Results with randomly initialization SWAG are bad

[tlatlbtle]

Thanks for releasing the code.

I find that Fig. 2 (i) from the paper, shows that both pre-trained and random networks, models with SWAG could produce good results. Thus, I just try to change line 260 into resnet = models.resnet50(pretrained=True). However, it seems the result are bad.

Results with p-ResNet (arch=resnet):

Results with p-ResNet∗ (arch=resnet_swag):

Results with r-ResNet∗ (arch=resnet_swag):

Can you please clarify this? Thanks in advance if I missed something.

[peiwang062]

Hi, I think your results are consistent with figure 2. This figure is showing p-resnet* better than p-resnet. r-resnet* better than r-resnet. For your results, it is clear the second one is better than the first in the sense that more style patterns are transferred, seeing for example, upper left regions, the nose, neck, collar. And I think you could run a r-resnet without swag. I believe it would be worse than your last image.