A working machine learning framework in pure Mojo 🔥
Full Changelog: https://github.com/Benny-Nottonson/voodoo/compare/v0.7.8-alpha...v0.7.9-alpha
Refactor and change to Vector struct, Graph and Node changes, Lifetime and Pointer changes. Memory pool and graph fuse optimized.
Full Changelog: https://github.com/Benny-Nottonson/voodoo/compare/v0.7.7-alpha...v0.7.8-alpha
Modified Node, Tensor, and Graph struct for simplicity and speed, modified MatMul implementation for speed with backpropagation. Overall bug fixes.
Full Changelog: https://github.com/Benny-Nottonson/voodoo/compare/v0.7.6-alpha...v0.7.7-alpha
Removed Pointers from Node and Graph where possible, improved memory access and allocation.
Full Changelog: https://github.com/Benny-Nottonson/voodoo/compare/v0.7.5b-alpha...v0.7.6-alpha
Fixed a segfault issue in the previous release.
Entirely restructured Graph, Node, Tensor, and Vector types to better utilizer Pointers / References (Preparing for Lifetime support from Mojo). Small speedups and bug fixes. Small overall logic improvements and algorithmic changes.
Full Changelog: https://github.com/Benny-Nottonson/voodoo/compare/v0.7.0-alpha...v0.7.5-alpha
Updated for Mojo v0.7, from now on release tags will try and follow the Mojo release tags.
Improvements on Convolution / Pooling layers using tiling. General speed improvements and optimizations with vectorize_unroll and other related functions. Benchmark added to compare speed to similar projects, MNIST dataset / test updated & scaled out.
Convolution / Pooling (1D and 2D) added and tested with MNIST dataset, best so far 29% loss with 1440 labeled images. Still necessary to optimize convolution and pooling more.
Full Changelog: https://github.com/Benny-Nottonson/voodoo/compare/v0.5.0-alpha...v0.6.0-alpha
Rewrote Matmul resulting in a >300% speedup from earlier versions, general code improvements.