SLADD
SLADD copied to clipboard
Difficulties reproducing the results
Hello,
Thank you for your exciting work. I have been trying to reproduce the results of the method, but I was unsuccessful. I have run it on a single GPU because of my limited setup and followed the instructions according to the readme. But after 30 epochs the results are still bad, do you have an idea to why this is the case?
Testing from FF-DF ... Len of data_loader= 219 Acc 0.4821 (~0.00) |AUC 0.4614 (~0.10) |EER 0.5286 (~0.10) |AP 0.5142 (~0.11)
Testing from FF-NT ... Len of data_loader= 219
|Test Loss 14.85385it/s]| Acc 0.4847 (~0.00) |AUC 0.4335 (~0.11) |EER 0.5360 (~0.11) |AP 0.5005 (~0.11)
[3/4]Testing from FF-FS ...
Len of data_loader= 219 |Test Loss 1.436065it/s]|Acc 0.9536 (~0.00) |AUC 0.9916 (~0.01) |EER 0.0077 (~0.03) |AP 0.9975 (~0.01)
[4/4]Testing from FF-F2F ...
Len of data_loader= 219
219it [00:19, 11.46it/s]
|Test Loss 9.582998it/s]|Acc 0.6704 (~0.00) |AUC 0.6656 (~0.11) |EER 0.3854 (~0.13) |AP 0.7459 (~0.10)
@NamelessGaki Could you please tell me what your training batch size is.
I couldn't train on 32 batch size and had to go with 16, but also didn't get good results. I tried halving the learning rate but that didn't work either.
I didn't keep tracking of accuracy in each epoch, I cared more about AUC, especially for celebdf dataset.
anyone could provide pretrained model weight? thanks a lot