Download
SimSiam is a PyTorch implementation of “Exploring Simple Siamese Representation Learning” by Xinlei Chen and Kaiming He. The project introduces a minimalist approach to self-supervised learning that avoids negative pairs, momentum encoders, or large memory banks—key complexities of prior contrastive methods. SimSiam learns image representations by maximizing similarity between two augmented views of the same image through a Siamese neural network with a stop-gradient operation, preventing feature collapse. This elegant yet effective design achieves strong results in unsupervised learning benchmarks such as ImageNet without requiring contrastive losses. The repository provides scripts for both unsupervised pre-training and linear evaluation, using a ResNet-50 backbone by default. It is compatible with multi-GPU distributed training and can be fine-tuned or transferred to downstream tasks like object detection following the same setup as MoCo.
Features
- Minimal self-supervised learning framework without negative pairs or momentum encoders
- PyTorch-based implementation optimized for distributed multi-GPU training
- Fully reproducible training pipeline for ImageNet using default hyperparameters from the paper
- Includes both unsupervised pre-training and linear evaluation scripts
- LARS optimizer support via NVIDIA Apex for large-batch training
- Compatible with object detection transfer setups from MoCo
Project Samples
