Home

Awesome

Block • License PyPi

An intelligent block matrix library for numpy, PyTorch, and beyond. Crafted by Brandon Amos with significant contributions by Eric Wong.


Why do we need an intelligent block matrix library?

Let's try to construct the KKT matrix from Mattingley and Boyd's CVXGEN paper in numpy and PyTorch:

Without block, there is no way to infer the appropriate sizes of the zero and identity matrix blocks. It is an inconvenience to think about what size these matrices should be.

What does block do?

Block acts a lot like np.bmat and replaces:

Isn't constructing large block matrices with a lot of zeros inefficient?

Yes, block is meant to be a quick prototyping tool and there's probably a more efficient way to solve your system if it has a lot of zeros or identity elements.

How does block handle numpy and PyTorch with the same interface?

I wrote the logic to handle matrix sizing to be agnostic of the matrix library being used. numpy and PyTorch are just backends. More backends can easily be added for your favorite Python matrix library.

class Backend(metaclass=ABCMeta):

    @abstractmethod
    def extract_shape(self, x): pass

    @abstractmethod
    def build_eye(self, n): pass

    @abstractmethod
    def build_full(self, shape, fill_val): pass

    @abstractmethod
    def build(self, rows): pass

    @abstractmethod
    def is_complete(self, rows): pass

Getting Started

Issues and Contributions

I'd be happy to hear from you about any issues or features you add, please file an issue or send in a PR.

Licensing

This repository is Apache-licensed.