uek-1/puffpastry

basic rust neural network framework

puffpastry

puffpastry is a very basic feedforward neuron network library with a focus on parity with mathematical representations. It can be used to create and train simple models.

Usage

puffpastry is used very similarly to keras - stack layers and fit to training data.

Learning XOR

let mut model : Model<f64> = Model::new(Loss::MeanSquaredError);

// Dense::from_size(input_neurons, output_neurons, activation) -> Dense
model.push_layer(Dense::from_size(2, 2, Activation::Sigmoid));
model.push_layer(Dense::from_size(2, 1, Activation::None));


let train_inputs = vec![
    tensor!([[0.0], [0.0]]),
    tensor!([[1.0], [0.0]]),
    tensor!([[0.0], [1.0]]),
    tensor!([[1.0], [1.0]]),
];

let train_outputs = vec![
    tensor!([[0.0]]),
    tensor!([[1.0]]),
    tensor!([[1.0]]),
    tensor!([[0.0]]),
];

// fit(&mut self, inputs, outputs, epochs, learning_rate) -> Result
model.fit(train_inputs, train_outputs, 100, 1.2).unwrap();  

// evaluate(&self, input: Tensor) -> Result<Tensor>
model.evaluate(&tensor!([[1.0], [0.0]]).unwrap();
// stdout: Tensor {shape: [1], data: [0.9179620463347642]}

Features

Activation functions: [softmax, relu, sigmoid, linear]
Loss functions: [categorical cross entropy, mean squared error]
Layers: [dense, conv2d, maxpool2d, flatten]

Roadmap

1. Improve convolution layer performance
2. Make variations more explicit i.e. CategoricalCrossentropy vs ClippedCategoricalCrossEntropy
3. Documentation
4. Add pptimizers, weight initializers
5. Add more losses, layers, metrics