Rust Neural Network

I’m done!!

New magic:

• Demo website to run the model! (recognize digits)
• Better demo model (trained longer)
• and yeah github stuff like readme :p

Thank you for following me during this journey! Learned so much!!! :D

Oh my! ✨

• Back-propagation! You can now train models!
• Optimizers! SGD or AdamBuilder (Adam is better)
• Batch training! Using accumulators and apply (really cool, makes training faster)
• LOTS of optimizations (got down from 204s to 111s per 10 iters, with lots of memory optims lol)
• lots of sweating lol
• Dropout layers: makes some neural nets more robust… works bad with really small nets tho
• super simple network builder!

What noooooww 🚀

• I guess a usable demo would be awesome, like a web interface to try out the networks? I mean that would def be fire :p
• Maybe also more demos, will see

🦀 learned so much stuff!! like wow that was such a good project idea! :DD

🦀 Whats new

kinda everything new lol

• Activation functions (implemented ReLU, Sigmoid, Tanh, Linear) & Softmax
• CLEAN MATRIX OPERATIONS, type safe! Compiler will say nuh uuhh if the dimensions do not match :DD (that’s so fricking awseomse)
• you can use operators + * - for matrix operations directly (operator overloading is awesome)
• Layer structure with weight, bias and activation
• Network structure, allows you to add layers (type safe! cant add a layer whose input doesn’t match the previous output)
• Network allows you to run(input) and gives you the output of the neural network

And now…. 😅

TIME FOR TRAININGG!! (backpropagation gonna be hardddd)

Sorry for basically no screenshots but there’s not much to show… lol

Woohoo!! I wrote useful functions for matrices operation:

• multiplication, scalar, addition, subtraction, transposition and mapping with a function

holy moly i got multiplication on my own!!

-> I tried not to use AI & Internet AT ALL, and I MADE IT YIIIPEEE

oh and the time in the benchmark is sometimes for small matrices in multiplication and HUGE ones for addition and stuff