{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.6.6","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[],"dockerImageVersionId":28755,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session\n\n# Use the kagglehub client library to attach Kaggle resources like competitions, datasets, and models to your session\n# Learn more about kagglehub: https://github.com/Kaggle/kagglehub/blob/main/README.md\n\nimport kagglehub\n# kagglehub.dataset_download('<owner>/<dataset-slug>')","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\"\"\"Standalone Kaggle Notebook script for Diabetic Retinopathy classification (EfficientNet-B0).\n\nCan be executed in a single Kaggle code cell with GPU enabled.\n\"\"\"\n\nfrom __future__ import annotations\n\nimport json\nimport os\nimport pathlib\nimport platform\nimport random\nimport re\nimport shutil\nimport sys\nimport time\nimport zipfile\nfrom dataclasses import asdict, dataclass, field\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nfrom PIL import Image\nimport torch\nimport torch.nn as nn\nfrom torch.utils.data import DataLoader, Dataset, Subset, WeightedRandomSampler\nfrom torchvision import datasets, models, transforms\n\n# ---------------------------------------------------------------------------\n# CONSTANTS & CONFIGURATION\n# ---------------------------------------------------------------------------\nDATASET_DIR = \"/kaggle/input/aptos2019-blindness-detection\"\nMETADATA_CSV = \"train.csv\"\nIMAGE_SUBDIRS = [\"train_images\", \".\"]\n\nID_COLUMN = \"id_code\"\nLABEL_COLUMN = \"diagnosis\"\n\nCLASS_NAMES = [\"No_DR\", \"Mild\", \"Moderate\", \"Severe\", \"Proliferative_DR\"]\nLABEL_MAP = {\n    0: \"No_DR\", \"0\": \"No_DR\",\n    1: \"Mild\", \"1\": \"Mild\",\n    2: \"Moderate\", \"2\": \"Moderate\",\n    3: \"Severe\", \"3\": \"Severe\",\n    4: \"Proliferative_DR\", \"4\": \"Proliferative_DR\",\n}\n\nBACKBONE = \"efficientnet_b0\"\nIMAGE_SIZE = 224\nEPOCHS = 6\nBATCH_SIZE = 32\nLEARNING_RATE = 3e-4\nMAX_IMAGES_PER_CLASS = 400\nOUTPUT_DIR = \"/kaggle/working/diabetic_retinopathy/efficientnet_b0\"\n\nIMAGENET_MEAN = [0.485, 0.456, 0.406]\nIMAGENET_STD = [0.229, 0.224, 0.225]\nIMAGE_SUFFIXES = {\".png\", \".jpg\", \".jpeg\", \".bmp\", \".tif\", \".tiff\", \".webp\"}\n\n\ndef working_root() -> Path:\n    kaggle = Path(\"/kaggle/working\")\n    try:\n        kaggle.mkdir(parents=True, exist_ok=True)\n        probe = kaggle / \".write_probe\"\n        probe.touch()\n        probe.unlink()\n        return kaggle\n    except OSError:\n        local = Path.cwd() / \"_work\"\n        local.mkdir(parents=True, exist_ok=True)\n        return local\n\n\ndef rebase_to_working(path: str | Path) -> Path:\n    candidate = Path(path)\n    try:\n        return working_root() / candidate.relative_to(\"/kaggle/working\")\n    except ValueError:\n        return candidate\n\n\n@dataclass\nclass TrainingConfig:\n    disease_id: str = \"diabetic_retinopathy\"\n    dataset_dir: str = DATASET_DIR\n    class_names: list[str] = field(default_factory=lambda: list(CLASS_NAMES))\n    backbone: str = BACKBONE\n    image_size: int = IMAGE_SIZE\n    epochs: int = EPOCHS\n    batch_size: int = BATCH_SIZE\n    learning_rate: float = LEARNING_RATE\n    val_split: float = 0.15\n    seed: int = 42\n    num_workers: int = 2\n    max_images_per_class: int | None = MAX_IMAGES_PER_CLASS\n    output_dir: str | None = OUTPUT_DIR\n    model_version: str = \"demo-0.1.0\"\n\n    def resolved_output_dir(self) -> Path:\n        raw = (Path(self.output_dir) if self.output_dir\n               else Path(\"/kaggle/working\") / self.disease_id / self.backbone)\n        return rebase_to_working(raw)\n\n\ndef organise_from_metadata_csv(csv_path: Path, image_dirs: list[Path], id_column: str,\n                               label_column: str, class_names: list[str], destination: Path,\n                               label_map: dict | None = None) -> dict:\n    if not csv_path.exists():\n        print(f\"[data] no metadata CSV at {csv_path}\")\n        return {}\n\n    frame = pd.read_csv(csv_path)\n    if id_column not in frame.columns or label_column not in frame.columns:\n        print(f\"[warn] Required columns ({id_column}, {label_column}) not found in {list(frame.columns)}\")\n        return {}\n\n    index: dict[str, Path] = {}\n    for directory in image_dirs:\n        if not directory.exists():\n            continue\n        for path in directory.rglob(\"*\"):\n            if path.is_file() and path.suffix.lower() in IMAGE_SUFFIXES:\n                index.setdefault(path.stem, path)\n\n    if not index:\n        print(f\"[warn] no images found in {[str(d) for d in image_dirs]}\")\n        return {}\n\n    if destination.exists():\n        shutil.rmtree(destination)\n\n    linked = {name: 0 for name in class_names}\n    for _, row in frame.iterrows():\n        raw_label = row[label_column]\n        label = label_map.get(raw_label, label_map.get(str(raw_label))) if label_map else str(raw_label)\n        if label not in linked:\n            continue\n        source = index.get(str(row[id_column]))\n        if source is None:\n            continue\n        class_dir = destination / label\n        class_dir.mkdir(parents=True, exist_ok=True)\n        target = class_dir / source.name\n        if not target.exists():\n            try:\n                target.symlink_to(source.resolve())\n            except (OSError, NotImplementedError):\n                shutil.copy2(source, target)\n        linked[label] += 1\n\n    found = {name: destination / name for name, count in linked.items() if count > 0}\n    print(f\"[data] organised {csv_path.name}: \" + \", \".join(f\"{n}={linked[n]}\" for n in class_names))\n    return found\n\n\ndef build_transforms(image_size: int):\n    train_tf = transforms.Compose([\n        transforms.Resize((image_size, image_size)),\n        transforms.RandomHorizontalFlip(),\n        transforms.RandomVerticalFlip(),\n        transforms.RandomRotation(15),\n        transforms.ColorJitter(brightness=0.2, contrast=0.2),\n        transforms.ToTensor(),\n        transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD),\n    ])\n    eval_tf = transforms.Compose([\n        transforms.Resize((image_size, image_size)),\n        transforms.ToTensor(),\n        transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD),\n    ])\n    return train_tf, eval_tf\n\n\ndef build_class_view(mapping: dict[str, Path], class_names: list[str], destination: Path) -> Path:\n    if destination.exists():\n        shutil.rmtree(destination)\n    destination.mkdir(parents=True, exist_ok=True)\n    present = [c for c in class_names if c in mapping]\n    for index, name in enumerate(present):\n        link = destination / f\"{index:02d}_{name}\"\n        source = mapping[name].resolve()\n        try:\n            link.symlink_to(source, target_is_directory=True)\n        except (OSError, NotImplementedError):\n            shutil.copytree(source, link)\n    return destination\n\n\ndef _cap_per_class(dataset: datasets.ImageFolder, cap: int | None, seed: int) -> list[int]:\n    indices = list(range(len(dataset)))\n    if cap is None:\n        return indices\n    rng = random.Random(seed)\n    by_class: dict[int, list[int]] = {}\n    for i, (_, label) in enumerate(dataset.samples):\n        by_class.setdefault(label, []).append(i)\n    kept: list[int] = []\n    for label, items in by_class.items():\n        rng.shuffle(items)\n        kept.extend(items[:cap])\n    kept.sort()\n    return kept\n\n\ndef stratified_split(dataset: datasets.ImageFolder, indices: list[int],\n                     val_split: float, seed: int) -> tuple[list[int], list[int]]:\n    rng = random.Random(seed)\n    by_class: dict[int, list[int]] = {}\n    for i in indices:\n        by_class.setdefault(dataset.samples[i][1], []).append(i)\n\n    train_idx: list[int] = []\n    val_idx: list[int] = []\n    for label, items in sorted(by_class.items()):\n        rng.shuffle(items)\n        n_val = max(1, int(round(len(items) * val_split))) if len(items) > 1 else 0\n        val_idx.extend(items[:n_val])\n        train_idx.extend(items[n_val:])\n    return sorted(train_idx), sorted(val_idx)\n\n\ndef make_balanced_sampler(dataset: datasets.ImageFolder, indices: list[int]) -> WeightedRandomSampler:\n    counts: dict[int, int] = {}\n    for i in indices:\n        label = dataset.samples[i][1]\n        counts[label] = counts.get(label, 0) + 1\n    weights = [1.0 / counts[dataset.samples[i][1]] for i in indices]\n    return WeightedRandomSampler(weights, num_samples=len(indices), replacement=True)\n\n\ndef build_model(backbone: str, num_classes: int) -> tuple[nn.Module, bool]:\n    pretrained = True\n    try:\n        model = models.efficientnet_b0(weights=models.EfficientNet_B0_Weights.DEFAULT)\n    except Exception as exc:\n        print(f\"[warn] pretrained weights unavailable ({exc}); initializing random weights.\")\n        pretrained = False\n        model = models.efficientnet_b0(weights=None)\n\n    in_features = model.classifier[-1].in_features\n    model.classifier[-1] = nn.Linear(in_features, num_classes)\n    return model, pretrained\n\n\ndef score(true: np.ndarray, pred: np.ndarray, class_names: list[str]) -> dict:\n    n = len(class_names)\n    matrix = np.zeros((n, n), dtype=int)\n    for t, p in zip(true, pred):\n        matrix[int(t), int(p)] += 1\n    support = matrix.sum(axis=1)\n    predicted = matrix.sum(axis=0)\n    correct = np.diag(matrix)\n\n    with np.errstate(divide=\"ignore\", invalid=\"ignore\"):\n        recall = np.where(support > 0, correct / np.maximum(support, 1), 0.0)\n        precision = np.where(predicted > 0, correct / np.maximum(predicted, 1), 0.0)\n        f1 = np.where((precision + recall) > 0, 2 * precision * recall / np.maximum(precision + recall, 1e-12), 0.0)\n\n    return {\n        \"accuracy\": float(correct.sum() / max(len(true), 1)),\n        \"macro_f1\": float(f1.mean()),\n        \"macro_recall\": float(recall.mean()),\n        \"per_class\": {\n            name: {\n                \"precision\": round(float(precision[i]), 4),\n                \"recall\": round(float(recall[i]), 4),\n                \"f1\": round(float(f1[i]), 4),\n                \"support\": int(support[i]),\n            }\n            for i, name in enumerate(class_names)\n        },\n        \"confusion_matrix\": matrix.tolist(),\n        \"confusion_matrix_axes\": {\"rows\": \"true\", \"columns\": \"predicted\", \"order\": class_names},\n    }\n\n\ndef run_epoch(model, loader, device, criterion, optimiser=None) -> tuple[float, np.ndarray, np.ndarray]:\n    training = optimiser is not None\n    model.train(training)\n    total_loss, seen = 0.0, 0\n    trues, preds = [], []\n\n    with torch.set_grad_enabled(training):\n        for images, labels in loader:\n            images, labels = images.to(device), labels.to(device)\n            outputs = model(images)\n            loss = criterion(outputs, labels)\n            if training:\n                optimiser.zero_grad()\n                loss.backward()\n                optimiser.step()\n            total_loss += float(loss.item()) * labels.size(0)\n            seen += labels.size(0)\n            trues.append(labels.detach().cpu().numpy())\n            preds.append(outputs.detach().argmax(1).cpu().numpy())\n\n    return (total_loss / max(seen, 1),\n            np.concatenate(trues) if trues else np.array([]),\n            np.concatenate(preds) if preds else np.array([]))\n\n\ndef write_artifacts(output_dir: Path, config: TrainingConfig, model: nn.Module,\n                    class_names: list[str], metrics: dict, extra: dict) -> None:\n    output_dir.mkdir(parents=True, exist_ok=True)\n\n    torch.save({\n        \"state_dict\": model.state_dict(),\n        \"backbone\": config.backbone,\n        \"num_classes\": len(class_names),\n        \"class_names\": class_names,\n        \"disease_id\": config.disease_id,\n        \"model_version\": config.model_version,\n    }, output_dir / \"model.pt\")\n\n    (output_dir / \"labels.json\").write_text(json.dumps({\n        \"disease_id\": config.disease_id,\n        \"class_names\": class_names,\n        \"index_to_label\": {str(i): name for i, name in enumerate(class_names)},\n    }, indent=2), encoding=\"utf-8\")\n\n    (output_dir / \"metrics.json\").write_text(json.dumps({\n        \"disease_id\": config.disease_id,\n        \"backbone\": config.backbone,\n        \"model_version\": config.model_version,\n        \"is_demo_model\": True,\n        **extra,\n        **metrics,\n        \"config\": asdict(config),\n        \"environment\": {\n            \"python\": platform.python_version(),\n            \"torch\": torch.__version__,\n            \"device\": extra.get(\"device\", \"unknown\"),\n        },\n    }, indent=2), encoding=\"utf-8\")\n\n    (output_dir / \"preprocessing.json\").write_text(json.dumps({\n        \"disease_id\": config.disease_id,\n        \"input\": \"rgb_image\",\n        \"resize\": [config.image_size, config.image_size],\n        \"interpolation\": \"bilinear\",\n        \"scale\": \"divide_by_255\",\n        \"normalize\": {\"mean\": IMAGENET_MEAN, \"std\": IMAGENET_STD},\n        \"channel_order\": \"RGB\",\n        \"tensor_layout\": \"NCHW\",\n        \"notes\": \"Apply resize, then to-tensor (0-1), then normalize. No centre crop.\",\n    }, indent=2), encoding=\"utf-8\")\n\n\ndef package_zip(output_dir: Path, zip_name: str = \"retinopathy-effnet.zip\") -> Path:\n    zip_path = output_dir.parent / zip_name\n    files_to_pack = [\"model.pt\", \"labels.json\", \"metrics.json\", \"preprocessing.json\"]\n    with zipfile.ZipFile(zip_path, \"w\", zipfile.ZIP_DEFLATED) as zipf:\n        for fname in files_to_pack:\n            fpath = output_dir / fname\n            if fpath.exists():\n                zipf.write(fpath, arcname=fname)\n                print(f\"[zip] Packed: {fname} ({fpath.stat().st_size:,} bytes)\")\n    print(f\"\\n[zip] Created archive at: {zip_path} ({zip_path.stat().st_size:,} bytes)\")\n    return zip_path\n\n\ndef main() -> None:\n    started = time.time()\n    random.seed(42)\n    np.random.seed(42)\n    torch.manual_seed(42)\n    torch.cuda.manual_seed_all(42)\n\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    print(f\"[diabetic_retinopathy/efficientnet_b0] Device: {device}\")\n\n    config = TrainingConfig()\n    output_dir = config.resolved_output_dir()\n\n    # Organise from metadata\n    class_dirs = organise_from_metadata_csv(\n        csv_path=Path(DATASET_DIR) / METADATA_CSV,\n        image_dirs=[Path(DATASET_DIR) / d for d in IMAGE_SUBDIRS],\n        id_column=ID_COLUMN,\n        label_column=LABEL_COLUMN,\n        class_names=CLASS_NAMES,\n        destination=working_root() / \"_organised_dr_train\",\n        label_map=LABEL_MAP,\n    )\n\n    found_names = [n for n in config.class_names if n in class_dirs]\n    if not found_names:\n        raise RuntimeError(f\"No dataset images found under {DATASET_DIR}\")\n\n    view = build_class_view(class_dirs, config.class_names, working_root() / \"_view_dr\")\n    train_tf, eval_tf = build_transforms(config.image_size)\n\n    base = datasets.ImageFolder(str(view), transform=train_tf)\n    eval_base = datasets.ImageFolder(str(view), transform=eval_tf)\n    base.classes = eval_base.classes = found_names\n\n    indices = _cap_per_class(base, config.max_images_per_class, config.seed)\n    train_idx, val_idx = stratified_split(base, indices, config.val_split, config.seed)\n    n_train, n_val = len(train_idx), len(val_idx)\n    print(f\"[data] {n_train} train / {n_val} val across {len(found_names)} classes: {found_names}\")\n\n    train_loader = DataLoader(\n        Subset(base, train_idx),\n        batch_size=config.batch_size,\n        sampler=make_balanced_sampler(base, train_idx),\n        num_workers=config.num_workers,\n        drop_last=False,\n    )\n    val_loader = DataLoader(\n        Subset(eval_base, val_idx or train_idx),\n        batch_size=config.batch_size,\n        shuffle=False,\n        num_workers=config.num_workers,\n    )\n\n    model, pretrained = build_model(config.backbone, len(found_names))\n    model.to(device)\n    criterion = nn.CrossEntropyLoss()\n    optimiser = torch.optim.AdamW(model.parameters(), lr=config.learning_rate)\n    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimiser, T_max=max(config.epochs, 1))\n\n    print(\"\\n--- Starting Training ---\")\n    history = []\n    for epoch in range(1, config.epochs + 1):\n        train_loss, t_true, t_pred = run_epoch(model, train_loader, device, criterion, optimiser)\n        val_loss, v_true, v_pred = run_epoch(model, val_loader, device, criterion)\n        scheduler.step()\n\n        train_acc = float((t_true == t_pred).mean()) if len(t_true) else 0.0\n        val_acc = float((v_true == v_pred).mean()) if len(v_true) else 0.0\n        history.append({\n            \"epoch\": epoch,\n            \"train_loss\": round(train_loss, 4),\n            \"val_loss\": round(val_loss, 4),\n            \"train_accuracy\": round(train_acc, 4),\n            \"val_accuracy\": round(val_acc, 4),\n        })\n        print(f\"[epoch {epoch}/{config.epochs}] train_loss={train_loss:.4f} \"\n              f\"val_loss={val_loss:.4f} val_acc={val_acc:.4f}\")\n\n    print(\"\\n--- Evaluating Model ---\")\n    _, v_true, v_pred = run_epoch(model, val_loader, device, criterion)\n    metrics = score(v_true, v_pred, found_names)\n\n    write_artifacts(output_dir, config, model, found_names, metrics, {\n        \"data_source\": \"kaggle-aptos2019\",\n        \"dataset_root\": DATASET_DIR,\n        \"pretrained_backbone\": pretrained,\n        \"device\": str(device),\n        \"train_size\": n_train,\n        \"val_size\": n_val,\n        \"history\": history,\n        \"duration_seconds\": round(time.time() - started, 1),\n    })\n\n    print(f\"\\n[done] Accuracy: {metrics['accuracy']:.4f}\")\n    print(f\"[done] Macro F1: {metrics['macro_f1']:.4f}\")\n    print(f\"[done] Macro Recall: {metrics['macro_recall']:.4f}\")\n    print(\"\\nPer-class Metrics:\")\n    for cls_name, cls_metrics in metrics[\"per_class\"].items():\n        print(f\"  {cls_name}: Precision={cls_metrics['precision']:.4f}, \"\n              f\"Recall={cls_metrics['recall']:.4f}, F1={cls_metrics['f1']:.4f}\")\n\n    print(\"\\nConfusion Matrix:\")\n    print(np.array(metrics[\"confusion_matrix\"]))\n\n    package_zip(output_dir)\n\n\nif __name__ == \"__main__\":\n    main()\n","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}