lucidrains/vit-pytorch

This system is a Python library for implementing and training Vision Transformer (ViT) models and their variants. It provides a collection of architectural modules, including CaiT, CCT, and CrossViT, along with utilities for self-supervised learning and video processing. The library supports distributed training via Accelerate and includes example notebooks and scripts for common datasets like CIFAR-100 and Cats & Dogs.

A new training script (train\_vit\_decorr.py) is provided to demonstrate training a Vision Transformer with decorrelation auxiliary losses. This script uses the accelerate library for mixed-precision and distributed training, wandb for experiment tracking, and CIFAR-100 as the dataset. It also includes a MANIFEST.in file to ensure tests are included in the package distribution.

(repo-wide) · high confidence

This release introduces a wide range of new Vision Transformer (ViT) variants and specialized modules, including CaiT, CCT, CrossViT, CrossFormer, CvT, DeepViT, and Adaptive Token Sampling (ATS) ViT. It also adds self-supervised learning implementations for DINO and MAE, a video processing wrapper (AcceptVideoWrapper) for handling temporal sequences, and a 3D CCT variant. These additions expand the library's support for diverse transformer architectures and multi-modal inputs.

_vit\pytorch · high confidence

Added a new Jupyter notebook example (\examples/cats\_and\_dogs.ipynb\) demonstrating how to train a Visual Transformer using Linformer on the Dogs vs. Cats dataset. The notebook includes setup for dependencies (\vit\_pytorch\, \linformer\), data loading, and model training configuration.

examples · high confidence

Added new test coverage for the DINO distillation loss function, specifically verifying that loss values and gradients remain finite and numerically stable when using low-precision data types (float16 and bfloat16), and confirming gradient correctness against a reference implementation. Additionally, added a basic functional test for the ViT model to ensure it produces correctly shaped output logits.

tests · high confidence

The package has been updated to version 1.26.6. The build system now requires setuptools\>=61 and wheel, and the project specifies a minimum Python version of 3.8. Core dependencies include einops\>=0.8.2 and torch\>=2.5, with torchvision as a dependency. Test dependencies are pinned to pytest, torch==2.5.0, and torchvision==0.20.0.

(dependencies) · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• First survey — no prior run to compare against. CAI 52.

• Code Health 100
• Architecture 100
• Maturity 36
• Readiness 33
• Security 100

• 300 commits — 267 feature/other, 33 fixes

• (root) — 248 commits
• (repo) — 13 commits
• vit_pytorch/mae.py — 5 commits
• .github/workflows — 4 commits
• vit_pytorch/vit.py — 3 commits
• vit_pytorch/crossformer.py — 2 commits
• vit_pytorch/mobile_vit.py — 2 commits
• vit_pytorch/na_vit.py — 2 commits
• vit_pytorch/vivit.py — 2 commits
• .github/FUNDING.yml — 1 commit
• examples/cats_and_dogs.ipynb — 1 commit
• images/mbvit.png — 1 commit
• images/xcit.png — 1 commit
• tests/test_dino_loss.py — 1 commit
• tests/test_vit_with_decorr.py — 1 commit
• vit_pytorch/cct.py — 1 commit
• vit_pytorch/cct_3d.py — 1 commit
• vit_pytorch/mpp.py — 1 commit
• vit_pytorch/na_vit_nested_tensor.py — 1 commit
• vit_pytorch/na_vit_nested_tensor_3d.py — 1 commit

• fix: Fix ViViT Transformer not passing use_flash_attn to Attention and duplicate mask reshape (#360)
• fix: Merge pull request #101 from zankner/mpp-fix
• fix: Merge pull request #153 from developer0hye/fix-example
• fix: double down on dual patch norm, fix MAE and Simmim to be compatible with dual patchnorm
• fix: fix block repeats in readme example for Nest
• fix: fix distill
• fix: fix feature maps in Nest, thanks to @MarkYangjiayi
• fix: fix hard distillation, thanks to @CiaoHe
• fix: fix hidden dimension in MaxViT thanks to @arquolo
• fix: fix linear head in simple vit, thanks to @atkos
• fix: fix max pool in nest
• fix: fix maxvit - need feedforwards after attention
• fix: fix mbconv residual block
• fix: fix mpp
• fix: fix mpp
• fix: fix multiheaded qk rmsnorm in nViT
• fix: fix positional embed for mean pool case and cleanup
• fix: fix pypi
• fix: fix recorder in data parallel situation
• fix: fix small bug
• …and 280 more

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

lucidrains/vit-pytorch was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.