3 unstable releases
0.2.1 | Mar 19, 2019 |
---|---|
0.2.0 | Mar 19, 2019 |
0.1.0 | Mar 17, 2019 |
#682 in Machine learning
13KB
198 lines
A dead simple neural network built as a learning exercise.
Getting Started
use artha::{
NeuralNetwork,
neural_net::{
normaize_val,
mean_loss,
find_max,
}
};
use ndarray::array;
fn main() {
let mut xs = array![[2.,9.],[1.,5.],[3.,6.]];
normaize_val(find_max(&xs), &mut xs);
let mut ys = array![[92.], [86.], [89.]];
normaize_val(vec![100.], &mut ys);
let mut nn = NeuralNetwork::new(2,1,vec![3]);
let predicted = nn.train(&xs, &ys, 10000);
let loss = mean_loss(&ys, &predicted);
use artha::logln;
logln!("Input: ", xs);
logln!("Actual Output: ", ys);
logln!("Predicted Output: ", predicted);
logln!("Loss: ", loss);
}
This program is a direct translation of https://dev.to/shamdasani/build-a-flexible-neural-network-with-backpropagation-in-python into rust.
Also checko 3Blue1Browns's excellent series on Neural Network https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi
I found this network to be significantly slower than the tutorial. Perhaps ndarray is not as fast as numpy or perhaps my rust code is not optimiz. I'll definitely look into it.
Besides optimization, I am also hoping to implmenting a network that recognized handwritten digits and who knows what else from there. But for now, this is a fairly inaccurate rookie version that I could build on my own.
- If you have any questions or suggestions, feel free to submit issues, or contact me in other ways.
- If you found my sub-par rust skills offensive, please do provide some constructive criticism.
Dependencies
~3MB
~52K SLoC