trellixvulnteam/lxmert-vqacp_w0cc

This is an adaptation of LXMERT on both VQA-CP and VQA dataset.

LXMERT for VQA-CP and VQA.

This repo made a few modifications to support both VQA-CP and VQA datasets. Please find more details at the original LXMERT code.

We mainly use this repo to implement our paper - Loss Re-scaling VQA: Revisiting the Language Prior Problem from a Class-imbalance View.

Pre-trained models

The pre-trained model (870 MB) is available at http://nlp.cs.unc.edu/data/model_LXRT.pth, and can be downloaded with:

mkdir -p snap/pretrained 
wget https://nlp.cs.unc.edu/data/model_LXRT.pth -P snap/pretrained

Fine-tuning on VQA-CP or VQA

1. Please make sure the LXMERT pre-trained model is either downloaded or pre-trained.

2. Note that we DO NOT use the re-distributed json file provided by LXMERT authors. We use the official splits in this repo. Make sure that these data are in the right position according to the src/config.py!

3. Download faster-rcnn features for MS COCO train2014 (17 GB) and val2014 (8 GB) images (VQA 2.0 is collected on MS COCO dataset).

   mkdir -p data/mscoco_imgfeat
   wget https://nlp.cs.unc.edu/data/lxmert_data/mscoco_imgfeat/train2014_obj36.zip -P data/mscoco_imgfeat
   unzip data/mscoco_imgfeat/train2014_obj36.zip -d data/mscoco_imgfeat && rm data/mscoco_imgfeat/train2014_obj36.zip
   wget https://nlp.cs.unc.edu/data/lxmert_data/mscoco_imgfeat/val2014_obj36.zip -P data/mscoco_imgfeat
   unzip data/mscoco_imgfeat/val2014_obj36.zip -d data && rm data/mscoco_imgfeat/val2014_obj36.zip
   

4. We convert the image features from tsv to h5 first:

   python src/tools/detection_feature_converter.py
   

   We fold the train and val image features together for supporting both VQA-CP and VQA.

5. Process answers and question types:

   python src/tools/compute_softscore.py 
   

6. Fine-tuning on VQA-CP or VQA (set this on the src/config.py):

   PYTHONPATH=$PYTHONPATH:./src \
   python -u src/tasks/vqa.py \
   --train train --valid val  \
   --llayers 9 --xlayers 5 --rlayers 5 \
   --loadLXMERTQA snap/pretrained/model \
   --batchSize 32 --optim bert --lr 5e-5 --epochs 4 \
   --tqdm
   --name vqa-cp-test
   

7. Evaluating on the validation set (according to the official implementation):

   PYTHONPATH=$PYTHONPATH:./src \
   python -u src/tasks/vqa.py \
   --train train --test val  \
   --llayers 9 --xlayers 5 --rlayers 5 \
   --loadLXMERTQA snap/pretrained/model \
   --batchSize 32 --load output/vqa-cp-test.pth \
   --tqdm
   

   python acc_per_type.py output/val_predict.json
   

Performance on VQA-CP test

Loss Function	Model	Y/N	Num.	Others	All
BCE	LXMERT	46.70	27.14	61.20	51.78
BCE	LXMERT+Ours	79.77	59.06	61.41	66.40
CE	LXMERT	-	-	-	58.07
CE	LXMERT+Ours	-	-	-	69.37

Citation

If you found this repo useful, please cite the following paper:

@article{rescale-vqa,
  title={Loss Re-scaling VQA: Revisiting the Language Prior Problem from a Class-imbalance View},
  author={Guo, Yangyang and Nie, Liqiang and Cheng, Zhiyong and Tian, Qi and Zhang, Min},
  journal={IEEE TIP},
  year={2021}
}