//! Scaffolded by `cargo incin new mnist`, a 784 -> 128 -> 10 MLP training
//! loop with live telemetry pre-wired. `cargo run` plus `cargo incin watch`
//! gives you a live loss curve in the terminal with no extra setup.
//!
//! This trains on synthetic random data (not real MNIST), deliberately, so
//! the scaffold builds and runs immediately with no external download. Swap
//! in the real dataset by following incin's own
//! `crates/incin/examples/mnist_training.rs`, which loads
//! `incin_data::vision::mnist::MnistDataset` from a `./data/mnist`
//! directory instead of the `Tensor::rand`/`Tensor::from_slice` calls below.
use incin::Flatten;
use incin::prelude::*;
use incin_telemetry::emitter::Emitter;
use incin_telemetry::reporter::Reporter;
type Backend = incin::DefaultBackend;
fn main() -> incin::Result<()> {
let (reporter, run) = Emitter::to_run_dir(None).expect("failed to start telemetry");
println!(
"Telemetry run `{}` started. Watch it live in another terminal with:\n cargo incin watch --run-id {}",
run.run_id, run.run_id
);
let model = seq![
FlattenAxes::new(1, -1),
Linear::<Dyn, Backend>::build((784, 128))?,
ReLU,
Linear::<Dyn, Backend>::build((128, 10))?
];
let mut optim = incin::AdamW::<Backend>::from_module(&model, 0.001)?;
let batch_size = 32;
println!("Starting training on synthetic data...");
// Kept small so a debug (`cargo run`) build finishes in well under 30s;
// bump this up (and add `--release`) once you swap in real data.
for step in 0..30 {
let images = Tensor::<Dyn, Backend>::rand(vec![batch_size, 1, 28, 28])?;
// Synthetic class labels (0..10) -- not learnable signal, just enough
// to exercise the full forward/backward/optimizer/telemetry loop.
let labels_data: Vec<f32> = (0..batch_size).map(|i| ((step + i) % 10) as f32).collect();
let labels = Tensor::<Dyn, Backend>::from_slice(&labels_data, vec![batch_size])?;
let output = model.forward(images)?;
let loss = output.cross_entropy_loss(&labels)?;
let grads = loss.backward()?;
optim.step(&grads)?;
let loss_value = loss.to_scalar::<f32>()?;
reporter.scalar("loss", step, loss_value as f64);
if step % 5 == 0 {
println!("step {step}: loss = {loss_value:.4}");
}
}
println!("Done. See the module doc comment for how to swap in real MNIST data.");
Ok(())
}