Backbones

A library of backbone architectures for neural networks, standardized for easy use in research and development.

This library was created to solve the following problems.

1. Reduce boilerplate that comes with (efficiently) extracting features from some pre-trained network.

2. Implement compatability layers to port weights between different distributors of pre-trained weights.

3. Allow both training and inference of the sub-networks.

Alternatives

If you are looking for a more general-purpose library, consider using Torchvision or Timm. These libraries are geared towards production, and are more feature-rich than this library.

See Also

• Pre-trained models
• Backbones Documentation