Keras seq2seq 模型如何在训练时从验证中屏蔽填充零？

Question

我正在根据这个很棒的教程开展一个项目。 https://machinelearningmastery.com/develop-encoder-decoder-model-sequence-sequence-prediction-keras/

我不得不用零填充我的输入和输出序列的末尾以保持它们相同的长度，例如

[72 1 62 0 68 4 72 0 63 0 68 5 83 3 87 1 86 1 84 3 86 13 74 0 71 2 87 5 90 3 63 0 66 0 76 2 36 1 38 1 67 0 34 0 61 4 89 4 62 0 40 0 63 0 31 1 39 5 88 4 68 0 68 0 72 3 71 0 78 3 67 1 66 0 64 5 63 1 67 2 61 0 61 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]

[77 77 5 76 2 77 78 71 1 79 1 77 76 79 71 71 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]

但这意味着在训练/验证期间，模型将获得比应有的更高的验证结果，因为它很容易学会匹配序列的零元素。

我在encoder_inputs和decoder_inputs中添加了下面的Masking函数，但是修改后我得到了错误；

encoder_inputs = Input(shape=(None, n_in))
encoder_inputs = Masking(mask_value=0)(encoder_inputs) #****** TEST *****

....

decoder_inputs = Input(shape=(None, n_out))
decoder_inputs = Masking(mask_value=0)(decoder_inputs) #****** TEST *****

ValueError: Graph disconnected: 无法在“input_14”层获得张量 Tensor("input_14:0", shape=(None, None, 80), dtype=float32) 的值。访问以下之前的图层没有问题：['input_13']

def define_models(n_in, n_out, n_units):
# define encoder
encoder_inputs = Input(shape=(None, n_in))
encoder_inputs = Masking(mask_value=0)(encoder_inputs) #****** TEST *****
encoder = LSTM(n_units, return_state=True)
encoder_outputs, state_h, state_c = encoder(encoder_inputs)
encoder_states = [state_h, state_c]
# define decoder
decoder_inputs = Input(shape=(None, n_out))
decoder_inputs = Masking(mask_value=0)(decoder_inputs) #****** TEST *****
decoder_lstm = LSTM(n_units, return_sequences=True, return_state=True)
decoder_outputs, _, _ = decoder_lstm(decoder_inputs, initial_state=encoder_states)
decoder_dense = Dense(n_out, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
# define inference encoder
encoder_model = Model(encoder_inputs, encoder_states)
# define inference decoder
decoder_state_input_h = Input(shape=(n_units,))
decoder_state_input_c = Input(shape=(n_units,))
decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs)
decoder_states = [state_h, state_c]
decoder_outputs = decoder_dense(decoder_outputs)
decoder_model = Model([decoder_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states)
# return all models
return model, encoder_model, decoder_model

知道如何调整它以使新的遮罩功能起作用吗？

谢谢

Answer 1

在定义层时要注意，以免覆盖它们，尤其是输入

def define_models(n_in, n_out, n_units):

    # define encoder
    enc_inputs = Input(shape=(None, n_in))
    encoder_inputs = Masking(mask_value=0)(enc_inputs) #****** TEST *****
    encoder = LSTM(n_units, return_state=True)
    encoder_outputs, state_h, state_c = encoder(encoder_inputs)
    encoder_states = [state_h, state_c]

    # define decoder
    dec_inputs = Input(shape=(None, n_out))
    decoder_inputs = Masking(mask_value=0)(dec_inputs) #****** TEST *****
    decoder_lstm = LSTM(n_units, return_sequences=True, return_state=True)
    decoder_outputs, _, _ = decoder_lstm(decoder_inputs, initial_state=encoder_states)
    decoder_dense = Dense(n_out, activation='softmax')
    decoder_outputs = decoder_dense(decoder_outputs)
    model = Model([enc_inputs, dec_inputs], decoder_outputs)

    # define inference encoder
    encoder_model = Model(enc_inputs, encoder_states)

    # define inference decoder
    decoder_state_input_h = Input(shape=(n_units,))
    decoder_state_input_c = Input(shape=(n_units,))
    decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
    decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs)
    decoder_states = [state_h, state_c]
    decoder_outputs = decoder_dense(decoder_outputs)
    decoder_model = Model([dec_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states)

    # return all models
    return model, encoder_model, decoder_model