Cant find files

[joaqo]

For some reason I am getting assertion errors of the kind ckpt_file exp/eanet/test_paper_models/PAP_ST_PS_SPGAN_CFT/duke_to_market1501/ckpt.pth not found, but I am following the instructions pretty carefully. I downloaded all the models to exp/test_paper_models as instructed.

I am not understanding why it looks for the models in in exp/eanet/test_paper_models/ instead of exp/test_paper_models/. Even if I fix this I get a different need more than 0 values to unpack error, which looking at other issues here seems to be an error with how I downloaded the data. I just want to make sure though.

[joaqo]

Also I can't download cuhk03_np_detected_jpg and msmt17. In the first one the site is asking me to download some weird software to download the files, which I cant really understand cause I dont speak chinese, and in the second one they are asking me to prove I work in academy.

I would just ignore these datasets, but for some reason I cant get to run an eval on a single dataset, it asks me to have all the datasets downloaded before I can run any eval.

[huanghoujing]

Thank you for your feedback!

1. It should be exp/eanet/test_paper_models instead of exp/test_paper_models. I just fixed the path in README as well as net disk folder.
2. You can just test on one dataset. E.g. test the GlobalPool model that was trained on Market1501. You can download dataset Market1501 and place it in the position as described in README. Then download the ckpt.pth corresponding to GlobalPool model that was trained on Market1501, it is this link. Place the ckpt.pth in the exp_dir directory of the following command. Then run

   cd ${project_dir}
   CUDA_VISIBLE_DEVICES=0 python -m package.optim.eanet_trainer --exp_dir exp/eanet/GlobalPool/market1501 --cfg_file package/config/default.py --ow_file paper_configs/GlobalPool.txt --ow_str "cfg.dataset.train.name = 'market1501'; cfg.dataset.test.names = ['market1501']; cfg.only_test = True"

   The item cfg.dataset.test.names in configuration file package/config/default.py specifies which datasets to test on.
3. The links for CUHK03-NP, Partial-REID and Partial-iLIDS are Baidu Cloud. Are these datasets that you find unable to download? If you need to use these datasets, I may upload them to Google Drive as well. (For MSMT17 dataset, I have uploaded it to the link in README.)

[joaqo]

Hey @huanghoujing thanks a lot! Tried your command, worked perfectly!

One question, which of the models would you recommend for someone that is just trying to use this as a generic Person-ReId algorithm? Is there any one model that generalizes better to unseen domains than the other ones? I read your paper and noticed you mentioned that the Duke dataset has people dressed in coats and trousers, I guess this would be more similar to the use case I want to test, but still the difference in the domains may be huge, so I don't know if this makes a real difference.

Also, regarding your last question. Personally I am not planning to train the models, so I am not sure I would use those datasets, but I can say that most people who don't speak Chinese would have a very hard time navigating the site and downloading them. Specially considering that Baidu Cloud asks the user to download a software downloader in order to download the dataset, I think this may put some people off.

Thanks again for the prompt reply and the awesome project!

[joaqo]

Also, if there is some guide on how to use the project just for inference, I'd be very thankful if you could point me in its direction. I understand that this was probably no the intended use of the project, so there probably is no such guide, buy asking just in case I missed something!

[huanghoujing]

For cross-domain exploitation, model trained on MSMT17 seems to be the best. I think I have to provide an example to extract feature for single images recently.

[joaqo]

Hey I am using the inference api you just pushed, its great! One question, what is the most precise configuration I can use given that I plan to use MSMT17 and pose detection ?