sirmarcel/cmlkit

tools for machine learning in condensed matter physics and quantum chemistry

Hello! 👋

I'm Marcel, a postdoc in the COSMO Lab at EPFL in Lausanne, Switzerland. Previously, I was a doctoral student at the Fritz Haber Institute (in the NOMAD Laboratory) and TU Berlin (in the Machine Learning group).

I work on applying machine learning to materials science problems. Recently, I've particularly focused on machine-learning potentials and the computation of thermal conductivities with the Green-Kubo method. To this end, I've written the glp and gkx packages, both using jax.

Here are some other places on the internet you can have a look at:

• I'm on twitter
• I'm also on gitlab
• marcel.science is my science website
• marcel.computer is my personal website
• here's my google scholar profile page

Thanks for stopping by 🚀

cmlkit's Issues

yaml dumping with numpy arrays is ugly

Need to rethink this. Currently, anything that is a numpy array gets dumped as ugly binary junk.

Change unit of duration argument in Run.run() to minutes

It makes very little sense to configure this on a second basis, when the basic assumption is that a single loop takes about ~5s to run. And the jobro default unit is minutes.

Drop yaml dumper for son, replace with json

JSON is comically faster. Here is a benchmark for dumping/loading 2000 small dicts (3 repeats):

JSON:
{'times': array([0.38685818, 0.34857985, 0.34886317]), 'mean': 0.36143373133333334, 'min': 0.3485798470000001, 'max': 0.3868581750000001}
Yaml:
{'times': array([ 9.26613551, 10.8498244 ,  9.88484843]), 'mean': 10.000269449333334, 'min': 9.266135509, 'max': 10.849824405}

This really hurts when doing Run.checkout().

Edge case in Dataset.compute_dataset_info

This is a minor concern, especially if Dataset is being phased out. When creating a Dataset with only single-atom geometries, dists is a list of empty arrays and min_distance and max_distance fail. A real-world example of this might be energy-volume curves with FCC primitive cells.

cmlkit/cmlkit/dataset/dataset.py

Lines 548 to 552 in 5390215

	dists = [
	qmmlpack.lower_triangular_part(qmmlpack.distance_euclidean(rr), -1) for rr in r
	]
	i["min_distance"] = min([min(d) for d in dists if len(d) > 0])
	i["max_distance"] = max([max(d) for d in dists if len(d) > 0])

Toy example:

from ase import Atoms
from cmlkit.dataset import Dataset

geometries = [Atoms('Au', cell=[3, 3, 3], pbc=True),
              Atoms('Au', cell=[4, 4, 4], pbc=True),
              Atoms('Au', cell=[5, 5, 5], pbc=True),]
Dataset.from_Atoms(geometries)

---

ValueError                                Traceback (most recent call last)
<ipython-input-64-25473c66c8a8> in <module>
      5               Atoms('Au', cell=[3, 3, 3], pbc=True),
      6               Atoms('Au', cell=[4, 4, 4], pbc=True),]
----> 7 Dataset.from_Atoms(geometries)

~/PycharmProjects/cmlkit/cmlkit/dataset/dataset.py in from_Atoms(cls, atoms, p, name, desc, splits)
    288             name=name,
    289             desc=desc,
--> 290             splits=splits,
    291         )
    292 

~/PycharmProjects/cmlkit/cmlkit/dataset/dataset.py in __init__(self, z, r, b, p, name, desc, splits, _info, _hash, _geom_hash)
    169             self.info = _info
    170         else:
--> 171             self.info = self.get_info()
    172 
    173         # compute auxiliary info that we need to convert properties

~/PycharmProjects/cmlkit/cmlkit/dataset/dataset.py in get_info(self)
    218     def get_info(self):
    219         """Compute information on dataset."""
--> 220         return compute_dataset_info(self)
    221 
    222     def get_hash(self):

~/PycharmProjects/cmlkit/cmlkit/dataset/dataset.py in compute_dataset_info(dataset)
    549         qmmlpack.lower_triangular_part(qmmlpack.distance_euclidean(rr), -1) for rr in r
    550     ]
--> 551     i["min_distance"] = min([min(d) for d in dists if len(d) > 0])
    552     i["max_distance"] = max([max(d) for d in dists if len(d) > 0])
    553 

ValueError: min() arg is an empty sequence

Run context should also be on tape?

title says it all.

Upsides:

• Once you ran prepare(), restore and run will always do (approximately) the same
• If you encounter a tape.son in the wild, the information in it is slightly more complete

Downsides:

• Need additional functionality to overwrite the context in restore if required
• More complex restore

Also, it might be worth thinking about including more general (write-only) metadata on the tape.

Hash stability failure in `Dataset` with newer `numpy`

test_hash_stable in tests/test_dataset.py has started failing when numpy is upgraded somewhere above 1.16.5. Most likely this is due subtly different treatment of object arrays. Luckily, they should become obsolete once we transition to the new Data backend, so I'll close this issue then. Until then, it'll remain open in case anyone else notices!

Undefined behaviour in Dataset when filename is specified but not name

When a Dataset is created with no name but saved with a filename, references to that filename are ignored by certain other Components like tune.Run using a tune.TuneEvaluatorHoldout that was initialized with filenames for the train/test kwargs. Interestingly, EvaluatorHoldout.evaluate() does not suffer, while Run.run() does.

`tid`s for `quippy` interface appear to not be always unique

Occasionally execution halts with file ready exists due to the quippy scratch folder tid not being unique. This should not happen. It Should Be Tested whether this is actually due to not enough entropy in generating these + the timing being exactly the same, or whether there is a more fundamental problem (which seems a lot more likely.)

If it's actually a randomness problem, fixes could be:

• Switching to time_ns
• Adding geom_hash
• Adding a context-supplied random seed