cics-nd / cnn-surrogate Goto Github PK

Hello, is the dataset still available to download as i have failed to download using the command
bash ./scripts/download_dataset.sh 4225

Hi,
thanks for the easy-to-use code!
I had a question about the use of priors for the weights when calculating the log probability to be used inside the stein gradient update equation: is that something without which things would not work in your experience? I was wondering since the stein gradient just needs us to specify a log_prob that we're interested in maximizing, and so doing svgd with just the model likelihood (wrt ground truth data) as the log_p is also correct right?
I just wanted to clarify if the weight priors are something desirable that we can choose to add t the log_p term because of desired regularisation(as specified in your accompanying paper) or if that's something essential that if not included, renders the math/theory wrong.
Thanks,
Gunshi

Hi @bmmi - thanks for your code. Would you be able to explain why you are negating the gradients here?

https://github.com/cics-nd/cnn-surrogate/blob/master/models/svgd.py#L163

Doesn't pytorch take care of this for you?