python - 在 Keras 中加载保存的模型(双向 LSTM)

Question

我在 Keras 中成功训练并保存了双向 LSTM 模型:

model = Sequential()
model.add(Bidirectional(LSTM(N_HIDDEN_NEURONS,
                        return_sequences=True,
                        activation="tanh",
                        input_shape=(SEGMENT_TIME_SIZE, N_FEATURES))))
model.add(Bidirectional(LSTM(N_HIDDEN_NEURONS)))
model.add(Dropout(0.5))
model.add(Dense(N_CLASSES, activation='sigmoid'))
model.compile('adam', 'binary_crossentropy', metrics=['accuracy'])

model.fit(X_train, y_train,
          batch_size=BATCH_SIZE,
          epochs=N_EPOCHS,
          validation_data=[X_test, y_test])

model.save('model_keras/model.h5')

但是，当我想加载它时:

model = load_model('model_keras/model.h5')

我收到错误:

ValueError: You are trying to load a weight file containing 3 layers into a model with 0 layers.

我还尝试了不同的方法，例如分别保存和加载模型架构和权重，但它们都不适合我。另外，之前，当我使用普通(单向)LSTM 时，加载模型效果很好。

Answer 1

正如@mpariente所述和 @today ，input_shape 是双向的参数，而不是 LSTM，请参阅 Keras documentation 。我的解决方案:

# Model
model = Sequential()
model.add(Bidirectional(LSTM(N_HIDDEN_NEURONS,
                             return_sequences=True,
                             activation="tanh"), 
                        input_shape=(SEGMENT_TIME_SIZE, N_FEATURES)))
model.add(Bidirectional(LSTM(N_HIDDEN_NEURONS)))
model.add(Dropout(0.5))
model.add(Dense(N_CLASSES, activation='sigmoid'))
model.compile('adam', 'binary_crossentropy', metrics=['accuracy'])

model.fit(X_train, y_train,
          batch_size=BATCH_SIZE,
          epochs=N_EPOCHS,
          validation_data=[X_test, y_test])

model.save('model_keras/model.h5')

然后，要加载，只需执行以下操作:

model = load_model('model_keras/model.h5')