use flodl::nn::Module;
use flodl::nn::optim::Optimizer;
use flodl::tensor::Result;
use flodl::{Tensor, Variable};
use crate::engine::fitness::{Fitness, Metric};
use crate::graph::network::Network;
pub fn seed_step_randomness(net_seed: u64, step: u64, call_index: u64) {
let mixed = net_seed
.wrapping_add(step.wrapping_mul(0x9E37_79B9_7F4A_7C15))
.wrapping_add(call_index.wrapping_mul(0xBF58_476D_1CE4_1FF5));
fastrand::seed(mixed);
}
pub fn train_one_step(
net: &mut Network,
optimizer: &mut dyn Optimizer,
loss_fn: &dyn Fn(&Variable, &Variable) -> Result<Variable>,
batch: &(Tensor, Tensor),
grad_clip: f32,
) -> Result<f32> {
let (inputs, targets) = batch;
let x = Variable::new(inputs.clone(), true);
let y = Variable::new(targets.clone(), false);
let pred = net.forward(&x)?;
let loss = loss_fn(&pred, &y)?;
let loss_val = loss.item().unwrap_or(0.0) as f32;
loss.set_requires_grad(true)?;
optimizer.zero_grad();
loss.backward()?;
if grad_clip > 0.0 {
flodl::clip_grad_norm(&net.parameters(), grad_clip as f64)?;
}
optimizer.step()?;
Ok(loss_val)
}
pub struct EvalReport {
pub eval_loss: Option<f32>,
pub fitness: f32,
pub metrics: Vec<f32>,
}
pub fn eval_one_step(
net: &mut Network,
loss_fn: &dyn Fn(&Variable, &Variable) -> Result<Variable>,
fitness: &Fitness,
informative: &[Metric],
batch: &(Tensor, Tensor),
) -> Result<EvalReport> {
let (inputs, targets) = batch;
let y = Variable::new(targets.clone(), false);
net.eval();
let x = Variable::new(inputs.clone(), false);
let pred = net.forward(&x)?;
let loss = loss_fn(&pred, &y)?;
let loss_val = loss.item().unwrap_or(0.0) as f32;
net.train();
let score = fitness.score(&pred, &y)?;
let informative_vals = informative
.iter()
.map(|m| crate::utils::score::score_by_label(m.label(), &pred, &y))
.collect::<Result<Vec<f32>>>()?;
Ok(EvalReport {
eval_loss: Some(loss_val),
fitness: score,
metrics: informative_vals,
})
}