Regarding Validation Code for PASCAL VOC

Alibaba-MIIL / ASL

Official Pytorch Implementation of: "Asymmetric Loss For Multi-Label Classification"(ICCV, 2021) paper

MIT License

732 stars 102 forks source link

Regarding Validation Code for PASCAL VOC #62

Closed indersingh17188 closed 3 years ago

indersingh17188 commented 3 years ago

Hello again,

Thank you so much for your quick feedback on my last issue. Appreciate it.

I would like to know if you there is any validation code for PASCAL-VOC dataset in order to reproduce the same results.

Thanks, Inder

mrT23 commented 3 years ago

i released the model of pascal.

regarding validation code, it is very similar to MS-COCO code https://github.com/Alibaba-MIIL/ASL/blob/main/validate.py just load the PASCAL dataset instead

indersingh17188 commented 3 years ago

Ok, thank you so much for your quick response.

indersingh17188 commented 3 years ago

Hello again,

I tried to load the PASCAL VOC dataset and train TResNet-XL using ImageNet Pretrained weights but after 3 or 4 epochs, the Loss started to go negative by a huge number and max accuracy achieved was 82% (using the same mAP function as for MS-COCO). Could you please, from a high level perspective, share your thoughts on what is going wrong here?

Thanks in advance.

Regards, Inder

mrT23 commented 3 years ago

to be honest, i don't know. but you can try several things (do all of them)

move to tresnet_m. dont start training with a large model. make everything work first on a smaller model
move to 21K pretraining https://github.com/Alibaba-MIIL/ImageNet-21K-P
use adam with a learning rate of 5e-5 instead 3e-4
make sure your code has proper warmup, and that you are using EMA

indersingh17188 commented 3 years ago

Hello again,

There was a problem with the dataloader itself. Now it is resolved hence closing the issue.

Thanks for your quick feedback.

Regards, Inder

indersingh17188 commented 3 years ago

Hello again,

Congratulations on the ICCV acceptance. Great work indeed.

I would like to know if you could share the dataloader source code for loading VOC dataset as the one I am using I am unable to reproduce the 96% using XL.

Thanks in advance.

Best regards, Inder

mrT23 commented 3 years ago

Hi Inder i am sorry, but i won't engage code issues further, i have answered more than 60 issues.

the results on Pascal Voc were reproduced and even surpassed by a later article (which used ASL): https://arxiv.org/pdf/2107.10834.pdf maybe check their repo for code or code-related questions