原文链接: infogan 生成mnist 手写数字

上一篇: teamviewer 远程操控

下一篇: ros 发布订阅模型 自定义消息

InfoGAN介绍

尝试加深生成器网络结构以便于生成更加逼真的信息，结果效果反而很差

生成器中使用relu比leaky_relu效果也好，不知道为什么。。。。

判别器加深后效果也不怎么好，貌似这个结构是 经过多次试验之后才确定的，修改反而效果很差

效果

由于其中采用了bn层，所以如果只对一张图片进行可视化操作会出现下面的问题，但是多张图片就没有 问题

不同的激活函数和网络结构训练得到的结果也不一样

输入,z表示噪声，由两部分组成，一部分蕴含信息，一部分是随机噪声

    x = tf.placeholder(tf.float32, [None, n_input])in_y = tf.placeholder(tf.int32, [None])z_con = tf.placeholder(tf.float32, (None, con_dim))z_rand = tf.placeholder(tf.float32, (None, rand_dim))z = tf.concat([tf.one_hot(in_y, n_class), z_con, z_rand], axis=1)print('z ', z.shape)  # z  (?, 50)

生成器和判别器

生成器输入数据，n*（2+38+10） 输出 n*28*28*1

判别器输入图片返回三个结果

1，真实图片概率 0---1.0

2，分类信息n*10

3，隐含信息n*2

def generator(x):reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0with tf.variable_scope('generator', reuse=reuse):x = slim.fully_connected(x, 1024)x = slim.batch_norm(x, activation_fn=tf.nn.relu)x = slim.fully_connected(x, 7 * 7 * 128)x = slim.batch_norm(x, activation_fn=tf.nn.relu)x = tf.reshape(x, [-1, 7, 7, 128])x = slim.conv2d_transpose(x, 64, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.relu)x = slim.batch_norm(x, activation_fn=tf.nn.relu)x = slim.conv2d_transpose(x, 1, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.sigmoid)print('x ', x.shape)  # x  (?, 28, 28, 1)return xdef discriminator(x):reuse = len([t for t in tf.global_variables() if t.name.startswith('discriminator')]) > 0with tf.variable_scope('discriminator', reuse=reuse):x = tf.reshape(x, shape=[-1, 28, 28, 1])x = slim.conv2d(x, num_outputs=64, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.leaky_relu)x = slim.conv2d(x, num_outputs=128, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.leaky_relu)print('x ', x.shape)  # x  (?, 7, 7, 128)x = slim.flatten(x)  # x  (?, 6272)print('x ', x.shape)shared_tensor = slim.fully_connected(x, num_outputs=512, activation_fn=tf.nn.leaky_relu)recog_shared = slim.fully_connected(shared_tensor, num_outputs=128, activation_fn=tf.nn.leaky_relu)disc = slim.fully_connected(recog_shared, num_outputs=1, activation_fn=tf.nn.leaky_relu)print('disc ', disc.shape)  # disc  (?, 1)disc = tf.squeeze(disc, -1)print('disc ', disc.shape)  # disc  (?,)recog_cat = slim.fully_connected(recog_shared, num_outputs=n_class, activation_fn=tf.nn.leaky_relu)print('recog_cat ', recog_cat.shape)  # recog_cat  (?, 10)recog_cont = slim.fully_connected(recog_shared, num_outputs=con_dim, activation_fn=tf.nn.sigmoid)print('recog_cont ', recog_cont.shape)  # recog_cont  (?, 2)return disc, recog_cat, recog_cont

将输入放入生成器中，然后等到判别器的输出用于计算loss

    gen = generator(z)# labels for discriminatory_real = tf.ones(batch_size)  # 真y_fake = tf.zeros(batch_size)  # 假print('y_real ', y_real.shape)  # y_real  (64,)# 判别器disc_real, class_real, _ = discriminator(x)disc_fake, class_fake, con_fake = discriminator(gen)

loss

判别器中，判别结果的loss有两个：真实输入的结果与模拟输入的结果，将两个结合在一起生成loss_d，生成器的loss为自己输出的模拟数据，让他在判别器中为真，定义为loss_g

定义网络中共有的loss，真实的标签与输入真实样0本判别出的标签，真实的标签与输入模拟样0本判别出的标签，隐含信息的重构误差，然后创建两个优化器，将他们放到对应的优化器中

loss_cf 分类正确，但生成的样本错了

loss_cr 分类与样本正确，但与输入的分类对不上

loss_con 隐含变量的loss

loss_g 生成器的loss目标是判别结果全为真

loss_d 判别器的loss目标是将假的判为假，真的判为真

所谓的AC-GAN就是将loss_cr加入到loss_c中，如果没有loss_cr，令loss_c=loss_cf，对于网络生成的模拟数据是不影响的，但是会损失真实分类与模拟数据间的对应关系。

由于损失函数中的判别为长度为n的数组，所以使用sigmoid交叉熵，softmax用于进行单分类，即一个物体只能归属于一个类别这种情况。

    # 判别器 loss 使用sigmoid函数因为判别器输出的是一个数字表示真实的概率# softmax在输出为one hot编码时使用，在这里由于只有一个数字所以计算结果永远为0loss_d_r = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real, labels=y_real))loss_d_f = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake, labels=y_fake))loss_d = (loss_d_r + loss_d_f) / 2# generator lossloss_g = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake, labels=y_real))loss_cf = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=class_fake, labels=in_y))loss_cr = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=class_real, labels=in_y))loss_c = (loss_cf + loss_cr) / 2loss_con = tf.reduce_mean((con_fake - z_con) ** 2)

训练

小技巧: 将判别器的学习率设小，将生成器的学习率设大，这么做是为了 让生成器有更快的进化速度来模拟真实数据，优化同样是adam

    # 获得各个网络中各自的训练参数t_vars = tf.trainable_variables()d_vars = [var for var in t_vars if 'discriminator' in var.name]g_vars = [var for var in t_vars if 'generator' in var.name]train_disc = tf.train.AdamOptimizer(0.0001).minimize(loss_d + loss_c + loss_con, var_list=d_vars,global_step=disc_global_step)train_gen = tf.train.AdamOptimizer(0.001).minimize(loss_g + loss_c + loss_con, var_list=g_vars,global_step=gen_global_step)

完整代码

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import tensorflow.contrib.slim as slim
from tensorflow.examples.tutorials.mnist import input_data
from scipy.stats import normn_class = 10batch_size = 64  # 获取样本的批次大小64
con_dim = 2  # total continuous factor
n_input = 784
training_epochs = 20000
display_step = 500
rand_dim = 38def generator(x):reuse = len([t for t in tf.global_variables() if t.name.startswith('generator')]) > 0with tf.variable_scope('generator', reuse=reuse):x = slim.fully_connected(x, 1024)x = slim.batch_norm(x, activation_fn=tf.nn.relu)x = slim.fully_connected(x, 7 * 7 * 128)x = slim.batch_norm(x, activation_fn=tf.nn.relu)x = tf.reshape(x, [-1, 7, 7, 128])x = slim.conv2d_transpose(x, 64, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.relu)x = slim.batch_norm(x, activation_fn=tf.nn.relu)x = slim.conv2d_transpose(x, 1, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.sigmoid)print('x ', x.shape)  # x  (?, 28, 28, 1)return xdef discriminator(x):reuse = len([t for t in tf.global_variables() if t.name.startswith('discriminator')]) > 0with tf.variable_scope('discriminator', reuse=reuse):x = tf.reshape(x, shape=[-1, 28, 28, 1])x = slim.conv2d(x, num_outputs=64, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.leaky_relu)x = slim.conv2d(x, num_outputs=128, kernel_size=[4, 4], stride=2, activation_fn=tf.nn.leaky_relu)print('x ', x.shape)  # x  (?, 7, 7, 128)x = slim.flatten(x)  # x  (?, 6272)print('x ', x.shape)shared_tensor = slim.fully_connected(x, num_outputs=512, activation_fn=tf.nn.leaky_relu)recog_shared = slim.fully_connected(shared_tensor, num_outputs=128, activation_fn=tf.nn.leaky_relu)disc = slim.fully_connected(recog_shared, num_outputs=1, activation_fn=tf.nn.leaky_relu)print('disc ', disc.shape)  # disc  (?, 1)disc = tf.squeeze(disc, -1)print('disc ', disc.shape)  # disc  (?,)recog_cat = slim.fully_connected(recog_shared, num_outputs=n_class, activation_fn=tf.nn.leaky_relu)print('recog_cat ', recog_cat.shape)  # recog_cat  (?, 10)recog_cont = slim.fully_connected(recog_shared, num_outputs=con_dim, activation_fn=tf.nn.sigmoid)print('recog_cont ', recog_cont.shape)  # recog_cont  (?, 2)return disc, recog_cat, recog_contdef main():x = tf.placeholder(tf.float32, [None, n_input])in_y = tf.placeholder(tf.int32, [None])z_con = tf.placeholder(tf.float32, (None, con_dim))z_rand = tf.placeholder(tf.float32, (None, rand_dim))z = tf.concat([tf.one_hot(in_y, n_class), z_con, z_rand], axis=1)print('z ', z.shape)  # z  (?, 50)gen = generator(z)# labels for discriminatory_real = tf.ones(batch_size)  # 真y_fake = tf.zeros(batch_size)  # 假print('y_real ', y_real.shape)  # y_real  (64,)# 判别器disc_real, class_real, _ = discriminator(x)disc_fake, class_fake, con_fake = discriminator(gen)disc_global_step = tf.Variable(0, trainable=False)gen_global_step = tf.Variable(0, trainable=False)# 判别器 loss 使用sigmoid函数因为判别器输出的是一个数字表示真实的概率# softmax在输出为one hot编码时使用，在这里由于只有一个数字所以计算结果永远为0loss_d_r = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real, labels=y_real))loss_d_f = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake, labels=y_fake))loss_d = (loss_d_r + loss_d_f) / 2# generator lossloss_g = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake, labels=y_real))loss_cf = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=class_fake, labels=in_y))loss_cr = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=class_real, labels=in_y))loss_c = (loss_cf + loss_cr) / 2loss_con = tf.reduce_mean((con_fake - z_con) ** 2)# softmax在输出为one hot编码时使用，在这里由于只有一个数字所以计算结果永远为0# loss_d_r = tf.reduce_mean((disc_real - y_real) ** 2)# loss_d_f = tf.reduce_mean((disc_fake - y_fake) ** 2)# loss_d = (loss_d_r + loss_d_f) / 2# # generator loss# loss_g = tf.reduce_mean((disc_fake - y_real) ** 2)# 获得各个网络中各自的训练参数t_vars = tf.trainable_variables()d_vars = [var for var in t_vars if 'discriminator' in var.name]g_vars = [var for var in t_vars if 'generator' in var.name]train_disc = tf.train.AdamOptimizer(0.0001).minimize(loss_d + loss_c + loss_con, var_list=d_vars,global_step=disc_global_step)train_gen = tf.train.AdamOptimizer(0.001).minimize(loss_g + loss_c + loss_con, var_list=g_vars,global_step=gen_global_step)mnist = input_data.read_data_sets(r"D:\code\python\tf\gandemo\MNIST_data")for i in tf.global_variables():print(i.name, i.shape)with tf.Session() as sess:sess.run(tf.global_variables_initializer())for epoch in range(1, 1 + training_epochs):# 遍历全部数据集batch_xs, batch_ys = mnist.train.next_batch(batch_size)  # 取数据# print('batch_xs ys', batch_xs.shape, batch_ys.shape) # batch_xs ys (64, 784) (64,)batch_con = np.random.randn(batch_size, con_dim)batch_rand = np.random.randn(batch_size, rand_dim)feeds = {x: batch_xs, z_con: batch_con, z_rand: batch_rand, in_y: batch_ys}# Fit training using batch datal_disc, _, disc_step = sess.run([loss_d, train_disc, disc_global_step], feeds)l_gen, _, gen_step = sess.run([loss_g, train_gen, gen_global_step], feeds)# 显示训练中的详细信息if not epoch % display_step:batch_xs, batch_ys = mnist.train.next_batch(batch_size)  # 取数据batch_con = np.random.randn(batch_size, con_dim)batch_rand = np.random.randn(batch_size, rand_dim)feeds = {x: batch_xs, z_con: batch_con, z_rand: batch_rand, in_y: batch_ys}disc_fake_val, disc_real_val, loss_d_val, loss_g_val = sess.run([disc_fake, disc_real, loss_d, loss_g], feeds)print(epoch, loss_d_val, loss_g_val)col = []for i in range(n_class):t = norm.ppf(np.linspace(.5, .95, n_class))batch_con = np.stack([t] * con_dim).Tbatch_rand = np.random.randn(n_class, rand_dim)feeds = {z_con: batch_con, z_rand: batch_rand, in_y: np.arange(0, n_class)}img_val = sess.run(gen, feeds)img_val = np.reshape(img_val, (-1, 28, 28))col.append(np.concatenate(img_val))col = np.concatenate(col, axis=1)col2 = []for i in range(n_class):batch_con = np.random.randn(n_class, con_dim)batch_rand = np.random.randn(n_class, rand_dim)feeds = {z_con: batch_con, z_rand: batch_rand, in_y: np.arange(0, n_class)}img_val = sess.run(gen, feeds)img_val = np.reshape(img_val, (-1, 28, 28))col2.append(np.concatenate(img_val))col2 = np.concatenate(col2, axis=1)img = np.concatenate([col, col2])plt.imshow(img)plt.show()if __name__ == '__main__':main()