrt(1.0 / fan_in) # he init
shape = [fan_in, self.fan_out]
W, b = weight_bias(shape, stddev=stddev, bias_init=0.0)
self.h = tf.matmul(x, W) + b
return self.h
def apply(self, x, index, model):
with tf.name_scope(self.name):
input_shape = x.get_shape()
fan_in = input_shape[-1].value
stddev = math.sqrt(1.0 / fan_in) # he init
shape = [fan_in, self.fan_out]
W, b = weight_bias(shape, stddev=stddev, bias_init=0.0)
self.h = tf.matmul(x, W) + b
return self.h
def apply(self, x, index, model):
with tf.name_scope(self.name):
input_shape = x.get_shape()
input_channels = input_shape[-1].value
k_w, k_h = self.filter_shape
stddev = math.sqrt(2.0 / ((k_w * k_h) * input_channels)) # he init
shape = self.filter_shape + [input_channels, self.output_channels]
W, b = weight_bias(shape, stddev=stddev, bias_init=0.0)
self.h = tf.nn.conv2d(x, W, self.strides, self.padding) + b
return self.h
def apply(self, x, index, model):
with tf.name_scope(self.name):
input_shape = x.get_shape()
input_channels = input_shape[-1].value
k_w, k_h = self.filter_shape
stddev = math.sqrt(2.0 / ((k_w * k_h) * input_channels)) # he init
shape = self.filter_shape + [input_channels, self.output_channels]
W, b = weight_bias(shape, stddev=stddev, bias_init=0.0)
self.h = tf.nn.conv2d(x, W, self.strides, self.padding) + b
return self.h
