Crate kerasty

Crate kerasty 

Source
Expand description

Keras for Candle (Rust ML framework) with support for Web Assembly.

§Roadmap of Supported Layers

LayerStateExample
Dense✅Dense
Convolution🏗️CNN
Normalization🏗️Norm
Flatten🏗️Flatten
Pooling🏗️Pool
Recurrent🏗️RNN
Attention🏗️Attn
Bert🏗️BERT
Llama🏗️LLAMA

§Examples

Solution to the classic XOR problem

use kerasty::{Dense, Device, Loss, Metric, Model, Optimizer, Result, Sequential, Tensor};

fn test_xor_problem() -> Result<()> {
  // Define the XOR input and output data
  let x_data = vec![0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0];
  let x = Tensor::from_slice(&x_data, (4, 2), &Device::Cpu)?;
  let y_data = vec![0.0, 1.0, 1.0, 0.0];
  let y = Tensor::from_slice(&y_data, (4, 1), &Device::Cpu)?;

  // Build the neural network model
  let mut model = Sequential::new();
  model.add(Dense::new(2, 2, "relu"));
  model.add(Dense::new(1, 2, "sigmoid"));

  // Compile the model
  model.compile(
      Optimizer::Adam(0.001, 0.9, 0.999, 1e-8, 0.0),
      Loss::BinaryCrossEntropyWithLogit,
    vec![Metric::Accuracy],
  )?;

  // Train the model
  model.fit(x.clone(), y.clone(), 10000)?;

  // Make predictions
  let predictions = model.predict(&x);
  let predictions = predictions.reshape(4)?.to_vec1::<f64>()?;
  let predictions: Vec<i32> = predictions
      .iter()
      .map(|&p| if p >= 0.5 { 1 } else { 0 })
      .collect();

  println!("Predictions:");
  for i in 0..4 {
      println!(
          "Input: {:?} => Predicted Output: {}, Actual Output: {}",
          &x_data[i * 2..i * 2 + 2],
          predictions[i],
          y_data[i]
      );
  }
 
  // Evaluate the model
  let score = model.evaluate(&x, &y);
  println!("Average loss: {}", score[0]);
  println!("Accuracy: {}", score[1]);

  Ok(())
}

The expected output is as follows:

Predictions:
Input: [0.0, 0.0] => Predicted Output: 0, Actual Output: 0
Input: [0.0, 1.0] => Predicted Output: 1, Actual Output: 1
Input: [1.0, 0.0] => Predicted Output: 1, Actual Output: 1
Input: [1.0, 1.0] => Predicted Output: 0, Actual Output: 0

§License

MIT

Copyright © 2025-2035 Homero Roman Roman
Copyright © 2025-2035 Frederick Roman

Re-exports§

pub use crate::common::definitions::Activation;
pub use crate::common::definitions::Initializer;
pub use crate::common::definitions::Loss;
pub use crate::common::definitions::Metric;
pub use crate::common::definitions::Optimizer;
pub use crate::common::definitions::Regularizer;
pub use crate::common::traits::Layer;
pub use crate::common::traits::Model;
pub use crate::layer::dense::Dense;
pub use crate::sequential::Sequential;

Modules§

common
layer
sequential

Macros§

bail

Structs§

Tensor
The core struct for manipulating tensors.

Enums§

DType
The different types of elements allowed in tensors.
Device
Cpu, Cuda, or Metal

Type Aliases§

Result