Link

Model Evaluation Design

This document describes the design of model evaluation task for ElasticDL.

Minimal Viable Product

Definitions

  • Model evaluation: Computing metrics to judge the performance of the trained model.
  • Evaluation worker: The worker responsible for performing model evaluation task.
  • Multiprocessing: Executing tasks in multiple threads in parallel on the same pod.

Requirements

  • There‚Äôs only one evaluation worker without multiprocessing.
  • Master pod is responsible for creating the evaluation worker.
  • Evaluation worker is created by master pod together with the workers for training.
  • Evaluation starts after a specified warm-up period and on a given time interval. For example, we need to expose the following parameters to users:
    • start_delay_secs: Start evaluating after waiting for this many seconds.
    • throttle_secs: Do not re-evaluate unless the last evaluation was started at least this many seconds ago.
  • The evaluation worker fetches the latest model from master pod.
  • Model can be evaluated by a specified number of steps or batches of evaluation samples. If None, evaluation will continue until reaching the end of input.
  • Model evaluation metrics can be defined by users together with the model definition.
  • The computed model evaluation metrics can be report back to master through RPC call.

Implementation Plan

  • Implement MasterServicer.ReportEvaluationMetrics() and additional proto definitions such as ReportEvaluationMetricsReply and ReportEvaluationMetricsRequest.
  • Extend Worker to support the following:
    • distributed_evaluate() that contains the main logic for model evaluation.
    • report_task_result() that reports evaluation task result (e.g. task id and error message) back to master through RPC call.
    • report_evaluation_metrics() that reports the computed evaluation metrics (e.g. accuracy, precision, recall, etc.) back to master through RPC call.
  • Add main CLI entry-point to Worker.distributed_evaluate() that will be used in WorkerManager.
  • Extend WorkerManager to support the following:
    • Instantiate a separate evaluation task queue from evaluation data directory.
    • Start an evaluation worker from evaluation task queue.
    • Update master.main() to support model evaluation task if user requested.

Future Development

A list of potential features we may want for model evaluation in the future:

  • num_parallel_processes: The number of children processes to run evaluation on each individual evaluation worker.
  • sample_weights: Optional Numpy array of weights for the test samples, used for weighting the loss function.

References

Some of the ideas are borrowed from existing solutions listed below: