This is an official implementation for ICCV 2021 paper "TRAR: Routing the Attention Spans in Transformers for Visual Question Answering". It currently includes the code for training TRAR on VQA2.0 and CLEVR dataset. Our TRAR model for REC task is coming soon.

• (2021/10/10) Release our TRAR-VQA project.
• (2021/08/31) Release our pretrained CLEVR TRAR model on train split: TRAR CLEVR Pretrained Models.
• (2021/08/18) Release our pretrained TRAR model on train+val split and train+val+vg split: VQA-v2 TRAR Pretrained Models
• (2021/08/16) Release our train2014, val2014 and test2015 data. Please check our dataset setup page DATA.md for more details.
• (2021/08/15) Release our pretrained weight on train split. Please check our model page MODEL.md for more details.
• (2021/08/13) The project page for TRAR is avaliable.

TRAR vs Standard Transformer

TRAR Overall

0. Installation
1. Dataset setup
2. Config Introduction
3. Training
4. Validation and Testing
5. Models

• Clone this repo

git clone https://github.com/rentainhe/TRAR-VQA.git
cd TRAR-VQA

• Create a conda virtual environment and activate it

conda create -n trar python=3.7 -y
conda activate trar

• Install CUDA==10.1 with cudnn7 following the official installation instructions
• Install Pytorch==1.7.1 and torchvision==0.8.2 with CUDA==10.1:

conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.1 -c pytorch

• Install Spacy and initialize the GloVe as follows:

pip install -r requirements.txt
wget https://github.com/explosion/spacy-models/releases/download/en_vectors_web_lg-2.1.0/en_vectors_web_lg-2.1.0.tar.gz -O en_vectors_web_lg-2.1.0.tar.gz
pip install en_vectors_web_lg-2.1.0.tar.gz

see DATA.md

In trar.yml config we have these specific settings for TRAR model

ORDERS: [0, 1, 2, 3]
IMG_SCALE: 8 
ROUTING: 'hard' # {'soft', 'hard'}
POOLING: 'attention' # {'attention', 'avg', 'fc'}
TAU_POLICY: 1 # {0: 'SLOW', 1: 'FAST', 2: 'FINETUNE'}
TAU_MAX: 10
TAU_MIN: 0.1
BINARIZE: False

• ORDERS=list, to set the local attention window size for routing.0 for global attention.
• IMG_SCALE=int, which should be equal to the image feature size used for training. You should set IMG_SCALE: 16 for 16 × 16 training features.
• ROUTING={'hard', 'soft'}, to set the Routing Block Type in TRAR model.
• POOLING={'attention', 'avg', 'fc}, to set the Downsample Strategy used in Routing Block.
• TAU_POLICY={0, 1, 2}, to set the temperature schedule in training TRAR when using ROUTING: 'hard'.
• TAU_MAX=float, to set the maximum temperature in training.
• TAU_MIN=float, to set the minimum temperature in training.
• BINARIZE=bool, binarize the predicted alphas (alphas: the prob of choosing one path), which means during test time, we only keep the maximum alpha and set others to zero. If BINARIZE=False, it will keep all of the alphas and get a weight sum of different routing predict result by alphas. It won't influence the training time, just a small difference during test time.

Note that please set BINARIZE=False when ROUTING='soft', it's no need to binarize the path prob in soft routing block.

TAU_POLICY visualization

For MAX_EPOCH=13 with WARMUP_EPOCH=3 we have the following policy strategy:

Train model on VQA-v2 with default hyperparameters:

python3 run.py --RUN='train' --DATASET='vqa' --MODEL='trar'

and the training log will be seved to:

results/log/log_run_<VERSION>.txt

Args:

• --DATASET={'vqa', 'clevr'} to choose the task for training
• --GPU=str, e.g. --GPU='2' to train model on specific GPU device.
• --SPLIT={'train', 'train+val', train+val+vg'}, which combines different training datasets. The default training split is train.
• --MAX_EPOCH=int to set the total training epoch number.

Resume Training

Resume training from specific saved model weights

python3 run.py --RUN='train' --DATASET='vqa' --MODEL='trar' --RESUME=True --CKPT_V=str --CKPT_E=int

• --CKPT_V=str: the specific checkpoint version
• --CKPT_E=int: the resumed epoch number

Multi-GPU Training and Gradient Accumulation

1. Multi-GPU Training: Add --GPU='0, 1, 2, 3...' after the training scripts.

python3 run.py --RUN='train' --DATASET='vqa' --MODEL='trar' --GPU='0,1,2,3'

The batch size on each GPU will be divided into BATCH_SIZE/GPUs automatically.

2. Gradient Accumulation: Add --ACCU=n after the training scripts

python3 run.py --RUN='train' --DATASET='vqa' --MODEL='trar' --ACCU=2

This makes the optimizer accumulate gradients for n mini-batches and update the model weights once. BATCH_SIZE should be divided by n.

Warning: The args --MODEL and --DATASET should be set to the same values as those in the training stage.

Validate on Local Machine Offline evaluation only support the evaluations on the coco_2014_val dataset now.

1. Use saved checkpoint

python3 run.py --RUN='val' --MODEL='trar' --DATASET='{vqa, clevr}' --CKPT_V=str --CKPT_E=int

2. Use the absolute path

python3 run.py --RUN='val' --MODEL='trar' --DATASET='{vqa, clevr}' --CKPT_PATH=str

Online Testing All the evaluations on the test dataset of VQA-v2 and CLEVR benchmarks can be achieved as follows:

python3 run.py --RUN='test' --MODEL='trar' --DATASET='{vqa, clevr}' --CKPT_V=str --CKPT_E=int

Result file are saved at:

results/result_test/result_run_<CKPT_V>_<CKPT_E>.json

You can upload the obtained result json file to Eval AI to evaluate the scores.

Here we provide our pretrained model and log, please see MODEL.md

• openvqa
• grid-feats-vqa

if TRAR is helpful for your research or you wish to refer the baseline results published here, we'd really appreciate it if you could cite this paper:

@InProceedings{Zhou_2021_ICCV,
    author    = {Zhou, Yiyi and Ren, Tianhe and Zhu, Chaoyang and Sun, Xiaoshuai and Liu, Jianzhuang and Ding, Xinghao and Xu, Mingliang and Ji, Rongrong},
    title     = {TRAR: Routing the Attention Spans in Transformer for Visual Question Answering},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {2074-2084}
}