Fields

Fields are additional pieces of data, including labels (the thing being predicted).

At a most basic level, a field has a name, and depending on the type of field it may take additional options to do more work.

Once a field is declared the it is added to the data dictionary that is returned when processing sequences. See the example in the PyTorch integration section.

Types of Fields

As stated different kinds of fields afford different functionality.

LabelField

A LabelField ascribes a particular label to a sequence.

FileMetaDataField

A FileMetaDataField can be used in cases where the filenames contain label information.

For example, imagine there's a list of files:

from pathlib import Path
paths = [Path("a/f1.fasta"), Path("a/f2.fasta"), Path("b/f3.fasta")]

In this case there are two files, f1 and f2, that are under label "a" and f3 is under label b. In order to create a FileMetaDataField we'll write a function that takes in a Path and returns a string (the label).

def preprocess(p: Path) -> str:
    # Path("a/b") -> Path("a") -> "a"
    return str(p.parent)

Then create the FileMetaDataField.

from gcgc.fields import FileMetaDataField
label = FileMetaDataField.from_paths(name='label', paths=paths, preprocess=preprocess)

After that, the field has an encoding dictionary and, more importantly, it will add the label name and corresponding integer encoded class to the dataset.

label.encoding_dict
{'a': 0, 'b': 1}

AnnotationField

On the SeqRecord BioPython object exists an .annotations property, which potentially contains labels that contain labels or other features. To help access these the AnnotationField works in a similar fashion to the FileMetaDataField: write a processing function and create an exemplar dataset.

.annotations is just a dictionary, so this is the simplest example.

annotations = [{"a": "a"}, {"a": "b"}]

Because .annotations is a Dict, the preprocess signature is slightly different than FileMetaDataField.

def preprocess(a: Dict) -> str:
    return a["a"]

After this, the field behaves similarly to the other fields.

from gcgc.fields import AnnotationField

af = AnnotationField.from_annotations("a", annotations, preprocess)
af.encoding_dict
# {"a": 0, "b": 1}