Vision-SSL

This repository contains clean, modular implementations of popular self-supervised learning (SSL) methods for computer vision.

---

Navigation

Methods are organized into the following categories:

• Generative: Models that reconstruct masked parts of the input image (e.g., MAE, iGPT).

• Invariance-based: Models that enforce consistent embeddings across augmented views of the same image (e.g., SimCLR, MoCo).

• Other: Methods that do not explicitly fit into the above categories but are part of the broader SSL landscape (e.g., I-JEPA).

---

Implemented Methods

Currently implemented methods include the following models:

Model	Year	Category
MoCo	2019	Invariance
SimCLR	2020	Invariance
BYOL	2020	Invariance
Barlow Twins	2021	Invariance
MAE	2021	Generative
I-JEPA	2023	Other

Note

All models are structured such that the forward() method returns the encoder output used in downstream tasks.
Pretraining components (e.g., projection heads, momentum encoders) are implemented separately in training logic.

More methods coming soon...

Planned implementations include the following:

Model	Year	Category
SimMIM	2021	Generative
BEiT	2022	Generative
DINO	2021	Other
DINOv2	2023	Other

📁 Project Structure

. 
├── generative/ 
│   └── <method>/ 
├── invariance/ 
│   └── <method>/ 
├── other/ 
│   └── <method>/ 
└── images/ 
    └── <method>.png

Each method includes its own:

• train.py — for pretraining
• model.py — with clean forward() for downstream use
• misc.py — with other required functions