question about the needed iteration number

[yfpeng1234]

Dear author,
Your work is really interesting and insightful.
When reproduce your experiment results, I found that iteration number of each experiment is set to 2k.
However, I guess the needed iteration number for each dataset (mnist, cifar...) varies and 2k iteration is a little bit time consuming.
Could you please share your opinion on it.
Also it would be better if you could share the GPU hours of each experiment.

Thanks a lot!

[Yanqing0327]

Thank you for your interest in our work! You are right, on many datasets, we do not need 2k iterations to achieve the desired performance. Generally speaking, the best results we claim in the paper are usually produced within 1.2k iterations. The main consideration of the default setting of 2k iterations is to ensure that a better extreme point is found. If you are more concerned about distillation efficiency than performance, you can choose to stop early : )
As for GPU hours, it may be different on different devices. And welcome to try our enhanced version DREAM+!

[yfpeng1234]

Thanks for your prompt reply, it helps me a lot!

[yfpeng1234]

Dear author,
I have another question about the GPU device.
My server is a 4GPU A6000, and I hope to use gpu:2 for the experiment.
Then I add the following code:

However, I found that the process is still running on gpu:0.
Could you please give me some advice on how to deal with it?

Thanks a lot!

[Yanqing0327]

You can add CUDA_VISIBLE_DEVICES=N to the running command to make it.