The dataset and checkpoint is available at Huggingface, Baidu Cloud(key: 3iqf).
Repo Structure
src/:
|--setup.py
|--pmc_clip/
| |--loss/
| |--model/: PMC-CLIP model and variants
| |--model_configs/
| |--factory.py: Create model according to configs
| |--transform.py: data augmentation
|--training/
| |--main.py
| |--scheduler.py: Learning rate scheduler
| |--train.py
| |--evaluate.py
| |--data.py
| |--params.py
docs/: project pages
conda create -n pmc_clip python=3.8
conda activate pmc_clip
pip install -r requirements.txt
# pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt
python setup.py develop # install pmc_clip with dev mode
Download from Huggingface, Baidu Cloud(key: 3iqf). Or follow the Pipeline of PMC-OA Development if you want to start from scratch.
Single GPU
python -m training.main \
--dataset-type "csv" --csv-separator "," --save-frequency 5 \
--report-to tensorboard \
--train-data="path/to/train.csv" --val-data="path/to/valid.csv" \
--csv-img-key image --csv-caption-key caption \
--warmup 500 --batch-size=8 --lr=1e-4 --wd=0.1 --epochs=100 --workers=8 \
--model RN50_fusion4 --hugging-face --mlm --crop-scale 0.5
Multi GPU
CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 --rdzv_endpoint=$HOSTE_NODE_ADDR -m training.main \
--dataset-type "csv" --csv-separator "," --save-frequency 5 \
--report-to tensorboard \
--train-data="path/to/train.csv" --val-data="path/to/valid.csv" \
--csv-img-key image --csv-caption-key caption \
--warmup 500 --batch-size=128 --lr=1e-4 --wd=0.1 --epochs=100 --workers=8 \
--model RN50_fusion4 --hugging-face --mlm --crop-scale 0.5
Load checkpoint and eval on 2k samples from testset.
python -m training.main \
--dataset-type "csv" --csv-separator "," --report-to tensorboard \
--val-data="path/to/test.csv" \
--csv-img-key image --csv-caption-key caption \
--batch-size=32 --workers=8 \
--model RN50_fusion4 --hugging-face --mlm --crop-scale 0.1 \
--resume /path/to/checkpoint.pt \
--test-2000
The code is based on OpenCLIP and M3AE. We thank the authors for their open-sourced code and encourage users to cite their works when applicable.
Note that our code don't supported tools like horovod, wandb in OpenCLIP. But we keep the code from OpenCLIP for consistency.
Please raise an issue if you need help, any contributions are welcomed.
- Compatibility testing on more env settings
- Support for horovod, wandb
@article{lin2023pmc,
title={PMC-CLIP: Contrastive Language-Image Pre-training using Biomedical Documents},
author={Lin, Weixiong and Zhao, Ziheng and Zhang, Xiaoman and Wu, Chaoyi and Zhang, Ya and Wang, Yanfeng and Xie, Weidi},
journal={arXiv preprint arXiv:2303.07240},
year={2023}
}