tfx_bsl.public.beam.run_inference.RunInferencePerModel

Vectorized variant of RunInference (useful for ensembles).

tfx_bsl.public.beam.run_inference.RunInferencePerModel(
    examples: beam.pvalue.PCollection,
    inference_spec_types: Iterable[tfx_bsl.public.proto.model_spec_pb2.InferenceSpecType],
    load_override_fn: Optional[run_inference.LoadOverrideFnType] = None
) -> beam.pvalue.PCollection

examples A PCollection containing examples of the following possible kinds, each with their corresponding return type.

  • PCollection[Example] -> PCollection[ Tuple[PredictionLog, ...]]

    • Works with Classify, Regress, MultiInference, Predict and RemotePredict.

  • PCollection[SequenceExample] -> PCollection[ Tuple[PredictionLog, ...]]

    • Works with Predict and (serialized) RemotePredict.

  • PCollection[bytes] -> PCollection[ Tuple[PredictionLog, ...]]

    • For serialized Example: Works with Classify, Regress, MultiInference, Predict and RemotePredict.
    • For everything else: Works with Predict and RemotePredict.

  • PCollection[Tuple[K, Example]] -> PCollection[ Tuple[K, Tuple[PredictionLog, ...]]]

    • Works with Classify, Regress, MultiInference, Predict and RemotePredict.

  • PCollection[Tuple[K, SequenceExample]] -> PCollection[ Tuple[K, Tuple[PredictionLog, ...]]]

    • Works with Predict and (serialized) RemotePredict.

  • PCollection[Tuple[K, bytes]] -> PCollection[ Tuple[K, Tuple[PredictionLog, ...]]]

    • For serialized Example: Works with Classify, Regress, MultiInference, Predict and RemotePredict.
    • For everything else: Works with Predict and RemotePredict.
inference_spec_types A flat iterable of Model inference endpoints. Inference will happen in a fused fashion (ie without data materialization), sequentially across Models within a Beam thread (but in parallel across threads and workers).
load_override_fn Optional function taking a model path and sequence of tags, and returning a tf SavedModel. The loaded model must be equivalent in interface to the model that would otherwise be loaded. It is up to the caller to ensure compatibility. This argument is experimental and subject to change.

A PCollection (possibly keyed) containing a Tuple of prediction logs. The Tuple of prediction logs is 1-1 aligned with inference_spec_types.