examples

RusTorch Examples 🧪

This directory contains examples demonstrating the capabilities of RusTorch.

Runnable Examples

To run these examples, you should use cargo run -p <package> --example <example_name>.

🧠 Neural Networks (rustorch-nn)

Located in rustorch-nn/examples:

• full_training_demo: A complete CNN training loop (Conv2d -> ReLU -> MaxPool -> Linear) on dummy data.

  cargo run -p rustorch-nn --example full_training_demo

• serialization: Demonstrates saving and loading models using Serde.

  cargo run -p rustorch-nn --example serialization

• tracing: Shows how to use the JIT tracer to optimize a model.

  cargo run -p rustorch-nn --example tracing

⚡ Core & JIT (rustorch-core)

Located in rustorch-core/examples:

• jit_demo: Demonstrates graph construction, operator fusion (Conv+ReLU), and execution.

  cargo run -p rustorch-core --example jit_demo

---

Featured Code: CNN Training

Here is a snippet of what the training code looks like (from full_training_demo.rs):

// Define a simple CNN
struct CNN {
    conv1: Conv2d,
    fc: Linear,
}

impl Module for CNN {
    fn forward(&self, x: &Tensor) -> Tensor {
        let x = self.conv1.forward(x).relu().max_pool2d((2,2), (2,2), (0,0));
        let x = x.reshape(&[x.shape()[0], 16 * 14 * 14]);
        self.fc.forward(&x)
    }
}

// Training Loop
let model = CNN::new();
let optimizer = SGD::new(model.parameters(), 0.01);

for epoch in 0..10 {
    let output = model.forward(&input);
    let loss = criterion.forward(&output, &target);
    optimizer.zero_grad();
    loss.backward();
    optimizer.step();
}