在 /home/your_name/TensorFlow/DCGAN/ 下新建文件 utils.py,輸入以下代碼:html
import scipy.misc import numpy as np # 保存圖片函數 def save_images(images, size, path): """ Save the samples images The best size number is int(max(sqrt(image.shape[0]),sqrt(image.shape[1]))) + 1 example: The batch_size is 64, then the size is recommended [8, 8] The batch_size is 32, then the size is recommended [6, 6] """ # 圖片歸一化,主要用於生成器輸出是 tanh 形式的歸一化 img = (images + 1.0) / 2.0 h, w = img.shape[1], img.shape[2] # 產生一個大畫布,用來保存生成的 batch_size 個圖像 merge_img = np.zeros((h * size[0], w * size[1], 3)) # 循環使得畫布特定地方值爲某一幅圖像的值 for idx, image in enumerate(images): i = idx % size[1] j = idx // size[1] merge_img[j*h:j*h+h, i*w:i*w+w, :] = image # 保存畫布 return scipy.misc.imsave(path, merge_img)
這個函數的做用是在訓練的過程當中保存採樣生成的圖片。git
在 /home/your_name/TensorFlow/DCGAN/ 下新建文件 model.py,定義生成器,判別器和訓練過程當中的採樣網絡,在 model.py 輸入以下代碼:github
import tensorflow as tf from ops import * BATCH_SIZE = 64 # 定義生成器 def generator(z, y, train = True): # y 是一個 [BATCH_SIZE, 10] 維的向量,把 y 轉成四維張量 yb = tf.reshape(y, [BATCH_SIZE, 1, 1, 10], name = 'yb') # 把 y 做爲約束條件和 z 拼接起來 z = tf.concat(1, [z, y], name = 'z_concat_y') # 通過一個全鏈接,BN 和激活層 ReLu h1 = tf.nn.relu(batch_norm_layer(fully_connected(z, 1024, 'g_fully_connected1'), is_train = train, name = 'g_bn1')) # 把約束條件和上一層拼接起來 h1 = tf.concat(1, [h1, y], name = 'active1_concat_y') h2 = tf.nn.relu(batch_norm_layer(fully_connected(h1, 128 * 49, 'g_fully_connected2'), is_train = train, name = 'g_bn2')) h2 = tf.reshape(h2, [64, 7, 7, 128], name = 'h2_reshape') # 把約束條件和上一層拼接起來 h2 = conv_cond_concat(h2, yb, name = 'active2_concat_y') h3 = tf.nn.relu(batch_norm_layer(deconv2d(h2, [64,14,14,128], name = 'g_deconv2d3'), is_train = train, name = 'g_bn3')) h3 = conv_cond_concat(h3, yb, name = 'active3_concat_y') # 通過一個 sigmoid 函數把值歸一化爲 0~1 之間, h4 = tf.nn.sigmoid(deconv2d(h3, [64, 28, 28, 1], name = 'g_deconv2d4'), name = 'generate_image') return h4 # 定義判別器 def discriminator(image, y, reuse = False): # 由於真實數據和生成數據都要通過判別器,因此須要指定 reuse 是否可用 if reuse: tf.get_variable_scope().reuse_variables() # 同生成器同樣,判別器也須要把約束條件串聯進來 yb = tf.reshape(y, [BATCH_SIZE, 1, 1, 10], name = 'yb') x = conv_cond_concat(image, yb, name = 'image_concat_y') # 卷積,激活,串聯條件。 h1 = lrelu(conv2d(x, 11, name = 'd_conv2d1'), name = 'lrelu1') h1 = conv_cond_concat(h1, yb, name = 'h1_concat_yb') h2 = lrelu(batch_norm_layer(conv2d(h1, 74, name = 'd_conv2d2'), name = 'd_bn2'), name = 'lrelu2') h2 = tf.reshape(h2, [BATCH_SIZE, -1], name = 'reshape_lrelu2_to_2d') h2 = tf.concat(1, [h2, y], name = 'lrelu2_concat_y') h3 = lrelu(batch_norm_layer(fully_connected(h2, 1024, name = 'd_fully_connected3'), name = 'd_bn3'), name = 'lrelu3') h3 = tf.concat(1,[h3, y], name = 'lrelu3_concat_y') # 全鏈接層,輸出覺得 loss 值 h4 = fully_connected(h3, 1, name = 'd_result_withouts_sigmoid') return tf.nn.sigmoid(h4, name = 'discriminator_result_with_sigmoid'), h4 # 定義訓練過程當中的採樣函數 def sampler(z, y, train = True): tf.get_variable_scope().reuse_variables() return generator(z, y, train = train)
能夠看到,生成器由 7 × 7 變爲 14 × 14 再變爲 28 × 28大小,每一層都加入了約束條件 y,完美的詮釋了論文所給出的網絡,之因此要加入 is_train 參數,是因爲 Batch_norm 層中訓練和測試的時候的過程是不一樣的,用這個參數區分訓練和測試,生成器的最後一層,用了一個 sigmoid 函數把值歸一化到 0~1 之間,若是是不加約束的網絡,則用 tanh 函數,因此在 save_images 函數中要用到語句:img = (images + 1.0) / 2.0。網絡
sampler 函數的做用是在訓練過程當中對生成器生成的圖片進行採樣,因此這個函數必須指定 reuse 可用,關於 reuse 說明,請看:http://www.cnblogs.com/Charles-Wan/p/6200446.html。函數
參考資料:測試
1. https://github.com/carpedm20/DCGAN-tensorflowspa