Hi,
Could I ask if the Oscar VQA task was updated for the use of VinVL image features?
I have this concern for several reasons:
(1)The script for finetuing VQA task is still using the mask-rcnn and faster-rcnn features: https://github.com/microsoft/Oscar/blob/master/VinVL_MODEL_ZOO.md#VQA
(2) See the Github history, the "VinVL Release" version committed on 8 Mar, I see only four lines of updates for the "run_vqa.py", and they are all commented.
(3) I searched for the ".tsv" and only see "{}_img_frcnn_feats.tsv" is mentioned. Not "predictions.tsv"/"labels.tsv"/"features.tsv" which are the VinVL image features.
Could you teach me how to use VinVL features on the Oscar VQA task? I know how to prepare VinVL features (.yaml, predictions.tsv,lineidx, etc) but am not sure how to integrate it into the run_vqa.py.
Thanks a lot in advance!
Hi, Could I ask if the Oscar VQA task was updated for the use of VinVL image features? I have this concern for several reasons: (1)The script for finetuing VQA task is still using the mask-rcnn and faster-rcnn features: https://github.com/microsoft/Oscar/blob/master/VinVL_MODEL_ZOO.md#VQA (2) See the Github history, the "VinVL Release" version committed on 8 Mar, I see only four lines of updates for the "run_vqa.py", and they are all commented.
(3) I searched for the ".tsv" and only see "{}_img_frcnn_feats.tsv" is mentioned. Not "predictions.tsv"/"labels.tsv"/"features.tsv" which are the VinVL image features.
Could you teach me how to use VinVL features on the Oscar VQA task? I know how to prepare VinVL features (.yaml, predictions.tsv,lineidx, etc) but am not sure how to integrate it into the run_vqa.py. Thanks a lot in advance!