Pruning and Sparsemax Methods for Hierarchical Attention Networks

Deep Structured Learning Course, Computer Science PhD Program, Instituto Superior Técnico

To install dependencies:

1. Install Pytorch (pick config at will):

   https://pytorch.org/get-started/locally/

2. Install latest Pytorch NLP version:

   $ pip install git+https://github.com/PetrochukM/PyTorch-NLP.git

3. Install remaining requirements:

   $ pip install -r requirements.txt

   $ python -m spacy download en

To run:

$ python run.py model dataset

e.g. - Hierarchical Attention Network on IMDB dataset

$ python run.py han imdb

Available models - han, hpan, hsan, lstm, hn

Available datasets - imdb, yelp, yahoo, amazon

FAQ

1. I'm getting the following error:

    RuntimeError: "view size is not compatible with input tensor's size and stride 
    (at least one dimension spans across two contiguous subspaces). 
    Use .reshape(...) instead."
   

   Solution: Update pytorch-nlp to its latest version:

    $ pip uninstall pytorch-nlp
    $ pip install git+https://github.com/PetrochukM/PyTorch-NLP.git