A collection of vision models implemented in tinygrad, in a similar vein to timm.
Mostly targeting models trained on Imagenet-1k, and other models that are fast on resource-constrained devices.
- ShuffleNetV2 - paper code
- GhostNetV2 - paper code
- FocalNet - paper code
- FastViT - paper code
- RepViT - paper code
- For models that can be reparameterized, add that functionality
- Training
See LICENSE.
Certain parts of the code are adapted from the original implementations, but they should all be under permissive licenses.