Deterministic runs of train.py

Open manisoftwartist opened this issue 4 years ago • 1 comments

Is your feature request related to a problem? Please describe. It is hard to reproduce training results between two different runs when using tools/train.py

Describe the solution you'd like to propose. The mmdetection backend has 2 arguments seed and deterministic which if we set through train.py, can solve this issue.

There is a related question I have created which if answered, can help me implement this feature myself. And I will give a pull request.

Oct 19 '21 01:10 manisoftwartist

@manisoftwartist is this relevant now? If now, I will close the issue.

Feb 16 '22 10:02 morkovka1337

Close this issue, please reopen it if there is any problem.

Apr 20 '23 04:04 sungmanc