ColumbiaDVMM / AutoGAN

96 stars 22 forks source link

About the place of pretrained CycleGAN #9

Open IItaly opened 3 years ago

IItaly commented 3 years ago

Thanks for your work. I noticed that there are 18 pretrained model to make the dataset. And the script to download the models have failed. So I want to know if I should put them to the "autogan/scripts" and then run the "run_test.py"? I got the error when I try to "python run_test.py --dataset Cyclegan" : NameErroe: "parser" is not defined

IItaly commented 3 years ago

And How can I get the frequency data to train the classifier ?just like the pics referenced in your paper

spongezhang commented 3 years ago

Hi IItaly,

Sorry for the delayed reply. I left Columbia University and have less time to maintain this package.

For your first question, I'm not fully sure what models you are referring. There are two types of models, 1) autogan model(generate autoencoder like images) the models are still here https://drive.google.com/file/d/1IZZ5QCLbzAmrPDHh7EdOigPCdz46zPc6/view. I mentioned this "Remove it once training is done. Once training is done, change all the autogan models' name in ./autogan/checkpoints/{category}_auto/latest_net_G_A.pth to ./autogan/checkpoints/{category}_auto/latest_net_G.pth" in the repo. 2) the Classifier. The classifiers are missing, but training the classifiers with autogan/cyclegan images are relatively easy.

For the spectrum, here is how I calculate that. https://github.com/ColumbiaDVMM/AutoGAN/blob/master/code/GAN_Detection_Train.py#L204

IItaly commented 3 years ago

Thanks for your reply.Does your spectrum classifier use the color spectrum image shown in your paper? I tried to run your code for spectrum generation(just you have quoted ), and it generated a pure black image. Is it related to the args.mode parameter? I found that the parameter defaults to 0.

spongezhang commented 3 years ago

Haha, the color map is just the heatmap for the grayscale spectrum for visualization. Three channels are processed separately as shown in the code. :)

IItaly commented 3 years ago

Haha, the color map is just the heatmap for the grayscale spectrum for visualization. Three channels are processed separately as shown in the code. :)

I get it.I was very confused by the color map, because I got the grayscale spectrum. Thank you for your answers.