We provide access to our preprocessed data and preprocessing scripts to replicate our setup.
prepare_data.sh provides sample execution for data preparation. It calls prepare_data.py.
Note that the output corpus directory might look different than the final one (see below).
The corpus directory looks as follows:
corpus/
|-- coco2014/
| |-- annotations/
| | |-- captions_train2014.json
| | |-- captions_val2014.json
| | |-- instances_train2014.json
| | |-- instances_val2014.json
| |-- images/
| | |-- coco_detections.hdf5 (used by M2-TRM)
| | |-- image_features.hdf5
| | |-- images.hdf5
|
|-- compgen/
| |-- annotations/
| | |-- captions.json
| | |-- ccg_captions.json
| | |-- chunk_captions.json
| | |-- dep_captions.json
| | |-- head_captions.json
| | |-- idle_captions.json
| | |-- lemma_captions.json
| | |-- pos_captions.json
| |-- concepts/
| | |-- occurrences/
| | | |-- big_bird.json
| | | |-- big_cat.json
| | | |-- ...
| | | |-- white_truck.json
| | |-- synonyms/
| | | |-- adjectives/
| | | | |-- big.json
| | | | |-- black.json
| | | | |-- ...
| | | | |-- young.json
| | | |-- nouns/
| | | |-- verbs/
| |-- datasets/
| | |-- coco_heldout_1/
| | | |-- captions.json
| | | |-- dataset_splits.json
| | | |-- encoded_captions.json
| | | |-- word_map.json
| | |-- coco_heldout_1_ccg_inter/
| | |-- coco_heldout_1_ccg_multi/
| | |-- coco_heldout_1_ccg_seq/
| | |-- coco_heldout_1_chunk_inter/
| | |-- ...