3.4. Extractors and Comparators¶

The previous section described how an arbitrary linear combination of properties extracted from jobs \(P(J)\) can be added to a Data Set for fitting. In this chapter, we are going to describe what happens when entries are added and evaluated by the Data Set and how the user can extend the number of properties that can be fitted with ParAMS.

We have briefly described the extractors in the Data Set: They are a collection of Python code snippets available to the DataSet instance which define how to extract an individual property \(P\) from a job. The extractors available to each DataSet can be checked with the extractors attribute:

>>> ds = DataSet()
>>> ds.extractors
{'angles', 'vibfreq', 'charges', 'distance', 'stresstensor', 'energy', 'forces', 'hessian', 'dihedral'}

Any expression passed to the DataSet.add_entry() method can be constructed from the available extractors, for example "angles('myJob1')".

The design of the Data Set allows for an easy extension of the extractors to suit personal needs beyond what is already provided in the package. Users are encouraged to define additional extractors whenever a new property becomes relevant for the fitting process. In the following example, we will add the functionality to extract and evaluate elastic tensors to the Data Set:

>>> 'elastictensor' in ds.extractors
False

Start by creating an elastictensor.py in an empty directory of your choice with the following contents:

from numpy import ndarray, asarray

sigma = 1e-2
def extract(amsresults) -> ndarray:
  return asarray(amsresults.get_engine_results('ElasticTensor'))

This is a minimal and complete definition of our extractor. The code snippet can be saved under any accessible path and used by providing DataSet with the corresponding directory. The base file name will be used as the extractor’s name (make sure not include extractors with the same names). Here, we have saved the above under path/to/extractors/:

>>> ds = DataSet(more_extractors='path/to/extractors/')
>>> 'elastictensor' ds.extractors
True

Because the Data Set natively works with plams.AMSResults, we were able to take a shortcut in the creation of our extractor: All we needed to do in this case is wrap a PLAMS method around the extract() function. Note that in addition to the function definition, we also provided a sigma variable. This serves as a default value when a related entry is added with through DataSet.add_entry() and should roughly be in the same order of magnitude as the “accepted accuracy” for this property. ParAMS will evaluate all entries according to \((w/\sigma)(y-\hat{y})\), where \(w\) is the weight. Providing a sigma is not mandatory but highly recommended, any extractor without a sigma will set it’s value to 1.

However, because the definition of extract() is completely up to the user, it can be made to read and process data from any other source as well. Assuming that a property is stored in a text file, an extractor that is independent of plams.AMSResults could look like this:

def extract(_, filepath):
  # AMSResults instance: `_` is ignored.
  with open(filepath) as f:
    return float(f.read())

# dictextractor.py
from typing import Dict

sigma = 0.1
def extract(amsresults) -> Dict:
  return amsresults.some_dict_property()

def comparator(y:Dict, yhat:Dict, au_to_ref:float) -> float:
  y    = list(y.values())
  yhat = list(yhat.values())

  # Unit conversion must be handled by the custom comparator:
  yhat = [i*au_to_ref for i in yhat]

  for i,ihat  in zip(y, yhat):
    ...
  return ...