RNN word language model example

[adamlerer]

@apaszke for your perusal.

If you uncomment the profiling code, you can then run
CUDA_LAUNCH_BLOCKING=1 python main.py -model LSTM -cuda -nlayers 2

[adamlerer]

FYI the perplexity isn't as good for this model as an "equivalent" rnnlib3 model, I'm looking at why now - probably has to do with different initialization.

[adamlerer]

My latest commit fixes @colesbury and @apaszke comments, and some other bugs. I also fixed a few discrepancies with the torch model (remove biases, fix lr) so it now reaches the same perplexity as the torch model (~114).

Requires pytorch/pytorch#106

[adamlerer]

I updated this to use the torch RNN library (the monolithic one, that uses cudnn). It's within 5% of the speed of the lua-torch version under the standard parameters. Ready to merge.

P.S. Should I also include the version that builds an RNN from scratch? It might be instructive for people who want to do something different than what's supported by the monolith.

[adamlerer]

Depends on pytorch/pytorch#129