Kraft is a pythonic deep learning framework with a built-in GPU acceleration. It is following an early PyTorch, slightly altering its API where necessary.
kraft.autograd is an automatic differentiation framework built upon NumPy and CuPy.
It defines Variable, Function, and some basic functions in the kraft.autograd.ops module.
kraft.optim contains simple, easy to follow implementations of several popular optimization algorithms, such as Adam and SGD.
kraft.nn provides a Module class.
Every neural network built with kraft must inherit this class.
kraft.nn also provides some basic layers, such as Linear, Conv2d, MaxPool2d, and AvgPool2d.
Kraft uses chain rule under the hood.
There's a simplified implementation of this rule in the _overview/one_file_framework.py file, Kraft mostly follows ideas expressed in that file.
There's a file called mnist_example.py.
It contains an example of this framework application to a MNIST problem, with precision ~= 97-98%.
It features convolutional layers, regularization mechanisms (Dropout, L1/L2 regularization), and Cross Entropy loss