OrdinalCLIP

Created by Wanhua Li*, Xiaoke Huang*, Zheng Zhu, Yansong Tang, Xiu Li, Jiwen Lu†, Jie Zhou

This repository contains PyTorch implementation of "OrdinalCLIP: Learning Rank Prompts for Language-Guided Ordinal Regression".

[Project Page] [arXiv] [Zhihu]

Prerequisite

Clone repository

git clone --recursive https://github.com/xk-huang/OrdinalCLIP.git

Dataset

Download links: [Google Drive] [Baidu Drive]

Download data and save them to data/MORPH according to the config files.

Checkpoint

Download checkpoints (Pytorch/CLIP/Custom) and save them to .cache/ accordingly.

Weights download links: [Google Drive] [Baidu Drive]

.cache
├── clip
│   ├── RN50.pt
│   └── ViT-B-16.pt
├── resnet
│   └── resnet50_imagenet.pt
└── vgg
    ├── vgg16_imagenet.pt
    └── vgg_imdb_wiki.pt

Installation

There are two options to set up environment.

Recommend Docker environment.

Docker Environment

[Docker installation guideline]

---

1. Install Docker:
   1. Ubuntu >= 18.04: https://docs.docker.com/engine/install/ubuntu/
   2. Ubuntu == 16.04: https://www.digitalocean.com/community/tutorials/how-to-install-and-use-docker-on-ubuntu-16-04
2. Then install nvidia-docker: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#setting-up-nvidia-container-toolkit

---

docker build -t ordinalclip:latest .

Local Environment

By .yaml file:

conda env create -f environment.yaml
conda activate ordinalclip
pip install -r requirements.txt
# git submodule update --init # git submodule add [email protected]:openai/CLIP.git
pip install -e CLIP/
pip install -e .

Or manually install:

conda create --name ordinalclip python=3.8
conda activate ordinalclip
conda install pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.3 -c pytorch -c conda-forge
pip install -r requirements.txt
# git submodule update --init # git submodule add [email protected]:openai/CLIP.git
pip install -e CLIP/
pip install -e .
# pip install setuptools==59.5.0 # https://github.com/pytorch/pytorch/pull/69904

For dev:

pip install bandit==1.7.0 black==22.3.0 flake8-docstrings==1.6.0 flake8==3.9.1 flynt==0.64 isort==5.8.0 mypy==0.902 pre-commit==2.13.0 pytest ipython
pre-commit install

Usage

[If you run scripts in docker, click me]

---

Start Docker container

docker run -itd --gpus all \
  -v $(realpath .cache/):/workspace/OrdinalCLIP/.cache \
  -v $(realpath data/):/workspace/OrdinalCLIP/data \
  -v $(realpath results/):/workspace/OrdinalCLIP/results \
  -v $(realpath configs/):/workspace/OrdinalCLIP/configs \
  --name ordinalclip \
  --shm-size 8gb \
  ordinalclip bash
docker exec -it ordinalclip bash
# In the container, run `python ...`

After running, remove container and release resources:

exit # or Ctrl^D
docker rm -f ordinalclip

---

Single-run mode:

python scripts/run.py --config configs/default.yaml --config configs/base_cfgs/*.yml --config ...

Multi-run mode:

python scripts/experiments/meta_config_generator.py -c $meta_config_file
python scripts/experiments/config_sweeper.py --max_num_gpus_used_in_parallel 8 --num_jobs_per_gpu 1 -d $sweep_config_dir 

Visualizing and quantifying ordinality:

CUDA_VISIBLE_DEVICES=-1 find results/ -name 'config.yaml' -exec python scripts/vis_ordinality.py -c {} \;

Parsing results:

python scripts/experiments/parse_results.py -d $result_dir -p 'test_stats.json'
python scripts/experiments/parse_results.py -d $result_dir -p 'ordinality.json'
# or
python scripts/experiments/parse_results.py -T <(find -name 'test_stats.json') -p 'test_stats.json'
python scripts/experiments/parse_results.py -T <(find -name 'ordinality.json') -p 'ordinality.json'

Experiments Structure

Each experiment has its own name, with several config component names split by "_". The experiment folder includes:

• ckpt/: checkpoints
• *_logger/: the logs from pytorch_lightning
• config.yaml: config
• run.log: log the status of model during running
• val/test_stats.json: metrics to be parsed
• val/val_stats.json: metrics to be parsed
• val/ordinality.json: metrics to be parsed

Training and Evaluation Logs

Please check out releases: https://github.com/xk-huang/OrdinalCLIP/releases/tag/train_eval_logs

The Code Structure

Structure of the Codebase (click to expand)

Structure of ordinalclip

• ordinalclip

  • models attributes: image_encoder (torchvision model: fp32, CLIP image encoder: fp16), text_encoder (float32, but layer norm are computed in float32), all converted to float32
    • prompt learner
      • plain prompt learner args: num_ranks, num_tokens_per_rank, num_tokens_for_context, rank_tokens_position, init_rank_path, init_context, rank_specific_context attributes: context_embeds, rank_embeds, pseudo_sentence_tokens,
      • rank prompt learner (inherited from plain prompt learner) args: num_ranks, num_tokens_per_rank, num_tokens_for_context, rank_tokens_position, init_rank_path, init_context, rank_specific_context,interpolation_type attributes: weights for interpolation
  • runner
    • runner: A wrapper using pl.LightningModule, defines: loss_computation, metrics computation, create_optimizer, lr_scheduler
    • data: pl.LightningDataModule
    • utils: model io, parameter (un)freeze
  • utils
    • logging & registry from MMCV

• scripts

  • run.py To prepare args, use OmegaConf, logging and wandb logger, train/val/test dataloader, model (runner), setup trainer

• configs

License

MIT License

Acknowledgement

Many thanks to the following repositories:

• CoOp, CLIP, PyTorch Lightning, MMCV, MMDetection

What's More

Check out these amazing works leveraging CLIP for number problems!

• CrowdCLIP: Unsupervised Crowd Counting via Vision-Language Model
• L2RCLIP: Learning-to-Rank Meets Language: Boosting Language-Driven Ordering Alignment for Ordinal Classification

Citation

If you find this codebase helpful, please consider to cite:

@article{Li2022OrdinalCLIP,
  title={OrdinalCLIP: Learning Rank Prompts for Language-Guided Ordinal Regression},
  author={Wanhua Li and Xiaoke Huang and Zheng Zhu and Yansong Tang and Xiu Li and Jiwen Lu and Jie Zhou},
  journal={ArXiv},
  year={2022},
  volume={abs/2206.02338}
}