Tensorflow/tfjs "When inputs is an array, neither initialState or constants should be provided"

Question

我正在尝试使用 python (tensorflow/keras) 制作 ChatBot 来制作、训练和转换神经网络，然后在我的 Angular 应用程序中使用它 tensorflow/tfjs.我正在按照此处找到的示例进行操作：https://github.com/tensorflow/tfjs-examples/tree/master/translation 但也尝试添加嵌入层。

正在创建模型：

latent_dim = 200
encoder_inputs = Input(shape=(encoder_max_length, ), dtype='int32', )
encoder_embedding = Embedding(num_tokens,
                             embedd_size,
                             weights=[word2em],
                             input_length=encoder_max_length,
                             mask_zero=True,
                             trainable=False
                             )(encoder_inputs)
encoder_outputs, state_h, state_c = LSTM(latent_dim, return_state=True)(encoder_embedding)
encoder_states = [state_h, state_c]

decoder_inputs = Input(shape=(decoder_max_length, ), dtype='int32', )
decoder_embedding = Embedding(num_tokens,
                             embedd_size,
                             weights=[word2em],
                             input_length=decoder_max_length,
                             mask_zero=True,
                             trainable=False
                             )(decoder_inputs)
decoder_LSTM = LSTM(latent_dim, return_state=True, return_sequences=True)
decoder_outputs, _, _ = decoder_LSTM(decoder_embedding, initial_state=encoder_sates)
decoder_dense = Dense(num_tokens, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)

model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

创建编码器/解码器模型 (tfjs)：

如果我从 LSTM.apply() 中删除 'initialState'，错误就会消失，但无论输入如何，结果都是相同的句子。

prepareEncoder(model) {
    const encInputs = model.input[0];
    const stateH = model.layers[4].output[1];
    const stateC = model.layers[4].output[2];
    const encoderStates = [stateH, stateC];
    this.encoder = tf.model({inputs: encInputs, outputs: encoderStates});
  }

  prepareDecoder(model) {
    const tmp = model.layers[4].output[1];
    const latentDim = tmp.shape[tmp.shape.length - 1];

    const decoderStateInputH = tf.input({shape: [latentDim], name: 'decoder_state_input_h'});
    const decoderStateInputC = tf.input({shape: [latentDim], name: 'decoder_state_input_c'});
    const decoderStateInputs = [decoderStateInputH, decoderStateInputC];

    const decoderLSTM = model.layers[5];
    const decoderInputs = model.input[1];
    const decoderEmbedding = decoderLSTM.input[0];

    const applyOutputs = decoderLSTM.apply(decoderEmbedding, {initialState: decoderStateInputs});
    let decoderOutputs = applyOutputs[0];
    const decoderStateH = applyOutputs[1];
    const decoderStateC = applyOutputs[2];

    const decoderStates = [decoderStateH, decoderStateC];
    const decoderDense = model.layers[6];
    decoderOutputs = decoderDense.apply(decoderOutputs);

    this.decoder = tf.model({
      inputs: [decoderInputs].concat(decoderStateInputs),
      outputs: [decoderOutputs].concat(decoderStates)
    });
  }

使用模型进行预测：

  botReply(input_seq) {
    let states_value = this.encoder.predict(input_seq);
    let target_seq = tf.buffer([1, data['dec_max_length']], 'int32');
    target_seq.set(dict['<START>'], 0, 0);
    let stop_condition = false;
    let decoded_sentence = '';
    let word_count = 1;
    while (!stop_condition) {
      let predict_outputs = this.decoder.predict([target_seq.toTensor()].concat(states_value));

这就是一切失败的地方，我收到以下错误：

ERROR Error: When inputs is an array, neither initialState or constants should be provided
    at standardizeArgs (recurrent.js:54)
    at LSTM.apply (recurrent.js:465)
    at execute (executor.js:275)
    at training.js:856
    at engine.js:307
    at Engine.scopedRun (engine.js:317)
    at Engine.tidy (engine.js:306)
    at Module.tidy (globals.js:166)
    at training.js:839
    at engine.js:307

可能值得一提的事情。

该问题只出现在tfjs。在 python 中，它工作正常，考虑到数据集和训练时期，结果也不错。推理模型的制作方式与 tfjs
以前我没有使用嵌入层并且只有输入和输出的单热编码向量时它起作用了所以我怀疑这是版本问题还是什么。
在与模型训练相同的数据集上使用 word2vec 进行嵌入。
我用tfjsconverter转换keras模型

任何帮助将不胜感激，因为我真的运行没有想法。

Answer 1

试试这个 - 从嵌入层中删除 'mask_zero=True'，看看是否能解决问题。

Tensorflow/tfjs "When inputs is an array, neither initialState or constants should be provided"

Tensorflow/tfjs "When inputs is an array, neither initialState or constants should be provided"

python

tensorflow

keras

tensorflow.js