pretrained model checkpoint of ema_lsun.ckpt

[rnqm]

Hello, I appreciate your work. Could you please provide ema_lsun.ckpt? I don't know where I can download it. Thanks!

[GuanlinLee]

Please try to use the model from https://huggingface.co/fusing/ddpm-lsun-bedroom-ema. I think the original Google Drive link is missing now.

[rnqm]

Thank you very much!

[rnqm]

I'm sorry, there is another mistake. self.env = lmdb.open(root, max_readers=1, readonly=True, lock=False, readahead=False, meminit=False) lmdb.Error: ./data/bedroom_train_lmdb: No such file or directory. When I try to download the dataset myself. I visit http://www.yf.io/p/lsun and got HTTP Error 404: Not Found. Where can I download it? Thanks again! I successfully ran warfare on 64×64 celeba dataset, and now I want to try it on 256×256 dataset.

[GuanlinLee]

Could you try to use HuggingFace for dataset downloading? The dataset can be found in https://huggingface.co/datasets/pcuenq/lsun-bedrooms.

[rnqm]

I downloaded the dataset and renamed it bedroom_train_lmdb in the project's data folder, but still reported the error lmdb.Error: ./data/bedroom_train_lmdb: No such file or directory. I found lmdb is a kind of database, now I found a link to download the lmdb, I will try it again.

[rnqm]

I can run successfully. I found that it needs 22 days to diffusion purify 10,000 images, so I intend to use the fewshot setting of 500 images. I have read the paper that warfare has strong generalization ability of small samples, which can meet the practical application.

[GuanlinLee]

Yes, you could try to use the few-shot setting. But I am not very sure how it works on different image domains. We only test it on the same image domain but with different watermarks.

BTW, the speed of the diffusion model depends on your GPUs. We adopt A100 in our experiments.

[rnqm]

I use 3090 to run on 256×256 LSUN dataset. Last time it ran much faster on 64×64 CelebA dataset because of image size. By the way, I wonder if the result in Table 5 is trained on StegaStamp, and then few shot on WGAN-div and EDM watermark schemes, or few shot on different watermark binary information of StegaStamp?

[GuanlinLee]

They are few-shot results on different watermark binary information of StegaStamp. The results of WGAN-div and EDM are in Table 6.

[rnqm]

OK, thanks for your early reply!