Struct LossContext 

pub struct LossContext<'a> {
    pub output: &'a Variable,
    pub batch: &'a Batch,
    pub tags: &'a HashMap<String, Variable>,
    pub traces: &'a HashMap<String, Vec<Variable>>,
}

Context passed to the per-batch loss closure during El Che distributed training. All fields carry live autograd graphs, so the returned loss scalar can be backpropagated immediately.

model.set_loss_fn(|ctx: &LossContext| {
    let cls  = cross_entropy_loss(&ctx.tags["head"], &ctx.batch["label"])?;
    let rec  = mse_loss(&ctx.tags["recon"], &ctx.batch["image"])?;
    Ok(cls + rec)
});

Fields

output: &'a Variable

Forward output (live autograd).

batch: &'a Batch

The per-device batch with all named fields (inputs + targets).

tags: &'a HashMap<String, Variable>

Tagged outputs from this forward pass (live autograd).

traces: &'a HashMap<String, Vec<Variable>>

Loop traces keyed by tag name (live autograd).