本文使用的tensorflow版本：1.4
tensorflow安装：pip install tensorflow
1、先来目睹一下效果吧
这篇文章讲解了如何使用GAN来生成我们的手写数字，我们首先来看看生成的效果吧：
10轮：
50轮：
100轮：
200轮：
可以看到，在10轮的时候，我们的Generator生成的图片非常模糊，几乎是无法用肉眼来分别数字的，到了第50轮的时候，已经初见雏形了，有一些数字比如1、4、5这些都可以很清楚的分辨出来，不过还并不是十分完美。到了100轮的时候，像数字8和9这些也基本能准确的生成了，而到200轮的时候，除去个别的以外，基本上都能正确的手写出来了，由于时间的原因，没有继续训练下去，如果大家感兴趣，可以训练更多轮，看看更好的效果。
2、思路解析
设定参数
本文设定的参数是，图片的大小是28*28，这是mnist图片的标准大小，后面是一些保存模型的设定。我们总共的训练轮数是500轮。在我们的Generator和Discriminator中，我们设定的是一个简单的有两层隐藏层的全链接神经网络。对于Generator来说，输入的的大小是[batch_size,z_size]，第一个隐藏层有150个神经元，第二个隐藏层有300个神经元，输出的大小就是图片的size28*28。而对于Discriminator来说，我们的输入的大小是[batch_size * 2,img_size]，因为我们要掺杂真实的img和Generator生成的img。第一个隐藏层有300个神经元，第二个隐藏层有150个神经元，输出层只有1个数，表示该图片为真实图片的概率。关于神经网络的结构我们会在后面详细讲解。
img_height =
28
img_width =
28
img_size = img_height * img_width
to_train = Trueto_restore = Falseoutput_path = "output"# 总迭代次数500max_epoch = 500h1_size = 150h2_size = 300z_size = 100batch_size = 256
创建Generator
刚才也讲到了，对于Generator来说，输入的的大小是[batch_size,z_size]，第一个隐藏层有150个神经元，第二个隐藏层有300个神经元，输出的大小就是图片的size28*28。总的来说，经过Generator，由[batch_size,z_size] 变为 [batch_size,img_size]
# generate (model 1)
def build_generator(z_prior):
w1 = tf.Variable(tf.truncated_normal([z_size, h1_size], stddev=0.1), name="g_w1", dtype=tf.float32)
b1 = tf.Variable(tf.zeros([h1_size]), name="g_b1", dtype=tf.float32)
h1 = tf.nn.relu(tf.matmul(z_prior, w1) + b1)
w2 = tf.Variable(tf.truncated_normal([h1_size, h2_size], stddev=0.1), name="g_w2", dtype=tf.float32)
b2 = tf.Variable(tf.zeros([h2_size]), name="g_b2", dtype=tf.float32)
h2 = tf.nn.relu(tf.matmul(h1, w2) + b2)
w3 = tf.Variable(tf.truncated_normal([h2_size, img_size], stddev=0.1), name="g_w3", dtype=tf.float32)
b3 = tf.Variable(tf.zeros([img_size]), name="g_b3", dtype=tf.float32)
h3 = tf.matmul(h2, w3) + b3
x_generate = tf.nn.tanh(h3)
g_params = [w1, b1, w2, b2, w3, b3] return x_generate, g_params
创建Discrminator
而对于Discriminator来说，我们的输入的大小是[batch_size * 2,img_size]，因为我们要掺杂真实的img和Generator生成的img。第一个隐藏层有300个神经元，第二个隐藏层有150个神经元，输出层只有1个数，表示该图片为真实图片的概率。要注意，我们的输入和输出是严格对应的，所以对于输出成的输出h3来说，前batch_size个代表着对真实图片的判别概率，而后batch_size代表着对Generator生成的图片的判别概率。这里是用了一个tf.slice()函数，之前没有接触过，故在这里做一下记录：
1，函数原型 tf.slice(inputs,begin,size,name='')
2，用途：从inputs中抽取部分内容
inputs：可以是list,array,tensor
begin：n维列表，begin[i] 表示从inputs中第i维抽取数据时，相对0的起始偏移量，也就是从第i维的begin[i]开始抽取数据
size：n维列表，size[i]表示要抽取的第i维元素的数目
所以可以看到，最终y_data保存的是真实图片的判别概率，这些值要越接近于1越好，而y_generated保存的是Generator生成的图片的判别概率，这些值要越接近于0越好。
def build_discriminator(x_data, x_generated, keep_prob):
# tf.concat
x_in = tf.concat([x_data, x_generated], 0)
w1 = tf.Variable(tf.truncated_normal([img_size, h2_size], stddev=0.1), name="d_w1", dtype=tf.float32)
b1 = tf.Variable(tf.zeros([h2_size]), name="d_b1", dtype=tf.float32)
h1 = tf.nn.dropout(tf.nn.relu(tf.matmul(x_in, w1) + b1), keep_prob)
w2 = tf.Variable(tf.truncated_normal([h2_size, h1_size], stddev=0.1), name="d_w2", dtype=tf.float32)
b2 = tf.Variable(tf.zeros([h1_size]), name="d_b2", dtype=tf.float32)
h2 = tf.nn.dropout(tf.nn.relu(tf.matmul(h1, w2) + b2), keep_prob)
w3 = tf.Variable(tf.truncated_normal([h1_size, 1], stddev=0.1), name="d_w3", dtype=tf.float32)
b3 = tf.Variable(tf.zeros([1]), name="d_b3", dtype=tf.float32)
h3 = tf.matmul(h2, w3) + b3
y_data = tf.nn.sigmoid(tf.slice(h3, [0, 0], [batch_size, -1], name=None))
y_generated = tf.nn.sigmoid(tf.slice(h3, [batch_size, 0], [-1, -1], name=None))
d_params = [w1, b1, w2, b2, w3, b3] return y_data, y_generated, d_params
保存图片
关于保存图片的代码，我们这里就不讲了，这也不是重点，大家有兴趣的话可以研究下：
def show_result(batch_res, fname, grid_size=(8, 8), grid_pad=
5
):
batch_res = 0.5 * batch_res.reshape((batch_res.shape[0], img_height, img_width)) + 0.5
img_h, img_w = batch_res.shape[1], batch_res.shape[2]
grid_h = img_h * grid_size[0] + grid_pad * (grid_size[0] - 1)
grid_w = img_w * grid_size[1] + grid_pad * (grid_size[1] - 1)
img_grid = np.zeros((grid_h, grid_w), dtype=np.uint8) for i, res in enumerate(batch_res): if i >= grid_size[0] * grid_size[1]: break
img = (res) * 255
img = img.astype(np.uint8)
row = (i // grid_size[0]) * (img_h + grid_pad)
col = (i % grid_size[1]) * (img_w + grid_pad)
img_grid[row:row + img_h, col:col + img_w] = img
imsave(fname, img_grid)
设定训练目标
对于Discriminator来说，他希望能够使二分类的结果越准确越好，即越能准确判别真实图片和Generator生成的图片越好，所以我们这里使用类似于逻辑回归中的损失函数，而对于Generator来说，它希望的是Discriminator无法分辨它生成的图片，所以它希望Discriminator能将它生成的图片越多的分类为真实图片，所以我们设定的训练目标如下：
# 损失函数的设置
d_loss = - (tf.log(y_data) + tf.log(1 - y_generated))
g_loss = - tf.log(y_generated)
optimizer = tf.train.AdamOptimizer(0.0001)
# 两个模型的优化函数
d_trainer = optimizer.minimize(d_loss, var_list=d_params)
g_trainer = optimizer.minimize(g_loss, var_list=g_params)
训练
训练其实很简单，每次得到batch_size大小的样本，先训练一次Discriminator，然后再训练我们的Generator
z_sample_val = np.random.normal(0, 1, size=(batch_size, z_size)).astype(np.float32)
steps = 60000 / batch_size for i in range(sess.run(global_step), max_epoch): for j in np.arange(steps): # for j in range(steps): print("epoch:%s, iter:%s" % (i, j)) # 每一步迭代，我们都会加载256个训练样本，然后执行一次train_step
x_value, _ = mnist.train.next_batch(batch_size)
x_value = 2 * x_value.astype(np.float32) - 1
z_value = np.random.normal(0, 1, size=(batch_size, z_size)).astype(np.float32) # 执行生成
sess.run(d_trainer,
feed_dict={x_data: x_value, z_prior: z_value, keep_prob: np.sum(0.7).astype(np.float32)}) # 执行判别 if j % 1 == 0:
sess.run(g_trainer,
feed_dict={x_data: x_value, z_prior: z_value, keep_prob: np.sum(0.7).astype(np.float32)})
3、完整代码
本文涉及到的完整代码如下：
import
tensorflow
as
tf
from
tensorflow.examples.tutorials.mnist
import
input_data
import
numpy
as
np
from
skimage.io
import
imsave
import
os
import
shutil
img_height = 28img_width = 28img_size = img_height * img_width
to_train = Trueto_restore = Falseoutput_path = "output"# 总迭代次数500max_epoch = 500h1_size = 150h2_size = 300z_size = 100batch_size = 256# generate (model 1)def build_generator(z_prior):
w1 = tf.Variable(tf.truncated_normal([z_size, h1_size], stddev=0.1), name="g_w1", dtype=tf.float32)
b1 = tf.Variable(tf.zeros([h1_size]), name="g_b1", dtype=tf.float32)
h1 = tf.nn.relu(tf.matmul(z_prior, w1) + b1)
w2 = tf.Variable(tf.truncated_normal([h1_size, h2_size], stddev=0.1), name="g_w2", dtype=tf.float32)
b2 = tf.Variable(tf.zeros([h2_size]), name="g_b2", dtype=tf.float32)
h2 = tf.nn.relu(tf.matmul(h1, w2) + b2)
w3 = tf.Variable(tf.truncated_normal([h2_size, img_size], stddev=0.1), name="g_w3", dtype=tf.float32)
b3 = tf.Variable(tf.zeros([img_size]), name="g_b3", dtype=tf.float32)
h3 = tf.matmul(h2, w3) + b3
x_generate = tf.nn.tanh(h3)
g_params = [w1, b1, w2, b2, w3, b3] return x_generate, g_params# discriminator (model 2)def build_discriminator(x_data, x_generated, keep_prob): # tf.concat
x_in = tf.concat([x_data, x_generated], 0)
w1 = tf.Variable(tf.truncated_normal([img_size, h2_size], stddev=0.1), name="d_w1", dtype=tf.float32)
b1 = tf.Variable(tf.zeros([h2_size]), name="d_b1", dtype=tf.float32)
h1 = tf.nn.dropout(tf.nn.relu(tf.matmul(x_in, w1) + b1), keep_prob)
w2 = tf.Variable(tf.truncated_normal([h2_size, h1_size], stddev=0.1), name="d_w2", dtype=tf.float32)
b2 = tf.Variable(tf.zeros([h1_size]), name="d_b2", dtype=tf.float32)
h2 = tf.nn.dropout(tf.nn.relu(tf.matmul(h1, w2) + b2), keep_prob)
w3 = tf.Variable(tf.truncated_normal([h1_size, 1], stddev=0.1), name="d_w3", dtype=tf.float32)
b3 = tf.Variable(tf.zeros([1]), name="d_b3", dtype=tf.float32)
h3 = tf.matmul(h2, w3) + b3
y_data = tf.nn.sigmoid(tf.slice(h3, [0, 0], [batch_size, -1], name=None))
y_generated = tf.nn.sigmoid(tf.slice(h3, [batch_size, 0], [-1, -1], name=None))
d_params = [w1, b1, w2, b2, w3, b3] return y_data, y_generated, d_params#def show_result(batch_res, fname, grid_size=(8, 8), grid_pad=5):
batch_res = 0.5 * batch_res.reshape((batch_res.shape[0], img_height, img_width)) + 0.5
img_h, img_w = batch_res.shape[1], batch_res.shape[2]
grid_h = img_h * grid_size[0] + grid_pad * (grid_size[0] - 1)
grid_w = img_w * grid_size[1] + grid_pad * (grid_size[1] - 1)
img_grid = np.zeros((grid_h, grid_w), dtype=np.uint8) for i, res in enumerate(batch_res): if i >= grid_size[0] * grid_size[1]: break
img = (res) * 255
img = img.astype(np.uint8)
row = (i // grid_size[0]) * (img_h + grid_pad)
col = (i % grid_size[1]) * (img_w + grid_pad)
img_grid[row:row + img_h, col:col + img_w] = img
imsave(fname, img_grid)def train(): # load data（mnist手写数据集）
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
x_data = tf.placeholder(tf.float32, [batch_size, img_size], name="x_data")
z_prior = tf.placeholder(tf.float32, [batch_size, z_size], name="z_prior")
keep_prob = tf.placeholder(tf.float32, name="keep_prob")
global_step = tf.Variable(0, name="global_step", trainable=False) # 创建生成模型
x_generated, g_params = build_generator(z_prior) # 创建判别模型
y_data, y_generated, d_params = build_discriminator(x_data, x_generated, keep_prob) # 损失函数的设置
d_loss = - (tf.log(y_data) + tf.log(1 - y_generated))
g_loss = - tf.log(y_generated)
optimizer = tf.train.AdamOptimizer(0.0001) # 两个模型的优化函数
d_trainer = optimizer.minimize(d_loss, var_list=d_params)
g_trainer = optimizer.minimize(g_loss, var_list=g_params)
saver = tf.train.Saver() # 启动默认图
init = tf.global_variables_initializer()
sess = tf.Session() # 初始化
sess.run(init) if to_restore:
chkpt_fname = tf.train.latest_checkpoint(output_path)
saver.restore(sess, chkpt_fname) else: if os.path.exists(output_path):
shutil.rmtree(output_path)
z_sample_val = np.random.normal(0, 1, size=(batch_size, z_size)).astype(np.float32)
os.mkdir(output_path)
steps = 60000 / batch_size for i in range(sess.run(global_step), max_epoch): for j in np.arange(steps): # for j in range(steps):
print("epoch:%s, iter:%s" % (i, j)) # 每一步迭代，我们都会加载256个训练样本，然后执行一次train_step
x_value, _ = mnist.train.next_batch(batch_size)
x_value = 2 * x_value.astype(np.float32) - 1
z_value = np.random.normal(0, 1, size=(batch_size, z_size)).astype(np.float32) # 执行生成
sess.run(d_trainer,
feed_dict={x_data: x_value, z_prior: z_value, keep_prob: np.sum(0.7).astype(np.float32)}) # 执行判别 if j % 1 == 0:
sess.run(g_trainer,
feed_dict={x_data: x_value, z_prior: z_value, keep_prob: np.sum(0.7).astype(np.float32)})
x_gen_val = sess.run(x_generated, feed_dict={z_prior: z_sample_val})
show_result(x_gen_val, "output/sample{0}.jpg".format(i))
z_random_sample_val = np.random.normal(0, 1, size=(batch_size, z_size)).astype(np.float32)
x_gen_val = sess.run(x_generated, feed_dict={z_prior: z_random_sample_val})
show_result(x_gen_val, "output/random_sample{0}.jpg".format(i))
sess.run(tf.assign(global_step, i + 1))
saver.save(sess, os.path.join(output_path, "model"), global_step=global_step)if __name__ == '__main__':
train()
原文发布时间为：2018-08-23
本文作者：文文
本文来自云栖社区合作伙伴“
Python爱好者社区”，了解相关信息可以关注“
Python爱好者社区”。
