hawk_eye.inference.benchmark_inference

A script to benchmark inference for both detector and classification models. This is useful for seeing how a new model performs on a device. For example, this might be run on the Jetson to see if the model meets performance cutoffs.

hawk_eye.inference.benchmark_inference.benchmark(timestamp: str, model_type: str, batch_size: int, run_time: float) → None[source]

Benchmarks a model.

This function will load the specified model, create a random tensor from the model’s internal height and width and the given batch then perform forward passes through the model for run_time seconds.

Parameters

  • timestamp – The model’s specific timestamp.

  • model_type – Which type of model this is.

  • batch_size – The batch size to benchmark the model on.

  • run_time – How long to run the benchmark in seconds.

Example

$ PYTHONPATH=. hawk_eye/inference/benchmark_inference.py \
  --timestamp 2020-09-05T15.51.57 \
  --model_type classifier \
  --batch_size 50 \
  --run_time 10

$ PYTHONPATH=. hawk_eye/inference/benchmark_inference.py \
  --timestamp 2020-10-10T14.02.09 \
  --model_type detector \
  --batch_size 10 \
  --run_time 15