search

naoto0804 / pytorch-AdaIN

Unofficial pytorch implementation of 'Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization' [Huang+, ICCV2017]
MIT License
1.08k stars 208 forks source link

vgg_normalized.pth vs torchvision pretrained vgg #43

Closed ZeroAct closed 3 years ago

ZeroAct commented 3 years ago

Hi.

I wonder the difference between 'vgg_normalized.pth' you uploaded and pretrained vgg19 from torchvision.

I tried to train AdaIN with pretrained vgg19 from torchvision many times, but I've failed. (I've succeed with 'vgg_normalized.pth')

Sample of my AdaIN inferences that trained with torchvision vgg19 for so many epochs looking so bad... epoch_761

Thank you : )

wjjjjyourFA commented 3 years ago

Hi, may I know how you did it？ I put a image in ./input/content and another in ./input/style, using "vgg_normalized.pth" CMD:python train.py --content_dir input/content --style_dir input/style --batch_size 4 --max_iter 160 Though I got a xx.pth.tar in ./experiments, I use it "test.py", getting a gray image result.

naoto0804 commented 3 years ago

@ZeroAct I guess the VGG model provided is a bit modified. I am still not 100% sure, but it seems that Gatys et al. normalized the VGG network for style transfer, according to this.

ZeroAct commented 3 years ago

@wjjjjyourFA I wrote my new AdaIN code. So, I can't answer you. Sorry!

ZeroAct commented 3 years ago

@naoto0804 Thank you for your answer. It was very helpful.