AndersonStra/MuKEA

How to acquire the file of ' vqa_img_feature_train.pickle'?

Closed this issue · 3 comments

coral3333 commented

How to acquire the file of ' vqa_img_feature_train.pickle'?

AndersonStra commented

Image features are recommended to download from the LXMERT's repo or bottom-up attention 's repo, We reorganized the storage structure of features as

vqa_img_feature_train.pickle{
"image_id":{'feats': features, 'sp_feats': spatial features}
}

The pre-trained LXMERT model expects these spatial features to be normalized bounding boxes on a scale of 0 to 1

The normalized process can refer to the prepare_img.py

we will provide the shell to help generate the pickle file from the original bottom-up attention file further.

  • 👍1
AndersonStra commented

I have updated the script to help generate the pickle file from the original bottom-up attention file.

coral3333 commented