# Keras 中的變分自編碼器 在 Keras 中，構建變分自編碼器更容易，并且代碼行更少。 Keras 變分自編碼器最好使用函數式風格構建。到目前為止，我們已經使用了在 Keras 中構建模型的順序樣式，現在在這個例子中，我們將看到在 Keras 中構建 VAE 模型的函數式風格。在 Keras 建立 VAE 的步驟如下： 1. 定義隱藏層和潛在變量層中的超參數和神經元數量： ```py import keras from keras.layers import Lambda, Dense, Input, Layer from keras.models import Model from keras import backend as K learning_rate = 0.001 batch_size = 100 n_batches = int(mnist.train.num_examples/batch_size) # number of pixels in the MNIST image as number of inputs n_inputs = 784 n_outputs = n_inputs # number of hidden layers n_layers = 2 # neurons in each hidden layer n_neurons = [512,256] # the dimensions of latent variables n_neurons_z = 128 ``` 1. 構建輸入層： ```py x = Input(shape=(n_inputs,), name='input') ``` 1. 構建編碼器層，以及潛在變量的均值和方差層： ```py # build encoder layer = x for i in range(n_layers): layer = Dense(units=n_neurons[i], activation='relu',name='enc_{0}'.format(i))(layer) z_mean = Dense(units=n_neurons_z,name='z_mean')(layer) z_log_var = Dense(units=n_neurons_z,name='z_log_v')(layer) ``` 1. 創建噪聲和后驗分布： ```py # noise distribution epsilon = K.random_normal(shape=K.shape(z_log_var), mean=0,stddev=1.0) # posterior distribution z = Lambda(lambda zargs: zargs[0] + K.exp(zargs[1] * 0.5) * epsilon, name='z')([z_mean,z_log_var]) ``` 1. 添加解碼器層： ```py # add generator / probablistic decoder network layers layer = z for i in range(n_layers-1,-1,-1): layer = Dense(units=n_neurons[i], activation='relu', name='dec_{0}'.format(i))(layer) ``` 1. 定義最終輸出層： ```py y_hat = Dense(units=n_outputs, activation='sigmoid', name='output')(layer) ``` 1. 最后，從輸入層和輸出層定義模型并顯示模型摘要： ```py model = Model(x,y_hat) model.summary() ``` 我們看到以下摘要： ```py _________________________________________________________________________ Layer (type) Output Shape Param # Connected to ========================================================================= input (InputLayer) (None, 784) 0 _________________________________________________________________________ enc_0 (Dense) (None, 512) 401920 input[0][0] _________________________________________________________________________ enc_1 (Dense) (None, 256) 131328 enc_0[0][0] _________________________________________________________________________ z_mean (Dense) (None, 128) 32896 enc_1[0][0] _________________________________________________________________________ z_log_v (Dense) (None, 128) 32896 enc_1[0][0] _________________________________________________________________________ z (Lambda) (None, 128) 0 z_mean[0][0] z_log_v[0][0] _________________________________________________________________________ dec_1 (Dense) (None, 256) 33024 z[0][0] _________________________________________________________________________ dec_0 (Dense) (None, 512) 131584 dec_1[0][0] _________________________________________________________________________ output (Dense) (None, 784) 402192 dec_0[0][0] ========================================================================= Total params: 1,165,840 Trainable params: 1,165,840 Non-trainable params: 0 _________________________________________________________________________ ``` 1. 定義一個計算重建和正則化損失之和的函數： ```py def vae_loss(y, y_hat): rec_loss = -K.sum(y * K.log(1e-10 + y_hat) + (1-y) * K.log(1e-10 + 1 - y_hat), axis=-1) reg_loss = -0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1) loss = K.mean(rec_loss+reg_loss) return loss ``` 1. 使用此損失函數來編譯模型： ```py model.compile(loss=vae_loss, optimizer=keras.optimizers.Adam(lr=learning_rate)) ``` 1. 讓我們訓練 50 個周期的模型并預測圖像，正如我們在前面的部分中所做的那樣： ```py n_epochs=50 model.fit(x=X_train_noisy,y=X_train,batch_size=batch_size, epochs=n_epochs,verbose=0) Y_test_pred1 = model.predict(test_images) Y_test_pred2 = model.predict(test_images_noisy) ``` 讓我們顯示結果圖像： ```py display_images(test_images.reshape(-1,pixel_size,pixel_size),test_labels) display_images(Y_test_pred1.reshape(-1,pixel_size,pixel_size),test_labels) ``` 我們得到如下結果：  ```py display_images(test_images_noisy.reshape(-1,pixel_size,pixel_size), test_labels) display_images(Y_test_pred2.reshape(-1,pixel_size,pixel_size),test_labels) ``` 我們得到以下結果：  這很棒！！生成的圖像更清晰，更清晰。