COMP 551 - PROJECT 4 - Reproducible Machine Learning

In this project we attempt to reproduce the results of the following paper:

Reference:

Training RNNs as Fast as CNNs

@article{lei2017sru,
  title={Training RNNs as Fast as CNNs},
  author={Lei, Tao and Zhang, Yu},
  journal={arXiv preprint arXiv:1709.02755},
  year={2017}
}

The code base is a fork of the source code provided by the authors; https://github.com/taolei87/sru

Progress

• Setup GCE instance for training
  • Obtain GCP approval for additional GPUs
• Reproduce author's SRU implementation
• Reproduce classification model
• Reproduce question answering model
• Reproduce langauge model
• Reproduce speech model (Attempted but failed to reproduce)

Requirements

• GPU and CUDA 8 are required
• PyTorch
• CuPy
• pynvrtc

Install requirements via pip install -r requirements.txt. CuPy and pynvrtc needed to compile the CUDA code into a callable function at runtime. Only single GPU training is supported.

Check individual tests for steps taken to reproduce the results