修改的wgan代码

最新推荐文章于 2024-04-16 12:15:21 发布

Jack_Weng

最新推荐文章于 2024-04-16 12:15:21 发布

阅读量909

点赞数

文章标签：深度学习生成对抗网络

本文链接：https://blog.csdn.net/wj310298/article/details/85328533

版权

基于《Python深度学习》中的GAN代码，作者进行了个人理解的修改，实验结果显示性能显著提升。

摘要由CSDN通过智能技术生成

对《Python深度学习》的GAN源码，根据自己的理解做了些许修改，发现性能提高很多。

import keras
from keras.constraints import min_max_norm
from keras import layers, initializers
import numpy as np
import os
from keras.preprocessing import image
from keras import backend as K


def wasserstein(y_true, y_pred):
    return K.mean(y_true * y_pred)


batch_size = 64
latent_dim = 32
height = 32
width = 32
s = 32
channels = 3
Citers = 5

start_dim = 2

c = 0.01
lr_G = 5e-5
lr_D = 5e-5

f = 512
reshape_shape = (start_dim, start_dim, f)

generator_input = keras.Input(shape=(latent_dim,))

x = layers.Dense(f * start_dim * start_dim, use_bias=False)(generator_input)
x = layers.BatchNormalization(axis=-1)(x)
x = layers.LeakyReLU()(x)
x = layers.Reshape(reshape_shape)(x)

for i in range(3):
    nb_filters = int(f / (2 ** (i + 1)))
    x = layers.Deconv2D(nb_filters, (3, 3),