Twenty SSL methods and one ruler to measure them all

What it does MMSelfSup is a PyTorch-based toolbox for self-supervised representation learning in computer vision. It implements two dozen algorithms—from classic contrastive methods like MoCo and SimCLR to masked-image models like MAE and BEiT—under a single modular framework. The project also standardizes evaluation pipelines, including linear probing, semi-supervised classification, and downstream detection and segmentation, so pre-trained backbones can be compared on equal footing.

The interesting bit The real value is the standardization. Most papers report numbers using their own hidden hyperparameters and hardware; MMSelfSup forces everyone into the same gymnasium, making it possible to see which self-supervised method actually generalizes. It also inherits the OpenMMLab plugin architecture, so a backbone trained here drops directly into object-detection or segmentation pipelines without rewriting boilerplate.

Key highlights

• Implements two dozen algorithms including MoCo, SimCLR, BYOL, SwAV, MAE, BEiT, EVA, and PixMIM.
• Standardizes benchmarks: linear classification, low-shot SVM, semi-supervised learning, and downstream detection/segmentation.
• Modular design shared with other OpenMMLab projects (MMCV, MMEngine, MMClassification).
• Provides a model zoo with pre-trained weights and benchmark results.
• Version 1.0.0 (April 2023) added DINO support and refactored file I/O interfaces.

Caveats

• The README prominently flags MMPreTrain as the “newly upgraded” visual pre-training framework in the OpenMMLab family, so the long-term maintenance split between MMSelfSup and MMPreTrain is unclear.
• Migration is required between 0.x and 1.x branches.

Verdict Researchers and engineers who need to benchmark or reproduce self-supervised vision results should look here; casual users who just want a quick pre-trained backbone might find the OpenMMLab dependency stack heavier than necessary.