How to Calculate grid features from given pretrained X152pp model

[ifmaq1]

I am trying to extract the features for test split of Coco dataset. I used the pretrained model of Backbone X-152++

to extract features I used
python3.6 extract_grid_feature.py --config-file configs/X-152-challenge.yaml --dataset coco_2015_test MODEL.WEIGHTS /home/grid-feats-vqa-master/output/X-152pp.pth

However, when I matched the calculated features with https://github.com/facebookresearch/mmf features for movie_mcan model. I found a huge file size gap between this repository features and MMF's

i-e

The test split image features for X-152pp

The one I calculated myself is of size
3.9M 1024.pth

The one I downloaded from mmf for Movie_MCAN is of size
8.1M 1024.pth

the features that I calculated for the whole test split of vqa2 for X-152pp are of size approx. 320gb
please let me know how there could be huge file size difference

[endernewton]

Hmm, I am not sure. Is it because of padding?

[ifmaq1]

padding in MMF features?
Also due to this there is a huge accuracy difference on MOVIE_MCAN model. Can you please somehow tell where the problem lies. Maybe confirming from https://github.com/facebookresearch/mmf!!