practical-pytorch

practical-pytorch copied to clipboard

Hi,

Thanks for the wonderful explanation. I am using this code as guideline to build a speech recognition network. I am giving speech frames (sequence of 40-dim feature vectors) as input to the encoder and trying to predict characters as output of the decoder. Speech frames can be very large in number (>1000) as compared to output length (<100).

So, I have set MAX_LENGTH of attention to be 5000. Unfortunately, it never predicts <eos> token and keeps predicting till 5000 characters.

I am using bidirectional LSTM as encoder. I concatenate the output and hidden from encoder to feed in the decoder, which is just LSTM.

I would highly appreciate any directions.

Thanks Brij