Skip to content

The objective here is to study the plausibility of attention mechanisms in automatic language processing on an NLI (natural naguage inference) task, in transformers (BERT) architecture

Notifications You must be signed in to change notification settings

lolofo/stage_4_gm

Repository files navigation

Table of content

Introduction

Natural Language Inference (NLI) task

The data (SNLI dataset)

Command lines (How to use this git)

First of all make sure to use the environnement.

Virtualenv - pip environment (recommended)

Path to $VENV should be saved in ~/.bashrc

# Specify path to venv
export VENV=path/to/venv
echo $VENV

# Create venv
python -m venv $VENV/bert

# Activate venv
source $VENV/bert/bin/activate

# Replicate on cpu
pip install -r python_env/requirements.cpu.txt --no-cache-dir

# Replicate on gpu
pip install -r python_env/requirements.gpu.txt --no-cache-dir

# Exit venv
deactivate

Virtualenv - conda environment

  • if you are using conda you can use the two following command :
conda env create -f python_env/environment.yml
conda activate nlp
conda create --name nlp --file requirements.txt
conda activate nlp

WARNING: All the environments were exported on windows 11 -64 bits.

Download the data

To download the snli and e-snli data the command line is the following :

python data_download.py

All the data downloaded in this part will be stored in the folder : .cache\raw_data

Pytorch lightning training script

To run the training_bert.py for some tests we used the following command line :

python training_bert.py --epoch 3 --batch_size 4 --nb_data 16 --experiment bert --version 0

# Or by shorthand
python training_bert.py -e 3 -b 4 -n 16 --experiment bert --version 0

The objective was only to see the behaviour of the training with a small amount of data. (Spot some mistakes and see the behaviour of the loss)

To visualize our training performance we used the tool tensorboard. The default logdir in in .cache/logs/$EXPERIMENT where $EXPERIMENT is specified in --experiment. The log could be changed using flag --logdir or shorthand -s

tensorboard --logdir .cache/logs/$EXPERIMENT

About

The objective here is to study the plausibility of attention mechanisms in automatic language processing on an NLI (natural naguage inference) task, in transformers (BERT) architecture

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages