TensorFlow.js Benchmark Tools

This is a micro benchmark to measure the performance of TensorFlow.js kernel ops.

Benchmarks headless usage

Benchmarks will run nightly on browserstack automatically via yarn test-travis. When running benchmarks locally, use:

yarn benchmark

This will not write to firebase, it will simply log what would have been written to firebase.

If you want to push to firebase locally, add --travis to the test script. Be careful, as this will overwrite other runs of the day.

On travis, during crons, we run:

yarn benchmark-travis

This will run the benchmarks with the --travis flag flipped and write to firebase. Note, this only happens on a cron.

Benchmarks UI usage

While inside the ui directory, install all dependencies.

$ yarn

Launch the server to host benchmark application.

$ yarn server

http://localhost:8080 shows the benchmark tool for various kind of kernel ops.

• Batch Normalization 3D
• Matrix multiplication
• Convolutional Ops
• Pooling Ops
• Unary Ops
• Reduction Ops

Each benchmark suite runs kernel ops with specific size of input in target backend implementation. This benchmark tools support following backends for now.

• CPU
• WebGL