MCG

Implementation of the MCG model

Multi-granularity Contrastive Cross-modal Collaborative Generation for End-to-End Long-term Video Question Answering

Code structure

We present the official PyTorch code for MCG, with the complete code directory structured as follows:

./
├── config/	# Configuration files
│   ├── pretrain_mcg.json
│   ├── msvd_qa.json
│   ├── ...
│   └── timesformer_divst_8x32_224_k600.json
├── env/		# Environment requirements and setup scripts	
│   ├── install_pkg.sh
│   └── requirements.txt
├── src/		# MCG source code
│   ├── configs/
│   ├── datasets/
│   ├── __init__.py
│   ├── modeling/
│   ├── optimization/
│   ├── pretrain/
│   ├── tasks/
│   └── utils/
├── run_scripts/	# Pre-training and fine-tuning scripts
│   ├── pt_mcg.sh
│   ├── ...
│   └── ft_msvd_qa.sh
└── README.md

Note: We haven't explicitly labeled the dataset directory in the structure. We encourage to keep dataset separate from the code and store it in a specified data disk. Once you have downloaded the dataset, you can configure it in the configuration file in the '/config' directory.

Setup & Data Preparation

Install Dependencies

1. Creating conda environment

   conda create -n mcg python=3.8
   conda activate mcg

2. Run setup scripts

   cd env && bash install_pkg.sh

   Note: We utilize Horovod as our distributed deep learning training framework. Initial installation may pose some challenges; please refer to the official Horovod GitHub repository for guidance.

Pre-training Data Preparation

• WebVid2M
  • Download WebVid2M.
  • Put WebVid2M videos under your data path.
  • Change your config/pretrain_mcg.json
• CC3M
  • Download CC-3M.
  • Change cc3m.json with local image paths.
  • Change your config/pretrain_mcg.json

Fine-tuning Data Preparation

• MSRVTT-QA

  • Download train_val_videos.zip and test_videos.zip from here.

  • Check md5sum:

    51f2394d279cf84f1642defd9a651e6f  train_val_videos.zip
    0af68454cec9d586e92805739f3911d0  test_videos.zip

  • Unzip all the videos to your data path. (10k in total).

    unzip train_val_videos.zip -d /path/to/your/data/msrvtt/videos
    unzip test_videos.zip -d /path/to/your/data/msrvtt/videos

  • Download QA annotations from here

• MSVD-QA

  • Download Video from official release:

    wget -nc https://www.cs.utexas.edu/users/ml/clamp/videoDescription/YouTubeClips.tar
    

  • Check md5sum:

    9bdb20fcf14d59524a6febca9f6a8d89  YouTubeClips.tar
    

  • Unzip all the videos to your data path. (1,970 videos in total).

    tar xvf YouTubeClips.tar -C /path/to/your/data/msvd/videos --strip-components=1
    

  • Download QA annotations from here

• NExT-QA

  • Download the raw videos from NExTVideo.

  • Download the QA annotations fromhere and map video file from here.

    • ['nextqa.zip'] contains annotations of QAs and GloVe Embeddings.
    • As NExT-QA's videos are sourced from VidOR, you may need the map file ['nextqa/map_vid_vidorID.json']).

All the text annotation can be downloaded from this Link, with the complete annotions directory structured as follows:

./
├── fintune_data
│   ├── msrvttqa
│   │   ├── test.jsonl
│   │   ├── train_ans2label.json
│   │   ├── train.jsonl
│   │   └── val.jsonl
│   ├── msvdqa
│   │   ├── test.jsonl
│   │   ├── train_ans2label.json
│   │   ├── train.jsonl
│   │   └── val.jsonl
│   └── nextqa
│       ├── add_reference_answer_test.json
│       ├── glove_embed.npy
│       ├── map_vid_vidorID.json
│       ├── multi_vocab.pkl
│       ├── test.csv
│       ├── train.csv
│       ├── val.csv
│       └── vocab.pkl
└── pretrain_data
    ├── cc3m
    │   └── cc3m.json
    └── webvid2m
        ├── train.pkl
        └── val.pkl

Pretraining

1. Download WebVid2M and CC-3M.

2. Configure your pretraining configuration file config/pretrain_mcg.json with your dataset path and other hyper-parametes.

3. Modify your pre-training run scripts run_scripts/pretrain_mcg.sh

   cd ..
   
   export PYTHONPATH="$PYTHONPATH:$PWD"
   echo $PYTHONPATH
   
   CONFIG_PATH='config/pretrain_mcg.json'
   
   horovodrun -np 8 python src/pretrain/run_pretrain_sparse.py \		# change -np to GPUs numbers.
         --config $CONFIG_PATH \
         --output_dir /path/to/output_dir/pretrain/$(date '+%Y%m%d%H%M%S')

4. Training video-language model

   cd run_scripts && bash pretrain_mcg.sh

Downstream Task Finetuning

Once you have completed the model pre-training, you can use the downstream datasets to fine-tune your weights.

• We provide the fine-tuning and inference code for MSRVTT-QA, MSVD-QA, NExT-QA, you can run the following script for model training:

  cd run_scripts
  bash ft_msrvtt_qa.sh
  bash ft_msvd_qa.sh
  bash ft_next_qa.sh

  For example, with MSVD-QA:

  cd .. 
  
  export PYTHONPATH="$PYTHONPATH:$PWD"
  echo $PYTHONPATH
  
  CONFIG_PATH='config/msvd_qa.json'
  
  horovodrun -np 8 python src/tasks/run_video_qa.py \
        --debug 0 \
        --config $CONFIG_PATH \
        --output_dir /path/to/output_dir/finetune/msvd_qa/$(date '+%Y%m%d%H%M%S')
  

• Run inference with locally-finetuned checkpoints.

  cd run_scripts
  bash inf_msrvtt_qa.sh
  bash inf_msvd_qa.sh
  bash inf_next_qa.sh
  

  For example, with MSVD-QA:

  cd ..
  
  export PYTHONPATH="$PYTHONPATH:$PWD"
  echo $PYTHONPATH
  
  STEP='the_best_step'
  
  CONFIG_PATH='config/msvd_qa.json'
  OUTPUT_DIR='/path/to/output_dir/finetune/msvd_qa/the_finetuning_path/'
  
  TXT_DB='/path/to/dataset/msvd/txt/test.jsonl'
  IMG_DB='/path/to/dataset/msvd/train_video/'
  
  horovodrun -np 8 python src/tasks/run_video_qa.py \
        --do_inference 1 \
        --inference_split test \
        --inference_model_step $STEP \
        --inference_txt_db $TXT_DB \
        --inference_img_db $IMG_DB \
        --inference_batch_size 128 \
        --output_dir OUTPUT_DIR \
        --config $CONFIG_PATH

  • OUTPUT_DIR is the path after the --output_dir option in the finetuning script.
  • $STEP is a string, which tells the script to use the checkpoint $OUTPUT_DIR/ckpt/model_step_$STEP.pt for inference.

Acknowledgement

The implementation of MCG relies on ALPRO, ClipBERT, TimesFormer, Transformer. We use PyTorch as our deep learning framework, with Horovod and gradient-checkpoint as our distributed multi-GPU training frameworks. We thank the original authors for their work and open source code.