def __init__(self,
input_shape,
kernel_size,
kernel_initializer="glorot_uniform",
using_bias=False,
activation="relu",
**kwargs):
"""
input_shape: list [None,shape] or [shape]
kernel_size: int
using_bias: bool
activation: string like "relu","sigmoid","leaky_relu","tanh"
"""
super(Dense, self).__init__()
self.initializer = Reconstruction.initializer(
kernel_initializer) #定义初始化采用的方法
self.input_shape_zp = Reconstruction.remake_shape(shape=input_shape,
dims=2)
self.w = tf.Variable(initial_value=self.initializer(
shape=[self.input_shape_zp[-1], kernel_size], dtype=tf.float32),
trainable=True)
self.using_bias = using_bias
if using_bias:
self.b = tf.Variable(
initial_value=tf.zeros(shape=(1, kernel_size),
dtype=tf.float32),
trainable=True) #节点的偏置也是行向量 才可以正常计算 即对堆叠的batch 都是加载单个batch内
self.activation = Reconstruction.activation(activation)
def build(self,*args,input_shape=None,**kwargs):
out_shape,padding,padding_vect = Reconstruction.ConvCalculation(input_shape=input_shape[1:-1],
filters=self.filters_zp,
kernel_size=self.kernel_size_zp,
strides=self.strides_zp,
padding=self.padding_zp)
self.padding_zp = padding
self.padding_vect_zp = padding_vect
super(Conv2D,self).build(*args,input_shape=input_shape,**kwargs)
def __init__(self,
input_shape,
filters,
kernel_size,
strides=[1, 1, 1],
padding="SAME",
using_bias=False,
sn=False):
"""
将卷积统一为
计算参数量
自行padding
然后使用VALID卷积
这样就可以嵌入任意的padding模式了
则该层的输入shape对于SN而言，就需要加上padding_vect变成padding后的值
"""
super(Conv3D, self).__init__()
self.input_shape_zp = Reconstruction.remake_shape(input_shape,
dims=5) #5维的输出张量
self.kernel_size_zp = Reconstruction.remake_kernel_size(
kernel_size, dims=3) #3维卷积核 变为[1,x,x,x,1]5D形式
self.strides_zp = Reconstruction.remake_strides(
strides, dims=3) #3维步长 变为[1,x,x,x,1]5D形式
self.output_shape_zp, self.padding, self.padding_vect = Reconstruction.ConvCalculation(
self.input_shape_zp[1:-1], filters, self.kernel_size_zp[1:-1],
self.strides_zp[1:-1], padding)
weight_shape = self.kernel_size_zp[1:-1] + [
self.input_shape_zp[-1]
] + [filters] #[D,H,W,IC,OC]
self.initializer = Reconstruction.initializer("random_normal",
mean=0.0,
stddev=0.02)
self.w = tf.Variable(
initial_value=self.initializer(shape=weight_shape),
trainable=True,
dtype=tf.float32)
self.using_bias = using_bias
if using_bias == True:
self.b = tf.Variable(tf.zeros(self.output_shape_zp,
dtype=tf.float32),
trainable=True)
self.sn = sn
def __init__(self,
input_shape,
output_shape,
filters,
kernel_size,
strides=[2, 2, 2],
padding="SAME",
using_bias=False,
sn=False): #输出的深度已经给出 只要给出每个通道的H,W即可
super(UpSampaleConv3D, self).__init__()
self.input_shape_zp = Reconstruction.remake_shape(
input_shape, dims=5) #变成5维的输出张量 None补充不足的维度(batch)
self.output_shape_zp = Reconstruction.remake_shape(output_shape +
[filters],
dims=5) #变成5维的输出张量
self.kernel_size_zp = Reconstruction.remake_kernel_size(
kernel_size, dims=3) #3维卷积核 变为[1,x,x,x,1]5D形式
self.strides_zp = Reconstruction.remake_strides(
strides, dims=3) #3维步长 变为[1,x,x,x,1]4D形式
self.padding = padding
self.padding, self.padding_vect, self.cut_flag = Reconstruction.Trans2UpsampleCal(
self.input_shape_zp[1:-1], self.output_shape_zp[1:-1], filters,
self.kernel_size_zp[1:-1], self.strides_zp[1:-1], self.padding)
self.up_size = self.strides_zp[1:-1]
self.up_op = tf.keras.layers.UpSampling3D(
size=self.up_size) #上采样只对非通道和非batch进行 剔除冗余维度
self.strides_zp = Reconstruction.remake_strides([1, 1, 1], dims=3)
self.initializer = Reconstruction.initializer("random_normal",
mean=0.0,
stddev=0.02)
weight_shape = self.kernel_size_zp[1:-1] + [
self.input_shape_zp[-1]
] + [filters] #[D,H,W,IC,OC]
self.w = tf.Variable(
initial_value=self.initializer(shape=weight_shape),
trainable=True,
dtype=tf.float32)
self.using_bias = using_bias
if using_bias == True:
self.b = tf.Variable(tf.zeros(self.output_shape_zp[1:],
dtype=tf.float32),
trainable=True)
self.sn = sn
def __init__(self,
input_shape,
output_shape,
filters,
kernel_size,
strides=[1, 1],
padding="SAME",
using_bias=False,
sn=False):
super(Conv2DTranspose, self).__init__()
self.input_shape_zp = Reconstruction.remake_shape(
input_shape, dims=4) #变成4维的输出张量 None补充不足的维度(batch)
self.output_shape_zp = Reconstruction.remake_shape(output_shape +
[filters],
dims=4) #变成4维的输出张量
self.kernel_size_zp = Reconstruction.remake_kernel_size(
kernel_size, dims=2) #2维卷积核 变为[1,x,x,1]4D形式
self.strides_zp = Reconstruction.remake_strides(
strides, dims=2) #2维步长 变为[1,x,x,1]4D形式
self.padding = padding
Reconstruction.ConvTransCheck(input_shape=self.input_shape_zp[1:-1],
output_shape=self.output_shape_zp[1:-1],
filters=filters,
kernel_size=self.kernel_size_zp[1:-1],
strides=self.strides_zp[1:-1],
padding=self.padding)
weight_shape = self.kernel_size_zp[1:-1] + [filters] + [
self.input_shape_zp[-1]
] #[D,H,W,OC,IC]
self.initializer = Reconstruction.initializer("random_normal",
mean=0.0,
stddev=0.02)
self.w = tf.Variable(
initial_value=self.initializer(shape=weight_shape),
trainable=True,
dtype=tf.float32)
self.using_bias = using_bias
if using_bias == True:
self.b = tf.Variable(tf.zeros(self.output_shape_zp[1:],
dtype=tf.float32),
trainable=True)
self.sn = sn