2 releases

✓ Uses Rust 2018 edition

0.2.1 Apr 23, 2019
0.2.0 Apr 7, 2019

Custom license

27KB
579 lines

pyrus-nn

Build Status Dependabot Status crates.io

Rust API Documentation

Lightweight neural network framework written in Rust, with thin python bindings.

  • Features:

    • Serialize networks into/from YAML & JSON!
      • Rust -> serde compatible
      • Python -> network.to_dict() & Sequential.from_dict()
    • Python install requires zero dependencies
    • No external system libs to install
  • Draw backs:

    • Only supports generic gradient descent.
    • Fully connected (Dense) layers only so far
    • Activation functions limited to linear, tanh, sigmoid and softmax
    • Cost functions limited to MSE, MAE, Cross Entropy and Accuracy

Install:

Python:

pip install pyrus-nn  # Has ZERO dependencies!

Rust:

[dependencies]
pyrus-nn = "0.2.0"

From Python

from pyrus_nn.models import Sequential
from pyrus_nn.layers import Dense

model = Sequential(lr=0.001, n_epochs=10)
model.add(Dense(n_input=12, n_output=24, activation='sigmoid'))
model.add(Dense(n_input=24, n_output=1, activation='sigmoid'))

# Create some X and y, each of which must be 2d
X = [list(range(12)) for _ in range(10)]
y = [[i] for i in range(10)]  

model.fit(X, y)
out = model.predict(X)


From Rust

use ndarray::Array2;
use pyrus_nn::{network::Sequential, layers::Dense};


// Network with 4 inputs and 1 output.
fn main() {
    let mut network = Sequential::new(0.001, 100, 32, CostFunc::CrossEntropy);
    assert!(
        network.add(Dense::new(4, 5)).is_ok()
    );
    assert!(
        network.add(Dense::new(5, 6)).is_ok()
    );
    assert!(
        network.add(Dense::new(6, 4)).is_ok()
    );
    assert!(
        network.add(Dense::new(4, 1)).is_ok()
    );
    
    let X: Array2<f32> = ...
    let y: Array2<f32> = ...
    
    network.fit(X.view(), y.view());
    
    let yhat: Array2<f32> = network.predict(another_x.view());
}

Dependencies

~6.5MB
~131K SLoC