MLPerf reference convergence point (RCP) checker
Run Reference Convergence Point checks for a submission directory. This consists of testing whether a submission does not converge statistically faster than the reference.
For training, RCPs are loaded from directory mlperf_logging/rcp_checker/training_4.1.0/*.json
The RCP checker supports only the 1.0.0 version onwards. The current training version is 4.1.0.
From the base directory of the repo:
python3 -m mlperf_logging.rcp_checker FOLDERFOLDER is a results directory, i.e. it contains the correct number of result log files for a benchmark.
You need numpy and scipy python3 -m pip install numpy scipy
python 3.9.2 python 3.9.10