BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation

This is the PyTorch code of the BLIP paper [blog]. The code has been tested on PyTorch 1.10. To install the dependencies, run

pip install -r requirements.txt

Catalog:

• Inference demo
• Pre-trained and finetuned checkpoints
• Finetuning code for Image-Text Retrieval, Image Captioning, VQA, and NLVR2
• Pre-training code
• Zero-shot video-text retrieval
• Download of bootstrapped pre-training datasets

Inference demo:

Run our interactive demo using Colab notebook (no GPU needed). The demo includes code for:

1. Image captioning
2. Open-ended visual question answering
3. Multimodal / unimodal feature extraction
4. Image-text matching

Try out the Web demo, integrated into Huggingface Spaces 🤗 using Gradio.

Replicate web demo and Docker image is also available at

Pre-trained checkpoints:

Num. pre-train images	BLIP w/ ViT-B	BLIP w/ ViT-B and CapFilt-L	BLIP w/ ViT-L
14M	Download	-	-
129M	Download	Download	Download

Finetuned checkpoints:

Task	BLIP w/ ViT-B	BLIP w/ ViT-B and CapFilt-L	BLIP w/ ViT-L
Image-Text Retrieval (COCO)	Download	-	Download
Image-Text Retrieval (Flickr30k)	Download	-	Download
Image Captioning (COCO)	-	Download	Download
VQA	Download	Download	-
NLVR2	Download	-	-

Image-Text Retrieval:

1. Download COCO and Flickr30k datasets from the original websites, and set 'image_root' in configs/retrieval_{dataset}.yaml accordingly.
2. To evaluate the finetuned BLIP model on COCO, run:

python -m torch.distributed.run --nproc_per_node=8 train_retrieval.py \
--config ./configs/retrieval_coco.yaml \
--output_dir output/retrieval_coco \
--evaluate

3. To finetune the pre-trained checkpoint using 8 A100 GPUs, first set 'pretrained' in configs/retrieval_coco.yaml as "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base.pth". Then run:

python -m torch.distributed.run --nproc_per_node=8 train_retrieval.py \
--config ./configs/retrieval_coco.yaml \
--output_dir output/retrieval_coco 

Image-Text Captioning:

1. Download COCO and NoCaps datasets from the original websites, and set 'image_root' in configs/caption_coco.yaml and configs/nocaps.yaml accordingly.
2. To evaluate the finetuned BLIP model on COCO, run:

python -m torch.distributed.run --nproc_per_node=8 train_caption.py --evaluate

3. To evaluate the finetuned BLIP model on NoCaps, generate results with: (evaluation needs to be performed on official server)

python -m torch.distributed.run --nproc_per_node=8 eval_nocaps.py 

4. To finetune the pre-trained checkpoint using 8 A100 GPUs, first set 'pretrained' in configs/caption_coco.yaml as "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_capfilt_large.pth". Then run:

python -m torch.distributed.run --nproc_per_node=8 train_caption.py 

VQA:

1. Download VQA v2 dataset and Visual Genome dataset from the original websites, and set 'vqa_root' and 'vg_root' in configs/vqa.yaml.
2. To evaluate the finetuned BLIP model, generate results with: (evaluation needs to be performed on official server)

python -m torch.distributed.run --nproc_per_node=8 train_vqa.py --evaluate

3. To finetune the pre-trained checkpoint using 16 A100 GPUs, first set 'pretrained' in configs/vqa.yaml as "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_capfilt_large.pth". Then run:

python -m torch.distributed.run --nproc_per_node=16 train_vqa.py 

NLVR2:

1. Download NLVR2 dataset from the original websites, and set 'image_root' in configs/nlvr.yaml.
2. To evaluate the finetuned BLIP model, run

python -m torch.distributed.run --nproc_per_node=8 train_nlvr.py --evaluate

3. To finetune the pre-trained checkpoint using 16 A100 GPUs, first set 'pretrained' in configs/nlvr.yaml as "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base.pth". Then run:

python -m torch.distributed.run --nproc_per_node=16 train_nlvr.py 

Finetune with ViT-L:

In order to finetune a model with ViT-L, simply change the config file to set 'vit' as large. Batch size and learning rate may also need to be adjusted accordingly (please see the paper's appendix for hyper-parameter details). Gradient checkpoint can also be activated in the config file to reduce GPU memory usage.

Pre-train:

1. Prepare training json files where each json file contains a list. Each item in the list is a dictonary with two key-value pairs: {'image': path_of_image, 'caption': text_of_image}.
2. In configs/pretrain.yaml, set 'train_file' as the paths for the json files .
3. Pre-train the model using 8 A100 GPUs:

python -m torch.distributed.run --nproc_per_node=8 pretrain.py --config ./configs/Pretrain.yaml --output_dir output/Pretrain 

Zero-shot video-text retrieval:

1. Download MSRVTT dataset following the instructions from https://github.com/salesforce/ALPRO, and set 'video_root' accordingly in configs/retrieval_msrvtt.yaml.
2. Install decord with

   pip install decord

3. To perform zero-shot evaluation, run

python -m torch.distributed.run --nproc_per_node=8 eval_retrieval_video.py

Pre-training datasets download:

We provide bootstrapped pre-training datasets as json files. Each json file contains a list. Each item in the list is a dictonary with two key-value pairs: {'url': url_of_image, 'caption': text_of_image}.

Image source	Filtered web caption	Filtered synthetic caption by ViT-B	Filtered synthetic caption by ViT-L
CC3M+CC12M+SBU	Download	Download	Download
LAION115M	Download	Download	Download

Citation

If you find this code to be useful for your research, please consider citing.

@inproceedings{li2022blip,
      title={BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation}, 
      author={Junnan Li and Dongxu Li and Caiming Xiong and Steven Hoi},
      year={2022},
      booktitle={ICML},
}

Acknowledgement

The implementation of BLIP relies on resources from ALBEF, Huggingface Transformers, and timm. We thank the original authors for their open-sourcing.