annaferrari02
/

yolo8

+{
+ "metadata": {
+  "kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"},
+  "language_info": {"name": "python", "version": "3.12.12"},
+  "accelerator": "GPU",
+  "kaggle": {"accelerator": "gpu", "isInternetEnabled": true, "language": "python", "isGpuEnabled": true}
+ },
+ "nbformat_minor": 4,
+ "nbformat": 4,
+ "cells": [
+  {
+   "cell_type": "code",
+   "source": "# ============================================================================\n# CONFIGURATION\n# ============================================================================\n\nDATASET_NAME = 'your-username/surgical-tool-ism'\nDATASET_PATH = f'/kaggle/input/{DATASET_NAME.split(\"/\")[-1]}'\nDATA_YAML_PATH = f'{DATASET_PATH}/data.yaml'\n\nprint(f\"Dataset path: {DATASET_PATH}\")\nprint(f\"Data YAML path: {DATA_YAML_PATH}\")",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# Check GPU\n!nvidia-smi",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# Install ultralytics\n!pip install -q ultralytics",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "import os\n\n# Check dataset structure\nprint(\"=== ROOT DATASET ===\")\nprint(os.listdir(DATASET_PATH))\n\nprint(\"\\n=== STRUCTURE ===\")\nfor root, dirs, files in os.walk(DATASET_PATH):\n    level = root.replace(DATASET_PATH, '').count(os.sep)\n    if level < 3:\n        indent = ' ' * 2 * level\n        print(f'{indent}{os.path.basename(root)}/')\n        subindent = ' ' * 2 * (level + 1)\n        for file in files[:3]:\n            print(f'{subindent}{file}')\n        if len(files) > 3:\n            print(f'{subindent}... ({len(files)} files total)')",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# ===== CREATE TRAIN/VAL SPLIT =====\n\nimport os\nimport shutil\nfrom collections import defaultdict\n\nprint(\"Creating train/val split...\")\n\nSOURCE_IMAGES = f'{DATASET_PATH}/images'\nSOURCE_LABELS = f'{DATASET_PATH}/labels'\n\nDEST_BASE = '/kaggle/working/dataset_split'\nDEST_IMAGES_TRAIN = f'{DEST_BASE}/images/train'\nDEST_IMAGES_VAL = f'{DEST_BASE}/images/val'\nDEST_LABELS_TRAIN = f'{DEST_BASE}/labels/train'\nDEST_LABELS_VAL = f'{DEST_BASE}/labels/val'\n\nfor path in [DEST_IMAGES_TRAIN, DEST_IMAGES_VAL, DEST_LABELS_TRAIN, DEST_LABELS_VAL]:\n    os.makedirs(path, exist_ok=True)\n\nall_images = [f for f in os.listdir(SOURCE_IMAGES) if f.endswith('.jpg')]\nprint(f\"Found {len(all_images)} total images\")\n\n# Stratified split by background condition\nconditions = ['b00', 'b01', 'b02', 'b03']\ntrain_images = []\nval_images = []\n\nprint(\"\\nSplitting by condition (80/20):\")\nprint(f\"{'Condition':<12} | {'Total':<8} | {'Train':<8} | {'Val':<8}\")\nprint(\"-\" * 50)\n\nfor cond in conditions:\n    cond_images = [f for f in all_images if f.startswith(cond)]\n    n_total = len(cond_images)\n    n_train = int(n_total * 0.8)\n    \n    train_cond = cond_images[:n_train]\n    val_cond = cond_images[n_train:]\n    \n    train_images.extend(train_cond)\n    val_images.extend(val_cond)\n    \n    print(f\"{cond:<12} | {n_total:<8} | {len(train_cond):<8} | {len(val_cond):<8}\")\n\nprint(f\"\\nFinal split: {len(train_images)} train, {len(val_images)} val\")\n\n# Copy files\nprint(\"\\nCopying files...\")\nfor split_name, split_images in [('train', train_images), ('val', val_images)]:\n    dest_img_dir = DEST_IMAGES_TRAIN if split_name == 'train' else DEST_IMAGES_VAL\n    dest_lbl_dir = DEST_LABELS_TRAIN if split_name == 'train' else DEST_LABELS_VAL\n    \n    for img_name in split_images:\n        # Copy image\n        shutil.copy(\n            os.path.join(SOURCE_IMAGES, img_name),\n            os.path.join(dest_img_dir, img_name)\n        )\n        \n        # Copy label\n        lbl_name = img_name.replace('.jpg', '.txt')\n        src_lbl = os.path.join(SOURCE_LABELS, lbl_name)\n        if os.path.exists(src_lbl):\n            shutil.copy(src_lbl, os.path.join(dest_lbl_dir, lbl_name))\n\nprint(\"Split complete!\")",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# ===== CRITICAL FIX: CREATE CORRECT data.yaml =====\n\nimport yaml\n\n# CORRECTED data.yaml structure - names MUST be a list, not a dictionary!\ndata_config = {\n    'path': DEST_BASE,\n    'train': 'images/train',\n    'val': 'images/val',\n    'nc': 3,\n    'names': ['Large Needle Driver', 'Prograsp Forceps', 'Monopolar Curved Scissors']  # ✅ LIST format!\n}\n\ndata_yaml_output = os.path.join(DEST_BASE, 'data.yaml')\n\nwith open(data_yaml_output, 'w') as f:\n    yaml.dump(data_config, f, default_flow_style=False, sort_keys=False)\n\nprint(\"✅ Created corrected data.yaml\")\nprint(\"\\nContents:\")\nwith open(data_yaml_output) as f:\n    print(f.read())\n\n# Verify the format\nwith open(data_yaml_output) as f:\n    loaded = yaml.safe_load(f)\n    assert isinstance(loaded['names'], list), \"❌ ERROR: names should be a list!\"\n    assert len(loaded['names']) == 3, \"❌ ERROR: should have exactly 3 classes!\"\n    print(\"\\n✅ Verification passed: names is a list with 3 classes\")",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# ===== CHECK CLASS DISTRIBUTION =====\n\nimport os\nfrom collections import Counter\n\ndef check_class_distribution(labels_dir, split_name):\n    class_counts = Counter()\n    total_boxes = 0\n    \n    for label_file in os.listdir(labels_dir):\n        if not label_file.endswith('.txt'):\n            continue\n        with open(os.path.join(labels_dir, label_file)) as f:\n            for line in f:\n                parts = line.strip().split()\n                if len(parts) >= 5:\n                    class_id = int(parts[0])\n                    class_counts[class_id] += 1\n                    total_boxes += 1\n    \n    print(f\"\\n{'='*60}\")\n    print(f\"{split_name.upper()} CLASS DISTRIBUTION\")\n    print(f\"{'='*60}\")\n    \n    class_names = ['Large Needle Driver', 'Prograsp Forceps', 'Monopolar Curved Scissors']\n    \n    for class_id in sorted(class_counts.keys()):\n        count = class_counts[class_id]\n        percentage = (count / total_boxes) * 100\n        name = class_names[class_id] if class_id < 3 else f\"Unknown-{class_id}\"\n        print(f\"{name:<30}: {count:5d} ({percentage:5.1f}%)\")\n    \n    if len(class_counts) > 1:\n        max_count = max(class_counts.values())\n        min_count = min(class_counts.values())\n        ratio = max_count / min_count\n        print(f\"\\nImbalance ratio: {ratio:.2f}:1\")\n        if ratio > 3:\n            print(\"⚠️  SEVERE IMBALANCE!\")\n        elif ratio > 1.5:\n            print(\"⚠️  Moderate imbalance\")\n        else:\n            print(\"✅ Well balanced\")\n\ncheck_class_distribution(DEST_LABELS_TRAIN, 'training')\ncheck_class_distribution(DEST_LABELS_VAL, 'validation')",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# ===== TRAIN MODEL =====\n\nfrom ultralytics import YOLO\n\nOUTPUT_DIR = '/kaggle/working'\n\nprint(\"Initializing YOLOv8m model...\")\nmodel = YOLO('yolov8m.pt')\n\nprint(\"Starting training...\")\nprint(f\"Using data.yaml: {data_yaml_output}\")\n\nresults = model.train(\n    data=data_yaml_output,\n    epochs=60,\n    imgsz=640,\n    batch=10,\n    patience=15,\n    save=True,\n    project=OUTPUT_DIR,\n    name='yolov8m_kaggle_run',\n    exist_ok=True,\n    pretrained=True,\n    optimizer='AdamW',\n    lr0=0.001,\n    lrf=0.01,\n    momentum=0.937,\n    weight_decay=0.00046875,\n    warmup_epochs=3.0,\n    warmup_momentum=0.8,\n    warmup_bias_lr=0.1,\n    box=7.5,\n    cls=0.5,\n    dfl=1.5,\n    hsv_h=0.015,\n    hsv_s=0.3,\n    hsv_v=0.2,\n    degrees=0.0,\n    translate=0.05,\n    scale=0.3,\n    shear=0.0,\n    perspective=0.0,\n    flipud=0.0,\n    fliplr=0.5,\n    mosaic=0.5,\n    mixup=0.0,\n    copy_paste=0.0,\n    close_mosaic=20,\n    amp=True,\n    fraction=1.0,\n    profile=False,\n    freeze=None,\n    workers=2,\n    device=0,\n    plots=False,\n    val=True,\n    rect=False,\n    cos_lr=False,\n    overlap_mask=True,\n    mask_ratio=4,\n    dropout=0.0,\n    cache=False\n)\n\nprint(\"\\n\" + \"=\"*70)\nprint(\"TRAINING COMPLETE!\")\nprint(\"=\"*70)",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# ===== VALIDATE BEST MODEL =====\n\nbest_model_path = f'{OUTPUT_DIR}/yolov8m_kaggle_run/weights/best.pt'\n\nprint(f\"Loading best model from: {best_model_path}\")\nmodel = YOLO(best_model_path)\n\nmetrics = model.val()\n\nprint(\"\\n\" + \"=\"*70)\nprint(\"VALIDATION METRICS\")\nprint(\"=\"*70)\nprint(f\"mAP@0.5: {metrics.box.map50:.4f}\")\nprint(f\"mAP@0.5:0.95: {metrics.box.map:.4f}\")\nprint(f\"Precision: {metrics.box.mp:.4f}\")\nprint(f\"Recall: {metrics.box.mr:.4f}\")\n\n# Per-class metrics\nif hasattr(metrics.box, 'maps'):\n    print(\"\\nPer-class mAP@0.5:\")\n    class_names = ['Large Needle Driver', 'Prograsp Forceps', 'Monopolar Curved Scissors']\n    for i, (name, map_val) in enumerate(zip(class_names, metrics.box.maps)):\n        print(f\"  {name}: {map_val:.4f}\")",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "code",
+   "source": "# ===== LIST OUTPUT FILES =====\n\nprint(\"Output files available for download:\")\n!ls -lh {OUTPUT_DIR}/yolov8m_kaggle_run/weights/",
+   "metadata": {"trusted": true},
+   "outputs": [],
+   "execution_count": null
+  },
+  {
+   "cell_type": "markdown",
+   "source": "## Key Changes Made:\n\n### 1. **Fixed data.yaml structure** (CRITICAL)\n```yaml\n# ❌ WRONG (your old version):\nnames:\n  0: Large Needle Driver\n  1: Prograsp Forceps\n  2: Monopolar Curved Scissors\n\n# ✅ CORRECT (new version):\nnames: ['Large Needle Driver', 'Prograsp Forceps', 'Monopolar Curved Scissors']\n```\n\n### 2. **Added class distribution check**\nVerifies your data is balanced before training starts\n\n### 3. **Added verification step**\nConfirms the YAML file is in correct format\n\n### 4. **Added per-class validation metrics**\nSo you can see if all 3 classes are being detected\n\n## Expected Results:\n\nWith the corrected data.yaml:\n- All 3 classes should train properly\n- Per-class mAP should be similar across classes\n- Confusion matrix should show diagonal dominance for all classes\n\n## If Still Only Detecting One Class:\n\n1. Check the class distribution output - if one class has >70% of boxes, you have data imbalance\n2. Look at per-class mAP - if two classes show 0.000, there's still a data/label issue\n3. Verify label files have class IDs 0, 1, 2 (not 1, 2, 3)",
+   "metadata": {}
+  }
+ ]
+}