Introduction

This repo presents some example codes to reproduce some results in GIT: A Generative Image-to-text Transformer for Vision and Language.

Installation

• Install azfuse. The tool is used to automatically download the data. The configuration of AzFuse has already been in this repo.

• Download the source code by

  git clone https://github.com/microsoft/GenerativeImage2Text.git
  cd GenerativeImage2Text

• Install the package

  pip install -r requirements.txt
  python setup.py build develop

Inference

• Inference on a single image or multiple frames:

  # single image, captioning
  AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_image', \
        'image_path': 'aux_data/images/1.jpg', \
        'model_name': 'GIT_BASE', \
        'prefix': '', \
  }"
  # single image, question answering
  AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_image', \
        'image_path': 'aux_data/images/1.jpg', \
        'model_name': 'GIT_BASE_VQAv2', \
        'prefix': 'what is it?', \
  }"
  # multiple images, captioning
  AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_image', \
        'image_path': ['aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg'], \
        'model_name': 'GIT_BASE_VATEX', \
        'prefix': '', \
  }"
  # multiple images, question answering
  AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_image', \
        'image_path': ['aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg', 'aux_data/images/1.jpg'], \
        'model_name': 'GIT_BASE_MSRVTT_QA', \
        'prefix': 'what is it?', \
  }"

  • If prefix is empty, it is effectively the captioning task.

  • If prefix is a question, it is effectively the visual question answering task.

  • Use a list for image_path if it is for video. The example here is 6 identical images, only for a demo purpose. It should be different image frames from a video.

  • model_name here can be the following. Performance details can be found in the reference paper.

    model_name	Information	Performance
    GIT_BASE	pretrained on 4M images
    GIT_BASE_COCO	fine-tuned on COCO	CIDEr: 131.4
    GIT_BASE_TEXTCAPS	fine-tuned on TextCaps for captioning	val/CIDEr: 64.9
    GIT_BASE_VQAv2	fine-tuned on VQAv2	test-dev: 72.72
    GIT_BASE_TEXTVQA	fine-tuned on TextVQA	val/acc: 18.81
    GIT_BASE_VATEX	fine-tuned on VATEX for captioning	public/test/CIDEr: 60.0
    GIT_BASE_MSRVTT	fine-tuned on MSRVTT for captioning	test/CIDEr: 57.8
    GIT_BASE_MSRVTT_QA	fine-tuned on MSRVTT for question answering	acc: 41.0
    GIT_LARGE	pretrained on 14M images
    GIT_LARGE_COCO	fine-tuned on COCO	CIDEr: 138.5
    GIT_LARGE_TEXTCAPS	fine-tuned on TextCaps for captioning	val/CIDEr: 106.3
    GIT_LARGE_VQAv2	fine-tuned on VQAv2	test-dev: 75.51
    GIT_LARGE_TEXTVQA	fine-tuned on TextVQA	val/acc: 37.47
    GIT_LARGE_VATEX	fine-tuned on VATEX for captioning	public/test/CIDEr: 72.5
    GIT_LARGE_MSRVTT	fine-tuned on MSRVTT for captioning	test/CIDEr: 64.1
    GIT_LARGE_MSRVTT_QA	fine-tuned on MSRVTT for question answering	acc: 42.7

  • In the dataset of cc12m, the caption may contain some special tags to hide person names and the model might also predict such special tokens. To eliminate this issue, we remove these captions (around 25% in cc12m), and re-trained the large-sized model. The base-sized model is not affected as cc12 is not part of the training data.

    model_name	Information	Performance
    GIT_LARGE_R	pretrained on 14M images with special tag removed
    GIT_LARGE_R_COCO	fine-tuned on COCO	CIDEr: 137.6
    GIT_LARGE_R_TEXTCAPS	fine-tuned on TextCaps for captioning	val/CIDEr: 105.3

• Inference on a TSV file, which is a collection of multiple images.

  • Data format (for information only)
    • image TSV: Each row has two columns. The first is the image key; the second is base64-encoded jpg or png bit string.
    • caption or question tsv: Each row has two columns. The first is the image key; the second is a list of dictionaries in the json format. For caption TSV, the dictionary should contain at least the field of 'caption'. For the question answering TSV, it should contain at least question_id and question.
  • inference on COCO Karpathy test.
    1. Inference.

       # base
       AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_tsv', \
             'image_tsv': 'data/coco_caption/test.img.tsv', \
             'model_name': 'GIT_BASE_COCO', \
             'question_tsv': null, \
             'out_tsv': 'inference/GIT_BASE_COCO/coco.tsv', \
       }"
       # GIT_LARGE_COCO. If there are 8 GPUs, it can parallel by mpirun -n 8
       AZFUSE_TSV_USE_FUSE=1 mpirun -n 8 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_tsv', \
             'image_tsv': 'data/coco_caption/test.img.tsv', \
             'model_name': 'GIT_LARGE_COCO', \
             'question_tsv': null, \
             'out_tsv': 'inference/GIT_LARGE_COCO/coco.tsv', \
       }"

    2. Calculate the evaluation metric

       # base
       AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'evaluate_on_coco_caption', \
             'res_file': 'inference/GIT_BASE_COCO/coco.tsv', \
             'label_file': 'data/coco_caption/test.caption.tsv', \
       }"

       The CIDEr score should be 131.35 for GIT_BASE_COCO and 138.45 for GIT_LARGE_COCO. If you get lower score (e.g. 126 for the base model), the reason could be the misalignment of the environment, e.g. pytorch version.
    3. (optional) To exactly reproduce the number, please run the following:

       nvidia-docker run --ipc=host amsword/setup:py38pt19u20cu11 \
           bash -c "mkdir -p /tmp/code \
                   && cd /tmp/code \
                   && pip install git+https://github.com/microsoft/azfuse.git \
                   && git clone https://github.com/amsword/generativeimage2text.git \
                   && cd generativeimage2text \
                   && pip install -r requirements.txt \
                   && python setup.py build develop \
                   && AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_tsv', \
                            'image_tsv': 'data/coco_caption/test.img.tsv', \
                            'model_name': 'GIT_BASE_COCO', \
                            'question_tsv': null, \
                            'out_tsv': 'inference/GIT_BASE_COCO/coco.tsv', \
                      }" \
                   &&  AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'evaluate_on_coco_caption', \
                       'res_file': 'inference/GIT_BASE_COCO/coco.tsv', \
                       'label_file': 'data/coco_caption/test.caption.tsv', \
                       'outfile': 'inference/GIT_BASE_COCO/coco.score.json', \
                       }" \
                   && cat inference/GIT_BASE_COCO/coco.score.json \
                   "

  • Inference on vqa test

    1. Inference

       # base model
       AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_tsv', \
             'image_tsv': 'data/TaxVQAv2/test.tsv', \
             'model_name': 'GIT_BASE_VQAv2', \
             'question_tsv': 'data/TaxVQAv2/test.caption.tsv', \
             'out_tsv': 'inference/GIT_BASE_VQAv2/snapshot/vqav2.tsv', \
       }"
       # GIT_LARGE_VQAv2 with 8 GPUs.
       AZFUSE_TSV_USE_FUSE=1 mpirun -n 8 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_tsv', \
             'image_tsv': 'data/TaxVQAv2/test.tsv', \
             'model_name': 'GIT_LARGE_VQAv2', \
             'question_tsv': 'data/TaxVQAv2/test.caption.tsv', \
             'out_tsv': 'inference/GIT_LARGE_VQAv2/snapshot/vqav2.tsv', \
       }"

    2. Convert the output tsv to the json format for submission to evalai

       # base model
       AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'convert_tsv_to_vqa_json', \
             'predict_file': 'inference/GIT_BASE_VQAv2/snapshot/vqav2.tsv', \
             'out_json': 'inference/GIT_BASE_VQAv2/snapshot/vqav2.json', \
       }"
       # large model
       AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'convert_tsv_to_vqa_json', \
             'predict_file': 'inference/GIT_LARGE_VQAv2/snapshot/vqav2.tsv', \
             'out_json': 'inference/GIT_LARGE_VQAv2/snapshot/vqav2.json', \
       }"

       Submit the file of inference/GIT_BASE_VQAv2/snapshot/vqav2.json to evalai and you should get 72.72 on test-dev. If it is GIT_LARGE_VQAv2, the accuracy is 75.51.

    3. (optional) To exactly reproduce the number, you can use the following:

       # base model
       nvidia-docker run --ipc=host amsword/setup:py38pt19u20cu11 \
           bash -c "mkdir /tmp/code \
                   && cd /tmp/code \
                   && pip install git+https://github.com/microsoft/azfuse.git \
                   && git clone https://github.com/amsword/generativeimage2text.git \
                   && cd generativeimage2text \
                   && pip install -r requirements.txt \
                   && python setup.py build develop \
                   && AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'test_git_inference_single_tsv', \
                       'image_tsv': 'data/TaxVQAv2/test.tsv', \
                       'model_name': 'GIT_BASE_VQAv2', \
                       'question_tsv': 'data/TaxVQAv2/test.caption.tsv', \
                       'out_tsv': 'inference/GIT_BASE_VQAv2/snapshot/vqav2.tsv', \
                   }" \
                   &&  AZFUSE_TSV_USE_FUSE=1 python -m generativeimage2text.inference -p "{'type': 'convert_tsv_to_vqa_json', \
                       'predict_file': 'inference/GIT_BASE_VQAv2/snapshot/vqav2.tsv', \
                       'out_json': 'inference/GIT_BASE_VQAv2/snapshot/vqav2.json', \
                   }" \
       }"

       Note that, please modify the docker command properly so that the output file can be saved permanently to the host machine. It is also recommended to run it inside the docker container by

       nvidia-docker run --ipc=host amsword/setup:py38pt19u20cu11 sleep infinity
       docker ps # get the docker container ID
       docker exec -it container_id /bin/bash # attach inside the docker container
       # all other commands to run the inference.

Training

The repo shows the key code path of constructing the network input with transformations and forward/backward. The code can be plugged into any trainer easily. Here is the example for the base model.

• Pretraining/captioning

  python -m generativeimage2text.train -p "{'type': 'forward_backward_example', \
                  'image_files': ['aux_data/images/1.jpg', 'aux_data/images/2.jpg'], \
                  'captions': ['a couple of boats in a large body of water.', 'a view of a mountain with a tree'], \
              }"
  

• VQA

  python -m generativeimage2text.train -p "{'type': 'forward_backward_example', \
                  'image_files': ['aux_data/images/1.jpg', 'aux_data/images/2.jpg'], \
                  'prefixs': ['what is this?', 'how many trees?'], \
                  'captions': ['several boats in a large body of water', '1'], \
              }"
  

ImageNet

Class ID to unique readable names

• Save the file of LOC_synset_mapping.txt from Kaggle. under aux_data/imagenet/

• Convert the wordnet ID to readable names as follows

  python -m generativeimage2text.data_prepare -p "{'type': 'generate_imagenet_unique_names'}"

  The input file is hard coded as ./aux_data/imagenet/LOC_synset_mapping.txt and the output file is ./aux_data/imagenet/imagenet_unique_readable_names.txt

Citation

Please consider to cite the following reference if it helps.

@article{wang2022git,
  title={GIT: A Generative Image-to-text Transformer for Vision and Language},
  author={Wang, Jianfeng and Yang, Zhengyuan and Hu, Xiaowei and Li, Linjie and Lin, Kevin and Gan, Zhe and Liu, Zicheng and Liu, Ce and Wang, Lijuan},
  journal={arXiv preprint arXiv:2205.14100},
  year={2022}
}

Misc

The model is now available in 🤗 Transformers. You can also find a fine-tuning guide on image captioning with GIT here. Thanks to Niels Rogge for contributing the model to 🤗 Transformers and Sayak Paul for the fine-tuning guide.

Acknowledgement

Part of the code is based on transformers, clip, maskrcnn-benchmark, oscar, virtex.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.