dtorch 0.0.7

A scientific package made a fellow student

DTorch

A direct link to the package : here

DTorch is a package made by a student at Epitech to improve his understanding of pytorch and better his knowledge of ia. It is built on top of numpy which is a scientific framework for math between matrices.

This project run on cpu but still have decent computation time making it usable for model building, optimization and saving/loading.

The tensors created can work with numpy but remember that the gradients will not be calculated if the operation are not in the range of control of the library. Therefore, a usage of the packages methods on the tensors if preferable.

Cuda support may appear in the future and similarly for the dkldnn library that seems to be an excellent when working on CPU.

A direct advantage of using dtorch is it's lightness. The package is currently close to 14 Ko and is fast to load in any project while the use of torch often lead to a slow start.

More about DTorch

The package is organised in the following manner :

Module	Description
dtorch	the base module where other are regrouped
functionnal	a gather of all the functions that can be used on jtensors
jtensors	a simple module containing the JTensors declaration
loss	loss classes for model implementation
nn	a neural network layer system and its multiple layers (see doc for more details).
einops	a basic reproduction of some einops library methods that support autograd
optim	A library contaning the optimizers that can be used on the networks parameters.
dtorchvision	A lib containing visual datasets and models (pretrained too)
dtorchtext	A lib containing textual datasets and models (pretrained too)

About Autograd

The implementation in this library may not be the most effective computationnaly but the goal was to make it simple and make it work effectively by relying on numpy matrices operations for computation speed concerns.