Sigma level for models with input pixels between 0 and 1 about smoothing

[jmcohen]

It should be sigma = 0.5. (Yeah, it's a lot of noise)

[Zephyrasgit]

Thanks for the reply! I see. My problem is that my certified accuracy is as low as 0.1. My original accuracy is close to 0.8. Do you reckon that it is because my classifier had not seen enough noisy images? if so, may I know how extensive should I go about training the base classifier with the noisy images? Can you guide me on how you did it with your base classifier?

[jmcohen]

Hmm 0.1 seems very low. Are you training the base classifier with the same level of noise that you are later using for smoothing? This is the script I used to train with noise: https://github.com/locuslab/smoothing/blob/master/code/train.py. The exact commands I ran are here: https://github.com/locuslab/smoothing/blob/master/experiments.MD

[Zephyrasgit]

Yeah, 0.1 is about the standard of random guessing I would presume. Yes, I am training the base classifier with sigma=0.5 of noise, which is the sigma I use for RS as well. If you wouldn't mind, can I take more time to triple check my code to ensure there is no error, rerun RS on newly generated dataset with Gaussian noise and get back to you again? Hopefully this is due to errors in my code.

[Zephyrasgit]

Ah, I have my results and it turned out to be an error in my own codes. Thanks a lot, pardon me for my clumsiness.