Question

最近，我正在研究GAN网络，我用它来生成一个mnisit图像，我的电脑中的环境是ubuntu16.04，tensorflow，python3。

代码可以无任何错误地运行。但是结果显示网络学习没什么，通过训练，输出图像仍然是嘈杂的图像。

首先，我设计了一个生成器网络：输入是784维度的噪声数据，通过隐藏层并对其进行规则生成784维度的图像。

然后我设计了一个鉴别器网络：输入是真实图像和假图像，通过隐藏层并对其进行规则，输出为1维的logits。

然后我定义了generator_loss和discriminator_loss，然后训练生成器和鉴别器。它可以运行，但结果表明网络学习没什么，损失不能收敛。

import tensorflow as tf
import numpy as np 
import tensorflow.contrib.slim as slim
import matplotlib.pyplot as plt
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets("/home/zyw/data/tensor_mnist-master/MNIST_data/",one_hot=True)
batch_size = 100

G_in = tf.placeholder(tf.float32,[None,784])
G_h1 = tf.layers.dense(G_in, 128)
G_h1 = tf.maximum(0.01 * G_h1, G_h1)
G_out = tf.tanh(tf.layers.dense(G_h1, 784)) 

real = tf.placeholder(tf.float32,[None,784]) 
Dl0 = tf.layers.dense(G_out, 128)
Dl0 = tf.maximum(0.01 * Dl0, Dl0)
p0 = tf.layers.dense(Dl0, 1)

Dl1 = tf.layers.dense(real, 128)
Dl1 = tf.maximum(0.01 * Dl1, Dl1)
p1 = tf.layers.dense(Dl1, 1)

G_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits =p0,labels=tf.ones_like(p0)*0.9))
D_real_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits =p1,labels=tf.ones_like(p1)*0.9))
D_fake_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits =p0,labels=tf.zeros_like(p0)))
D_total_loss = tf.add(D_fake_loss,D_real_loss) 

G_train = tf.train.AdamOptimizer(0.01).minimize(G_loss)
D_train = tf.train.AdamOptimizer(0.01).minimize(D_total_loss)

init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)

    for i in range(1000):
        mnist_data,_ = mnist.train.next_batch(batch_size)

        # noise_org = tf.random_normal([batch_size,784],stddev = 0.1,dtype = tf.float32)
        noise_org = np.random.randn(batch_size, 784)
        a,b,dloss=  sess.run([D_real_loss,D_fake_loss,D_total_loss,G_train,D_train],feed_dict={G_in:noise_org,real:mnist_data})[:3]
        if i%100==0:
            print(a,b,dloss)
    #test_generative_image
    noise_org = np.random.randn(1, 784)
    image = sess.run(G_out,feed_dict ={G_in:noise_org})
    outimage = tf.reshape(image, [28,28])
    plt.imshow(outimage.eval(),cmap='gray')
    plt.show()
    print('ok')

结果是：

0.80509 0.63548 1.44057
0.33512 0.20223 0.53735
0.332536 0.97737 1.30991
0.328048 0.814452 1.1425
0.326688 0.411907 0.738596
0.325864 0.570807 0.896671
0.325575 0.970406 1.29598
0.325421 1.02487 1.35029
0.325222 1.34089 1.66612
0.325217 0.747129 1.07235

Answer 1

我添加了修改后的代码以及我做出更改的注释。此外，我已在下面描述了我的更改。

import tensorflow as tf
import numpy as np
import tensorflow.contrib.slim as slim
import matplotlib.pyplot as plt
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets("/home/zyw/data/tensor_mnist-master/MNIST_data/",one_hot=True)

batch_size = 100

#define the generator function
def generator(input):
    G_h1 = tf.layers.dense(input, 128)
    # G_h1 = tf.maximum(0.01 * G_h1, G_h1)
    G_out = tf.sigmoid(tf.layers.dense(G_h1, 784))  # sigmoid function added
    return G_out

#Define the discrminator function
def discriminator(input):
    Dl0 = tf.layers.dense(input, 128)
    # Dl0 = tf.maximum(0.01 * Dl0, Dl0)
    p0 = tf.layers.dense(Dl0, 1)
    return p0

#Generator
with tf.variable_scope('G'):
    G_in = tf.placeholder(tf.float32, [None, 784])
    G_out = generator(G_in)

real = tf.placeholder(tf.float32, [None, 784])

#Discrimnator that takes the real data
with tf.variable_scope('D'):
    D1 = discriminator(real)

#Discriminator that takes fake data
with tf.variable_scope('D', reuse=True):  # need to use the same copy of Discrminator
    D2 = discriminator(G_out)

G_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D2, labels=tf.ones_like(D2)))
D_real_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D1, labels=tf.ones_like(D1)))
D_fake_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D2, labels=tf.zeros_like(D2)))
D_total_loss = tf.add(D_fake_loss, D_real_loss)

vars = tf.trainable_variables() #all trainable variables
d_training_vars = [v for v in vars if v.name.startswith('D/')]  # varibles associated with the discrminator
g_training_vars = [v for v in vars if v.name.startswith('G/')]  # varibles associated with the generator

G_train = tf.train.AdamOptimizer(0.001).minimize(G_loss,var_list=g_training_vars)  # only train the variables associated with the generator
D_train = tf.train.AdamOptimizer(0.001).minimize(D_total_loss,var_list=d_training_vars)  # only train the variables associated with the discriminator

init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)

    for i in range(1000):
        mnist_data, _ = mnist.train.next_batch(batch_size)
        # noise_org = tf.random_normal([batch_size,784],stddev = 0.1,dtype = tf.float32)
        noise_org = np.random.randn(batch_size, 784)

        a, b, dloss = sess.run([D_real_loss, D_fake_loss, D_total_loss, G_train, D_train],feed_dict={G_in: noise_org, real: mnist_data})[:3]
        if i % 100 == 0:
            print(a, b, dloss)

    # test_generative_image
    noise_org = np.random.randn(1, 784)
    image = sess.run(G_out, feed_dict={G_in: noise_org})
    outimage = tf.reshape(image, [28, 28])

    plt.imshow(outimage.eval(), cmap='gray')
    plt.show()
    print('ok')

实施 GAN 时，您应该注意几点，

需要使用相同的鉴别器副本（即共享相同的副本权重）当实施鉴别器损失时（在您的情况下 Dl0 和 Dl1 应该共享相同的句子。）
生成器激活功能应为 sigmoid 而不是 tanh 因为生成器的输出应该只在0之间变化和（1.从它的图像）
在训练鉴别器时，您应该只训练与鉴别器相关的变量。同样，在训练生成器时，您只应训练与生成器相关联的变量。
有时确保鉴别器非常重要比生成器更强大，否则它就没有了足够的能力学习能够准确区分在生成的和真实的样本之间。

这些只是你应该注意的 GAN 的基本内容。但是，在开发 GAN 时，您应该考虑许多其他方面。通过阅读以下两篇文章，您可以了解 GAN 的基本概念。

希望这有帮助。

使用GAN来生成mnist，但什么都不学习

1 个答案: