MLX Image Models

Work in progress...

This repo is an attempt to design a MLX Deep Learning Model library equivalent to timm, also known as the Pytorch Image Models library.

Contrary to the MLX-image / MLX-vision packages, the goal is to get a fully compatible library with timm and models weights stored on the Hugging Face Hub.

The main design goals are:

• 100% coverage of the timm models
• Fully-tested models for exactness w.r.t timm

Things that won't be supported:

• Model training
• Optimizers, losses, etc.
• Export to ONNX, JIT, etc.
• Other niche things from timm

TODO

• Use FP32 for norms activations sensitive to low precision.
• Convert transient init to functional (i.e nn.MaxPool(2)(x)) in __call__
• Remove all dependencies except MLX.