Mentions légales du service

Skip to content

Support for per dataset split transforms

E. Madison Bray requested to merge embray/per-split-transforms into master

This adds support for a new way to configure transforms, though the old way is still supported.

The old way is the same:

dataset_transforms:
    - transform1
    - transform2

The same transforms are applied across the dataset. The new, optional format is like:

dataset_transforms:
    training:
        - transform1
        - transform2
    validation:
        - transform3
    test:
        - transform4

So as discussed in #68 (closed) / #80 (closed) it is possible to apply different lists of transforms to each dataset split. Note that if using this format, if you only list transforms for, say "training", then no transforms are applied to the "validation" set, and so on.

Also supports list of transforms for the test set, even though we don't use this capability, yet.

Merge request reports