Open
Description
Hi,
I tried to reproduce the r2b model results, but I couldn't reach 65.04% top-1 accuracy.
In fact, I couldn't even reproduce the first stage which is training original ResNet18 model (expected to achieve 70.32% validation accuracy (#196 (comment))).
It seems like the default hyper-parameter settings provided in the model zoo is not the optimal setting.
Can you provide the hyper-parameter settings for each stage in training r2b model?
P.S.) The training curve below is what we obtained by running the default script provided in larq model zoo. We could only achieve 66.03% top-1 val. accuracy with the default setting.
Metadata
Metadata
Assignees
Labels
No labels
Activity