Models

TorchTrainer accepts a native torch.nn.Module. Construct its optimizer and loss before creating the trainer; a basic supervised batch is (x, y) and the trainer calls model(x) and criterion(predictions, y).

class xflow.models.base.BaseModel

Optional abstract interface for an application-defined model. A subclass implements all of these methods:

• predict(inputs, **kwargs): run inference.

• save(path) and classmethod load(path, **kwargs): persist and restore.

• training_step(batch): perform one update and return a loss or metrics.

• validation_step(batch): return validation loss or metrics.

• configure_optimizers(): return application-specific optimizer objects.

set_train_mode(training=True) is an optional override. A metrics dictionary returned by training_step must contain "loss".

BaseModel cannot be instantiated directly. TorchTrainer uses the native PyTorch interface and does not dispatch to these abstract step methods or BaseModel.save. Use a custom trainer or ModelIO adapter when integrating that model contract.