readme.md

Documentation for variants

The variants specification should be as follows:

dataset_1:
  variant_1:
    ...
  variant_2:
    ...
dataset_2:
  variant_1:
    ...
  variant_2:
    ...

The default configs of the datasets dataset_1 and dataset_2 are expected to be available in corresponding dataset/dataset_1.yaml and dataset/dataset_2.yaml files.

For each variant, we can specify a base variant to inherit properties from. If the variant's parent is null then the defaults are taken from the dataset's standard definition.

The ... content above is itself a dictionary expressed in YAML and contains the updates to be applied to the base dataset. These updates are applied recursively in case of nested variants, which we can achieve with the parent field.

Shortcuts:

• If the parent key is missing, parent=null is inferred.
• Some variant names (crop, occlusion, object_color, object_style, object_shape) can be automatically filled with the appropriate transform and variant type, if not specified. Defaults are defined in _get_variant_defaults() in data/dataset_variants.py.

Usage example:

# Standard Tetrominoes dataset (no variant specified)
python train_object_discovery.py [...] dataset=tetrominoes

# Apply cropping to Tetrominoes
python train_object_discovery.py [...] dataset=tetrominoes +dataset.variant=crop

# Standard CLEVR-6 dataset
python train_object_discovery.py [...] dataset=clevr +dataset.variant=6

# Only images with more than 6 objects
python train_object_discovery.py [...] dataset=clevr +dataset.variant=greater_than_6

where [...] is to be replaced with other arguments. The + syntax in hydra will append the config key dataset.variant which is otherwise missing by default.