Frederic Rechtenstein d8f4e9a18e Quantization tests: compare TFLite model output agains pre-computed golden values.

Testing float16 quantization of MobileNet v1 and v2

PiperOrigin-RevId: 347966786
Change-Id: I4d984f237e81a3094a0d770bb6bad0eec6e8ac3a

2020-12-16 23:26:25 -08:00

TensorFlow Lite model coverage tests

Various conversion tests on popular mobile models.

Golden values

Some tests rely on pre-computed golden values. The main goal is to detect changes affecting unintended parts of TFLite.

Should a golden value test fail after an intended change, the golden values can be updated with the following command:

bazel run //third_party/tensorflow/lite/testing/model_coverage:<target> --test_output=all -- --update_goldens

Notice bazel run instead of bazel test and the addition of the --update_golden flag.

The updated golden data files must then be included in the change list.