Please upload the pretrained HiDDeN decoder weights.

[DI-LEE]

Hello, I'm DI in Korea univ. I'm very impressed with your research. So I want to reproduce various bits of it, but I failed to pretrain 32 bits HiDDeN decoder. Please upload the pretrained 32 bits decoder's weights or run code please.

Thank you. Best regards DI Lee

[pierrefdz]

Hi, The Hidden models are already uploaded (see the readme.md). The code to run watermark networks is in https://github.com/facebookresearch/stable_signature/tree/main/hidden. Tell me if I'm missing smthing.

[DI-LEE]

Thank you for fast reply. I'm reproducing 32 bits HiDDeN decoder but Its bit accuracy only goes up to 87%. I'm trying to get 100%, Can I have access to the decoder weights used in your code?

[pierrefdz]

I havent used 32bits decoder. The 48bits one is given in the reproduction: https://github.com/facebookresearch/stable_signature/tree/main/hidden/ckpts Or in https://github.com/facebookresearch/stable_signature/tree/main#watermark-models

What is the decoder you are referring to? You can also share the parameters used for your training (the optimization can be a bit unstable depending on them)

[DI-LEE]

Sorry, I mistakenly thought that you guys did the experiment on 32bits. After rechecking your paper, there was no 32 bits decoder.

If you know how to reach 100% in bit accuracy with 32 bits decoder, it would be thankful if you share with me.

Thank you for your time.

[pierrefdz]

What is your setup? Can you share the logs of your run?

This command should work directly if you have the same setup (and replacing 48 by 32). https://github.com/facebookresearch/stable_signature/tree/main/hidden#example

If you don't have the same setup (for instance number of gpus), you can try reducing the lr to lower values like 1e-3.

[DI-LEE]

I already tried replacing 48 with 32.

Command line is shown below. torchrun --nproc_per_node=8 main.py \ --val_dir path/to/coco/test2014/ --train_dir path/to/coco/train2014/ --output_dir output --eval_freq 5 \ --img_size 256 --num_bits 32 --batch_size 16 --epochs 300 \ --scheduler CosineLRScheduler,lr_min=1e-6,t_initial=300,warmup_lr_init=1e-6,warmup_t=5 --optimizer Lamb,lr=2e-2 \ --p_color_jitter 0.0 --p_blur 0.0 --p_rot 0.0 --p_crop 1.0 --p_res 1.0 --p_jpeg 1.0 \ --scaling_w 0.3 --scale_channels False --attenuation none \ --loss_w_type bce --loss_margin 1

[pierrefdz]

Thank you, I'll try to have a look at it in the following days.

[pierrefdz]

Hi DI, I've run the code on my side and accuracy gets to 1.00, here are the logs: https://drive.google.com/file/d/1qG21QKmVkikQlQ4c2mzqDIsLNe04bQqp/view?usp=drive_link

Could you share with me the logs of your console s.t. I can see if there is an issue? Otherwise there might be a difference either on the dataset, or in the code I run (but I don't think so since the code I've used is the one I pushed here...)

[DI-LEE]

I'm very appreciate it. But I can't access your drive, please give me access right.

[pierrefdz]

can you ask for it? otherwise I need an email to give you access rights

[DI-LEE]

My email address is dongeen1@gmail.com. And my log is below.

log.txt

[DI-LEE]

I'm really sorry to ask, but could you possibly upload the weight file for the 32bit pretrained model to the drive?

[pierrefdz]

Our logs do not share the same lr, so I think the issue comes from here (my guess is that your world size is different in your distributed setup, so the lr is scaled differently - I have not enough experiments on this and did the same scaling as image recognition, which might not be adapted).

You can try removing this line https://github.com/facebookresearch/stable_signature/blob/main/hidden/main.py#L218 and fix the lr to 0.005 instead of 2e-2.

I won't be able to share the networks (it would need to go through internal review), sorry for that.

[pierrefdz]

Closing since no activity, feel free to re-open!