Prediction

After a Scenario was trained and evaluated (LAV)<doc.evaluation:load and validate (LAV), a Predictor is used to apply the resulting model on new data. Similar to LAV, a prediction can be performed with the command line (tfaip-predict) or programmatically.

Each resulting Sample of the predictor passes the post_proc-Pipeline.

Command-Line

tfaip-predict --export_dir PATH_TO_EXPORTED_MODEL --data.DATA_GENERATOR_PARAMS --predict.PREDICTOR_PARAMS --pipeline.PIPELINE_PARAMS

tfaip-predict requires at least the path to the exported model (--export_dir). The --data flag is used to modify the input data, the --prediction flag sets up the predictor. Another optional flag is --pipeline which allow to specify the DataPipelineParams, e.g., the number of threads or batch size.

Programmatically

Create a Predictor by calling Scenario.create_predictor(model: str, params: PredictorParams). (Overriding Scenario.predictor_cls() can be used to customize a Predictor). The resulting object can be used to predict given DataGeneratorParams, raw input data predict_raw which is basically the output of a DataGenerator, to predict a DataPipeline (predict_pipeline), or to predict a tf.data.Dataset (predict_database). The output of each function is a generator of Samples.

A predicted Sample holds inputs and outputs. Optionally, if available in the data and if the predictor_params.include_targets flag is set, also the targets.

For example: