futurexiang / ddae Goto Github PK

Could you please provide the whole experiment environment to run the training script? I encountered distributed training error all the time and tried many ways regarding the training parameters but resulted in the same error. So I guess this error is caused by environment settings.
P.S: I installed recommended packages in readme and ran the training script under pytorch 1.12 with: python -m torch.distributed.launch --nproc_per_node=4 train.py --config config/DDPM_ddpm.yaml --use_amp

Hi,

Firstly, I'd like to express my appreciation for your excelent work and the provided implementation code. It's been incredibly helpful.

I'm particularly interested in the fine-tuning process mentioned in your paper. Could you kindly share the fine-tuning code used to achieve these results? Also, it would be immensely helpful if you could include the training configurations for the DDAE-EDM model (as referenced in Figure 5b of the paper).

Thank you for your time and consideration.
Best

Dear authors,

Thanks for open-sourcing your amazing and inspiring work! I am really interested in applying generative diffusion models to discriminative tasks, and I would really appreciate if you could help me through a couple of concerns regarding your paper.

1. If I have understood everything correctly, the experimental results for say CIFAR-10 in your paper follow the steps: 1) train the generative diffusion models using the images from CIFAR-10; 2) find the optimal layer feature; 3) linear probe or fine-tune for classification using both the images and labels from CIFAR-10. I believe that since the model already sees all training images during pre-training, it already learns the image features within the dataset. Have you tried pre-training on one dataset, e.g. CIFAR-10, and transferring to another dataset e.g. TinyImageNet for downstream classification?
2. Following up on my previous question, have you tried or considered trying to replicate such experiments on large-scale pretrained diffusion models, e.g., Stable Diffusion? If this is not the case, what is the major gap in your opinion?

Thanks again for your great work, looking forward to your response!

Hello,

Firstly, I'd like to express my appreciation for your outstanding work and the provided implementation code. It's been incredibly helpful.

I'm writing to kindly request if you could release the trained models mentioned in your paper. Access to these models would be highly beneficial for further research.

Thank you for considering this request.
Best,

Hi,

Thanks for your codes. But I have questions about FID when dataset is larger. If my dataset is either Tiny-ImageNet or ImageNet 64x64, how many images should I generate to calculate FID? The exact number of Tiny-ImageNet or ImageNet 64x64 (larger than 50k)? And I should change the batch number (125) and 400 (125*400=50k) in sample.py, right?

BTW, I see other codes use total_training_steps instead of epoch. What is the relationship between these?

I hope this message finds you well. First and foremost, I would like to express my sincere appreciation for the incredible work you have done on your project shared on GitHub.

I am particularly interested in your Tiny-ImageNet project and would greatly appreciate it if you could share some of the code related to it.

Best regards, kecheng