DepthGauger is a benchmarking library for deep learning models which runs on real Android devices, either locally or attached to a cloud test platform (like Microsoft App Center).

The generated dummy app APK needs to be signed to work with MS App Center. A guide for generating a signing key can be found here.

Once you have a signing key, you should define depthgauger_keystore_path, depthgauger_keystore_password, depthgauger_keystore_key_alias, and depthgauger_keystore_key_password in your ~/.gradle/gradle.properties file:

depthgauger_keystore_path=...
depthgauger_keystore_password=...
depthgauger_keystore_key_alias=...
depthgauger_keystore_key_password=...

If you're using Windows, remember to escape your slashes in depthgauger_keystore_path.

Add a compiled ONNX Runtime AAR file called onnxruntime-release.aar to the onnxruntime-release directory.

The deep learning framework you'd like to test is referred to as follows:

• <LOWERCASE_FRAMEWORK> is onnx, pytorch, or tensorflow, corresponding to ONNX Runtime, PyTorch Mobile and TensorFlow Lite
• <INTIAL_UPPERCASE_FRAMEWORK> is Onnx, Pytorch, or Tensorflow
• <CAMELCASE_FRAMEWORK> is Onnx, PyTorch, or TensorFlow
• <UPPERCASE_FRAMEWORK> is ONNX, PYTORCH, or TENSORFLOW

The dependencies for each framework contain a lot of functionality that aren't needed by all models. Consider using a custom build for each framework in order to produce a much slimmer library:

• ONNX Runtime
• PyTorch Mobile
• TensorFlow Lite

An example config file is shown below:

{
    "framework": "PYTORCH",
    "method_name": "train_step",
    "model_filename": "swiftkeyrnn.pt",
    "input_parameters": [
        {
            "name": "data",
            "parameter_type": "LongFlatTensorConstantParameter",
            "shape": [2, 3],
            "values": [1, 2, 3, 4, 5, 6]
        },
        {
            "name": "learning_rate",
            "parameter_type": "FloatConstantParameter",
            "shape": [1],
            "values": [0.01]
        }
    ],
    "output_parameters": [],
    "supplemental_files": [],
    "properties": {}
}

• framework is one of the values in <UPPERCASE_FRAMEWORK>.
• method_name is the name of the method to call in the model. If unknown, please use forward.
• model_filename is the filename of the model for which this config file applies.
• input_parameters are passed when calling the method specified in method_name. Possible parameter_type values and the schemas around them can be found in the com.microsoft.depthgauger.parameters package. The values for each input parameter are flat arrays. They can be rearranged into multiple dimensions via the shape property.
• output_parameters are similar to input_parameters, for the outputs produced when calling the method specified in method_name.
• supplemental_files is currently unused and should be an empty array.
• properties is currently unused and should be an empty map.

The path to your model file is referred to as <MODEL_FILE_PATH>. The filename should have a .ort, .pt, or .tflite extension.

The part to the config file for your model is referred to as <CONFIG_FILE_PATH>.

A dummy signed release app APK can be built for uploading to cloud test platforms. To generate a dummy app APK, run:

gradlew dummyapp:assembleRelease

The benchmark release APK contains only the necessary files and dependencies in order to keep the file size as small as possible. To generate the benchmark release APK, run:

gradlew clean benchmark:assemble<INTIAL_UPPERCASE_FRAMEWORK>ReleaseAndroidTest -Pmodel_file=<MODEL_FILE_PATH> -Pconfig_file=<CONFIG_FILE_PATH>

It is strongly recommended to clean before each build, especially if changing framework.

Install the benchmark APK:

adb install -r benchmark/build/outputs/apk/androidTest/<LOWERCASE_FRAMEWORK>/release/benchmark-<LOWERCASE_FRAMEWORK>-release-androidTest.apk

Then start the benchmark instrumentation test:

adb shell am instrument -w -e androidx.benchmark.profiling.mode none -e debug false -e class 'com.microsoft.depthgauger.benchmark.<CAMELCASE_FRAMEWORK>Benchmark' com.microsoft.depthgauger.benchmark.test/androidx.benchmark.junit4.AndroidBenchmarkRunner

Create an Espresso test, follow the setup instructions for the CLI, ensure that you're logged in, and then run:

appcenter test run espresso --app "<APP CENTER PROJECT NAME>" --devices <DEVICES> --app-path <FULL PATH TO DUMMY APP APK FILE> --test-series "master" --locale "en_US" --build-dir <FULL PATH TO BENCHMARK APK DIRECTORY>

The paths to the dummy app APK file and the benchmark APK directory must be absolute. The dummy app APK file is at dummyapp/build/outputs/apk/release/dummyapp-release.apk, and the benchmark APK directory is at benchmark/build/outputs/apk/androidTest/<LOWERCASE_FRAMEWORK>/release.

Results are output to the device logs, via both our own test time recorders and the Benchmark library too.

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.