Sam Koelle
Sam Koelle
Do we have to log _timestamp explicitly to make this happen?
We ran into this bug today - the fix worked!
fwiw im getting this error with transformers 4.27.1
That PR looks like it is releasing the code only MIT. A bit confused by the licensing as well. Audiocraft weights are not under commercial license, but HF transformers is?
I think the KL divergence may be with respect to a standard Gaussian so s_q is just 1?
Has there been any progress? Support for arbitrary DAGs would be fantastic.
I am also currently experiencing this issue and investigating
A few years later but saw this was still open - I wrote a tensor version here https://github.com/sjkoelle/montlake/blob/master/montlake/optimization/gradientgrouplasso.py
Was this ever knocked off?
I suppose this is just an issue of numerical precision and zeroing is a okay?