tree: 6ead4ab09ee40906bd6ee0a83dfd3aee8f1ca26b [path history] [tgz]
  1. README.md
ml_benchmark/model_zoo/README.md

ChromeOS ML Model Zoo

This is a collection of TFLite models that can be used to benchmark devices for typical ML use cases within ChromeOS. Where applicable, baseline figures are provided to indicate the minimum performance requirements for these models to meet the user experience goals of those use cases.

These models can be easily deployed to /usr/local/share/ml-test-assets on a DUT via the chromeos-base/ml-test-assets package:

emerge-${BOARD} ml-test-assets && cros deploy <DUT> ml-test-assets

The models can be downloaded directly here

Tools

Latency, Max Memory

Latency and maximum memory usage is measured by the TFLite Benchmark Model Tool.

This is installed by default on all ChromeOS test images.

Example usage:

benchmark_model --graph=${tflite_file} --min_secs=20 <delegate options>

Accuracy

Accuracy is measured by the TFLite Inference Diff Tool.

This is installed by default on all ChromeOS test images.

Example usage:

inference_diff_eval --graph=${tflite_file} <delegate options>

Use Cases

Video Conferencing

  • Note 1 : These models are CNN based.
  • Note 2 : selfie_segmentation_landscape_256x256 has an F16 and F32 variant, indicated by the filename.

The convolution_benchmark_*_144x256 models are production model graphs with randomized weights, so we don't measure the accuracy of these models.

ModelLatency (ms)AccuracyPower UsageMax Memory
selfie_segmentation_landscape_256x256<= 6avg_err <=0.0000003
std_dev<=5e-06
TBD<=100MB
convolution_benchmark_1_144x256<= 4-TBD<=100MB
convolution_benchmark_2_144x256<= 4-TBD<=100MB

Image Search

Note: These models are CNN based.

ModelLatency (ms)AccuracyPower UsageMax Memory
mobilenet_v2_1.0_224<= 5avg_err <=0.00005
std_dev <=6e-06
TBD<=150MB
mobilenet_v2_1.0_224_quant<= 5avg_err <=1.5
std_dev <=0.2
TBD<=150MB

Audio Models

Note: These models are running on CPU in production

Note2: While running benchmark_model with following models, add --run_delay=<secs> to simulate audio server behavior.

ModelLatency on CPU (ms)Extra argumentssha256
lstm<= 1--run_delay=0.01a1f1329501c0a87dff6a20d3b330cb73e85ccc23a5c36880c81476e2fb338fd2
seanet_wave<= 2--run_delay=0.027bb40d8e72471a13324491777e03207646f1641942d373d40478d237e87d032d
seanet_stft<= 2--run_delay=0.02a3ea8c3eae3373cb9ef4ac46d22ad5a254aa2e40d764a8f6dbee218be27f9b31