Self-Supervised Learning for Classification of Cell Line Images

This is a course project for the postgraduate level course of AI for Bioinformatics taught at DIEF, UniMoRe.

• Project Presentation: https://drive.google.com/file/d/14ohkZe7FEuB0umfB-gVuVjR5kEOa_waV/view?usp=sharing
• Project Report: https://drive.google.com/file/d/1_B437AIdNaT-8oJgbUSPmzNo9LGAtfom/view?usp=sharing

This repository contains implementations for training and evaluating models using the RxRx1-Wilds Cell-Level Dataset. The project is structured into two main workflows:

• standard processing
• crop-based processing (which also uses cell level dataset).

Quickstart

1. Download dataset

   sbatch download_dataset.slurm
   

2. Train and eval model

   sbatch run.slurm #standard processing
   sbatch run_crop.slurm #crop-based processing
   

Project Structure

Python Scripts

Script	Description
download_dataset.py	Script to download the RxRx1-Wilds dataset automatically.
dataset.py	Dataset class for standard processing.
train.py	Training script for the standard workflow.
eval.py	Evaluation script for standard models, including feature extraction and classification.
dataset_crop.py	Dataset class for crop-based processing.
train_crop.py	Training script for the crop-based workflow with advanced augmentations.
eval_crop.py	Evaluation script for crop-based models with feature visualization (PCA, t-SNE).
loss_crop.py	Implements NTXent loss for contrastive learning with distributed support.

SLURM Scripts

Script	Description
download_dataset.slurm	Download the dataset
run.slurm	SLURM job script for standard training.
run_crop.slurm	SLURM job script for crop-based training.