CCLM: Towards Unified Cross-Lingual Cross-Modal Pre-training

Creating symlink for GQA images on babel: ln -s /data/datasets/GQA/images gqa

**Cross-View Language Modeling: Towards Unified`` Cross-Lingual Cross-Modal Pre-training. Yan Zeng, Wangchunshu Zhou, Ao Luo, Xinsong Zhang arXiv 2022.

June 2023: release more powerful CCLM (base & large) pre-trained based on X²-VLM
May 2023: The paper has been accepted by ACL 2023
June 2022: release preprint in arXiv and official PyTorch implementation

Requirements

Install python3 environment

pip3 install -r requirements.txt

Download data from corresponding websites
If running pre-training scripts:
- install Apex
- download pre-trained models for parameter initialization
  - image encoder: swin-transformer-base
  - text encoder: xlm-roberta-large
Organize these files like this:

CCLM/
    data/
        xlm-roberta-large/...
        swin_base_patch4_window7_224_22k.pth
        multi30k/...
        mscoco/...
    
    iglue/
        datasets/...
    
    images/
        flickr30k-images/*.jpg
        
        coco/
            train2014/*.jpg
            val2014/*.jpg
            test2015/*.jpg
        
        marvl-images/
            id/
                images/...
                extra/...
            sw/...
            ta/...
            tr/...
            zh/...
        marvl_fewshot/
            id/
                3-Beo/...
                all/...
            tr/...
            zh/...

Pretrain

# 3M
python3 run.py --task "pretrain_cclm_3m" --dist "1" --output_dir "output/"

# 4M
python3 run.py --task "pretrain_cclm_4m" --dist "1" --output_dir "output/"

For distributed training across nodes, see run.py for more details.

Data

To facilitate research on multi-lingual multi-modal pre-training, we provide the text translation of COCO+VG+SBU+CC3M and the filtered subset of WikiMatrix.

🌟 We implement CCLM(4M) following X-VLM. We also provide the text translation of Object And Region Annotations.

Please cite the corresponding papers appropriately and download the images from their websites.

For more details, please read the code dataset/pretrain_dataset.py (more specifically ImageMultiTextDataset & RegionMultiTextDataset & ParaTextDataset) to see what format is needed.

Checkpoints

Following previous works, we pretrain the model for only 30 epochs on 8 A100 GPUs. The batch size is set to 1024.
CCLM (3M, 30 epochs)
CCLM (4M, 30 epochs)

Recently, we train CCLM based on X²-VLM and achieves better performances.
CCLM-X2VLM-base
CCLM-X2VLM-large
Refer this repo to load the new checkpoints.

Finetune

Please download MSCOCO, Multi30K, and IGLUE from the corresponding websites.

Data: MSCOCO and Multi30K

For these two datasets, you need to additionally reformulate the train json files like this:

[
    {
        "caption": "A woman wearing a net on her head cutting a cake. ",
        "image": "coco/val2014/COCO_val2014_000000522418.jpg",
        "image_id": 522418
    }, ...
]

and the valid and test files like this:

[
    {
        "image": "coco/val2014/COCO_val2014_000000391895.jpg",
        "caption": [
            "A man with a red helmet on a small moped on a dirt road. ",
            "Man riding a motor bike on a dirt road on the countryside.",
            "A man riding on the back of a motorcycle.",
            "A dirt path with a young person on a motor bike rests to the foreground of a verdant area with a bridge and a background of cloud-wreathed mountains. ",
            "A man in a red shirt and a red hat is on a motorcycle on a hill side."
        ],
        "image_id": 391895
    }, ...
]

Data: IGLUE

For IGLUE, you just need to clone this repo and place it in the root path of our repo as follows. Our code works on the original annotations of IGLUE without any preprocess.

CCLM/
    iglue/
        datasets/...

For MaRVL, please download the marvl-images.zip and few-shot.zip from the download page, and extract them to images with name of marvl-images and marvl_fewshot.

For WIT, please download the image_data_train.tar and test images from its kaggle webpage, and extract them to images, images/wit_test seperately.

Tips for WIT:

The download link of image_data_train.tar is in Data Description.
You need to extract the files again in images/image_data_train/image_pixels and iglue/datasets/wit/annotations/train_en.jsonl.zip)

Retrieval Tasks: Multi30K and MSCOCO

# English-only Fine-tune
## Multi30K
python3 run.py --dist 1 --task itr_multi30k --config configs/cclm-base-ft/Retrieval_multi30k_en_ft.yaml --output_dir output/path/to/save --bs 128 --seed 42 --epoch 10 --checkpoint pretrained_model.th

## MSCOCO
python3 run.py --dist 1 --task itr_coco --config configs/cclm-base-ft/Retrieval_coco_en_ft.yaml --output_dir output/path/to/save --bs 128 --seed 42 --epoch 10 --checkpoint pretrained_model.th


# Single-Language Fine-tune
## Multi30K, optional language: cs/de/fr
python3 run.py --dist 1 --task itr_multi30k --config configs/cclm-base-ft/Retrieval_multi30k_cs_ft.yaml --output_dir output/path/to/save --bs 128 --seed 42 --epoch 10 --checkpoint pretrained_model.th

## MSCOCO, optional config: ja/zh
python3 run.py --dist 1 --task itr_coco --config configs/cclm-base-ft/Retrieval_coco_ja_ft.yaml --output_dir output/path/to/save --bs 128 --seed 42 --epoch 10 --checkpoint pretrained_model.th


# All-Language Fine-tune
## Multi30K
python3 run.py --dist 1 --task itr_multi30k --config configs/cclm-base-ft/Retrieval_multi30k_all_ft.yaml --output_dir output/path/to/save --bs 128 --seed 42 --epoch 10 --checkpoint pretrained_model.th

## MSCOCO
python3 run.py --dist 1 --task itr_coco --config configs/cclm-base-ft/Retrieval_coco_all_ft.yaml --output_dir output/path/to/save --bs 128 --seed 42 --epoch 10 --checkpoint pretrained_model.th

IGLUE: Zero-Shot

We provide examples of fine-tuning on English train set and evaluating on the test sets of other languages.

# XVNLI
python3 run.py --dist 1 --task xvnli --output_dir output/path/to/save --output_hdfs hdfs://path/to/save --checkpoint pretrained_model.th --bs 128 --seed 42

# xGQA
python3 run.py --dist 1 --task gqa --output_dir output/path/to/save --checkpoint pretrained_model.th --bs 128 --seed 42

# MaRVL
python3 run.py --dist 1 --task nlvr --output_dir output/path/to/save --checkpoint pretrained_model.th --bs 128 --seed 42 

# xFlickr&CO
python3 run.py --dist 1 --task xflickrco --output_dir output/path/to/save --output_hdfs hdfs://path/to/save --checkpoint pretrained_model.th --bs 128 --seed 42

# WIT
python3 run.py --dist 1 --task wit --output_dir output/path/to/save --output_hdfs hdfs://path/to/save --bs 128 --seed 42 --checkpoint pretrained_model.th

IGLUE: Few-Shot

We also evaluate CCLM on IGLUE max-shot settings. Note that you need to finetune the pretrained model on English first, then load the checkpoints for few-shot learning.

# XVNLI, optional language: ar/es/fr/ru
python3 run.py --dist 1 --task xvnli --output_dir output/path/to/save --output_hdfs hdfs://path/to/save --checkpoint en_finetuned_model.th --bs 128 --seed 42 --fewshot ar,48 --lr 1e-6

# xGQA, optional language: bn/de/id/ko/pt/ru/zh
python3 run.py --dist 1 --task gqa --output_dir output/path/to/save --checkpoint en_finetuned_model.th --bs 128 --seed 42 --fewshot bn --lr 1e-6

# MaRVL, optional language: id/tr/zh
python3 run.py --dist 1 --task nlvr --output_dir output/path/to/save --checkpoint en_finetuned_model.th --bs 128 --seed 42 --fewshot id --lr 1e-6

# xFlickr&CO, optional language: de/es/id/ja/ru/tr/zh
python3 run.py --dist 1 --task xflickrco --output_dir output/path/to/save --output_hdfs hdfs://path/to/save --checkpoint en_finetuned_model.th --bs 128 --seed 42 --fewshot de,100 --lr 1e-6

The value after language in --fewshot settings of XVNLI and xFlickr&CO is the number of few-shot samples, where we always use the maximum values.

Customize Dataset

You can prepare your data as follows for customization:

# XVNLI (json line)
{
    "Flikr30kID": "4452450137",
    "gold_label": "neutral",
    "sentence2": "A man and woman are exiting a building after a dance.",
}


# xGQA (json)
[
    {
        "img_id": "2354786",
        "label": {
            "yes": 1.0
        },
        "question_id": "02930152",
        "sent": "Is the sky dark?"
    }, ...
]

# MaRVL (json)
[
    {
        "sentence": "左圖有好幾隻不同顏色的鉛筆，右圖則只有一支鉛筆。",
        "label": false,
        "images": [
            "images/marvl_official/zh/images/82/82-6.jpg",
            "images/marvl_official/zh/images/82/82-1.jpg"
        ]
    }, ...
]

# xFlickr&CO (json line)
{
    "sentences": [
        "The man with pierced ears is wearing glasses and an orange hat."
    ],
    "id": "1007129816",
    "img_path": "1007129816.jpg"
}

# WIT
{
    'question_id': '05515938',
    'question': 'What is this bird called?',
    'image': '2405722.jpg',
    'dataset': 'gqa',
    'answer': 'parrot'
}

Citation

If you find this repository useful, please considering giving ⭐ or citing:

@article{cclm,
  title={Cross-View Language Modeling: Towards Unified Cross-Lingual Cross-Modal Pre-training},
  author={Zeng, Yan and Zhou, Wangchunshu and Luo, Ao and Zhang, Xinsong},
  journal={arXiv preprint arXiv:2206.00621},
  year={2022}
}

@article{xvlm,
  title={Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts},
  author={Zeng, Yan and Zhang, Xinsong and Li, Hang},
  journal={arXiv preprint arXiv:2111.08276},
  year={2021}
}

Contact

For issues using this code, please submit a GitHub issue.

piyushkhanna00705 / cclm Goto Github PK

cclm's Introduction