wl-zhao/VPD

Reproducing 4k/8k segmentation results

manuelknott opened this issue · 3 comments

manuelknott commented

Hi,
Can you share the config you used to train 4k and 8k iters? Most importantly, the learning rate schedule you were using?

Trying to reproduce the results.

Thank you.

nkondapa commented

I have the same question

  • ❤️2
manuelknott commented

Using the 80k schedule that you provided in the repo and just reducing the total number of epochs gives ~39 mIoU while you are reporting 43.9 in the paper for 4k steps. So I assume you are using a different config for 4k and 8k iters. Would be very thankful if you could indicate what is different there. Thanks in advance.

  • ❤️1
manuelknott commented

From the paper:
For the fast schedule (8K iterations), we linear scale
the learning rate schedule and set the warming-up iterations
to 150.

Is it the same for 4k? And what does linear scale exactly mean?

  • ❤️1