Experimental readability model

srcmetrics has no built-in weights and does not compute a readability score by default: how each metric relates to human-perceived readability is not known in advance. If you have human ratings, you can fit a model on them.

Labels

A CSV file with the header path,function,score:

path,function,score
src/parser.py,parse_header,3.5
src/parser.py,parse_body,2
src/util.py,,4

• path is the path as it appears in the analysis result.
• function is the function name; leave it empty to rate a whole file. All labels of one model must be of the same kind.
• Each label must match exactly one file or function. Unknown paths, files that failed to parse and function names that are not unique in their file are errors.
• score must be a finite number, and not all scores may be equal.

Train

srcmetrics analyze src -o result.json
srcmetrics model train --labels labels.csv -o model.json result.json

The model is a ridge regression on standardized features:

• Features default to every metric of the labels' scope. Metrics that are null for some labelled unit, or constant over them, are excluded and listed with the reason in the model file. --features id1,id2 selects features explicitly; each must then be usable.
• --lambda sets the regularization strength (default 1.0).
• The model file records the features, coefficients, R² on the training data and the RMSE of 5-fold cross-validation (with at least 5 labels).

Predict

srcmetrics model predict --model model.json result.json

Prints a score for every file or function (matching the model's scope). A unit with a null feature, or a file that failed to parse, gets score: null and the reason.