3. prep_batch

trainer/my_trainer.py → cfg.main_runner

class Trainer(BaseTrainer):
    def prep_batch(self, batch, split="train", **kwargs) -> dict

This is normally the whole task-specific trainer. Engines, metrics, checkpointing, resume, logging, AMP and DDP are inherited.

examples/synthetic/trainer/synthetic_trainer.py

from gatle_ignite import BaseTrainer, to_device


class Trainer(BaseTrainer):
    def prep_batch(self, batch, split="train", **kwargs):
        x, y = batch
        return to_device({"model_input": {"x": x}, "targets": {"labels": y}})

• model_input is splatted into the model's forward, so its keys match that signature.
• targets is what the loss and metrics resolve their tgt_name paths against.

split is the engine's engine_type: "train", "valid", "test", or whatever an added engine calls itself. Read split, never kwargs: a mistyped kwargs key returns None and silently takes the wrong path.

The launcher constructs Trainer(local_rank, cfg) and drives the lifecycle, so you never call fit() yourself.

train_step and eval_step are the hooks below this one: override eval_step when eval is not a forward pass (a diffusion sampler, an autoregressive decode), and train_step when the step itself is not standard supervised. Both have working defaults, as do build_model, build_dataloaders, forward, backward, eval_specs, train_spec, eval_context and extra_to_save, which are documented in the Python API.

The loss receives prep_batch's whole return

Not x["targets"]. Every contract in this framework passes (y_pred, target) where target is what you returned here, which is why tgt_name paths start ("targets", ...).

Templates: trainer