HRVRL： An Efficient and Interpretable Foundation Model for Retinal Image Analysis in Disease Diagnosis
- Framework
- Environment Requirements
```shell
git clone https://github.com/sulab-wmu/HRVRL.git
cd HRVRL
conda create -n hrvrl python=3.10 -y
conda activate hrvrl
pip install poetry
poetry install
pip install tensorflow==2.9.1
pip install torch==1.13.1 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117 --force
```
> if you encounter the following problem:
> AttributeError: module 'tensorflow' has no attribute 'GraphKeys'
> `vim /home/xxx/.conda/envs/xxx/lib/python3.10/site-packages/tensorlayer/layers.py`
> `import tensorflow as tf ---> import tensorflow.compat.v1 as tf`
- ### Finetune on your data
Download the [pretrain weight](https://drive.google.com/drive/folders/1Hzxv36qyyqTgyE0jGICJ-1hlbg0DnSFh?usp=drive_link) and run
`cd finetune `
` sh run.sh or CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 --master_port=xxx main_finetune_rip.py --batch_size 16 --world_size 1 --epochs 100 --blr 8e-3 --layer_decay 0.9 --weight_decay 0.0005 --nb_classes [your class number] --data_path [your dataset] --task [your output dir] --cutmix 0.1 --warmup_epochs 10 --drop_path 0.8 --finetune G_pretrain.pth --min_lr 1e-6 --input_size 512 > [your log file name] 2>&1 &`
- ### Reproduction
Download the [pretrain weight](https://drive.google.com/drive/folders/1Hzxv36qyyqTgyE0jGICJ-1hlbg0DnSFh?usp=drive_link) and [reproduction weight](https://drive.google.com/drive/folders/1U9kxh7_d-dXtFw6b8Gkpl4ZXd0bMgtbo?usp=drive_link) and put the target dir:
- run `hrvrl_preprocess_augmentation.ipynb ` to get the demo and results of multi-level augmentation.
- run `hrvrl_pretrain_av_visulization.ipynb ` to get the demo and results of pretrained feature visulization.
- run `hrvrl_task_1_2_fig_plot_finetune.ipynb ` to get the result of Figure2.
- run `hrvrl_explain_example.ipynb ` to get the demo and results of hierarchical explainable method.
- ### Acknowledge
[RIP-AV](https://github.com/weidai00/RIP-AV) and [RETFound](https://github.com/rmaphoh/RETFound_MAE)
