GAN and CGAN on MNIST dataset

GAN and CGAN

关于GAN和CGAN的教程网上有好多，感兴趣的可以自己去找着看看。最重要的是要弄清楚，GAN是干嘛的？作者在论文和NIPS2016 tutorial里面都讲了很多设计GAN的初衷。简单来说，GAN就是用来拟合样本的分布的。

看了论文总觉得理解得不够透彻，就在网上找了一些简单的程序跑了跑，下面就直接来说说程序和实验结果吧。

---

GAN

GAN的结构如下图所示。

实验数据集采用了MNIST手写数字图像。代码参考github上的例子，使用TensorFlow1.2搭建了一个神经网络。准确来说是两个网络，判别器Discriminator和生成器Generator各是一个神经网络。两个网络的结构类似，都有两个全连接层，最后一个全连接层接输出的tanh或者sigmoid层。

code

代码参考了别人的，原作者是在ipython notebook里面写的代码，然后自己画图。我是使用的Tensorboard来进行可视化操作的，所以我把代码进行了相应的修改。下面只显示了主要的程序部分。请看完整代码，顺便说一句，这里还可以看到无水印的原图～

# reference https://github.com/NELSONZHAO/zhihu/tree/master/mnist_gan

def get_generator(noise_img, n_units, out_dim, reuse = False, alpha = 0.01):
    """
    generator

    noise_img: input of generator
    n_units: # hidden units
    out_dim: # output
    alpha: parameter of leaky ReLU
    """
    with tf.variable_scope("generator", reuse = reuse):
        # hidden layer
        hidden = tf.layers.dense(noise_img, n_units)
        # leaky ReLU
        relu = tf.maximum(alpha * hidden, hidden)
        # dropout
        drop = tf.layers.dropout(relu, rate = 0.5)

        # logits & outputs
        logits = tf.layers.dense(drop, out_dim)
        outputs = tf.tanh(logits)

        return logits, outputs


def get_discriminator(img, n_units, reuse = False, alpha = 0.01):
    """
    discriminator

    n_units: # hidden units
    alpha: parameter of leaky Relu
    """
    with tf.variable_scope("discriminator", reuse=reuse):
        # hidden layer
        hidden = tf.layers.dense(img, n_units)
        relu = tf.maximum(alpha * hidden, hidden)
        # dropout
        drop = tf.layers.dropout(relu, rate = 0.5)

        # logits & outputs
        logits = tf.layers.dense(drop, 1)
        outputs = tf.sigmoid(logits)

        return logits, outputs

...
with tf.Graph().as_default():

    real_img, noise_img = get_inputs(real_img_size, noise_img_size)

    # generator
    g_logits, g_outputs = get_generator(noise_img, g_units, real_img_size)

    sample_images = tf.reshape(g_outputs, [-1, 28, 28, 1])
    tf.summary.image("sample_images", sample_images, 10)

    # discriminator
    d_logits_real, d_outputs_real = get_discriminator(real_img, d_units)
    d_logits_fake, d_outputs_fake = get_discriminator(g_outputs, d_units, reuse = True)


    # discriminator loss
    d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = d_logits_real, 
        labels = tf.ones_like(d_logits_real)) * (1 - smooth))
    d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = d_logits_fake, 
        labels = tf.zeros_like(d_logits_fake)))
    # loss
    d_loss = tf.add(d_loss_real, d_loss_fake)

    # generator loss
    g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = d_logits_fake,
        labels = tf.ones_like(d_logits_fake)) * (1 - smooth) )

    ...

实验结果

我使用了TF的tensorboard来进行可视化，通过tensorboard可以看到生成的网络的结构。同时我也保存了Generator在训练的过程中生成的样本图片，还有Generator和Discriminator的Loss。

Graph

Loss

把最后30个epoch的Loss值显示出来

('Epoch 271/300', 'Discriminator loss: 0.8802(Real: 0.5291 + Fake: 0.3511)', 'Generator loss: 2.0067')
('Epoch 272/300', 'Discriminator loss: 0.8302(Real: 0.4576 + Fake: 0.3727)', 'Generator loss: 1.8738')
('Epoch 273/300', 'Discriminator loss: 0.7560(Real: 0.3527 + Fake: 0.4033)', 'Generator loss: 1.8327')
('Epoch 274/300', 'Discriminator loss