pip install subspyces
Pytorch is a very powerfull framework for developing ML algorithms, but implementing subspace methods in torch can be a bit repetitive when we don't have a starting point. For this, we developed this simple library which encapsulates some useful code that can be re-used and easily integrated with other torch codebases.
The core functionality of this library is the VectorSpace class.
This class encapsulates the core of all subspace methods: a simple vector space. Each VectorSpace contains a label, a dimension, and set of vectors, which are stored in torch.Tensor format.
Idealy, VectorSpace should handle all such operations regarding subspaces.
A generator is a class resposible for generating VectorSpace from a pytorch Dataset. As such, it will always receive a torch.utils.data.Dataset and outputs a list of VectorSpace, one vector space per label.
For example, the generator.IdentityGenerator class receives any torch dataset and reorganizes it in many vector spaces, without applying any other transformation (Identity).
Different from the generator class, a subspyce.transform transforms a VectorSpace into another VectorSpace. For example, the PCATransform will receive a VectorSpace and apply a PCA decomposition into it.
Finally, the metrics module implements some commonly used metric functions.