A from-scratch PyTorch implementation of a lightweight DINO-style self-supervised Vision Transformer, featuring multi-crop training, EMA teacher-student learning, optional RoPE positional encoding, and DINOv3-inspired Gram regularization.
computer-vision pytorch representation-learning dino self-supervised-learning tiny-imagenet multi-crop vision-transformer
-
Updated
Jun 8, 2026 - Python