Joshua V. Dillon

That's right. Needing tf.tile is actually fairly rare. This is good because using it is expensive. And if things aren't broadcasting then its more likely we have a bug than...

Having a CRF distribution is something Ive long wanted and seems generally useful.

Note: while this allows me to build the modules, I still cannot `insmod` them. ``` $ make make -C /lib/modules/`uname -r`/build M=$PWD modules make[1]: Entering directory '/usr/src/linux-headers-6.14.0-29-generic' make[2]: Entering directory...

I agree it would be nice to have this distribution. That said, the approach used here: https://github.com/tensorflow/probability/blob/master/tensorflow_probability/examples/jupyter_notebooks/Bayesian_Gaussian_Mixture_Model.ipynb is more efficient since matmul is faster than cholesky. (Basically, the trick is...

See also: tfb.CholeskyToInvCholesky