About parameter on Pascal.

[prettydong]

No description provided.

[prettydong]

Your work is very interesting !
would you share precision parameters on Pascal with me ? Thank you!

[prettydong]

I use rel 0.2, lr 1e-5, backbone imgaenetV2 of resnet50 , epoch 15.
then i get Test mAP : 89.124 😊

[jameslahm]

Hi, It's under "End-to-End" training setting?

[prettydong]

Hi, It's under "End-to-End" training setting?

Yes, 'end2end' is defualt setting.
I run the code on a PC with single RTX3060 12G not 4GPUs in paper.
And it seems that learning rate of last linear layer is different from resnet backbone.

[jameslahm]

It seems that the paper also uses a single GPU on pascal too and only use 4 GPUs on openimages.

And it seems that learning rate of last linear layer is different from resnet backbone.

Yes, the learning rate of last linear layer is 10 times larger than resnet backbone.

[jameslahm]

I use rel 0.2, lr 1e-5, backbone imgaenetV2 of resnet50 , epoch 15.
then i get Test mAP : 89.124 😊

Hi, @prettydong , did you use LL-R method？I try these hyperparameters under LL-R and only get 87.795 test mAP. Did you change somewhere else? Thanks!

[prettydong]

I use rel 0.2, lr 1e-5, backbone imgaenetV2 of resnet50 , epoch 15.
then i get Test mAP : 89.124 😊

Hi, @prettydong , did you use LL-R method？I try these hyperparameters under LL-R and only get 87.795 test mAP. Did you change somewhere else? Thanks!

I use LL-R under 0.2rel and 1e-5 lr ,other parameter is defualt. Then I can get 88 mAP. The torch warning me that I use old backbone , then I use IMAGENET1K_V2 which resize image to 232, you can find details on "pytorch/vision#3995 (comment)".
->models.py:22---feature_extractor = torchvision.models.resnet50(weights=ResNet50_Weights.DEFAULT).
This backbone spends more times. And best performance occures on 12-13th epoch. Then mAP get 89.1.
I plan to do some incremental work on this paper. 😂I'd be happy to discuss the details of this paper with you.

[youngwk]

Thanks for your interest.
For LL-R, I used delta_rel=0.1, lr=2.5E-03 with SGD optimizer.

[prettydong]

Thanks！

[jameslahm]

I use rel 0.2, lr 1e-5, backbone imgaenetV2 of resnet50 , epoch 15.
then i get Test mAP : 89.124 😊

Hi, @prettydong , did you use LL-R method？I try these hyperparameters under LL-R and only get 87.795 test mAP. Did you change somewhere else? Thanks!

I use LL-R under 0.2rel and 1e-5 lr ,other parameter is defualt. Then I can get 88 mAP. The torch warning me that I use old backbone , then I use IMAGENET1K_V2 which resize image to 232, you can find details on "pytorch/vision#3995 (comment)". ->models.py:22---feature_extractor = torchvision.models.resnet50(weights=ResNet50_Weights.DEFAULT). This backbone spends more times. And best performance occures on 12-13th epoch. Then mAP get 89.1. I plan to do some incremental work on this paper. 😂I'd be happy to discuss the details of this paper with you.

Thanks! I'd be happy too!