Issues in lstm.py -> AttentionLayer -> score #1
Description
Line 171 (within forward), self.score was called. encoder_out is in the shape of [batch_size, src_time_steps, output_dims]. Then, moving to the score function, projected_encoder_out is obtained by a linear layer [input_dims, output_dims]. However, the dimension of encoder_out is output_dims; it should be incorrect when pushing a tensor in the output_dims to a linear layer requiring the input of input_dims. Within the class of LSTMdecoder, the attention is built with the same input_dims and output_dims; hence, the file can run, but it is not logically correct.