Low results on i3d non local

[patykov]

Hi! I'm working on a pytorch implementation of the i3d non-local network and I'm facing a similar problem as yours. I've created the network and copied the weights, everything looks fine except for the low results on testing. I was wondering if you have any updates or ideas about your problem, since it's possible we are facing the same one.

[patykov]

Hi, removing the non-local blocks and coping the weights for the I3DResNet50 baseline also provided by the authors in https://github.com/facebookresearch/video-nonlocal-net I have achieved the same low results. I am getting 48%/76% (top1/top5) for the baseline and the non-local version. How much are you getting? Did you copy the weights from other source?

Em qua, 1 de mai de 2019 às 19:05, Tushar Nagarajan < [email protected]> escreveu:

@patykov https://github.com/patykov I'm headed down a similar direction, starting just at copying weights from a res-101 I3D model (without nonlocal blocks). Have you had any success setting that up (before you attempted the non-local version)?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/17Skye17/Non-local-Neural-Networks-Pytorch/issues/1#issuecomment-488449034, or mute the thread https://github.com/notifications/unsubscribe-auth/ABUJU2A6BWXYQA6HK4K3OLTPTIHZ7ANCNFSM4HIO6VBA .

[Tushar-N]

Oops, looks like my comment went missing. I managed to port the weights from the baseline I3D-ResNet50 model (not non-local) and got 72.2% top1. The converted model is here.

[17Skye17]

Hi, @patykov @Tushar-N, sorry for late reply, I've been busy the last two weeks. Thank you for your comments, I'll check my results again and find out what is wrong in my experiments this week. Haha, I'm getting excited to solve this problem!