BKG-CLEVR is a 2D CLEVR-like dataset generator that uses Answer Set Programming as the engine for generating positive and negative examples. These examples are divided between those who satisfy a given constraints and those who do not. Furthermore, a neural network model is provided to test the capacity of neural approaches to learn the aforementioned constraint.
- Install the requirements in requirements.txt.
- Use dataset/unified_generation.py or dataset/unified_generation_asp.py to generate images. They will be contained in a folder called test_dataset.
- Use utils/split_dataset_binary, split_dataset_odc.py and yolosify.py to split the image into training and validation sets.
- Use a models/binary_classification.py to train a binary classifier, or obj_detect_classification for training object detection and classification.