msultan

are we favoring kinetic mapping over commute mapping?

+1.

I propose v4.0+ given that we are doing version releases every quarter. That should be 15ish months from now. That should hopefully allow for a smoother transition.

Another option could be to identify the lowest populated states in your model to start new simulations from that region. To properly GMRQ a dataset under CV, you need both...

Thanks for raising this @ryjin . Could you let us know what order of magnitude in the number of states you get to before the mem issues arise. Few thoughts...

So that wont work here because the bad frames were causing the tica obj to become completely non-sensical(e+13 means, e-15 eigenvals etc). I think we need an intermediate step between...

like RMSD to start to traj? I kinda like that idea though i dont know what constitutes as an "un natural" change since we might randomly see something useful that...

wanna put that in as a pre-processing module to automatically rename those trajs to something like ```BAD.TRAJ```. Otherwise, I can look into it too.

Hmm, Jade and I were trying to unsuccessfully debug this problem this after noon with a custom 5*5 matrix but we were not able to pin point where the bug...

I think you should be able to specify subsets by adding a [boolean_mask](https://www.tensorflow.org/api_docs/python/tf/boolean_mask) to the network's first layer. A more involved alternate route might be to create a non-trainable affine...