@girafffeee Thanks for your work! However, I am confused about what this loss function to do: https://github.com/TencentYoutuResearch/PersonReID-NAFS/blob/71bf192f464a410845a88590e4660eeeba1a1e50/utils/metric.py#L265-L286

Why the criterionand t2i_similaritiesare not used in this function? Does this function do the same thing like CSAL in the paper?

I have downloaded pretrained Resnet50 model, bert-base-uncased model and vocabulary to project_directory/pretrained/.But when i run train.sh,it showed that '../pretrained/bert-base-uncased/bert-base-uncased-vocab.txt not found in model shortcut name list''.I attach the picture of the problem,please take a look at the problem. Is it a problem with the file directory structure?Great thanks if get your help!

Hi all, @X-ing @girafffeee

I found the memory increased with steps (about 20MB for several ten steps). This is not a good thing. Did you encounter the same phenomenon?'

BTW, would you please take a little time to clean this repo? The code is a little bit messy (too many unused variables and imports, no lint and so on)...

Thanks

Hello, thanks for sharing your code! :> I have two questions.

1. I found out that in the uploaded training log, arguments given to "--part2" and "--part3" are 3 and 2 (not 2 and 3). Is it just a typo?

2. As I have only TitanXP GPUs, when I run the code with 64 batches, I got 'runtime: CUDA out of memory' error. so I changed the batch size to 32 and it works. However, at 16 epochs suddenly loss increased and it did not decrease at all. Any idea about it? I set the environment the same as you specified in the readme.

First of all, thank you for sharing the code. But I have a bug running test.sh. I have a screenshot below. The bug is that best_model is being cast to an int. But I looked at model_path and there are only two files, model_best and best_model.pth.tar. There is nothing to convert to an int. Is there any problem with my file location?

Thanks for your code, but I have some confusion about the Jaccard distance API used in your testing code.

https://github.com/TencentYoutuResearch/PersonReID-NAFS/blob/5301b8d95bb711f88357d71cff1a9aaaac76f473/utils/metric.py#L482

As we can see in the document of SciPy, scipy.distance.jaccard is used to compute the Jaccard-Needham dissimilarity between two boolean 1-D arrays. It seems that it is not the expected implementation of the Jaccard Distance in your paper. Is it ok to directly use this API?

The last serveral layers of Resnet are not shared between the different scales, which seems like the final performance improvement mainly comes from random ensembling. May you report the results with the shared Resnet?

Please, what is the reason for this? File "/home/bobo/E/xxx/paper4_reid_2021/PersonReID-NAFS-main/models/model.py", line 76, in forward local_text_feat = local_text_feat.view(-1, n_sep, local_text_feat.size(1)) RuntimeError: shape '[-1, 2, 768]' is invalid for input of size 33024

Thanks for sharing your brilliant work! I see that in the train.log you shown, top1_t2i suddenly dropped to 0.244 when epoch was equal to 17, but top1_t2i returned to a normal value of 54.613 when epoch was equal to18. Could you please tell me what causes this problem in training?

I trained the model without any changes, with the default training strategy (set in run.sh), but obtained 54.2% R@1. Would you please tell me how to achieve the reported 61.5% R@1 performance as in paper? Should I change the training strategy or some hyper-parameters? Great thanks if get your help.