Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Co-attending Regions and Detections with Multi-modal Multiplicative Embedding for VQA.

The network has two attention branches with the proposed multiplicative feature embedding scheme: one branch attends free-form image regions, another branch attends detection boxes for encoding question-related visual features.

This current code can get 66.09 on Open-Ended and 69.97 on Multiple-Choice on test-standard split for the VQA 1.0 dataset.


  • Paper on arXiv:
  • Bounding box-based attention with Faster R-CNN pre-trained model
  • Co-attention with whole image and object detection
  • Multi-modal multiplicative embedding feature method
  • Torch implementation with multi-GPU acceleration

 | center


This main part of code is written in Lua and requires Torch. After installing torch, you can install these dependencies by running the following:

Installing related torch libraries
cd ~/torch
luarocks install loadcaffe
luarocks install hdf5
pip install h5py
luarocks install optim
luarocks install nn
luarocks install math
luarocks install image
luarocks install dp
Installing rnn dependencies
cd ~/torch
git clone
cd rnn
luarocks make rocks/rnn-scm-1.rockspec
Installing Cjson
cd /usr/local/
sudo wget
sudo tar -xzvf  lua-cjson-2.1.0.tar.gz
cd lua-cjson-2.1.0
sudo luarocks make
sudo rm ../lua-cjson-2.1.0.tar.gz
Installing NLTK
cd /usr/share/
sudo mkdir nltk_data
sudo pip install -U nltk
python -m nltk.downloader all
Installing GPU acceleration
  • If have an NVIDIA GPU and want to accelerate the model with CUDA, you'll also need to install torch/cutorch and torch/cunn; you can install these by running:
luarocks install cutorch
luarocks install cunn
luarocks install cudnn
Installing cuDNN (Optional)
  • If you want to use NVIDIA's cuDNN library, you'll need to register for the CUDA Developer Program (it's free) and download the library from NVIDIA's website. An running script just for an example:
cd ~/torch
# download the right cudnn file to cuda version
tar -xzvf  cudnn-7.5-linux-x64-v5.1.tgz
sudo cp cuda/lib64/libcudnn* /usr/local/cuda-7.5/lib64/
sudo cp cuda/include/cudnn.h /usr/local/cuda-7.5/include/
luarocks install cudnn
Installing nccl for Multi-GPU Optimization (Optional)
  • Optimized primitives for collective multi-GPU communication. Note that NVIDIA/nccl is an optimization tool, and you can run the codes on multi-GPUs without nccl.
cd torch/
git clone
# build the library
cd nccl/
make CUDA_HOME=/usr/local/cuda-7.5 test
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/plu/torch/nccl/build/lib
source ~/.bashrc
# test demo
$ ./build/test/single/all_reduce_test
$ ./build/test/single/all_reduce_test 10000000
luarocks install nccl

Caffe and faster-rcnn

Extracting and visualizing bounding boxes are supported by caffe and py-faster-rcnn. You can install Caffe and faster-rcnn following the instructions.

Then copy the faster-rcnn for vqa files to the target folder:

cp ~/dual-mfa-vqa/faster-rcnn-vqa/tools/*.py  ~/py-faster-rcnn/tools/

VQA Dataset

Download mscoco images

mkdir -p ~/VQA/Images/mscoco
cd ~/VQA/Images/mscoco
cd ~/VQA/Images/mscoco
cd ~/VQA/Images/mscoco
ln -s test2015 test-dev2015

Download VQA question and answers

mkdir -p ~/VQA/Annotations
cd ~/dual-mfa-vqa/data_train-val_test-dev_2k
python --download 1
cd ~/dual-mfa-vqa/data_train_test-dev_2k
cd ~/dual-mfa-vqa
th prepro/prepro_seconds.lua

Download resnet image model

mkdir -p ~/VQA/Images/Image_model
cd ~/VQA/Image_model

Download faster-rcnn model

cd ~/py-faster-rcnn/data/
mkdir faster_rcnn_models
cd faster_rcnn_models
mv coco_vgg16_faster_rcnn_final.caffemodel?dl=0  coco_vgg16_faster_rcnn_final.caffemodel

Download skipthoughts models

You can download the pretrained Skipthoughts models to folder skipthoughts_model/ for learning (See more details):

Download pretrained VQA models

This current code can get 66.01 on Open-Ended and 70.04 on Multiple-Choice on test-tev split for the VQA 1.0 dataset. Download the pre-trained model vqa_dual-mfa_model_6601.t7 (315M) from here into folder dual-mfa-vqa/model/save/.

Extracting Image Features

Free-form region based features

  • Image features for train dataset
cd prepro
th prepro_res_train.lua -batch_size 8
  • Image features for test-dev dataset
th prepro_res_test.lua -batch_size 8

Detection region based features

  • Image features for train dataset
  • Image features for test dataset
  • Bounding box coordinates for test dataset. You can directly download the result file faster-rcnn_box4_19_test.h5 from here.


Now, everything is ready, let's train the vqa network. Here are some common training ways for different needs.

  • Training the network on train dataset and validating on val dataset
th train.lua -phase 1 -val_nqs -1 -nGPU 4
  • Training the network on train-val dataset with multi-GPUs (4 GPUs)
th train.lua -phase 2 -nGPU 4 -batchsize 300
  • Training the network loading image features into memory (much faster, 200-300G memory is needed)
th train.lua -phase 1 -val_nqs 10000 -nGPU 4 -memory_ms -memory_frms
  • Training the network from the previous checkpoint
th train.lua -phase 2 -nGPU 4 -memory_ms -load_checkpoint_path model/save/vqa_model_dual-mfa_6601.t7 -previous_iters 350000
  • Main options useful training are listed as follows:
    • phase:training phase, 1: train on Train, 2: train on Train+Val
    • vqa_type: vqa dataset type, vqa or coco-qa
    • memory_ms: load image resnet feature to memory
    • memory_frms: load image fast-rcnn feature to memory
    • val: running validation
    • val_nqs: number of validation questions, -1 for all questions
    • batch_size: batch_size for each iterations, change it to smaller value if out of the memory
    • run_id: running model id
    • model_label: model label name
    • save_checkpoint_every: how often to save a model checkpoint
    • skip_save_model: skip saving t7 model
    • cg_every: How often do we collectgarbage in the training process, change it to smaller value if out of the memory
    • quick_check: quick check for code
    • quickquick_check: very quick check for code
    • nGPU: how many GPUs to use, 1 = use 1 GPU, change it to larger value if out of the memory


Evaluate the pre-trained model on VQA dataset:

cd ~/dual-mfa-vqa
th eval.lua -model_path model/vqa_model_dual-mfa_6601.t7 -output_model_name vqa_model_dual-mfa_6601 -batch_size 10

Then you can submit the result jsons and obtain the evaluation scores:

COCO-QA Dataset

  • Download the dataset
cd data_coco
python --download 1
  • Preprocess the dataset
  • Extract the free-form image features
cd prepro
th prepro_res_coco.lua -batch_size 8
  • Training the network
th train.lua -vqa_type coco-qa -learning_rate 4e-4 -nGPU 4 -batch_size 300 \
-model_id 1 -model_label dual-mfa
  • Evaluation based on WUPS
cd ~/dual-mfa-vqa/metric
python gt_ans_save.txt pd_ans_save.txt 0.9 
python gt_ans_save.txt pd_ans_save.txt 0.0 


  • Generate the attention maps. You can directly download the result file of attention map from here.
cd ~/dual-mfa-vqa
th eval_vis_att.lua -model_path model/vqa_model_dual-mfa_6601.t7 -output_model_name vqa_model_dual-mfa_6601 -batch_size 8
  • Preprocess the question data before visualization:
cd vis_att 
  • Run the matlab file vis_attention_demo.m to show the results of attention maps.
  • Run the matlab file vis_attention.m to save the results of attention maps.

 | center


If you use this code as part of any published research, please acknowledge the following paper.

	title={Co-attending Free-form Regions and Detections with Multi-modal Multiplicative Feature Embedding for Visual Question Answering.},
	author={Lu, Pan and Li, Hongsheng and Zhang, Wei and Wang, Jianyong and Wang, Xiaogang},
	booktitle={AAAI 2018},