kakaobrain / noc

Noise-aware Learning from Web-crawled Image-Text Data for Image Captioning

This is an official PyTorch Implementation of Noise-aware Learning from Web-crawled Image-Text Data for Image Captioning [arxiv]

• The method is a novel image captioning framework mainly for web-crawled image-text data. To be specific, the method addresses the noise issue in the web-crawled data by learning a quality controllable captioning model. Our model can generate high-quality captions of well-aligned by simply setting the control signal to desired alignment level at inference time and outperforms the filtering-based baseline algorithm.

1. Environmental Setup

We have tested the implementation on the following environment:

• Python 3.7.11 / PyTorch 1.10.0 / torchvision 0.11.0 / CUDA 11.3 / Ubuntu 18.04

Also, the code is based on pytorch-lightning (==1.6.0) and all necessary dependencies can be installed by running following command.

$ pip install -r requirements.txt

For evaluation using pycocotools, java is also required.

$ apt-get update
$ apt install default-jdk

2. Prepare Data

We provide a script to download data including (re-formatted) annotations for cc3m, coco, flickr30k and nocap datasets, and bucket file.

2-1. Prepare training and evaluation data

$ cd <path-to-root>
# download cc3m
$ bash scripts/download_cc3m.sh

# download mscoco
$ bash scripts/download_mscoco.sh

# download flickr30k
# For flickr30k data, you need to download raw images from (https://shannon.cs.illinois.edu/DenotationGraph/),
# the shell command below is for preparing only the annotation file
$ bash script/download_flickr30k.sh

# download nocaps
$ bash scripts/download_nocap.sh

Then, you need to download images from each dataset and locate the images within data folder whose structure should be as follows:

data
│─ bucket
│    │─ bucket_8bins.pickle  		 # bucketing file for cc3m data
│
│─ cc3m
│    │─ train_list_filtered_by_0.3.txt   # cc3m training annotation with filtering
│    │─ train_list.txt                   # cc3m training annotation
│    │─ val_list.txt                     # cc3m validation annotation
│    │─ images                           # link to root directory for images
│         │─ train                       # link to directory for training images
│         │─ val                         # link to directory for validation images
│
│─ coco
│    │─ captions_train2017.json          # training annotation
│    │─ captions_val2017.json            # validation annotation
│    │─ captions_karpathy_test.json      # Download on your own. Please check the scripts/download_mscoco.sh
│    │─ images                           # link to root directory for images
│         │─ val2017                     # link to directory for validation images
│         │─ val2014                     # link to directory for karpathy test split
│
│─ flickr30k
│    │─ ann_test.json                    # validation annotation
│    │─ images                           # link to root directory for images
│
│─ nocap
│    │─ val_4500_captions.json           # validation annotation
│    │─ images                           # link to root directory for images
│         │─ validation                  # link to directory for validation images
│
│─ pretrained_lm                         # will be automatically generated for gpt tokenizer
│    │─ gpt2_tokenizer                   # gpt2 tokenizer
│
│─ pretrained_clip                       # will be automatically generated for pre-trained clip model

3. Training

We use Hydra to provide flexible training configurations. Below examples explain how to modify each training parameter for your use cases. We assume that you are in <path-to-root> (i.e., root of this repository).

Our training is based on distributed environment (multi-GPUs training) using ddp environment supported by pytorch-lightning. The default setting requires 64-GPUs (of V100) with a total batch of 2048. However, you can set the parameter config.DISTRIBUTED.NUM_PROC_PER_NODE to the number of gpus you can use or change config.experiment.effective_batch_size_train.

(1) Training with noise-aware framework on CC3M

To run our NoC algorithm, you need a bucketing file of cosine similarities for all image-text pairs.

1. You can use the pre-computed bucketing file attached in data/bucket/bucket_8bins.pickle.

2. Or, you should generate the bucketing file with this following commands:

$ cd <path-to-root>
$ python3 preprocess/get_noc_bucket.py

Then, run a single command bash scripts/train_cc3m_noc.sh or following full command to train NoC model:

$ cd <path-to-root>
$ PYTHONPATH=./ python3 ./train.py \
		   distributed.num_nodes=16 distributed.num_proc_per_node=4 \
		   experiment.expr_name=cc3m_noc

Note that the checkpoints and log files (e.g., tensorboard) are automatically saved in results/<EXPR_NAME>.

(2) Training baselines on CC3M

$ cd <path-to-root>
$ bash scripts/train_cc3m_vanilla.sh    # for vanilla baseline

$ python3 preprocess/get_filtered_annotations.py  # generate filtered annotation for filtering strategy
$ bash scripts/train_cc3m_filtering.sh  # for vanilla (filtering) baseline

4. Evaluation

$ cd <path-to-root>
$ bash scripts/eval_all.sh <EXPR_NAME> 1  # our model, e.g., EXPR_NAME=cc3m_noc
$ bash scripts/eval_all.sh <EXPR_NAME> 0  # baselines, e.g., EXPR_NAME=cc3m_filtering, cc3m_vanilla

5. Model Zoo

We provide checkpoints to reproduce the results reported in our paper.

To predict captions from these checkpoints, decompress the .tar files and locate them in <path-to-root>/results .

Method	Checkpoint (pre-trained)
Vanilla	download
Vanilla (Filtering)	download
NoC (ours)	download

6. Citation

If you find this code helpful for your research, please cite our paper.

@InProceedings{Kang_2023_ICCV,
    author    = {Kang, Wooyoung and Mun, Jonghwan and Lee, Sungjun and Roh, Byungseok},
    title     = {Noise-Aware Learning from Web-Crawled Image-Text Data for Image Captioning},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {2942-2952}
}

7. Contact for Issues

Woo Young Kang, [e-mail] Jonghwan Mun, [e-mail]

8. License

This project is licensed under the terms of the MIT License.