BxTorch 0.3.4

High-level abstractions for PyTorch.

BxTorch

BxTorch is a high-level library for working with PyTorch. It is designed to make PyTorch much simpler in the most common cases. Yet, it is engineered to be highly extensible in order to preserve PyTorch's flexibility --- while relieving you from writing boilerplate code.

Have a look at BxModels if you want to use implementations of well-known machine learning models.

Installation

BxTorch is available on PyPi, so simply run the following command:

pip install bxtorch

The package will install the dependencies specified here. If you plan to use plotting features of BxTorch, make sure to also install the following packages:

matplotlib

Features

Generally, BxTorch provides an object-oriented approach to abstracting PyTorch's API. The core design objective is to provide an API both as simple and as extensible as possible --- usually at the expense of some milliseconds of execution time. Be aware that the goal of this library is not to maximize performance in cases where it is not needed.

This does not mean that BxTorch does not care about performance: in fact, the library has built-in support for multi-GPU training, both within a single process and split over multiple processes.

It must be emphasized that BxTorch is not meant to be a wrapper for PyTorch as Keras is for TensorFlow, for example. It only provides extensions for PyTorch.

Documentation

Examples of the usage of BxTorch can be found in the docs folder. Method documentation is currently only available as docstrings.

License

BxTorch is licensed under the MIT License.

The logo is modified from thenounproject.com, "Torch by iconsmind.com from the Noun Project".