sthalles/SimCLR

Implements contrastive learning through momentum-based batch augmentation and a non-learnable memory bank, training ResNet encoders with NT-Xent loss across large minibatches to learn view-invariant representations. Supports mixed-precision training via PyTorch's native AMP, multi-GPU distributed training, and evaluation through linear probing on frozen features. Includes reference implementations for STL10 and CIFAR10 datasets with configurable projection head dimensionality and training hyperparameters.

2,480 stars. No commits in the last 6 months.