Autoencoderで訓練済みモデルから中間層encoderの出力方法

後で利用したい/出力したいレイヤーに名前を付ける。

サンプルコード

from keras.utils import plot_model
from keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D
from keras.models import Model,load_model

#変数定義
x_train = #(省略)
x_test = #(省略)
epoch = 50
batch_size = 100
model = define_model()

#モデル訓練
model.fit(x=x_train,
          y=x_train,
          epochs=epoch,
          batch_size=batch_size,
          shuffle=True
)

#モデル保存
model.save('aaa.h5')

#モデルロード
model = load_model('aaa.h5')

#中間層(encoder)出力
encoder = Model(inputs=model.input, outputs=model.get_layer('encoder_layer').output)

#評価
result = encoder.predict(x_test)

# CNN Autoencoderの定義
def define_model():
    input_img = Input(shape=(3, 256, 256))

    # 3x256x256 -> 8x128x128
    encoded = Conv2D(8, (3, 3), activation='relu', padding='same', data_format='channels_first')(input_img)
    encoded = MaxPooling2D(pool_size=(2, 2), padding='same', data_format='channels_first')(encoded)

    # 8x128x128 -> 8x64x64
    #(省略)

    # 8x64x64 -> 8x32x32
    #(省略)

    # 8x64x64 -> 8x16x16
    encoded = Conv2D(8, (3, 3), activation='relu', padding='same', data_format='channels_first')(encoded)
    # ★この層を後ほど出力したいので、nameを追加★
    encoded = MaxPooling2D(pool_size=(2, 2), padding='same', data_format='channels_first', name='encoder_layer')(encoded)

    # 8x16x16 -> 8x32x32
    #(省略)

    # 8x32x32 -> 8x64x64
    #(省略)

    # 8x64x64 -> 8x128x128
    #(省略)

    # 8x128x128 -> 8x256x256
    decoded = Conv2D(8, (3, 3), activation='relu', padding='same', data_format='channels_first')(decoded)
    decoded = UpSampling2D(size=(2, 2), data_format='channels_first')(decoded)

    # 8x256x256 -> 3x256x256
    decoded = Conv2D(8, (3, 3), activation='sigmoid', padding='same', data_format='channels_first')(decoded)

    # モデルの定義
    model = Model(inputs=input_img, outputs=decoded)
    model.compile(optimizer='adam', loss='binary_crossentropy')

    return model

参照：
https://keras.io/ja/getting-started/faq/#keras-model
http://nobunaga.hatenablog.jp/entry/2017/10/18/000827