Skip to content
/ ALBEF Public
forked from salesforce/ALBEF

Code for ALBEF: a new vision-language pre-training method

License

Notifications You must be signed in to change notification settings

ztang07/ALBEF

 
 

Repository files navigation

Align before Fuse: Vision and Language Representation Learning with Momentum Distillation, NeurIPS 2021 Spotlight (Salesforce Research).

Announcement: ALBEF is now officially integrated into LAVIS - a one-stop library for language-and-vision research and applications!

This is the official PyTorch implementation of the ALBEF paper [Blog]. This repository supports pre-training on custom datasets, as well as finetuning on VQA, SNLI-VE, NLVR2, Image-Text Retrieval on MSCOCO and Flickr30k, and visual grounding on RefCOCO+. Pre-trained and finetuned checkpoints are released.

Requirements:

  • pytorch 1.8.0
  • transformers 4.8.1
  • timm 0.4.9

Download:

Visualization:

We provide code in visualize.ipynb to visualize the important areas in an image for each word in a text. Here is an example visualization using the visual grounding checkpoint.

Try the Replicate demo here Replicate.

Pre-training on custom datasets:

  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 the paths for the json files.
  3. Pre-train the model using 8 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=8 --use_env Pretrain.py --config ./configs/Pretrain.yaml --output_dir output/Pretrain 

Image-Text Retrieval:

  1. Download MSCOCO or Flickr30k datasets from the original websites.
  2. Download and extract the provided dataset json files.
  3. In configs/Retrieval_coco.yaml or configs/Retrieval_flickr.yaml, set the paths for the json files and the image path.
  4. Finetune the pre-trained checkpoint using 8 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=8 --use_env Retrieval.py \
--config ./configs/Retrieval_flickr.yaml \
--output_dir output/Retrieval_flickr \
--checkpoint [Pretrained checkpoint]

VQA:

  1. Download VQA v2 dataset and Visual Genome dataset from the original websites.
  2. Download and extract the provided dataset json files.
  3. In configs/VQA.yaml, set the paths for the json files and the image paths.
  4. Finetune the pre-trained checkpoint using 8 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=8 --use_env VQA.py \
--config ./configs/VQA.yaml \
--output_dir output/vqa \
--checkpoint [Pretrained checkpoint]
  1. Evaluate the result using the official evaluation server.

Visual Entailment:

  1. Download SNLI-VE dataset from the original website.
  2. Download and extract the provided dataset json files.
  3. In configs/VE.yaml, set the paths for the json files and the image path.
  4. Finetune the pre-trained checkpoint using 8 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=8 --use_env VE.py \
--config ./configs/VE.yaml \
--output_dir output/VE \
--checkpoint [Pretrained checkpoint]

Visual Grounding on RefCOCO+:

  1. Download MSCOCO dataset from the original website.
  2. Download and extract the provided dataset json files.
  3. In configs/Grounding.yaml, set the paths for the json files and the image path.
  4. Finetune the pre-trained checkpoint using 8 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=8 --use_env Grounding.py \
--config ./configs/Grounding.yaml \
--output_dir output/RefCOCO \
--gradcam_mode itm \ 
--block_num 8 \
--checkpoint [Pretrained checkpoint]

NLVR2:

NLVR2 requires an additional pre-training step with text-assignment (TA) to adapt the model for image-pair inputs. In order to perform TA, first set the paths for the json training files in configs/NLVR_pretrain.yaml, then run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env Pretrain_nlvr.py \
--config ./configs/NLVR_pretrain.yaml \
--output_dir output/NLVR_pretrain \
--checkpoint [Pretrained checkpoint]

We provide the checkpoint after TA pre-training, which can be fine-tuned with the following steps.

  1. Download NLVR2 dataset from the original website.
  2. Download and extract the provided dataset json files.
  3. In configs/NLVR.yaml, set the paths for the json files and the image path.
  4. Finetune the pre-trained checkpoint using 8 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=8 --use_env NLVR.py \
--config ./configs/NLVR.yaml \
--output_dir output/NLVR \
--checkpoint [TA pretrained checkpoint]

Citation

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

@inproceedings{ALBEF,
      title={Align before Fuse: Vision and Language Representation Learning with Momentum Distillation}, 
      author={Junnan Li and Ramprasaath R. Selvaraju and Akhilesh Deepak Gotmare and Shafiq Joty and Caiming Xiong and Steven Hoi},
      year={2021},
      booktitle={NeurIPS},
}

About

Code for ALBEF: a new vision-language pre-training method

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 96.7%
  • Jupyter Notebook 3.3%