A Pytorch pipeline for Tweet Sentiment Extraction Kaggle competition

Introduction

This is a Pytorch training pipeline for a text span selection task. It also uses the Catalyst deep learning framework.

Installation

1. You need to have Anaconda installed
2. Clone the repo

git clone https://github.com/Kirill-Kravtsov/kaggle-tweet-sentiment-extraction

3. Create and activate provided Anaconda enviroment

conda env create -f tweet_env.yml
conda activate tweet_env

4. Download competition data and put in data dir in root of the project
5. Create folds by running

python create_folds.py

Project structure:

├── configs
│   ├── best_bertweet.yml
│   ├── best_roberta.yml
│   ├── experiments
│   └── optimization
├── create_folds.py
├── data
├── logs
├── scripts
├── src
│   ├── callbacks.py
│   ├── collators.py
│   ├── datasets.py
│   ├── data_utils.py
│   ├── hooks.py
│   ├── losses.py
│   ├── optimize_experiment.py
│   ├── tokenization.py
│   ├── train.py
│   ├── transformer_models.py
│   └── utils.py
└── tweet_env.yml

Running pipeline

To train tha basic Roberta and BERTweet models run:

python train.py --cv --config ../configs/best_roberta.yml
python train.py --cv --config ../configs/best_bertweet.yml

Note: the code is supposed to work with one gpu, so if you have multi-gpu system do not forget to specify CUDA_VISIBLE_DEVICE variable, e.g.:

CUDA_VISIBLE_DEVICES=0 python train.py --cv --config ../configs/best_roberta.yml