https://github.com/hsinyuan-huang/FusionNet-NLI

An example for applying FusionNet to Natural Language Inference
https://github.com/hsinyuan-huang/FusionNet-NLI

deep-learning machine-comprehension nlp

Last synced: 12 months ago
JSON representation

An example for applying FusionNet to Natural Language Inference

• Host: GitHub
• URL: https://github.com/hsinyuan-huang/FusionNet-NLI
• Owner: hsinyuan-huang
• Created: 2017-12-15T00:55:36.000Z (over 8 years ago)
• Default Branch: master
• Last Pushed: 2018-12-10T20:40:50.000Z (over 7 years ago)
• Last Synced: 2024-11-27T03:34:37.010Z (over 1 year ago)
• Topics: deep-learning, machine-comprehension, nlp
• Language: Python
• Homepage:
• Size: 22.5 KB
• Stars: 135
• Watchers: 9
• Forks: 39
• Open Issues: 5
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# FusionNet for Natural Language Inference

This is an example for applying FusionNet to natural language inference task.  

For more details on FusionNet, please refer to our paper:  

[FusionNet: Fusing via Fully-Aware Attention with Application to Machine Comprehension](https://arxiv.org/abs/1711.07341)  

Requirements

------------

+ Python (version 3.5.2)

+ PyTorch (0.2.0)

+ spaCy (1.x)

+ NumPy

+ JSON Lines

+ MessagePack

Since package update sometimes break backward compatibility, it is recommended to use Docker, which can be downloaded from [here](https://www.docker.com/community-edition#/download). To enable GPU, [nvidia-docker](https://github.com/NVIDIA/nvidia-docker) may also needs to be installed.  

After setting up Docker, simply perform `docker pull momohuang/fusionnet-docker` to pull the docker file. Note that this may take some time to download. Then we can run the docker image through  

`docker run -it momohuang/fusionnet-docker` (Only CPU)  

or  

`nvidia-docker run -it momohuang/fusionnet-docker` (GPU-enabled).  

Quick Start

-----------

`pip install -r requirements.txt`  

`bash download.sh`  

`python prepro.py`  

`python train.py`  

  

`train.py` supports an option `--full_att_type`, where  

`--full_att_type 0`: standard attention  

`--full_att_type 1`: fully-aware attention  

`--full_att_type 2`: fully-aware multi-level attention