Learning Better Visual Dialog Agents with Pretrained Visual-Linguistic Representation

Or READ-UP: Referring Expression Agent Dialog with Unified Pretraining.

This repo includes the training/testing code for our paper Learning Better Visual Dialog Agents with Pretrained Visual-Linguistic Representation that has been accepted by CVPR 2021.

Please cite the following paper if you use the code in this repository:

@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}
}

Repository Setup

Environment

The following environment is recommended:

Instance storage: > 800 GB
pytorch 1.4.0
cuda 10.0

Set up virtual environment and install pytorch:

$ 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 dependencies:

# 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

Install Apex for distributed training

# 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" ./

Dataset

Meta-data

Download the GuessWhat?! dataset:

$ 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/

Prepare dict.json:

Set up repo as instructed in https://github.com/GuessWhatGame/guesswhat
Generate the dict.json file:

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

Copy dict.json file to read-up repo:

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

Dataset for Oracle models

1. Dataset for baseline Oracle + Faster-RCNN visual features.

Under vqa-maskrcnn-benchmark/data/, download RCNN model and COCO images:

# 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

Copy the guesswhat.train/valid/test.jsonl to vqa-maskrcnn-benchmark/data/. Unzip the COCO images into a folder image_dir/COCO_2014/images/, and prepare a npy file for feature extraction later.


$ 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

Repeat the same process for val and test. The generated file looks like the following:

{
    {
        '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
    },
    ....
}

Extract features from the ground-truth bounding boxes generated before:

$ 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

Repeat this process for val and test data.

2. Dataset for our Oracle model.

Download the pretrained VilBERT model (both vanilla and 12-in-1 have similar performance in our experiments).

# 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

Download the features for COCO:

$ 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/

Dataset for Q-Gen models

1. Dataset for baseline Q-Gen model [1]

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

2. Dataset for VDST Q-Gen model [2]

$ 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

Dataset for Guesser models

1. Dataset for baseline Guesser model[1]

$ 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/

Model Training & Evaluation

Oracle

To train our Oracle model:

$ 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

To evaluate our Oracle model:

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

This repo also implements other Oracle models:

Baseline Oracle model [1]
Baseline Oracle model + Faster-RCNN visual features (our ablation model)

To train and evaluate this model, run the main.py with corresponding config file and command.

Guesser

To train our Guesser model:

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

To evaluate our Guesser model:

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

This repo also implements other Guesser models:

Baseline Guesser model [1]

To train and evaluate this model, run the main.py with corresponding config file and command.

Q-Gen

To train our Q-Gen model:

# 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

To evalaute our Q-Gen model:

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

This repo also implements other Q-Gen models:

Baseline Q-Gen model [1]
VDST Q-Gen model [2]

To train and evaluate these models, run the main.py with corresponding config file and command.

References

[1] Strub, F., De Vries, H., Mary, J., Piot, B., Courvile, A., & Pietquin, O. (2017, August). End-to-end optimization of goal-driven and visually grounded dialogue systems. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (pp. 2765-2771).

[2] Pang, W., & Wang, X. (2020, April). Visual dialogue state tracking for question generation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34, No. 07, pp. 11831-11838).

amazon-science / read-up