Imagenet training using Detectron

[vaishnavmohit]

I am trying to reproduce Imagenet classification accuracy using the Detectron pipeline but unable to do so. Certain things I have tried are:

• Created the pipeline github link
• Trained from scratch and got the max accuracy of ~71.4%
• But while loading the same weight and doing --eval-only I do not get back the same accuracy

Did anyone ever try to do something similar? Is it about the test time augmentation part which I am going wrong?
Looking forward to a discussion or solution.