Unable to train impweight GAN under a certain condition: bias=90_10 & perc=0.1

[fdirdesigner]

I've attempted to train impweight GAN with bias=90_10 and perc=0.1, but failed to get meaningful results. Please see the attached sample image file that was acquired during the training:

I have also tried other settings that include (1) no reweighting under bias=90_10 and perc=0.1; (2) impweight under bias=90_10 and perc=0.25; (3) impweight under bias=90_10 and perc=1.0. They all worked well and bahaved as I expected.

Note that I used your codes as they are with no modifications, and 'faithfully' followed the instructions described while executing.

[Bearwithchris]

I've also observed similar results. Upon further investigation, I realise that the density ratio classifier ratio is also not calibrated as per supplementary figure 5. Similar observation for perc=0.25 and 0.5 (with regards to the calibration). Im wondering if this could be a possible contribution to stability issue seen at perc =0.1

Please find attached the plots of

celeba_90_10_perc 0.1 , Bin 10 calibration curve

)

celeba_90_10_perc 0.25, Bin 10 calibration curve