Multiple errors when trying to run `ff_optimize()`

[Yoshanuikabundi]

Hi Trevor,

I've written the ff_optimize run I was trying to do last week up as a script. I've spent a fair whack of time trying to figure out what's going wrong and come up with obvious fixes, but each time I fix one thing another springs up. We discussed wrapping the periodicity in round(), and that fixed the initial error about it not being an integer, but I then got errors about an epsilon value being negative. I dug into this a little and tried to just set epsilons to zero if they were negative just to get something to work, but it just turned into other errors.

Here's the script I'm running and the datasets necessary to run it. If you feel like doing some bug fixing, it'd be great if you could have a look! Thanks

mre_ff_optimize.zip

[trevorgokey]

Hi Josh, can you please post your environment since the example uses non-besmarts dependencies?

[Yoshanuikabundi]

Oops, sorry: env.yml.txt

I've just added a .txt extension so that GH will let me upload it. The environment has Besmarts as a local pip/git dependency so you'll need to edit it for it to run on your machine.

[trevorgokey]

2GB to download over 430 packages!

There is a weird interaction happening somewhere, but the bottom line is that you didn't specify the symbols that you wanted to fit and chaos ensued. After obj_tier.fit_models = list(split_on_models), adding

obj_tier.fit_symbols = ['k']
obj_tier.fit_names = ['t17']

Should make things work with no other modifications (like the periodicity change). Looking further into this, the reason the periodicity was erroring was that you didn't specify the symbols to fit, defaulting to everything. It was trying to fit the periodicities too, causing it numerically evaluate the periodicity as a float, and OpenMM refuses to accept floating-point periodicities.

In the short term I can't see a world where I want to 1) allow non-integer periodicities and 2) break compatibility with OpenMM, so I've hard-coded pruning periodicities from ever being included as a fitting symbol. I've also updated the code to do the sane thing in your case, which was specifying fit_models=[2] without specifying symbols, which should mean all symbols in model 2. However, be advised that this will also include the phases, hence my recommendation above to use ['k']. The fit_names is optional and not related to this issue, but without it the fits to evaluate each split will occur on all torsions that matched the data which will be quite expensive.

I've tried loading the entire dataset and running through it. It crashes since you have a Tensor object in the energy data seomwhere.. you'll need to turn these into regular lists.