コード
- モデル定義
- モデルの訓練

バッチサイズを狭めながらオートエンコーダを訓練して、その損失関数の減少具合を確認する。緑色線がエポックごとのおおよその損失関数のlossの値、赤色がバッチごとのlossの値を全てプロットしたものです。

コード

モデル定義

from keras.layers import Input, Dense
from keras.models import Model
from sklearn.preprocessing import PolynomialFeatures
from keras.callbacks import Callback
class LossHistory(Callback):
    def on_train_begin(self, logs={}):
        self.losses = []

    def on_batch_end(self, batch, logs={}):
        self.losses.append(logs.get('loss'))

poly = PolynomialFeatures(2).fit(train_X)
train_pX, test_pX = poly.transform(train_X), poly.transform(test_X)

feature = train_pX.shape[1]
input_v = Input(shape=(feature,))
emd_dim = 50

encoded = Dense(int(feature/2), activation='relu')(input_v)
encoded = Dense(int(feature/4), activation='relu')(encoded)
encoded = Dense(emd_dim, activation='relu')(encoded)

decoded = Dense(int(feature/4), activation='relu')(encoded)
decoded = Dense(int(feature/2), activation='relu')(decoded)
decoded = Dense(feature, activation='sigmoid')(decoded)

autoencoder = Model(input=input_v, output=decoded)
autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy')
print(autoencoder.summary())

出力結果は

____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
input_6 (InputLayer)             (None, 253)           0                                            
____________________________________________________________________________________________________
dense_25 (Dense)                 (None, 126)           32004       input_6[0][0]                    
____________________________________________________________________________________________________
dense_26 (Dense)                 (None, 63)            8001        dense_25[0][0]                   
____________________________________________________________________________________________________
dense_27 (Dense)                 (None, 50)            3200        dense_26[0][0]                   
____________________________________________________________________________________________________
dense_28 (Dense)                 (None, 63)            3213        dense_27[0][0]                   
____________________________________________________________________________________________________
dense_29 (Dense)                 (None, 126)           8064        dense_28[0][0]                   
____________________________________________________________________________________________________
dense_30 (Dense)                 (None, 253)           32131       dense_29[0][0]                   
====================================================================================================
Total params: 86613

モデルの訓練

# train
plotdata = LossHistory()
total_step = 10
batch_size = 200
loss_history, avgloss_history, index_history = [], [], []

for i in range(total_step):
    # fit model
    autoencoder.fit(train_pX, train_pX,
                    nb_epoch=10,
                    batch_size=int(batch_size*(total_step-i)/total_step),
                    shuffle=True,
                    validation_data=(test_pX, test_pX),
                    verbose=2,
                    callbacks=[plotdata])
    
    index_history += [i]
    loss_history += list(plotdata.losses)
    avgloss_history += [sum(loss_history[-100:])/100.0]
    
    # plot loss
    plt.plot(loss_history)
    
    # show epoch num
    for j in index_history:
        plt.axvline(x=j, alpha=.5, color='k')
    
    # show avarage loss
    for j in avgloss_history:
        plt.axhline(y=j, alpha=.5, color='g')
    plt.yticks(avgloss_history, np.arange(i))
    
    # show
    plt.show()
    
    # save weight
    autoencoder.save_weights('../param/autoencoder_a.w')