I'm wondering what is the correct way to get distilled images and test performance on them as well check performance on a normal dataset after training on the distilled images. I'm confused since there are many parameters and I've already read the advanced docs. So, to get distilled data, for example on Cifar10, I need to run
So, to train the network on usual full dataset I need to set --mode train, and if I want to test network performance after training on distilled data I need to set --mode train --phase test ?
Or, in other words, how to get results like in your paper where it’s said that you get 80% when fully trained against 54% with distilled data on CIFAR10?
Hello!
I'm wondering what is the correct way to get distilled images and test performance on them as well check performance on a normal dataset after training on the distilled images. I'm confused since there are many parameters and I've already read the advanced docs. So, to get distilled data, for example on Cifar10, I need to run
Then distilled images are in file
result.pth.So, to train the network on usual full dataset I need to set
--mode train, and if I want to test network performance after training on distilled data I need to set--mode train --phase test?Or, in other words, how to get results like in your paper where it’s said that you get 80% when fully trained against 54% with distilled data on CIFAR10?
Looking forward for your response! Thanks!