D-VQA
We provide the PyTorch implementation for Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021).
Dependencies
- Python 3.6
- PyTorch 1.1.0
- dependencies in requirements.txt
- We train and evaluate all of the models based on one TITAN Xp GPU
Getting Started
Installation
-
Clone this repository:
git clone https://github.com/Zhiquan-Wen/D-VQA.git cd D-VQA -
Install PyTorch and other dependencies:
pip install -r requirements.txt
Download and preprocess the data
cd data
bash download.sh
python preprocess_features.py --input_tsv_folder xxx.tsv --output_h5 xxx.h5
python feature_preprocess.py --input_h5 xxx.h5 --output_path trainval
python create_dictionary.py --dataroot vqacp2/
python preprocess_text.py --dataroot vqacp2/ --version v2
cd ..
Training
- Train our model
CUDA_VISIBLE_DEVICES=0 python main.py --dataroot data/vqacp2/ --img_root data/coco/trainval_features --output saved_models_cp2/ --self_loss_weight 3 --self_loss_q 0.7
- Train the model with 80% of the original training set
CUDA_VISIBLE_DEVICES=0 python main.py --dataroot data/vqacp2/ --img_root data/coco/trainval_features --output saved_models_cp2/ --self_loss_weight 3 --self_loss_q 0.7 --ratio 0.8
Evaluation
- A json file of results from the test set can be produced with:
CUDA_VISIBLE_DEVICES=0 python test.py --dataroot data/vqacp2/ --img_root data/coco/trainval_features --checkpoint_path saved_models_cp2/best_model.pth --output saved_models_cp2/result/
- Compute detailed accuracy for each answer type:
python comput_score.py --input saved_models_cp2/result/XX.json --dataroot data/vqacp2/
Pretrained model
A well-trained model can be found here. The test results file produced by it can be found here and its performance is as follows:
Overall score: 61.91
Yes/No: 88.93 Num: 52.32 other: 50.39
Reference
If you found this code is useful, please cite the following paper:
@inproceedings{D-VQA,
title = {Debiased Visual Question Answering from Feature and Sample Perspectives},
author = {Zhiquan Wen,
Guanghui Xu,
Mingkui Tan,
Qingyao Wu,
Qi Wu},
booktitle = {NeurIPS},
year = {2021}
}
Acknowledgements
This repository contains code modified from SSL-VQA, thank you very much!
Besides, we thank Yaofo Chen for providing MIO library to accelerate the data loading.
下面是我的代码
def plot_rect(image, boxes):
img = Image.fromarray(np.uint8(image))
draw = ImageDraw.Draw(img)
for k in range(2):
box = boxes[k,:]
print(box)
drawrect(draw, box, outline='green', width=3)
img = np.asarray(img)
return img
def drawrect(drawcontext, xy, outline=None, width=0):
x1, y1, x2, y2 = xy
points = (x1, y1), (x2, y1), (x2, y2), (x1, y2), (x1, y1)
drawcontext.line(points, fill=outline, width=width)
7 May 29, 2022
1.3k Jan 05, 2023
8 May 22, 2022
1.8k Jan 08, 2023
4.5k Jan 03, 2023
577 Dec 17, 2022
9 Mar 29, 2022
9 Nov 28, 2022
46 Oct 16, 2022
68 Oct 18, 2022
243 Jan 07, 2023
6 Sep 30, 2022
47 Jul 26, 2022
9 Feb 23, 2022
143 Jan 05, 2023
125 Dec 31, 2022
1.1k Jan 03, 2023
2 Mar 13, 2022
22 Nov 03, 2022
2 Dec 14, 2022