带有自定义 RNN 单元的 keras 双向层

Question

下面的模型

lstm_model = Sequential()
lstm_model.add(embedding)
tensor_lstm_cell = TensorLSTMCell(hidden_size=lstm_size, num_units=4)
lstm_model.add(Bidirectional(RNN(tensor_lstm_cell, return_sequences=True)))

抛出以下错误:ValueError: Unknown layer: TensorLSTMCell，它似乎来自从config双向加载它。我想知道如何使用 model.add 功能将自定义 rnn 层添加到双向包装器

Answer 1

您可以使用 CustomObjectScope 包裹 Bidirectional 行，以便它可以识别您的自定义对象 TensorLSTMCell。例如，

from keras.utils.generic_utils import CustomObjectScope

class DummyLSTMCell(LSTMCell):
    pass

embedding = Embedding(10000, 32, input_shape=(None,))

lstm_model = Sequential()
lstm_model.add(embedding)
lstm_cell = DummyLSTMCell(32)
with CustomObjectScope({'DummyLSTMCell': DummyLSTMCell}):
    lstm_model.add(Bidirectional(RNN(lstm_cell, return_sequences=True)))