You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
With the NVlabs implementation (using --cfg paper512), the visualized training samples shows a smooth conversion of the car images from the pretrained model into the objective trucks. However, using the mmgen implementation, the visualized training samples shows that the initial car images are quickly converted into noise (unintelligible images), and after that they tried to capture the objective mode (the truck images) but with a worst image quality.
I suspect that the observed training difference is due to a different choice of hyperparameters, but the default hyperparameters from both implementations (NVlabs and mmgen) seems to be almost the same.
Am I missing some other important hyperparameters or the observed training difference is due to intrinsic implementation differences?
So I also modified the MMSegmentation config file accordingly:
# MMsegmentation configcfg.ema_half_life=paper512_cfg['ema'] # Defaults to 10.0cfg.optimizer.generator.lr=paper512_cfg['lrate'] # Defaults to 0.0016cfg.optimizer.discriminator.lr=paper512_cfg['lrate'] # Defaults to 0.0018823529411764706### Here I assumed that 'nvlabs_gamma'=='mmseg_loss_weight' / 2cfg.model.disc_auxiliary_loss.loss_weight=paper512_cfg['gamma'] *2# Defaults to 80.0.
In addition, I also modified some other hyperparameters as their default value from the MMSegmentation implementation differs from the NVlabs implementation:
I also checked that all other default hyperparameters were the same for both implementations (note that I removed the dict(type='Flip', keys=['real_img'], direction='horizontal') operation from data pipelines, since I used "xflip": false in the NVlabs implementation). However, with all the mentioned changes, the training performance is even worst compared to the previous hyperparameter configuration.
Could someone help me find the cause of the observed training difference between the MMsegmentation and NVlabs implementations? Is there a way to replicate the same training performance of the NVlabs implementation?
Hi everybody,
I tried to make transfer learning from the stylegan2_config-f_lsun-car_384x512 pretrained model on a custom dataset of trucks images (with the same aspect ratio), but the training seems to differ from what I got using the Pytorch StyleGAN2-ADA official implementation.
With the NVlabs implementation (using
--cfg paper512
), the visualized training samples shows a smooth conversion of the car images from the pretrained model into the objective trucks. However, using the mmgen implementation, the visualized training samples shows that the initial car images are quickly converted into noise (unintelligible images), and after that they tried to capture the objective mode (the truck images) but with a worst image quality.I suspect that the observed training difference is due to a different choice of hyperparameters, but the default hyperparameters from both implementations (NVlabs and mmgen) seems to be almost the same.
Am I missing some other important hyperparameters or the observed training difference is due to intrinsic implementation differences?
The text was updated successfully, but these errors were encountered: