Keras：设置Bi-RNN初始状态

Question

我正在为有关前提和假设的自然语言推理（蕴含）问题编写Keras模型。

我通过带有return_state = True的Bi-LSTM来运行前提和假设，然后使用通过这种最后状态向量初始化的另一个Bi-LSTM来独立地对前提和假设进行编码。在代码中，这是Bi-LSTM的创建：

LastStateLSTM = LSTM(hidden_units, implementation=2, return_sequences=False, name='laststate', return_state=True )
LastStateLSTM = Bidirectional(LastStateLSTM, name='bilstm1')
EncoderLSTM = LSTM(hidden_units, implementation=2, return_sequences=True, name='encoderlstm')
EncoderLSTM = Bidirectional(EncoderLSTM, name='bilstm2')

在这里，我要应用这些图层：

out1, forward_prem_h, forward_prem_c, backward_prem_h,backward_prem_c = LastStateLSTM(prem)
out2, forward_hyp_h, forward_hyp_c, backward_hyp_h, backward_hyp_c = LastStateLSTM(hyp)

u = EncoderLSTM(prem, initial_state=[forward_hyp_h, forward_hyp_c, backward_hyp_h, backward_hyp_c])
v = EncoderLSTM(hyp, initial_state=[forward_prem_h, forward_prem_c, backward_prem_h, backward_prem_c])

最后，有summary（）和fit（）方法：

final_model.compile(optimizer=Adam(amsgrad=True), loss='categorical_crossentropy', metrics=['accuracy'])

final_model.summary()

final_model.fit([X1_nli, X2_nli, overlapFeatures_nli, refutingFeatures_nli, polarityFeatures_nli, handFeatures_nli, cosFeatures_nli, bleu_nli, rouge_nli, cider_nli_train], \
                Y_nli, validation_data=([X1_test_nli, X2_test_nli, overlapFeatures_nli_test, refutingFeatures_nli_test, polarityFeatures_nli_test, handFeatures_nli_test, cosFeatures_nli_test, bleu_nli_test, rouge_nli_test, cider_snli_test], Y_test_nli), \
                callbacks=[checkpoint], epochs=1, batch_size=8)

但是，我发现，即使将模型摘要打印出来，将initial_state参数传递给EncoderLSTM时，训练也不会开始（也没有错误）。

我正在使用带有Theano后端的Keras 2.2.2。有人有类似的问题吗？

谢谢。