Models¶
Self-supervised methods with runnable examples in PyTorch, PyTorch Lightning and distributed PyTorch Lightning. Each one is built from the same losses, heads and transforms, so you can start from a method and swap the parts.
Pulls two augmented views of an image together and pushes other images away.
Contrastive learning with a momentum encoder and a memory bank of negatives.
Uses a view’s nearest neighbour from a support set as its positive.
SimCLR with a decoupled loss that drops the positive pair from the denominator.
Agreement between distorted views plus a covariance contrast term.
An online network predicts the output of a momentum target network, with no negatives.
A Siamese network with a stop-gradient: no negatives, large batches or momentum encoder.
A student network matches the outputs of a momentum-averaged teacher.
SimSiam that averages several target predictions to train with small batches.
Compares the cluster assignments of masked and unmasked views.
MSN with a custom clustering prior, a power law by default, for imbalanced data.
DINO with self-distillation at both the image and the patch level.
Masks most patches and reconstructs their pixels with a light decoder.
Like MAE, but the encoder sees both masked and visible patches.
Masked image modeling with an online tokenizer and DINO-style self-distillation.
Predicts the pixels of image patches autoregressively.
Predicts the cluster assignments of masked patches instead of their pixels.
MAE with a much deeper decoder and block-wise masking.
Drives the cross-correlation of two views’ embeddings toward the identity.
Variance, invariance and covariance terms prevent collapse without negatives.
VICReg on global and local features, for detection and segmentation.
Maximizes manifold capacity: the spread of each image’s mean view embedding.
Avoids collapse by maximizing embedding entropy through Frobenius norms.
Invariance across views, with embeddings regularized by SIGReg.
Clusters the data and swaps cluster assignments between views.
MoCo with a dense, pixel-level contrastive loss for detection and segmentation.
SimCLR on per-region embeddings from unsupervised segmentation masks.
Contrastive learning between groups of images instead of single images.
A latent world model: predicts the next frame’s embedding from past frames and actions.