paper-reproductions is a monorepo of from-scratch PyTorch reproductions of selected machine-learning papers. Each folder holds one paper's model, loss and dataset plus a short write-up of the method and results; the heavy lifting (training loop, validation, caching, component wiring, hook orchestration) is delegated to the torchestrator runner and the registry-pattern factory, so a reproduction is just domain code and a declarative config. Each paper FROMs a shared env.torchfile and runs with a single command; reproducible Docker and Apptainer environments are provided. It consolidates the earlier ot-reproduce and ssl-reproduce projects.
| Paper | Venue |
|---|---|
| SimCLR: contrastive learning of visual representations (Chen et al.) | ICML 2020 |
| BYOL: Bootstrap Your Own Latent (Grill et al.) | NeurIPS 2020 |
| DINO: emerging properties in self-supervised ViTs (Caron et al.) | ICCV 2021 |
| VICReg (Bardes, Ponce, LeCun) | ICLR 2022 |
| MAE: Masked Autoencoders (He et al.) | CVPR 2022 |
| I-JEPA (Assran et al.) | CVPR 2023 |
| DINOv2 (Oquab et al.) | TMLR 2024 |