UniTAB: Unifying Text and Box Outputs for Grounded VL Modeling

UniTAB: Unifying Text and Box Outputs for Grounded Vision-Language Modeling

by Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, Lijuan Wang

European Conference on Computer Vision, 2022, Oral Presentation

Introduction

We propose UniTAB, a vision-language (VL) model that unifies text generation and bounding box prediction into a single architecture. For more details, please refer to our paper.

Citation

@inproceedings{yang2022unitab,
  title={UniTAB: Unifying Text and Box Outputs for Grounded Vision-Language Modeling},
  author={Yang, Zhengyuan and Gan, Zhe and Wang, Jianfeng and Hu, Xiaowei and Ahmed, Faisal and Liu, Zicheng and Lu, Yumao and Wang, Lijuan},
  booktitle={ECCV},
  year={2022}
}

Installation

Clone the repository:

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

New conda env:

conda create -n unitab python=3.8
conda activate unitab

Install packages in requirements.txt (separately install numpy and pytorch (LTS 1.8.2) if fails):

pip install -r requirements.txt

AzCopy

We recommend using the following AzCopy command to download. AzCopy executable tools can be downloaded here.

Example command:

path/to/azcopy copy <folder-link> <target-address> --resursive"

# For example:
path/to/azcopy copy https://unitab.blob.core.windows.net/data/data <local_path> --recursive
path/to/azcopy copy https://unitab.blob.core.windows.net/data/weights <local_path> --recursive
path/to/azcopy copy https://unitab.blob.core.windows.net/data/annotations <local_path> --recursive

Distributed Training

We do not specify distributed training tool in the example commands below. Pytorch distributed python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py or submitit supported. Or update util/dist.py/init_distributed_mode() to fit your cluster setting.

Data

• Download the original Flickr30k image dataset from : Flickr30K webpage and update the flickr_img_path to the folder containing the images.
• Download the original Flickr30k entities annotations from: Flickr30k annotations and update the flickr_dataset_path to the folder with annotations.
• Download the gqa images at GQA images and update vg_img_path to point to the folder containing the images.
• Download COCO images Coco train2014. Update the coco_path to the folder containing the downloaded images.

Or download the cached data (~77G) (use AzCopy with the link).

• Download our pre-processed annotations (~3.7G) (use AzCopy with the link, or zip file) and update the flickr_ann_path, gqa_ann_path and refexp_ann_path to this folder with pre-processed annotations.

Pre-train

The config file for pretraining is configs/pretrain.json. Optionally starting from MDETR pretrain with --load https://zenodo.org/record/4721981/files/pretrained_resnet101_checkpoint.pth. Weights availble here.

Example command (ngpu=64):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/pretrain.json --batch_size 2 --lr_backbone 2e-5 --text_encoder_lr 2e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --unitab_pretrain --pretrain_seqcrop mixed --ema --output-dir weights/$exp_id --load https://zenodo.org/record/4721981/files/pretrained_resnet101_checkpoint.pth

Multi-task Finetuning

The config file for pretraining is configs/multitask.json. Weights availble here.

Example command (ngpu=32):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/multitask.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 1e-5 --lr 5e-5 --num_queries 200 --max_decoding_step 256 --load weights/pretrained_checkpoint.pth --ema --output-dir weights/$exp_id

Downstream tasks

Optionally, downloading all weights at once (~54G):

path/to/azcopy copy https://unitab.blob.core.windows.net/data/weights <local_path> --recursive

For model inference, use the input arguments --eval --test. For captioning tests (Flickr grounded captioning, COCO image captioning, VQAv2 visual question answering), the computed captioning metrics displayed is only for reference. For the final number, an output prediction json file will be automatically stored at weights/$exp_id/results/pred_dict_$CIDEr.json. Please follow the official evaluation for Flickr grounded captioning, COCO captioning, and VQAv2 evaluation. We will better intergrate the caption evaluations in future versions.

Grounded captioning

The config file for pretraining is configs/flickr_kp.json. For model inference, use the input arguments --eval --test.

For the final number, an output prediction json file will be automatically stored at weights/$exp_id/results/pred_dict_$CIDEr.json. Please follow the official evaluation for Flickr grounded captioning evaluation. We will better intergrate the caption evaluations in future versions.

Weights: Separate, Pre-finetuning.

Model	CIDEr	F1_all
Separate	65.6	11.46
Pre-finetuning	69.7	12.95

Example command (ngpu=8):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr_kp.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 1e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --ema --output-dir weights/$exp_id --load weights/pretrained_checkpoint.pth

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr_kp.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 1e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --ema --output-dir weights/$exp_id --load weights/prefinetune_flickrcaptionKP_checkpoint.pth --eval --test

Referring expression comprehension

The config file for pretraining is configs/refcoco/+/g.json. For model inference, use the input arguments --eval --test --test_type testA/testB/test.

Weights: Separate, Pre-finetuning (refcoco/refcoco+/refcocog).

Model	Refcoco	Refcoco+	Refcocog
Separate	86.32	78.70	79.96
Pre-finetuning	88.59	80.97	84.58

Example command (ngpu=8):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/refcoco.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 5e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --ema --output-dir weights/$exp_id --load weights/pretrained_checkpoint.pth

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/refcoco.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 5e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --ema --output-dir weights/$exp_id --load weights/prefinetune_refcoco_checkpoint.pth --eval --test --test_type testA

Phrase grounding

The config file for pretraining is configs/flickr.json. For model inference, use the input arguments --eval --test.

Weights: Separate, Pre-finetuning.

Model	Flickr
Separate	79.39
Pre-finetuning	79.58

Example command (ngpu=8):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 5e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --ema --do_flickrgrounding --output-dir weights/$exp_id --load weights/pretrained_checkpoint.pth

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr.json --batch_size 2 --lr_backbone 1e-5 --text_encoder_lr 5e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --ema --do_flickrgrounding --output-dir weights/$exp_id --load weights/prefinetune_flickrGrounding_checkpoint.pth --eval --test

COCO captioning

The config file for pretraining is configs/flickr_cococaption.json. For model inference, use the input arguments --eval --test.

For the final number, an output prediction json file will be automatically stored at weights/$exp_id/results/pred_dict_$CIDEr.json. Please follow the official evaluation for COCO captioning evaluation. We will better intergrate the caption evaluations in future versions.

Weights: Separate, Pre-finetuning.

Model	CIDEr
Separate	119.3
Pre-finetuning	119.8

Example command (ngpu=16):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr_cococaption.json --lr_backbone 2e-5 --text_encoder_lr 2e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --ema --output-dir weights/$exp_id --load weights/pretrained_checkpoint.pth

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr_cococaption.json --lr_backbone 2e-5 --text_encoder_lr 2e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --ema --output-dir weights/$exp_id --load weights/prefinetune_MScococaption_checkpoint.pth --eval --test

Visual question answering on VQAv2

The config file for pretraining is configs/flickr_vqav2caption.json and configs/flickr_vqav2captionKP.json. Adjust the GT_type between vqav2caption and vqav2captionKP for std and KP splits. For model inference, use the input arguments --eval --test.

For the final number, an output prediction json file will be automatically stored at weights/$exp_id/results/pred_dict_$CIDEr.json. Please follow the official evaluation for VQAv2 evaluation. We will better intergrate the caption evaluations in future versions.

Weights: Separate, Pre-finetuning. KP split: Separate, Pre-finetuning.

Model	test-dev	KP-test
Separate	69.9	66.6
Pre-finetuning	70.7	67.5

Example command (ngpu=16):

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr_vqav2caption.json --lr_backbone 2e-5 --text_encoder_lr 2e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --ema --output-dir weights/$exp_id --load weights/pretrained_checkpoint.pth

CUBLAS_WORKSPACE_CONFIG=:4096:8  python main.py --dataset_config configs/flickr_vqav2caption.json --lr_backbone 2e-5 --text_encoder_lr 2e-5 --lr 1e-4 --num_queries 200 --max_decoding_step 256 --do_caption --no_detection --ema --output-dir weights/$exp_id --load weights/prefinetune_VQAv2_checkpoint.pth --eval --test

Acknowledgement

The project is built based on the following repository:

• MDETR--Modulated Detection for End-to-End Multi-Modal Understanding,
• transformers.

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 opencode@microsoft.com 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.