雖然暫時還沒用到對抗生成網路，但是看過GAN以及WGAN、IRGAN後覺得非常有意思，將博弈論的思路放到了神經網路裡面來。主要的思路就是一個生成器與一個分類器，分類器的作用是區分這個資料來源是真實的還是生成器產生的，生成器的作用就是產生新的資料儘可能矇混分類器不讓區分開，在訓練過程中交替訓練分類器與生成器讓兩者在競爭中不斷進化提高自身的效能，訓練完成後分類器和生成器都可以使用，下面邊上程式碼邊講。

img_height = 28
img_width = 28
img_size = img_height * img_width

to_train = True
to_restore = False
output_path = "output"

max_epoch = 500

h1_size = 150
h2_size = 300
z_size = 100
batch_size = 256
 

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
 

build_generator定義了生成器，輸入引數為一個長度為100的先驗向量，然後通過三個全連線層（其實這裡可以使用任何形式不一定是全連線層，變種中有ＣＮＮ和ＬＳＴＭ的）對映到長度為784的向量也就是影象扁平化後的長度，返回了對映中用到的引數

def build_discriminator(x_data, x_generated, keep_prob):
    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 train():
    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)

    x_generated, g_params = build_generator(z_prior)
    y_data, y_generated, d_params = build_discriminator(x_data, x_generated, keep_prob)

    d_loss = - (tf.log(y_data) + tf.log(1 - y_generated))
    g_loss = - tf.log(y_generated)

    optimizer = tf.train.AdamOptimizer(0.0001)

    d_trainer = optimizer.minimize(d_loss, var_list=d_params)
    g_trainer = optimizer.minimize(g_loss, var_list=g_params)

    init = tf.initialize_all_variables()

    saver = tf.train.Saver()

    sess = tf.Session()

    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)
        os.mkdir(output_path)

    z_sample_val = np.random.normal(0, 1, size=(batch_size, z_size)).astype(np.float32)

    for i in range(sess.run(global_step), max_epoch):
        for j in range(60000 / batch_size):
            print "epoch:%s, iter:%s" % (i, j)
            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)
            sess.run(d_trainer,
                     feed_dict={x_data: x_value, z_prior: z_value, keep_prob: np.sum(0.7).astype(np.float32)})
            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, os.path.join(output_path, "sample%s.jpg" % 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, os.path.join(output_path, "random_sample%s.jpg" % i))
        sess.run(tf.assign(global_step, i + 1))
        saver.save(sess, os.path.join(output_path, "model"), global_step=global_step)

train函式中首先載入資料，然後定義了佔位符，接下來定義了交替訓練的損失函式，g_loss是生成器的損失函式，計算的就是生產資料的交叉熵，d_loss是整個資料的交叉熵，因為分類器要保證在所有資料上都能很好的區分，所以損失函式中包含了所有資料。之後做了一些儲存的資料夾操作，接著從0-1均勻分佈中抽取了ｚ（至於為什麼用這個分佈，可以去檢視一個概率論，幾乎所有重要的概率分佈都可以從均勻分佈Uniform(0,1)中生成出來），接著就是交替訓練以及生產一個訓練好的生成器生成的圖片了。

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)

GAN合集