- Python 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| Patch_CLIP | ||
| AUTHORS.md | ||
| CODE_OF_CONDUCT.md | ||
| CONTRIBUTING.md | ||
| LICENSE | ||
| MAINTAINERS 2.md | ||
| Patch_CLIP.zip | ||
| README.md | ||
| SECURITY 1.md | ||
| SETUP.md | ||
Patch-CLIP
🏥 Overview
Patch-CLIP is based on CLIP open source implementation. It introduces additional loss correlating patch token embeddings to text, similar to the global image-text contrastive loss.
To finetune
torchrun --nnodes=1 --nproc_per_node=1 finetune.py
--epochs 10
--batch_size 64
--optimizer "sgd"
--lr 0.00001
--context_length 77
--save_interval 100 --log_interval 10
--train_datapath ""
--train_csvpath ""
--val_datapath ""
--val_csvpath ""
--test_datapath ""
--test_csvpath ""
--checkpoint ""
--arch 'vitl14'
--summary_file ""
--model_name
--concept "Pleural Effusion"
To evaluate
torchrun --nnodes=1 --nproc_per_node=1 finetune.py
--epochs 10
--batch_size 64
--evaluate
--optimizer "sgd"
--lr 0.00001
--context_length 77
--save_interval 100 --log_interval 10
--train_datapath ""
--train_csvpath ""
--val_datapath ""
--val_csvpath ""
--test_datapath ""
--test_csvpath ""
--checkpoint ""
--arch 'vitl14'
--summary_file ""
--model_name
--concept "Pleural Effusion"
Prerequisites
- Install packages in the requirements.txt
Getting Help
- Please email sheethal.bhat@fau.de
License
This software is licensed under the GNU GENERAL PUBLIC LICENSE, Version 3, 29 June 2007 license.
Acknowledgments
We thank the authors of CheXZero https://github.com/rajpurkarlab/CheXzero/tree/main Tiu, E., Talius, E., Patel, P. et al. Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning. Nat. Biomed. Eng (2022). https://doi.org/10.1038/s41551-022-00936-9
Ready to get started? Run the installation script and visit the setup wizard to configure your installation!