search

airsplay / lxmert

PyTorch code for EMNLP 2019 paper "LXMERT: Learning Cross-Modality Encoder Representations from Transformers".
MIT License
923 stars 157 forks source link

usage of tsv file (image feature extracted) #74

Open coldtomatojuice opened 4 years ago

coldtomatojuice commented 4 years ago

Hello guys first of all, thank you for your amazing works

I tried to use the file that you uploaded on Google Cloud vg_gqa_obj36.tsv & gqa_testdev_obj36.tsv which contains image features extracted by RCNN

while those files are splitted into vg_gqa and gqa_testdev in vg_gqa, I guess images from all question split data (I mean train, val, test, challenge, submission) except testdev are included in one file therefore, I have no idea how I could know which split does the image feature comes from (train or challenge?)

Thank you!

airsplay commented 4 years ago

Here is a mapping for the split of GQA to their features:

gqa_train: vg_gqa_obj36.tsv
gqa_valid: vg_gqa_obj36.tsv
gqa_testdev: gqa_testdev_obj36.tsv
gqa_test: vg_gqa_obj36.tsv (the image ids for this split are started with 'n', e.g., n000XXXX)