We will need to make a few changes to the existing functions to cater for the embeddings layer. Firstly, process_input would process the input to have word indexes in different time steps, instead of the one-hot encoded vectors, as follows:
def process_input(self,input_texts,target_texts=None,verbose=True): encoder_input_data = np.zeros( (len(input_texts), self.max_encoder_seq_length), dtype='float32') decoder_input_data = np.zeros( (len(input_texts), self.max_decoder_seq_length), dtype='float32') decoder_target_data = np.zeros( (len(input_texts), self.max_decoder_seq_length,1), dtype='float32') if self.mode == 'train': for i, (input_text, target_text) in enumerate(zip(input_texts,target_texts)): for ...
