search

skynbe / pseudo-attributes

Official implementation of BPA (CVPR 2022)
12 stars 2 forks source link

Resume checkpoint baseline #2

Open Jackpepito opened 1 year ago

Jackpepito commented 1 year ago

Hi, I'm trying to replicate your experiments and I have a doubt: did you use the last baseline epoch or an intermediate one to extract the features? I found a rather significant difference in performance compared to the paper values ​​using the last training epoch of the baseline as a feature extractor (as described in the paper, I have trained models for 100 epochs).

Could you make the baseline checkpoints available for a fair comparison or to rule out bugs? Even just for a target, such as 'Blond Hair'.

Thanks, Giacomo

skynbe commented 1 year ago

Hi Giacomo,

Could you let me know what is the performance of your reproduced baseline? I used the last epochs' one, but there is no significant difference between intermediate and last ones.

I also uploaded my baseline in Blond Hair in the following link: https://drive.google.com/file/d/1pveU8cTbU-PMGwffPmH0kjfOGfQANWFc/view?usp=sharing

Note that this baseline model is only used for clustering to calculate the importance weight.

Jackpepito commented 1 year ago

Hi, thanks for the weights, I will check as soon as possible.

Yes, between the final and intermediate epochs I didn't notice any big differences, but I wasn't able to reach the results in the paper. In fact, the baseline you provided me will be very helpful in understanding the problem.

Thank you