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: These models are CNN based.

Model	Latency (ms)	Accuracy	Power Usage	Max Memory
selfie_segmentation_landscape_256x256	<= 6	avg_err <=0.0000003 std_dev<=5e-06	TBD	<=100MB
convolution_benchmark_1_144x256	<= 4	avg_err <=0.0003 std_dev <=5e-06	TBD	<=100MB
convolution_benchmark_2_144x256	<= 4	avg_err <=0.0003 std_dev <=5e-06	TBD	<=100MB

Image Search

Note: These models are CNN based.

Model	Latency (ms)	Accuracy	Power Usage	Max Memory
mobilenet_v2_1.0_224	<= 5	avg_err <=0.00005 std_dev <=6e-06	TBD	<=150MB
mobilenet_v2_1.0_224_quant	<= 5	avg_err <=1.5 std_dev <=0.2	TBD	<=150MB

ML accelerator requirements

The API for running ML workloads on ChromeOS is Tensorflow Lite. A discrete ML accelerator such as a TPU/NPU or a GPU can be made accessible through TFLite to improve the performance of ML workloads.

The following requirements apply to such accelerators:

Functional requirements

1. Any device kernel driver must be open source and integrated with upstream Linux or implemented in userspace through VFIO.
2. Direct dmabuf data sharing must be supported between the accelerator and other relevant IP blocks (e.g., GPU, ISP). Both buffer-user and exporter roles must be supported.

Security requirements

1. Sandboxing must be supported for isolating untrusted workloads and any binary-only driver components.
2. Only signed and verified firmware must be allowed to be loaded onto the accelerator. See Peripheral Firmware Security.
3. An IOMMU must control access to system memory from the accelerator.

Miscellaneous requirements

1. The driver's binary size (including dependent libraries and middleware) must be below 64 MB.
2. Tools should be provided for ChromeOS developers to analyze the performance of inference workloads (e.g. Perfetto and/or ftrace instrumentation).