CJJ2923 / RWSAN

Official implementation of "Deep Residual Weight-Sharing Attention Network with Low-Rank Attention for Visual Question Answering" (RWSAN) published in the IEEE Transactions on Multimedia (TMM), 2022.

RWSAN

This repo is the official implementation of "Deep Residual Weight-Sharing Attention Network with Low-Rank Attention for Visual Question Answering" published in the IEEE Transactions on Multimedia (TMM), 2022. (Paper) (Code)

Citation

If this repository is helpful to your research, we'd really appreciate it if you could cite the following paper:

@ARTICLE{9770348,
  author={Qin, Bosheng and Hu, Haoji and Zhuang, Yueting},
  journal={IEEE Transactions on Multimedia}, 
  title={Deep Residual Weight-Sharing Attention Network with Low-Rank Attention for Visual Question Answering}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMM.2022.3173131}}

Hardware Requirements

1. 1 Nvidia GPU (recommend >=6GB memory).
2. 50GB free disk space (recommend SSD).

Software Requirements

1. python >= 3.6 (Project works with Python 3.6~3.9)
2. cudatoolkit >= 10.1 (Project works with cudatookit 10.1, 10.2, and 11.1)
3. pytorch >= 1.7.0 with CUDA (Project works with PyTorch 1.7.0~1.9.1)
4. pytorch-nlp

Dataset

Please download the processed VQA-v2 and VG datasets in OneDrive or BaiduYun and place them in the root folder as follows:

|-- datasets
	|-- coco_extract
	|  |-- train2014.tar.gz
	|  |-- val2014.tar.gz
	|  |-- test2015.tar.gz
	|-- vqa
	|  |-- VG_questions.json
	|  |-- VG_annotations.json

After that, run the following script to download QA files and unzip the image features.

$ sh setup.sh

Finally, the datasets folders shall have the following structure:

|-- datasets
	|-- coco_extract
	|  |-- train2014
	|  |  |-- COCO_train2014_...jpg.npz
	|  |  |-- ...
	|  |-- val2014
	|  |  |-- COCO_val2014_...jpg.npz
	|  |  |-- ...
	|  |-- test2015
	|  |  |-- COCO_test2015_...jpg.npz
	|  |  |-- ...
	|-- vqa
	|  |-- v2_OpenEnded_mscoco_train2014_questions.json
	|  |-- v2_OpenEnded_mscoco_val2014_questions.json
	|  |-- v2_OpenEnded_mscoco_test2015_questions.json
	|  |-- v2_OpenEnded_mscoco_test-dev2015_questions.json
	|  |-- v2_mscoco_train2014_annotations.json
	|  |-- v2_mscoco_val2014_annotations.json
	|  |-- VG_questions.json
	|  |-- VG_annotations.json

Training and Evaluate on the Val Split

Run the following command to train the RWSAN on train split of the VQA-v2 dataset and evaluate on the val split.

python3 run.py --RUN='train' --VERSION='RWSAN_val' --GPU='0' --SPLIT='train' --ACCU=1 --NW=4

Command

1. --VERSION=str: The name of this experiment.
2. --GPU=str: Use the specific GPU device.
3. --SPLIT={'train', 'train+val', 'train+val+vg'}: The training set you want to use.
4. --ACCU=int: Gradient accumulation when GPU memory is not sufficient. 1 for not using gradient accumulation. Note that BATCH_SIZE must be divided by ACCU. (The default BATCH_SIZE is 64, so the --ACCU can be 1, 2, 4, 8 ...).
5. --NW=int: Number of processes to read the dataset. The pre-load data is stored in the memory, and a larger number results in more memory cost. During our experiment, 4 is optimal for both training speed and memory cost.

The checkpoints are stored in ./ckpts/ckpt_RWSAN_val/, and the log files for average training loss and performance on val split in every epoch are stored in ./results/log/.

Resume Training

If the training process is interrupted, run the following command to resume training.

python3 run.py --RUN='train' --VERSION=str --GPU='0' --SPLIT='train' --ACCU=1 --NW=4 --RESUME=True --CKPT_V=str --CKPT_E=int

Command

1. --VERSION=str --CKPT_V=str: The name of the model which is going to resume training.
2. --CKPT_E=int: The number of epoch you want to resume.

For example, if you want to resume training the model RWSAN_val from epoch 8, please run the following command.

python3 run.py --RUN='train' --VERSION='RWSAN_val' --GPU='0' --SPLIT='train' --ACCU=1 --NW=4 --RESUME=True --CKPT_V='RWSAN_val' --CKPT_E=8

Evaluation

If the log file is lost and you want to reevaluate the performance of RWSAN in a specific epoch on val split of the VQA-v2 dataset, run the following command.

python3 run.py --RUN='val' --VERSION=str --GPU='0' --SPLIT='train' --ACCU=1 --NW=4 --RESUME=True --CKPT_V=str --CKPT_E=int

Command

1. --VERSION=str --CKPT_V=str: The name of the model which is going to evaluate.
2. --CKPT_E=int: The specific number of epoch you want to evaluate.

For example, if you want to evaluate the performance RWSAN_val for epoch 16, please run the following command.

python3 run.py --RUN='val' --VERSION='RWSAN_val' --GPU='0' --SPLIT='train' --ACCU=1 --NW=4 --RESUME=True --CKPT_V='RWSAN_val' --CKPT_E=16

Training and Evaluate on the Test Split

Run the following command to train the RWSAN on train, val and vg splits of the VQA-v2 dataset.

python3 run.py --RUN='train' --VERSION='RWSAN_test' --GPU='0' --SPLIT='train+val+vg' --ACCU=1 --NW=4

After that, run the following code to generate the predictions on the test split.

python3 run.py --RUN='test' --VERSION='RWSAN_test' --GPU='0' --SPLIT='train+val+vg' --ACCU=1 --NW=4 --RESUME=True --CKPT_V='RWSAN_test' --CKPT_E=16

Prediction file is stored in ./results/result_test/

You could upload the result file to Eval AI to evaluate the model on test-dev and test-std splits.

Modifications

Modify the Path of Dataset

If you prefer to place the dataset in another path, you could modify the path of the dataset. The relating configuration file is ./cfgs/path_cfgs.py.

Modify the Hyper-Parameter

All training hyper-parameter is stored in the ./cfgs/base_cfgs.py.

Build Your Own Model

If you want to build your own model, please define the model under ./model/. Import your model to ./train/execution.py and define it as net. Then you could run your own model.

Note

1. The input of the network contains two parts, including img_feat_iter in [BATCH_SIZE, IMG_FEAT_PAD_SIZE, IMG_FEAT_SIZE] and ques_ix_iter in [BATCH_SIZE, MAX_TOKEN]. (Default BATCH_SIZE=64, IMG_FEAT_PAD_SIZE=100, IMG_FEAT_SIZE=2048, MAX_TOKEN=16.)
2. The network output is ans_iter in [BATCH_SIZE, answer_size]. The default answer_size=3129. (The program will calculate the answer_size automatically based on the provided ./data/answer_dict.json.)

Speed Up the Training Process

1. We already use the mix-precision training in the project to speed up the training process and save GPU memory. This strategy does not hurt the performance of RWSAN after testing.
2. This project supports multi-GPU training. However, please do not use multi-GPU as it does not speed up the training process.
3. A high-end CPU with fast single-core speed is recommended.

Acknowledgement

MCAN

bottom-up-attention