coco-train-words.p

[DesaleF]

I am trying to finetune Oscar(vinvl) on my own dataset using vinvl feature. I extracted the features using scene_graph_benchmark repo. My custom data is prepared using coco-caption format. Does anyone know how to prepare coco-train-words.p file for my own custom dataset?

[jontooy]

Hi DesaleF,

I am in the same situation, but have not had the time to look into coco-train-words.p yet, although I found this issue on the cider github, which may hint towards preparing your own coco-train-words.p Let me know how it goes!

My current understanding is that the pickle file (.p file) contains dataset document-frequencies that is used when calculating the cider-scores. This pickle file should be passed into the function pycocoevalcap.eval when evaluation your captions. Someone please correct me if I'm wrong.

[DesaleF]

Hello @jontooy! Thank you for the information. I was taking a look at the file a little bit. this is basically the highlight what is in the "coco-train-words.p". What I don't understand was how we can create the document_frequency in the pickle file.

`

import pickle
with open("datasets/coco_caption/coco-train-words.p", "rb") as f:
... words_p = pickle.load(f)
...
words_p.keys()
dict_keys(['document_frequency', 'ref_len'])
words_p["ref_len"]
113287
len(list(words_p["document_frequency"]))
3636892
type(words_p["document_frequency"])
<class 'collections.defaultdict'>
len(list(words_p["document_frequency"].keys()))
3636892
list(words_p["document_frequency"].items())[:5]
[(('knife',), 855.0), (('her', 'head', 'cutting'), 1.0), (('cutting', 'a'), 502.0), (('a', 'chefs', 'knife'), 7.0), (('on', 'her', 'head'), 54.0)]
`

I will take a look at the git issue that you pointed me above and I will reply what I found here.