MLCommons Algorithmic Efficiency is a benchmark and competition measuring neural network training speedups due to algorithmic improvements in both training algorithms and models.
The submission_runner module has an obscure check in train_once to set imagenet_v2_data_dir to None for non-imagenet workloads. It only performs the check for the external_tuning ruleset for some reason.
As a result the with the self-tuning ruleset other workloads are being passed the default str for imagenet_v2_data_dir as the test_dir and breaking in the test eval.
To fix, this just set the default value for imagenet_v2_data_dir in the flag definition to None.
The submission_runner module has an obscure check in train_once to set
imagenet_v2_data_dir
to None for non-imagenet workloads. It only performs the check for the external_tuning ruleset for some reason. As a result the with the self-tuning ruleset other workloads are being passed the default str for imagenet_v2_data_dir as the test_dir and breaking in the test eval.To fix, this just set the default value for
imagenet_v2_data_dir
in the flag definition to None.