@inproceedings{tu2021learning,
  title={Learning Better Visual Dialog Agents with Pretrained Visual-Linguistic Representation},
  author={Tu, Tao and Ping, Qing and Thattai, Govindarajan and Tur, Gokhan and Natarajan, Prem},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5622--5631},
  year={2021}
}

Instance storage: > 800 GB
pytorch 1.4.0
cuda 10.0

$ conda create -n read_up python=3.6
$ conda activate read_up
$ git clone https://github.com/amazon-research/read-up.git

# [IMPORTANT] pytorch 1.4.0 have no issue for parallel training
$ conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.0 -c pytorch

# Install general dependencies
$ sudo apt-get install build-essential libcap-dev
$ pip install -r requirement.txt

# Install vqa-maskrcnn-benchmark (for feature extraction only)
$ git clone https://gitlab.com/vedanuj/vqa-maskrcnn-benchmark.git
$ cd vqa-maskrcnn-benchmark
$ python setup.py build develop

# Apex is used for both `faster-rcnn feature extraction` & `distributed training`
$ git clone https://github.com/NVIDIA/apex
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

$ wget https://florian-strub.com/guesswhat.train.jsonl.gz -P data/
$ wget https://florian-strub.com//guesswhat.valid.jsonl.gz -P data/
$ wget https://florian-strub.com//guesswhat.test.jsonl.gz -P data/

$ python src/guesswhat/preprocess_data/create_dictionary.py -data_dir data -dict_file dict.json -min_occ 3

$ cd read-up
$ mkdir tf-pretrained-model
$ cp guesswhat/data/dict.json read-up/tf-pretrained-model/

# download RCNN model
$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/detectron_model.pth
$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/detectron_config.yaml

# download COCO data  
$ wget http://images.cocodataset.org/zips/train2014.zip
$ wget http://images.cocodataset.org/zips/val2014.zip
$ wget http://images.cocodataset.org/zips/test2014.zip
$ unzip -j train2014.zip 
$ unzip -j valid2014.zip 
$ unzip -j test2014.zip 

$ python bin/prepare_extract_gt_features_gw.py \
    --src vqa-maskrcnn-benchmark/data/guesswhat.train.jsonl \
    --img-dir vqa-maskrcnn-benchmark/image_dir/COCO_2014/images/ \
    --out vqa-maskrcnn-benchmark/image_dir/COCO_2014/npy_files/guesswhat.train.npy

{
    {
        'file_name': 'name_of_image_file',
        'file_path': '<path_to_image_file_on_your_disk>',
        'bbox': array([
                        [ x1, y1, width1, height1],
                        [ x2, y2, width2, height2],
                        ...
                    ]),
        'num_box': 2
    },
    ....
}

$ python bin/extract_features_from_gt.py \
    --model_file vqa-maskrcnn-benchmark/data/detectron_model.pth \
    --config_file vqa-maskrcnn-benchmark/data/detectron_config.yaml \
    --imdb_gt_file vqa-maskrcnn-benchmark/image_dir/COCO_2014/npy_files/guesswhat.train.npy \
    --output_folder data/rcnn/from_gt_gw_xyxy_scale/train

# download vanilla pretrained model
$ cd vilbert-pretrained-model
$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/pretrained_model.bin

# download 12-in-1 pretrained model
$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/multi_task_model.bin

$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/datasets/coco/features_100/COCO_trainval_resnext152_faster_rcnn_genome.lmdb/data.mdb && mv data.mdb COCO_trainval_resnext152_faster_rcnn_genome.lmdb/
$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/datasets/coco/features_100/COCO_test_resnext152_faster_rcnn_genome.lmdb/data.mdb && mv data.mdb COCO_test_resnext152_faster_rcnn_genome.lmdb/

$ wget www.florian-strub.com/github/ft_vgg_img.zip
$ unzip ft_vgg_img.zip -d img/

$ python bin/extract_features.py \
    --model_file vqa-maskrcnn-benchmark/data/detectron_model.pth \
    --config_file vqa-maskrcnn-benchmark/data/detectron_config.yaml \
    --image_dir vqa-maskrcnn-benchmark/image_dir/COCO_2014/images/ \
    --output_folder data/rcnn/from_rcnn/ \
    --batch_size 8

$ cd data/vilbert-multi-task
$ wget https://dl.fbaipublicfiles.com/vilbert-multi-task/datasets.tar.gz
$ tar -I pigz -xvf datasets.tar.gz datasets/guesswhat/

$ python -m torch.distributed.launch \
    --nproc_per_node=4 \
    --nnodes=1 \
    --node_rank=0 \
    main.py \
    --command train-oracle-vilbert \
    --config config_files/oracle_vilbert.yaml \
    --n-jobs 8

$ python main.py \
    --command test-oracle-vilbert \
    --config config_files/oracle_vilbert.yaml \
    --load ckpt/oracle_vilbert-sd0/epoch-3.pth

$ python main.py \
    --command train-guesser-vilbert \
    --config config_files/guesser_vilbert.yaml \
    --n-jobs 8

$ python main.py \
    --command test-guesser-vilbert \
    --config config_files/guesser_vilbert.yaml \
    --n-jobs 8 \
    --load ckpt/guesser_vilbert-sd0/best.pth

# Distributed training
$ python -m torch.distributed.launch \
    --nproc_per_node=4 \
    --nnodes=1 \
    --node_rank=0 \
    main.py \
    --command train-qgen-vilbert \
    --config config_files/qgen_vilbert.yaml \
    --n-jobs 8 

# Non-distributed training
$ python main.py \
    --command train-qgen-vilbert \
    --config config_files/qgen_vilbert.yaml \
    --n-jobs 8 

$ python main.py \
    --command test-self-play-all-vilbert \
    --config config_files/self_play_all_vilbert.yaml \
    --n-jobs 8