Skip to content

zuzannad2/PixelSum

Repository files navigation

Setup

This repository builts upon PIXEL and extends the linked repository with a summarisation model based on the Pixel-based Encoder of Language. This codebase is built on Transformers for PyTorch. The default font GoNotoCurrent.ttf that we used for all experiments is a merged Noto font built with go-noto-universal.

You can set-up this codebase in the following way:

  1. Clone repo and initialize submodules
git clone https://github.com/zuzannad2/PixelSum.git
cd PixelSum
git submodule update --init --recursive
  1. Create a fresh conda environment
conda create -n venv python=3.9
conda activate venv
  1. Install Python packages
conda install pytorch torchvision cudatoolkit=11.3 -c pytorch
conda install -c conda-forge pycairo pygobject manimpango
pip install --upgrade pip
pip install -r requirements.txt
pip install ./datasets
pip install -e .
  1. Download fallback fonts using python3 -m scripts.data.download_fallback_fonts fonts. Download the font GoNotoCurrent.ttf using the repo go-noto-universal into the folder called "fonts".

Pre-training PixelSum

The pretraining scripts for training PixelSum with Huggingface's trainer (run_pretraining.py) or without the trainer (run_pretrainer_no_trainer.py) are located in scripts/training. These are meant to be ran via bash scripts pretrain.sh / pretrain_no_trainer.sh and are configured for Slurm.

Finetuning PixelSum

The finetuning script for PixelSum is in scripts/training/run_finetuning.py. The script is ran via finetune.sh.

  • You need to pass the pretrained model path in the model_path argument.
  • The arguments train_encoder and train_decoder are set to False, meaning by default only the cross-attention layers will be trained. Adjust to desired setting.

Inference

The inference script is in scripts/training/run_inference.py. The script is ran via infer.sh.

  • Pass the pretrained model in model_path argument (leave empty string if zero-shot inference from model initialised from_encoder_decoder_pretrained).

Logging

By default scripts are set to log to WandB.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •