/R-net

R-net in PyTorch

Primary LanguagePythonMIT LicenseMIT

Work In Progress.

An unofficial implementation of R-net in PyTorch.

Natural Language Computing Group, MSRA: R-NET: Machine Reading Comprehension with Self-matching Networks

Python 3.5/3.6 and PyTorch 0.2

Usage

python main.py --batch_size 64

Current Problems

Char-Level embedding with RNN is too slow and requires much GPU memory, so I removed it from my implementation. I plan to try using CNN to do char-level embedding.

Performance

I would like to recommend https://github.com/HKUST-KnowComp/R-Net (Tensorflow) if you are looking for high performance. Currently my implementation has not fully reached the paper's performance.