Slow LSTM speed relative to theano.

[c0nn3r]

Recently, a benchmark was made comparing the rnn package to Theano, and found Theano to be much much faster (the benchmark can be found here). This is concerning. It's currently unknown if other implementation of RNNs in torch are faster (torch-rnn might be), but it seems some improvement to speed could be made. I'm hope to revisit this problem when I have less deadlines to deal with, but just wanted to hear other's thoughts. Thanks.

[uralik]

as I got, in this benchmark they did not use nngraph. (at least in the source code) With nngraph FastLSTM is faster.

[c0nn3r]

How do you set the variable? Do you do it before you build rnn? I will submit a pull request to make the change.

[uralik]

Should be something like in this simple example, so you just set the flag.

[nicholas-leonard]

@Torkoal Thanks for updating that benchmark :)

[nicholas-leonard]

@Torkoal Also feel free to try to improve the speed of the package. I know for sure that the LSTM module could benefit from a usenngraph = true implementation.

[nicholas-leonard]

I also submitted a PR with some improvements to the benchmark : glample/rnn-benchmarks#5