This is the code for the paper on CLEVR

Visual Question Reasoning on General Dependency Tree
Qingxing Cao, Xiaodan Liang, Bailin Li, Guanbin Li, Liang Lin
Presented at CVPR 2018 (Spotlight Presentation)

If you find this code useful in your research then please cite

@InProceedings{Cao_2018_CVPR,
author = {Cao, Qingxing and Liang, Xiaodan and Li, Bailing and Li, Guanbin and Lin, Liang},
title = {Visual Question Reasoning on General Dependency Tree},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}

• tensorboardX
• skimage
• scipy
• numpy
• torchvision
• h5py
• tqdm

Before you can train any models, you need to download the datasets; you also need to preprocess questions, and extract features for the images.

You can download CLEVR v1.0 (18 GB) with the common below.

$ sh data/clevr/download_dataset.sh

Codes for preprocessing would be available soon. For now you can download our preprocessed data with the following command:

$ sh data/clevr/download_preprocessed_questions.sh

You can extract image features with the command below.

$ sh scripts/extract_image_feature.sh

The extracted features features_train.h5, features_val.h5, features_test.h5 woulde be placed in ./data/clevr/clevr_res101/.

You can download the pretrained models with the command below. The model will take about 2.6 GB on disk.

$ sh data/clevr/download_pretrained_model.sh

It is trained on CLEVR-train and can be validate on CLEVR-val.

You can use the train_val.py script to train on CLEVR-train and validate the model on CLEVR-val.

$ python scripts/train_val.py --clevr_qa_dir=data/clevr/clevr_qa_dir/ --clevr_img_h5=data/clevr/clevr_res101/

The below script has the hyperparameters and settings to reproduce ACMN CLEVR results.

$ sh scripts/train_val.sh

You can use train_val.py to simply evaluate the model on CLEVR-val with --no_train option to skip the training process.

$ python scripts/train_val.py \
  --no_train=True \
  --clevr_qa_dir=data/clevr/clevr_qa_dir/ \
  --clevr_img_h5=data/clevr/clevr_res101/ \
  --resume=data/clevr/clevr_pretrained_model.pth

You can use test.py to generate CLEVR-test results in .json format so that you can upload to CLEVR official.

$ python scripts/test.py \
  --clevr_qa_dir=data/clevr/clevr_qa_dir/ \
  --clevr_img_h5=data/clevr/clevr_res101/ \
  --resume=data/clevr/clevr_pretrained_model.pth

You can use vis.py to visualize the attention maps discribed in Figure 4 of our paper.

$ python scripts/vis.py \
  --clevr_qa_dir=data/clevr/clevr_qa_dir/ \
  --clevr_img_h5=data/clevr/clevr_res101/ \
  --clevr_img_png=data/clevr/CLEVR_v1.0/ \
  --clevr_load_png=True \
  --logdir=logs/attmaps \
  --resume=data/clevr/clevr_pretrained_model.pth