损失函数: 基于SGAN的Loss 判别器输出为概率值需要sigmoid
网络结构: MLP 第一层的MLP的要加上 Concat One_Hot 条件
数据形式: 不带卷积 没有深度维 图片压缩到0 1 之间
生成器: sigmoid 映射到0 1 之间 迎合数据格式
判别器: sigmoid 映射到0 1 之间 迎合loss公式的约束
初始化: xavier初始化 即考虑输入输出维度的 glorot uniform
训练： 判别器和生成器同时训练 同步训练 不偏重任一方
"""
import my_mnist
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers
import time
(train_images, train_labels), (_, _) = my_mnist.load_data(get_new=False,
normalization=True,
one_hot=True,
detype=np.float32)
train_images = train_images.reshape(train_images.shape[0], 28,
28).astype('float32')
print(train_labels[0])
plt.imshow(train_images[0, :, :], cmap='gray')
plt.show()
class Dense(layers.Layer):
def __init__(self, input_dim, units):
super(Dense, self).__init__()
initializer = tf.initializers.glorot_uniform()
# initializer = tf.initializers.glorot_normal()
self.w = tf.Variable(initial_value=initializer(shape=(input_dim,
units),
conv_out = tf.nn.conv2d(
input=x,
filters=self.w,
strides=self.strides,
padding=self.pandding_way,
data_format='NHWC',
dilations=None,
name=None) #dilations是空洞卷积的一个系数 相当于对卷积核做上采样同时部分置零 这里不进行空洞卷积
l_out = conv_out + self.b
return l_out
if __name__ == "__main__":
(train_images, train_labels), (_,
_) = my_mnist.load_data(get_new=False,
normalization=False,
one_hot=True,
detype=np.float32)
train_images = (train_images.astype('float32') - 127.5) / 127.5
train_labels = (train_labels.astype('float32') - 0.5) / 0.5
train_images = train_images.reshape(train_images.shape[0], 28, 28, 1)
print(train_labels[0])
plt.imshow(train_images[0, :, :, 0], cmap='gray')
plt.show()
x = tf.random.normal(shape=(64, 784))
a = Dense(28 * 28, 128)
print(a(x))
print(len(a.trainable_variables))
y = train_images[0:1, :, :, 0:1]
C1 = Conv2D([28, 28, 1],
2, [5, 5],