Benchmarking

[thekevinscott]

We've had some queries around performance of the library.

I think that having public benchmarks, across different hardware, would be extremely useful.

I imagine we'd want to test images at different resolutions (say, 128, 256, 512); different models at different scales; different browsers; and differing devices (on-device, not simulators).

We could use something like BrowserStack (which I believe supports running on actual devices) and something like benchmark.js.

I think performance would be different from accuracy - I imagine the latter would be better handled in conjunction with how we develop new models for the library.