RETFound/examples/RETFound_MESSIDOR2_demo.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "76b39fb1",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## Jupyter notebook example - Classification task\n",
    "### Example using [MESSIDOR2](https://www.adcis.net/en/third-party/messidor2/) dataset\n",
    "**Application**: Using RETFound for five-category diabetic retinopathy classification\n",
    "\n",
    "**Author**: Yukun Zhou\n",
    "\n",
    "**Date**: 30 Nov 2025\n",
    "\n",
    "**Performance**:\n",
    "\n",
    "<table align=\"left\">\n",
    "<tr>\n",
    "  <th>Accuracy</th>\n",
    "  <th>Recall</th>\n",
    "  <th>F1 Score</th>\n",
    "  <th>ROC AUC</th>\n",
    "  <th>PR AUC</th>\n",
    "</tr>\n",
    "<tr>\n",
    "  <td>0.7091</td>\n",
    "  <td>0.5616</td>\n",
    "  <td>0.6078</td>\n",
    "  <td>0.9037</td>\n",
    "  <td>0.6863</td>\n",
    "</tr>\n",
    "</table>\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7ec435a7",
   "metadata": {},
   "source": [
    "## 1. Install environment\n",
    "1. Follow [RETFound README](https://github.com/rmaphoh/RETFound) to install environment\n",
    "2. Restart this Jupyter Notebook\n",
    "3. Select Kernel retfound"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7cbf5e93-6ca0-4401-88e6-64e39968e7cd",
   "metadata": {},
   "outputs": [],
   "source": [
    "import sys, torch\n",
    "from pathlib import Path\n",
    "import os\n",
    "\n",
    "PROJECT_ROOT = Path.cwd().resolve()\n",
    "\n",
    "if PROJECT_ROOT.name == 'examples': PROJECT_ROOT = PROJECT_ROOT.parent\n",
    "os.chdir(PROJECT_ROOT)\n",
    "\n",
    "print('Project root:', PROJECT_ROOT)\n",
    "print(\"sys.executable:\", sys.executable)\n",
    "print(\"torch version:\", torch.__version__)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ed67953f",
   "metadata": {},
   "source": [
    "## 2. Prepare MESSIDOR2 dataset\n",
    "1. Download from the [shared data pool](https://github.com/rmaphoh/RETFound/blob/main/BENCHMARK.md).\n",
    "2. Put the data folder under the project directory, e.g. \"RETFound/MESSIDOR2\"\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "357be2fa-a914-4d1f-8759-76b2b1c3f20f",
   "metadata": {},
   "source": [
    "## 3. Hyperparameter and path settings\n",
    "1. Can choose finetune or lp (linear probe)\n",
    "2. Model selection [info](https://github.com/rmaphoh/RETFound#:~:text=In%20train.sh%2C%20the%20model%20can%20be%20selected%20by%20changing%20the%20hyperparameters%20MODEL%2C%20MODEL_ARCH%2C%20FINETUNE%3A)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5f675843",
   "metadata": {},
   "outputs": [],
   "source": [
    "from pathlib import Path\n",
    "ADAPTATION='finetune'\n",
    "MODEL='RETFound_dinov2'\n",
    "MODEL_ARCH='retfound_dinov2'\n",
    "FINETUNE='RETFound_dinov2_meh'\n",
    "DATASET='MESSIDOR2'\n",
    "NUM_CLASS=5\n",
    "DATA_PATH=PROJECT_ROOT/DATASET\n",
    "BATCH_SIZE=24\n",
    "EPOCHS=50\n",
    "INPUT_SIZE=224\n",
    "WORLD_SIZE=1\n",
    "TASK=f\"{MODEL_ARCH}_{DATASET}_{ADAPTATION}\"\n",
    "OUTPUT_DIR=PROJECT_ROOT/'output_dir'/TASK\n",
    "print('DATA_PATH:',DATA_PATH)\n",
    "print('TASK:',TASK)\n",
    "print('OUTPUT_DIR:',OUTPUT_DIR)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6ac04845",
   "metadata": {},
   "source": [
    "## 4. Fine-tuning and testing RETFound on MESSIDOR2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d23ff751",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import sys\n",
    "\n",
    "!{sys.executable} main_finetune.py \\\n",
    "  --model {MODEL} \\\n",
    "  --model_arch {MODEL_ARCH} \\\n",
    "  --finetune {FINETUNE} \\\n",
    "  --savemodel \\\n",
    "  --global_pool \\\n",
    "  --batch_size {BATCH_SIZE} \\\n",
    "  --epochs {EPOCHS} \\\n",
    "  --nb_classes {NUM_CLASS} \\\n",
    "  --data_path {DATA_PATH} \\\n",
    "  --input_size {INPUT_SIZE} \\\n",
    "  --task {TASK} \\\n",
    "  --adaptation {ADAPTATION}"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "84ce93ac",
   "metadata": {},
   "source": [
    "## 5. Evaluation-only"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0af0f8a7",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import sys\n",
    "\n",
    "CKPT = OUTPUT_DIR / \"checkpoint-best.pth\"\n",
    "\n",
    "!{sys.executable} main_finetune.py \\\n",
    "  --model {MODEL} \\\n",
    "  --model_arch {MODEL_ARCH} \\\n",
    "  --finetune {FINETUNE} \\\n",
    "  --savemodel \\\n",
    "  --global_pool \\\n",
    "  --batch_size 128 \\\n",
    "  --nb_classes {NUM_CLASS} \\\n",
    "  --data_path {DATA_PATH} \\\n",
    "  --input_size {INPUT_SIZE} \\\n",
    "  --task {TASK} \\\n",
    "  --adaptation {ADAPTATION} \\\n",
    "  --eval \\\n",
    "  --resume {CKPT}\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "02d2dce7-31c2-48e2-87ce-9223b74cf94e",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "environment": {
   "kernel": "retfound",
   "name": "workbench-notebooks.m128",
   "type": "gcloud",
   "uri": "us-docker.pkg.dev/deeplearning-platform-release/gcr.io/workbench-notebooks:m128"
  },
  "kernelspec": {
   "display_name": "retfound_jupyter (Local)",
   "language": "python",
   "name": "retfound"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}