code for detection and colorization about visual_prompt_retrieval

[ZhangYuanhan-AI]

Already uploaded

[chmxu]

Can you please provide some detailed instruction on how to run the code on these two tasks? Thanks!

[chmxu]

Sorry to bother you again, I wonder if featextrater_det.py is used for unsupervised feature extraction and featextrater_det_cont.py is used after the SupCon model is trained?

[ZhangYuanhan-AI]

Sorry to bother you again, I wonder if featextrater_det.py is used for unsupervised feature extraction and featextrater_det_cont.py is used after the SupCon model is trained?

Hi, what is the “ featextrater_det.py?”

[chmxu]

https://github.com/ZhangYuanhan-AI/visual_prompt_retrieval/blob/det/tools/featextrater_det.py

Sorry to bother you again, I wonder if featextrater_det.py is used for unsupervised feature extraction and featextrater_det_cont.py is used after the SupCon model is trained?

Hi, what is the “ featextrater_det.py?”

[chmxu]

Hi Yuanhan, I am trying to reproducing the colorization task now. I find that the origin MAE-VQGAN randomly samples from ImageNet validation set for both the support and query samples. You paper mentioned 'For all experiments, in-context examples come from the training set'.

As I can understand, a reasonable pipeline would be training the SupCon model using support-query pairs from training set, and test it with pairs from validation set. I wonder which is the true setting for this experiment.

I would appreciate it if you could help me with this problem. Thank you very much!

[ZhangYuanhan-AI]

Hi Yuanhan, I am trying to reproducing the colorization task now. I find that the origin MAE-VQGAN randomly samples from ImageNet validation set for both the support and query samples. You paper mentioned 'For all experiments, in-context examples come from the training set'.

As I can understand, a reasonable pipeline would be training the SupCon model using support-query pairs from training set, and test it with pairs from validation set. I wonder which is the true setting for this experiment.

I would appreciate it if you could help me with this problem. Thank you very much!

Hi, support image comes from training set.

[chmxu]

So in your case, I have to calculate a 1.3M(or perhaps the randomly chosen 50000)*50000 similarity matrix, pick top-50 for each test sample, then use the trained SupCon model to choose the best support sample. Is this right?