generic

Adapt an estimator, callable, or torch.nn.Module to Anomx.

fit_fn(model, x, y) and predict_fn(model, x) explicitly override the underlying implementation. Estimators otherwise use their fit and predict methods; callable models support inference without fitting. The observation axis is always first. Prediction must preserve its length, while the remaining axes may change (for example for embeddings).

PyTorch modules receive float32 inputs and train with Adam/MSE unless an optimizer factory or loss function is provided. Classification uses integer class targets and cross-entropy; prediction returns class logits. Reconstruction models may omit targets; other families require them. No scaling, window creation, or forecast alignment is implicit. Use TorchForecastingModel for sliding-window forecasting. Torch remains an optional dependency.

Parameters

model

Estimator, callable, or torch module owned by this adapter.

family

predictive, forecasting, reconstruction, representation, or classification. Describes purpose without restricting tensor axes.

epochs, batch_size, learning_rate

Positive PyTorch training settings; ignored by estimator backends.