Jarrel Seah

Yes. Doesn't look like this is maintained anymore but I've forked it and made a few changes including this one.

Same thing - noticed that in myo.py if HAVE_SK and self.X.shape[0] >= 20: self.nn = neighbors.KNeighborsClassifier(n_neighbors=15, algorithm='kd_tree').fit(self.X[::3], self.Y[::3]) else: self.nn = None here self.X.shape[0] is 0 for me and al...

The error is NoneType because m.cls.nn is none, not initialized. If you look at my fork I've put a patch on it.

Hi I would like to check in on planning entity difficulty - is there a development branch with this enabled and / or is it possible to contribute to this?

what speeds are you getting @noamgat vs the outlines backend?

I think they said they use RMSProp.

Ah yeah, sorry just realised you'd already had that.

One thing I think we could do to make it faster is to use the fact that allowed_tokens is either almost all the tokens, or none of the tokens. Currently...

Hi From my reading of the paper especially in appendix C.1 I think they use different models. Currently I have not implemented classification. Jarrel On Mon., 1 Apr. 2019, 18:31...

Hi Jiachang - what batch size are you using? It may be related to learning rate / batch size. Also, the authors have released their official code at https://github.com/jhjacobsen/invertible-resnet -...