-
Notifications
You must be signed in to change notification settings - Fork 252
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Hific Colab codes are outdated #125
Comments
Hi Yifei, |
Hi Fabian, For LPIPS loss, Tensorflow 2 currently doesn't support it. It's only supported by Tensorflow 1. Please check this link: https://github.com/alexlee-gk/lpips-tensorflow. Would you mind to update you HiFic codes to Tensorflow 2 codes and provide LPIPS Tensorflow 2 pretrained model? I trained HiFic model using the Pytorch codes: https://github.com/Justin-Tan/high-fidelity-generative-compression. But I found the big gap between the bpp using entropy loss and the real bpp. I also tested the pretrained model provided by Justin using Kodak dataset on the Colab demo. But I still found a big gap between the theoretical bpp and the real bpp. Can you please check the issue? You might just use the pretrained model from the Github colab demo to check ithttps://colab.research.google.com/github/Justin-Tan/high-fidelity-generative-compression/blob/master/assets/HiFIC_torch_colab_demo.ipynb If you agree with the HiFic Pytorch codes https://github.com/Justin-Tan/high-fidelity-generative-compression, should I use the theoretical bpp or the real bpp for my comparison experiments? Yes, I can write the Codes in Tensorflow 2 but I am not sure whether I can replicate close results since my codes are not official codes. Thank you! Sincerely, |
Hi Fabrian, Sincerely, |
Hi, there are currently no plans to update the published code. Is it not possible to run it by pinning TF to the version from the requirements? You should always use real bpp (that's what we use for the paper). I do not know why the code from justin has a big difference there. |
Hi @yifeipet |
Great. Thank you! I will try this method to port LPIPS to Tensorflow 2. |
Can you please share with me the codes that works with TF2.x |
Hi @MahmoudAshraf97 .. For the compression project I am working on I have a gpu-tf1 compatibility issues. Can you please add the TF2 HiFiC codes to your repo. |
Hello Dr. Balle,
Please check the Colab codes. Can you change the codes and other necessary Tensorflow Compression codes (such as hyperprior models) to fully support Tensorflow 2? New NVIDIA cards do not support old CUDNN libraries that are for Tensorflow 1.15. That can be a big issue if more people buy new NVIDIA cards that cannot support Tensorflow version 1. They might give up Tensorflow Compression or even Tensorflow, and go completely to Pytorch.
Yes, there are Pytorch codes for Hyperprior and Hific. But they might not be official codes so the results by these codes might be not accurate.
Tensorflow version 2 doesn't support LPIPS loss. The pretrained neural network for LPIPS is not in Tensorflow 2. Can you also provide some codes for that?
I understand you are busy. But hope you can solve the most issues. Thank you!
Sincerely,
Yifei
The text was updated successfully, but these errors were encountered: