snucml / largelossmatters Goto Github PK

Thank you for your great work!
If it is convenient for you, could you share the training code under "LinearInit" training setting? Thanks a lot!

Congratulations on such an excellent job! Would you mind sharing the paper?
Thanks very much!
[email protected]

Hi, Thank you for your great work!
I try to download the rest of training images via its URL using https://github.com/snucml/LargeLossMatters/blob/main/data/openimages/downloader.py, but all the images are NOT FOUND. I generate images id file like these.

train/000007e5752f8f1c
train/000056beaaea0225
train/000059e96d42da48

Is there something wrong? Thanks very much!

Thank you for your great work!
Woud you mind sharing the best hyperparameters for coco, pascal, nuswide and cub under "End-to-End" training setting and "LinearInit" training setting? Thanks very much!

It looks like there is a correlation between the hyperparameter delta_rel and the noise of the dataset (the number of missing labels in the dataset). Can I ask how you determined delta_rel? Especially on a large dataset like OpenImages V3?

Thanks for your work!
It seems that the current code is only suited for one positive label. Could you share the code on the real partial label dataset training setting?

Hi,
in the paper supplementary you mention you "search the learning rate by dividing the range between values in {0.01, 0.001, 0.0001, 0.00001} into quarters". Does this mean you search over learning rates 0.01, 0.0075, 0.050, 0.025, 0.01 etc? Over all learning rates, do you pick the run with the best validation accuracy and run that one once on the test set? Or do you calculate the test set result for each learning rate and report the best value?

Thanks!