captions_to_gt_for_eval.py

import json

"""
Grungy script to convert clipcap-style caption files to groudn-truth coco-caption eval style

I use https://github.com/EricWWWW/image-caption-metrics plus a patch to filter out '\n' and '\r'
"""

filenames = [
  'data/coco/karpathy_split_annotations/test_caption.json',
  'data/coco/karpathy_split_annotations/train_caption.json',
  'data/coco/karpathy_split_annotations/val_caption.json',
  'data/coco/annotations/train_caption.json',
]

for filename in filenames:
  print(f'Processing {filename}...')
  d = json.load(open(filename))
  output = {}
  for i in d:
    image_id = i['image_id']
    output.setdefault(image_id, [])
    output[image_id].append(i['caption'])
  json.dump(output, open(filename + '_evaluation.json', 'w'))