eli-yili / pmm Goto Github PK

Hi @Eli-YiLi , Thanks for sharing your nice work!

I notice that you report the CAM result on ResNet38d. However, in your released code, you only use the resent38d to generate CAM at training multiscale stage. Then, you use the scalenet101 as backbone to train the network at multi-crop stage. So the CAM result on ResNet38d (57.32%) is achieved with a hybrid manner (First train on resnet38d, followed by scalenet101)? I think only train with the resnet38d should be more appropriate.

When I generate mask on COCO2017. It seems need to a lot of time to finish this work. I use 8 T4 or 2 RTX3090 to test the speed. Both of them get 200 .npy files for an hour. It means I need nearly 11 days to generate mask. Could you tell me. What can I do for speed up this code？Thanks.

Greetings! I am your truly sincere follower!
I found that the training cost of COCO14 classification is to heavy. Could you please share the trained COCO14 classification model?

Greetings!

We are making some incremental work on your PMM. The PPMG's performance is amazing! But we find that your provided pseudo masks for VOC2012 and COCO2014 do not contain pseudo masks of val set. And we wonder about more details of the PPMG pseudo masks on val set for these two datasets. We will appreciate it if you could answer some of our questions!

1. If the 'pseudo_mask.tar' in the onedrive correspond to the 61.49% val mIoU in Tab. 2 of the paper?
2. If the answer to 1) is not, what is the mIoU of the 'pseudo_mask.tar' on the val set?
3. what is the mIoU of the 'ppmg_coco.zip' in the baiduyun on the val set?
4. If you are busy, could you provide the pseudo masks on val sets of VOC2012 and COCO2014 to us?

Thanks again for your patience and help!

I saw the result of LIID reported in your paper, may I ask how to reconstruct it? There was no body help me. I was too tired.

Hello, I am recently reproducing your paper, where is the Pretended Under-fitting Strategy module reflected in the code? Thank you!

Thank you for sharing your great work!

Now, I gonna try to run your source code, however, the OneDrive link is not valid.

Could you update the link or share using other cloud services such as google-drive or dropbox?

I cannot access Baidu :(

Thanks for your perfect work! Could you provide the code for converting COCO2014 to VOC format?

Hi, Thanks for your code!

When I used your code to run experiments on the COCO dataset. We used the 8 v100 32g card. Although I set the bs=4,
When infer, eval cam and PPMG, I had met the following error,

THCudaCheck FAIL file=/pytorch/aten/src/THC/THCCachingHostAllocator.cpp line=296 error=2 : out of memory
Segmentation fault (core dumped)

Could you please give me some advice?

Best wishes to you!

Hi, thanks for your great work.

while I trained the scalenet101 with SEAM, it can't capture the person.

The model only localizes the corner of the image for the person's CAM.

And I used the pre-trained weight.

Thanks.

Where can I download this json file? I did not find it in your project

Hi, thanks for the great work. But when I predict the result based on the released weights, it is difficult to find the person category. And the IoU of person is 0, why?

Hello, I am very interested in your work. I want to do some further study. Can you share the model weights and semantic segmentation results on COCO dataset.
Thanks a lot !

Hello ，thanks to the code release.But I cannot find the models and coco14 document.I am looking for the ones .thank you.
Best wishes to you