Predict

talos.Predict selects a trained model from Scan or a RunResult and performs inference on caller-supplied features. Import it from talos. Construction stores the run; prediction happens when .predict() or .predict_classes() is called.

Prerequisites are a completed run with a recoverable model and its installed framework backend. Apply the same preprocessing and feature structure used for training; this wrapper does not reconstruct caller-owned preprocessing automatically.

Examples below use the held-out Iris setup in Scan → Minimal Example. Run that setup first; it defines scan_object, p, input_model, x, y, x_val, y_val, x_test, and y_test.

predictor = talos.Predict(scan_object)
probabilities = predictor.predict(x_test, metric='val_loss', asc=True)

Prediction methods

predict returns the trained model's raw outputs for x. These are probabilities only when the model produces probabilities; regression predictions and logits remain unchanged.

predictor.predict(x_test, metric='val_loss', asc=True)

---

predict_classes makes class predictions on x which has to be in the same form as the input data used in the Scan() experiment.

predictor.predict_classes(x_test, metric='val_loss', asc=True, task='multi_class')

Arguments

Parameter	Default	Description
x	required	Features in the model's expected input structure.
metric	required	Scan result column used to select a model.
asc	required	True for lower metric values, False for higher values.
model_id	None	Explicit result-row index; otherwise select by metric.
task	required for predict_classes	Class conversion rule described below.
saved	False	Recover from persisted artifacts rather than a retained live model.
custom_objects	None	Keras objects needed for reconstruction.
model_factory	None	Torch reconstruction factory when needed.
**kwargs	empty	.predict() only: forwarded to backend prediction, such as Keras batch_size.

Signatures and class conversion

The constructor is Predict(scan_object). Its methods are predict(x, metric, asc, model_id=None, saved=False, custom_objects=None, model_factory=None, **kwargs) and predict_classes(x, metric, asc, task, model_id=None, saved=False, custom_objects=None, model_factory=None).

Task	Class conversion
binary	argmax for two outputs; otherwise threshold each value at 0.5.
multi_class, multiclass, multi_label	argmax across the last axis for multidimensional output; one-dimensional output becomes integers.
multilabel, multi_label_independent	Independent 0.5 thresholds.
continuous, regression	Return the raw predicted values.

The historical multi_label spelling means exclusive class conversion here. Use multilabel or multi_label_independent for independent labels. Thresholds require suitably scaled outputs; Torch logits are not transformed into probabilities automatically. Structured multi-output predictions can be obtained through .predict(); class conversion expects a single array.

Prediction returns outputs and does not add columns to the result table or retrain the model. Tied selection metrics preserve trial order, and missing values are excluded. Missing metric columns raise KeyError; an unusable selection or unretained model raises ValueError. Unknown class tasks raise ValueError, and framework shape or restoration errors propagate.

Read next

See Evaluate for held-out scoring, Backends for native inference behavior, or Deploy to archive the selected trained model.