search

EPFL-VILAB / ViPer

Apache License 2.0
57 stars 2 forks source link

using the metric network #3

Closed wangkua1 closed 1 month ago

wangkua1 commented 1 month ago

Dear Authors,

Thank you for your excellent work on this project! I'm super interested in using the learned metric network. However, I’ve encountered an issue where the predicted scores seem to be independent of the input queries.

In the provided example script (run_metric.py), it requires images in the liked and disliked folders, which I couldn't locate. To work around this, I cropped the individual images from assets/scores.png. I used the 8 Liked images from User2 as the positive images, and 8 Liked images from User1 as the negative images. I tested the 4 query images, but they all scored around 0.5 (specifically, 0.47 for the top 2 images and 0.51 for the bottom 2).

I would greatly appreciate any guidance you could provide to help identify the source of this issue!

Thanks, Jackson

sogandstormesalehi commented 1 month ago

Hi Jackson,

Glad you like the project! I couldn't reproduce the issue, but I’ve uploaded example images (same scenario you mentioned) and updated run_metric.py with the correct paths. It should work now—please re-clone the repo and give it a try.

Best,

wangkua1 commented 1 month ago

Hey Storme!

Thanks so much for your quick reply! Okay, it works perfectly now ...

It's embarrassing to say what happened, but basically it was a bug I introduced. I wanted to loop over N query images for evaluations, just for convenience. The function calculate_score had this line

context_images.append(Image.open(query_image))

, which modifies context_images in-place. This means that in the 2nd iteration, my context_images has the query_image from the last iteration.

So, being a genius, I did this

context_images = context_images.append(Image.open(query_image))

which actually made context_images None, since append doesn't return anything. And tragically, processor didn't complain, and model failed silently. So now I do this

context_images = context_images.copy()
context_images.append(Image.open(query_image))

Anyways ... thanks a lot for your help!! I can now replicate the scores. I'll close the issue!