Fold Generator: Data Tables to Data Folds
The Fold Generator is responsible for partitioning the full set of examples into a group of folds, which are the atomic unit for assembling training/validation/testing data sets. Zero2Neuro provides a range of options, depending on the organization of the incoming data and the goals of the experiment.
The specific approach is determined by the argument:
--data_fold_split=YYY
where YYY is one of identity, group-by-file, group-by-example, and random
Identity
The identity split simply performs a one-to-one map of data tables to folds.
--data_fold_split=identity
--data_files
data_A.csv
data_B.csv
data_C.csv
For this example, data table A becomes fold 0, table B becomes fold 1, and table C becomes fold 2.
Group-by-File
Multiple data tables can be merge together for the purposes of creating a single fold. This mapping is explicitly defined in the data_files declaration.
--data_split=group-by-file
--data_files
data_A.csv,0
data_B.csv,0
data_C.csv,1
In this example, the data examples from data_A and data_B are merged into fold 0. The examples from data_C are assigned to fold 1.
Random Split
When there is no explicit relationship between data files and the examples that contain them, it is common to randomly split the examples into folds.
--data_split=random
--data_n_folds=3
--data_files
data_A.csv
data_B.csv
data_C.csv
In this example, all of the data examples contained in data_A, data_B, and data_C are randomly partitioned into 3 separate folds.
Notes:
- It is not appropriate to use this method if there is substantial autocorrelation between data examples within individual files (e.g., if one file contains all of the data examples derived from a single experimental subject). In this situation, it is more apprioriate to use identity or group-by-file.
- If one is performing a classification task and there is an imbalance between the number of examples between the classes, then it is appropriate to use the random-stratify split method.
Random-Stratify Split
Not yet supported.
Group-by-Example
The tabular files can also contain an explicit assignment of individual examples to folds. In this situation, a column is added to each of the tabular files; this column must be declared using the data_groups argument:
--data_fold_split=group-by-example
--data_groups=FIELD
where FIELD is the column name.