Equimo 0.1.3a5

Implementation of popular vision models in Jax

Equimo: Modern Vision Models in JAX/Equinox

WARNING: This is a research library implementing recent computer vision models. The implementations are based on paper descriptions and may not be exact replicas of the original implementations. Use with caution in production environments.

Equimo (Equinox Image Models) provides JAX/Equinox implementations of recent computer vision models, currently focusing (but not limited to) on transformer and state-space architectures.

Features

• Pure JAX/Equinox implementations
• Focus on recent architectures (2023-2024 papers)
• Modular design for easy experimentation
• Extensive documentation and type hints

Installation

From PyPI

pip install equimo

From Source

git clone https://github.com/clementpoiret/equimo.git
cd equimo
pip install -e .

Implemented Models

Model	Paper	Year	Status
FasterViT	FasterViT: Fast Vision Transformers with Hierarchical Attention	2023	✅
Castling-ViT	Castling-ViT: Compressing Self-Attention via Switching Towards Linear-Angular Attention During Vision Transformer Inference	2023	Partial*
MLLA	Mamba-like Linear Attention	2024	✅
PartialFormer	Efficient Vision Transformers with Partial Attention	2024	✅
SHViT	SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design	2024	✅
VSSD	VSSD: Vision Mamba with Non-Causal State Space Duality	2024	✅

*: Only contains the Linear Angular Attention module. It is straight forward to build a ViT around it, but may require an additional __call__ kwarg to control the sparse_reg bool.

Basic Usage

import jax

import equimo.models as em

# Create a model (e.g. `faster_vit_0_224`)
key = jax.random.PRNGKey(0)
model = em.FasterViT(
    img_size=224,
    in_channels=3,
    dim=64,
    in_dim=64,
    depths=[2, 3, 6, 5],
    num_heads=[2, 4, 8, 16],
    hat=[False, False, True, False],
    window_size=[7, 7, 7, 7],
    ct_size=2,
    key=key,
)

# Generate random input
x = jax.random.normal(key, (3, 224, 224))

# Run inference
output = model(x, enable_dropout=False, key=key)

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use Equimo in your research, please cite:

@software{equimo2024,
  author = {Clément POIRET},
  title = {Equimo: Modern Vision Models in JAX/Equinox},
  year = {2024},
  publisher = {GitHub},
  url = {https://github.com/clementpoiret/equimo}
}