UNITER: UNiversal Image-TExt Representation Learning

This is the official repository of UNITER (ECCV 2020). This repository currently supports finetuning UNITER on NLVR2, VQA, VCR, SNLI-VE, Image-Text Retrieval for COCO and Flickr30k, and Referring Expression Comprehensions (RefCOCO, RefCOCO+, and RefCOCO-g). Both UNITER-base and UNITER-large pre-trained checkpoints are released. UNITER-base pre-training with in-domain data is also available.

Some code in this repo are copied/modified from opensource implementations made available by PyTorch, HuggingFace, OpenNMT, and Nvidia. The image features are extracted using BUTD.

Requirements

We provide Docker image for easier reproduction. Please install the following:

• nvidia driver (418+),
• Docker (19.03+),
• nvidia-container-toolkit.

Our scripts require the user to have the docker group membership so that docker commands can be run without sudo. We only support Linux with NVIDIA GPUs. We test on Ubuntu 18.04 and V100 cards. We use mixed-precision training hence GPUs with Tensor Cores are recommended.

Quick Start

NOTE: Please run bash scripts/download_pretrained.sh $PATH_TO_STORAGE to get our latest pretrained checkpoints. This will download both the base and large models.

We use NLVR2 as an end-to-end example for using this code base.

1. Download processed data and pretrained models with the following command.

   bash scripts/download_nlvr2.sh $PATH_TO_STORAGE

   After downloading you should see the following folder structure:

   ├── ann
   │   ├── dev.json
   │   └── test1.json
   ├── finetune
   │   ├── nlvr-base
   │   └── nlvr-base.tar
   ├── img_db
   │   ├── nlvr2_dev
   │   ├── nlvr2_dev.tar
   │   ├── nlvr2_test
   │   ├── nlvr2_test.tar
   │   ├── nlvr2_train
   │   └── nlvr2_train.tar
   ├── pretrained
   │   └── uniter-base.pt
   └── txt_db
       ├── nlvr2_dev.db
       ├── nlvr2_dev.db.tar
       ├── nlvr2_test1.db
       ├── nlvr2_test1.db.tar
       ├── nlvr2_train.db
       └── nlvr2_train.db.tar
   

2. Launch the Docker container for running the experiments.

   # docker image should be automatically pulled
   source launch_container.sh $PATH_TO_STORAGE/txt_db $PATH_TO_STORAGE/img_db \
       $PATH_TO_STORAGE/finetune $PATH_TO_STORAGE/pretrained

   The launch script respects $CUDA_VISIBLE_DEVICES environment variable. Note that the source code is mounted into the container under /src instead of built into the image so that user modification will be reflected without re-building the image. (Data folders are mounted into the container separately for flexibility on folder structures.)

3. Run finetuning for the NLVR2 task.

   # inside the container
   python train_nlvr2.py --config config/train-nlvr2-base-1gpu.json
   
   # for more customization
   horovodrun -np $N_GPU python train_nlvr2.py --config $YOUR_CONFIG_JSON

4. Run inference for the NLVR2 task and then evaluate.

   # inference
   python inf_nlvr2.py --txt_db /txt/nlvr2_test1.db/ --img_db /img/nlvr2_test/ \
       --train_dir /storage/nlvr-base/ --ckpt 6500 --output_dir . --fp16
   
   # evaluation
   # run this command outside docker (tested with python 3.6)
   # or copy the annotation json into mounted folder
   python scripts/eval_nlvr2.py ./results.csv $PATH_TO_STORAGE/ann/test1.json

   The above command runs inference on the model we trained. Feel free to replace --train_dir and --ckpt with your own model trained in step 3. Currently we only support single GPU inference.

5. Customization

   # training options
   python train_nlvr2.py --help

   • command-line argument overwrites JSON config files
   • JSON config overwrites argparse default value.
   • use horovodrun to run multi-GPU training
   • --gradient_accumulation_steps emulates multi-gpu training

6. Misc.

   # text annotation preprocessing
   bash scripts/create_txtdb.sh $PATH_TO_STORAGE/txt_db $PATH_TO_STORAGE/ann
   
   # image feature extraction (Tested on Titan-Xp; may not run on latest GPUs)
   bash scripts/extract_imgfeat.sh $PATH_TO_IMG_FOLDER $PATH_TO_IMG_NPY
   
   # image preprocessing
   bash scripts/create_imgdb.sh $PATH_TO_IMG_NPY $PATH_TO_STORAGE/img_db

   In case you would like to reproduce the whole preprocessing pipeline.

Downstream Tasks Finetuning

VQA

NOTE: train and inference should be ran inside the docker container

1. download data

   bash scripts/download_vqa.sh $PATH_TO_STORAGE
   

2. train

   horovodrun -np 4 python train_vqa.py --config config/train-vqa-base-4gpu.json \
       --output_dir $VQA_EXP
   

3. inference

   python inf_vqa.py --txt_db /txt/vqa_test.db --img_db /img/coco_test2015 \
       --output_dir $VQA_EXP --checkpoint 6000 --pin_mem --fp16
   

   The result file will be written at $VQA_EXP/results_test/results_6000_all.json, which can be submitted to the evaluation server

VCR

NOTE: train and inference should be ran inside the docker container

1. download data

   bash scripts/download_vcr.sh $PATH_TO_STORAGE
   

2. train

   horovodrun -np 4 python train_vcr.py --config config/train-vcr-base-4gpu.json \
       --output_dir $VCR_EXP
   

3. inference

   horovodrun -np 4 python inf_vcr.py --txt_db /txt/vcr_test.db \
       --img_db "/img/vcr_gt_test/;/img/vcr_test/" \
       --split test --output_dir $VCR_EXP --checkpoint 8000 \
       --pin_mem --fp16
   

   The result file will be written at $VCR_EXP/results_test/results_8000_all.csv, which can be submitted to VCR leaderboard for evluation.

VCR 2nd Stage Pre-training

NOTE: pretrain should be ran inside the docker container

1. download VCR data if you haven't

   bash scripts/download_vcr.sh $PATH_TO_STORAGE
   

2. 2nd stage pre-train

   horovodrun -np 4 python pretrain_vcr.py --config config/pretrain-vcr-base-4gpu.json \
       --output_dir $PRETRAIN_VCR_EXP
   

Visual Entailment (SNLI-VE)

NOTE: train should be ran inside the docker container

1. download data

   bash scripts/download_ve.sh $PATH_TO_STORAGE
   

2. train

   horovodrun -np 2 python train_ve.py --config config/train-ve-base-2gpu.json \
       --output_dir $VE_EXP
   

Image-Text Retrieval

download data

bash scripts/download_itm.sh $PATH_TO_STORAGE

NOTE: Image-Text Retrieval is computationally heavy, especially on COCO.

Zero-shot Image-Text Retrieval (Flickr30k)

# every image-text pair has to be ranked; please use as many GPUs as possible
horovodrun -np $NGPU python inf_itm.py \
    --txt_db /txt/itm_flickr30k_test.db --img_db /img/flickr30k \
    --checkpoint /pretrain/uniter-base.pt --model_config /src/config/uniter-base.json \
    --output_dir $ZS_ITM_RESULT --fp16 --pin_mem

Image-Text Retrieval (Flickr30k)

• normal finetune

  horovodrun -np 8 python train_itm.py --config config/train-itm-flickr-base-8gpu.json
  

• finetune with hard negatives

  horovodrun -np 16 python train_itm_hard_negatives.py \
      --config config/train-itm-flickr-base-16gpu-hn.jgon
  

Image-Text Retrieval (COCO)

• finetune with hard negatives

  horovodrun -np 16 python train_itm_hard_negatives.py \
      --config config/train-itm-coco-base-16gpu-hn.json
  

Referring Expressions

1. download data

   bash scripts/download_re.sh $PATH_TO_STORAGE
   

2. train

   python train_re.py --config config/train-refcoco-base-1gpu.json \
       --output_dir $RE_EXP
   

3. inference and evaluation

   source scripts/eval_refcoco.sh $RE_EXP
   

   The result files will be written under $RE_EXP/results_test/

Similarly, change corresponding configs/scripts for running RefCOCO+/RefCOCOg.

Pre-tranining

download

bash scripts/download_indomain.sh $PATH_TO_STORAGE

pre-train

horovodrun -np 8 python pretrain.py --config config/pretrain-indomain-base-8gpu.json \
    --output_dir $PRETRAIN_EXP

Unfortunately, we cannot host CC/SBU features due to their large size. Users will need to process them on their own. We will provide a smaller sample for easier reference to the expected format soon.

Citation

If you find this code useful for your research, please consider citing:

@inproceedings{chen2020uniter,
  title={Uniter: Universal image-text representation learning},
  author={Chen, Yen-Chun and Li, Linjie and Yu, Licheng and Kholy, Ahmed El and Ahmed, Faisal and Gan, Zhe and Cheng, Yu and Liu, Jingjing},
  booktitle={ECCV},
  year={2020}
}

License

MIT