1、結構圖
2、知識點
生成器(G)：將噪音數據生成一個想要的數據
判別器(D)：將生成器的結果進行判別，
3、代碼及案例
# coding: utf-8 # ## 對抗生成網絡案例 ## # # # <img src="jpg/3.png" alt="FAO" width="590" > # - 判別器 : 火眼金睛，分辨出生成和真實的 <br /> # <br /> # - 生成器 : 瞞天過海，騙過判別器 <br /> # <br /> # - 損失函數定義 : 一方面要讓判別器分辨能力更強，另一方面要讓生成器更真 <br /> # <br /> # # <img src="jpg/1.jpg" alt="FAO" width="590" > # In[1]: import tensorflow as tf import numpy as np import pickle import matplotlib.pyplot as plt get_ipython().run_line_magic('matplotlib', 'inline') # # 導入數據 # In[2]: from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets('/data') # ## 網絡架構 # # ### 輸入層 ：待生成圖像（噪音）和真實數據 # # ### 生成網絡：將噪音圖像進行生成 # # ### 判別網絡： # - （1）判斷真實圖像輸出結果 # - （2）判斷生成圖像輸出結果 # # ### 目標函數： # - （1）對於生成網絡要使得生成結果通過判別網絡為真 # - （2）對於判別網絡要使得輸入為真實圖像時判別為真 輸入為生成圖像時判別為假 # # <img src="jpg/2.png" alt="FAO" width="590" > # ## Inputs # In[3]: #真實數據和噪音數據 def get_inputs(real_size, noise_size): real_img = tf.placeholder(tf.float32, [None, real_size]) noise_img = tf.placeholder(tf.float32, [None, noise_size]) return real_img, noise_img # ## 生成器 # * noise_img: 產生的噪音輸入 # * n_units: 隱層單元個數 # * out_dim: 輸出的大小（28 * 28 * 1） # In[4]: def get_generator(noise_img, n_units, out_dim, reuse=False, alpha=0.01): with tf.variable_scope("generator", reuse=reuse): # hidden layer hidden1 = tf.layers.dense(noise_img, n_units) # leaky ReLU hidden1 = tf.maximum(alpha * hidden1, hidden1) # dropout hidden1 = tf.layers.dropout(hidden1, rate=0.2) # logits & outputs logits = tf.layers.dense(hidden1, out_dim) outputs = tf.tanh(logits) return logits, outputs # ## 判別器 # * img：輸入 # * n_units：隱層單元數量 # * reuse：由於要使用兩次 # In[5]: def get_discriminator(img, n_units, reuse=False, alpha=0.01): with tf.variable_scope("discriminator", reuse=reuse): # hidden layer hidden1 = tf.layers.dense(img, n_units) hidden1 = tf.maximum(alpha * hidden1, hidden1) # logits & outputs logits = tf.layers.dense(hidden1, 1) outputs = tf.sigmoid(logits) return logits, outputs # ## 網絡參數定義 # * img_size：輸入大小 # * noise_size：噪音圖像大小 # * g_units：生成器隱層參數 # * d_units：判別器隱層參數 # * learning_rate：學習率 # In[6]: img_size = mnist.train.images[0].shape[0] noise_size = 100 g_units = 128 d_units = 128 learning_rate = 0.001 alpha = 0.01 # ## 構建網絡 # In[7]: tf.reset_default_graph() real_img, noise_img = get_inputs(img_size, noise_size) # generator g_logits, g_outputs = get_generator(noise_img, g_units, img_size) # 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) # ### 目標函數： # - （1）對於生成網絡要使得生成結果通過判別網絡為真 # - （2）對於判別網絡要使得輸入為真實圖像時判別為真 輸入為生成圖像時判別為假 # # <img src="jpg/2.png" alt="FAO" width="590" > # In[8]: # discriminator的loss # 識別真實圖片 d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_real, labels=tf.ones_like(d_logits_real))) # 識別生成的圖片 d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, labels=tf.zeros_like(d_logits_fake))) # 總體loss d_loss = tf.add(d_loss_real, d_loss_fake) # generator的loss g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=d_logits_fake, labels=tf.ones_like(d_logits_fake))) # ## 優化器 # In[9]: train_vars = tf.trainable_variables() # generator g_vars = [var for var in train_vars if var.name.startswith("generator")] # discriminator d_vars = [var for var in train_vars if var.name.startswith("discriminator")] # optimizer d_train_opt = tf.train.AdamOptimizer(learning_rate).minimize(d_loss, var_list=d_vars) g_train_opt = tf.train.AdamOptimizer(learning_rate).minimize(g_loss, var_list=g_vars) # # 訓練 # In[10]: # batch_size batch_size = 64 # 訓練迭代輪數 epochs = 300 # 抽取樣本數 n_sample = 25 # 存儲測試樣例 samples = [] # 存儲loss losses = [] # 保存生成器變量 saver = tf.train.Saver(var_list = g_vars) # 開始訓練 with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for e in range(epochs): for batch_i in range(mnist.train.num_examples//batch_size): batch = mnist.train.next_batch(batch_size) batch_images = batch[0].reshape((batch_size, 784)) # 對圖像像素進行scale，這是因為tanh輸出的結果介於(-1,1),real和fake圖片共享discriminator的參數 batch_images = batch_images*2 - 1 # generator的輸入噪聲 batch_noise = np.random.uniform(-1, 1, size=(batch_size, noise_size)) # Run optimizers _ = sess.run(d_train_opt, feed_dict={real_img: batch_images, noise_img: batch_noise}) _ = sess.run(g_train_opt, feed_dict={noise_img: batch_noise}) # 每一輪結束計算loss train_loss_d = sess.run(d_loss, feed_dict = {real_img: batch_images, noise_img: batch_noise}) # real img loss train_loss_d_real = sess.run(d_loss_real, feed_dict = {real_img: batch_images, noise_img: batch_noise}) # fake img loss train_loss_d_fake = sess.run(d_loss_fake, feed_dict = {real_img: batch_images, noise_img: batch_noise}) # generator loss train_loss_g = sess.run(g_loss, feed_dict = {noise_img: batch_noise}) print("Epoch {}/{}...".format(e+1, epochs), "判別器損失: {:.4f}(判別真實的: {:.4f} + 判別生成的: {:.4f})...".format(train_loss_d, train_loss_d_real, train_loss_d_fake), "生成器損失: {:.4f}".format(train_loss_g)) losses.append((train_loss_d, train_loss_d_real, train_loss_d_fake, train_loss_g)) # 保存樣本 sample_noise = np.random.uniform(-1, 1, size=(n_sample, noise_size)) gen_samples = sess.run(get_generator(noise_img, g_units, img_size, reuse=True), feed_dict={noise_img: sample_noise}) samples.append(gen_samples) saver.save(sess, './checkpoints/generator.ckpt') # 保存到本地 with open('train_samples.pkl', 'wb') as f: pickle.dump(samples, f) # # loss迭代曲線 # In[11]: fig, ax = plt.subplots(figsize=(20,7)) losses = np.array(losses) plt.plot(losses.T[0], label='判別器總損失') plt.plot(losses.T[1], label='判別真實損失') plt.plot(losses.T[2], label='判別生成損失') plt.plot(losses.T[3], label='生成器損失') plt.title("對抗生成網絡") ax.set_xlabel('epoch') plt.legend() # # 生成結果 # In[12]: # Load samples from generator taken while training with open('train_samples.pkl', 'rb') as f: samples = pickle.load(f) # In[13]: #samples是保存的結果 epoch是第多少次迭代 def view_samples(epoch, samples): fig, axes = plt.subplots(figsize=(7,7), nrows=5, ncols=5, sharey=True, sharex=True) for ax, img in zip(axes.flatten(), samples[epoch][1]): # 這里samples[epoch][1]代表生成的圖像結果，而[0]代表對應的logits ax.xaxis.set_visible(False) ax.yaxis.set_visible(False) im = ax.imshow(img.reshape((28,28)), cmap='Greys_r') return fig, axes # In[14]: _ = view_samples(-1, samples) # 顯示最終的生成結果 # # 顯示整個生成過程圖片 # In[15]: # 指定要查看的輪次 epoch_idx = [10, 30, 60, 90, 120, 150, 180, 210, 240, 290] show_imgs = [] for i in epoch_idx: show_imgs.append(samples[i][1]) # In[16]: # 指定圖片形狀 rows, cols = 10, 25 fig, axes = plt.subplots(figsize=(30,12), nrows=rows, ncols=cols, sharex=True, sharey=True) idx = range(0, epochs, int(epochs/rows)) for sample, ax_row in zip(show_imgs, axes): for img, ax in zip(sample[::int(len(sample)/cols)], ax_row): ax.imshow(img.reshape((28,28)), cmap='Greys_r') ax.xaxis.set_visible(False) ax.yaxis.set_visible(False) # # 生成新的圖片 # In[17]: # 加載我們的生成器變量 saver = tf.train.Saver(var_list=g_vars) with tf.Session() as sess: saver.restore(sess, tf.train.latest_checkpoint('checkpoints')) sample_noise = np.random.uniform(-1, 1, size=(25, noise_size)) gen_samples = sess.run(get_generator(noise_img, g_units, img_size, reuse=True), feed_dict={noise_img: sample_noise}) # In[18]: _ = view_samples(0, [gen_samples])
4、優化目標
