Greetings everyone...
Im happy to share with you my recent work on
Tensorium, a platform agnostic fast and intuitive tensor manipulation library written completely in pure pascal (and sometime SIMD assembly instructions for x86 CPUs, OpenCL for GPUs).
Hence the name
Tensorium for now im still considering to change the name, yet I didn't settle on a better one.
This library is supposed to be a replacement of my old library
LightNet which is another pascal neural network library that can run and train YOLO Object Detection models completely and self dependently in pure pascal.
The library is intended to be self dependant pascal (FPC or Delphi) tool for AI and Machine Learning algorithms.
The library contains
- Tensor arithmetic operations and permutations
- Basic Linear algebra
- Regressions and fitting (Linear and Polynomial)
- Neural networks (Examples provided for training MNIST, CIFAR, and inference for YOLO3 Object Detection)
- Console based tensor colorful visualization and plotting (works even with SSH terminals no need for a GUI)
- Lots of good stuff that I can't list them all here but if you like Python's NumPy this is my humble attempt to unreptile the world and dePythonize the machine learning hype and make pascal great again
The tensor types
var tensor:TTensor<T>; //Generic Type
or better optimized
var sTensor :TSingleTensor; iTensor: TInt32Tensor; //... etc
consist of managed records not classes so it should reside in the stack and any heap allocations or disposal operation will automatically follow the tensor variable life time within the code scope.
I have posted some initial instructions on how to use the library but there lots of good functionalities that needs to be documented with a code examples such as regressions and polynomial fitting,
If you find the library useful please don't forget to
star ⭐ the library.
due to my limited free time please note that I'm looking for good pascal enthusiast's contibuters to help with the WIKI page and to write more examples along with the library improvements, if anyone interested I can provide a quick tutorial to kickstart, any pull requests in this regards are welcome just let me know.
All the best
Recent
