https://github.com/maeotaku/pytorch_usm
A differentiable version of the well known unsharp masking filter for image processing. The filter is implemented as a trainable Pytorch layer and allows the user to learn its parameters during a model training, adjusting itself to any given dataset and task.
https://github.com/maeotaku/numpy_neural_net
A small deep learning framework made from scratch using only numpy. The framework allows to add new operations, but already support some basic building blocks such as fully connected, dropout, sigmoid, relu, softmax and cross-entropy. Each operation is differentiable and gradients are properly calculated by hand.
https://github.com/maeotaku/biomachina_hierarchical_softmax
Code used to participate in PlantCLEF 2022, where we developed a 2 layer hierarhical softmax for plant identification, to reduce model sizes up to 5.67x. The implementation work in Pytorch and can be connected to any classification model.
https://github.com/maeotaku/ray-tracer
Code in C to execute ray tracing from scratch. Several basic shapes have been modelled into it.