{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":117876,"databundleVersionId":14198377,"sourceType":"competition"},{"sourceId":13923389,"sourceType":"datasetVersion","datasetId":8872462},{"sourceId":666051,"sourceType":"modelInstanceVersion","isSourceIdPinned":true,"modelInstanceId":504148,"modelId":519193}],"dockerImageVersionId":31192,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Environment setup","metadata":{}},{"cell_type":"code","source":"# Install Ultralytics\n!pip install -q \"ultralytics==8.2.103\"\n!pip uninstall -y ray ray[tune]\n!pip install ensemble_boxes","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Core libs\nimport os\nimport random\nfrom pathlib import Path\nimport shutil\nimport ast\n\nimport numpy as np\nimport pandas as pd\nfrom tqdm.auto import tqdm\n\n# Disable Weights & Biases everywhere\nos.environ[\"WANDB_DISABLED\"] = \"true\"\nos.environ[\"WANDB_MODE\"] = \"disabled\"\n\nfrom ensemble_boxes import weighted_boxes_fusion\n\nfrom ultralytics import YOLO, settings\n\nprint(\"Ultralytics version:\", YOLO.__module__.split('.')[0])\n\nsettings.update({\"wandb\": False})","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:58:08.691503Z","iopub.execute_input":"2025-11-29T19:58:08.691850Z","iopub.status.idle":"2025-11-29T19:58:10.125373Z","shell.execute_reply.started":"2025-11-29T19:58:08.691816Z","shell.execute_reply":"2025-11-29T19:58:10.124416Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Dataset","metadata":{}},{"cell_type":"markdown","source":"### 1. Prepare editable working copy of the dataset\n\nThe original competition data is stored under `/kaggle/input`. \nHere I copy it once into `/kaggle/working/CottonWeedDetection` so that I can safely modify labels and files (e.g. after 3LC corrections) without touching the read-only input directory.\n","metadata":{}},{"cell_type":"code","source":"SRC_ROOT = Path(\"/kaggle/input/the-3lc-cotton-weed-detection-challenge/cotton_weed_competition_dataset\")\nWORK_ROOT = Path(\"/kaggle/working/CottonWeedDetection\") # editable copy\n\nif not WORK_ROOT.exists():\n print(\"Creating working dataset...\")\n WORK_ROOT.mkdir(parents=True, exist_ok=True)\n\n for split in [\"train\", \"val\", \"test\"]:\n src_split = SRC_ROOT / split\n dst_split = WORK_ROOT / split\n if src_split.exists():\n print(f\"Copying {split} ...\")\n shutil.copytree(src_split, dst_split)\n else:\n print(f\"WARNING: {src_split} not found\")\nelse:\n print(\"Working copy already exists:\", WORK_ROOT)\n\nprint(\"SRC_ROOT:\", SRC_ROOT)\nprint(\"WORK_ROOT:\", WORK_ROOT)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T18:43:27.059502Z","iopub.execute_input":"2025-11-29T18:43:27.059927Z","iopub.status.idle":"2025-11-29T18:44:56.552055Z","shell.execute_reply.started":"2025-11-29T18:43:27.059898Z","shell.execute_reply":"2025-11-29T18:44:56.550806Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 2. Applying 3LC-Corrected Labels (YOLO Format)\n\nThis cell imports the high-fidelity, corrected bounding box labels from the 3LC dashboard (ref: 3LC Tutorial). The plant and weed identification was meticulously corrected by cross-referencing external weed science resources (Cornell, Missouri, Purdue). The code parses the bounding box JSON and converts it to the standard YOLO format ( ) for the training splits.","metadata":{}},{"cell_type":"code","source":"DATA_ROOT = Path(\"/kaggle/working/CottonWeedDetection\")\n\nCSV_BY_SPLIT = {\n \"train\": Path(\"/kaggle/input/3cl-cotton-weed-det-final-labels/train_labels.csv\"),\n \"val\": Path(\"/kaggle/input/3cl-cotton-weed-det-final-labels/val_labels.csv\"),\n}\n\nfor split, csv_path in CSV_BY_SPLIT.items():\n img_dir = DATA_ROOT / split / \"images\"\n labels_dir = DATA_ROOT / split / \"labels\"\n labels_dir.mkdir(parents=True, exist_ok=True)\n\n print(f\"\\n=== Processing {split.upper()} ===\")\n print(\"CSV :\", csv_path)\n print(\"Images dir :\", img_dir)\n print(\"Labels dir :\", labels_dir)\n\n df = pd.read_csv(csv_path)\n\n required_cols = {\"image\", \"width\", \"height\", \"bbs\", \"weight\"}\n missing = required_cols - set(df.columns)\n if missing:\n raise ValueError(f\"{csv_path} missing required cols: {missing}\")\n\n # Wipe old labels so we only keep 3LC-corrected ones\n for txt in labels_dir.glob(\"*.txt\"):\n txt.unlink()\n\n num_files = 0\n num_boxes = 0\n\n # One image per row\n for _, row in tqdm(df.iterrows(), total=len(df)):\n # Normalize to just filename so it matches dataset layout\n file_name = Path(row[\"image\"]).name\n img_path = img_dir / file_name\n if not img_path.exists():\n # Label for an image that is not in this split – skip\n continue\n\n bbs_raw = row[\"bbs\"]\n if pd.isna(bbs_raw):\n continue\n\n try:\n bbs_dict = ast.literal_eval(bbs_raw)\n except Exception as e:\n print(f\" [WARN] Could not parse bbs for {file_name}: {e}\")\n continue\n\n bb_list = bbs_dict.get(\"bb_list\", [])\n if not bb_list:\n continue\n\n lines = []\n for bb in bb_list:\n # x0,y0 are centre; x1,y1 are width,height (already normalized 0–1)\n cls_id = int(bb[\"label\"])\n xc = float(bb[\"x0\"])\n yc = float(bb[\"y0\"])\n w = float(bb[\"x1\"])\n h = float(bb[\"y1\"])\n\n # Clamp to [0, 1] just in case\n xc = min(max(xc, 0.0), 1.0)\n yc = min(max(yc, 0.0), 1.0)\n w = min(max(w, 0.0), 1.0)\n h = min(max(h, 0.0), 1.0)\n\n lines.append(f\"{cls_id} {xc:.6f} {yc:.6f} {w:.6f} {h:.6f}\")\n num_boxes += 1\n\n if lines:\n label_path = labels_dir / f\"{img_path.stem}.txt\"\n label_path.write_text(\"\\n\".join(lines) + \"\\n\")\n num_files += 1\n\n print(f\"[{split}] wrote {num_files} label files, total boxes = {num_boxes}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T18:44:56.553276Z","iopub.execute_input":"2025-11-29T18:44:56.553558Z","iopub.status.idle":"2025-11-29T18:44:56.851490Z","shell.execute_reply.started":"2025-11-29T18:44:56.553526Z","shell.execute_reply":"2025-11-29T18:44:56.850641Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport random\nimport numpy as np\n\ndef set_seed(seed: int = 42):\n os.environ[\"PYTHONHASHSEED\"] = str(seed)\n random.seed(seed)\n np.random.seed(seed)\n\n try:\n import torch\n torch.manual_seed(seed)\n torch.cuda.manual_seed_all(seed)\n torch.backends.cudnn.deterministic = True\n torch.backends.cudnn.benchmark = False\n except ImportError:\n pass \n\nset_seed(42) ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:39:49.412389Z","iopub.execute_input":"2025-11-29T19:39:49.416583Z","iopub.status.idle":"2025-11-29T19:39:49.442928Z","shell.execute_reply.started":"2025-11-29T19:39:49.416461Z","shell.execute_reply":"2025-11-29T19:39:49.441879Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"ROOT = Path(\"/kaggle/working/CottonWeedDetection\")\nprint(\"Dataset root:\", ROOT)\n\nyaml_path = Path(\"/kaggle/working/cotton_weed.yaml\")\n# for final model we also included validation set in training\nyaml_text = f\"\"\"\npath: {ROOT} \n\ntrain:\n - train/images \n - val/images \n\nval: val/images \ntest: test/images \n\nnc: 3\nnames:\n 0: carpetweed\n 1: morningglory\n 2: palmer_amaranth\n\"\"\"\n\nyaml_path.write_text(yaml_text.strip() + \"\\n\")\nprint(\"==== cotton_weed.yaml ====\\n\")\nprint(yaml_path.read_text())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:41:01.109294Z","iopub.execute_input":"2025-11-29T19:41:01.110111Z","iopub.status.idle":"2025-11-29T19:41:01.121601Z","shell.execute_reply.started":"2025-11-29T19:41:01.110041Z","shell.execute_reply":"2025-11-29T19:41:01.120381Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def count_images(p: Path):\n return sum(1 for f in p.rglob(\"*\") if f.suffix.lower() in {\".jpg\", \".jpeg\", \".png\", \".bmp\"})\n\nfor split in [\"train\", \"val\", \"test\"]:\n img_dir = ROOT / split / \"images\"\n lbl_dir = ROOT / split / \"labels\"\n print(f\"\\n== {split.upper()} ==\")\n print(\" images:\", count_images(img_dir), \"in\", img_dir)\n print(\" labels:\", len(list(lbl_dir.glob(\"*.txt\"))), \"in\", lbl_dir)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:42:16.376062Z","iopub.execute_input":"2025-11-29T19:42:16.376916Z","iopub.status.idle":"2025-11-29T19:42:16.392544Z","shell.execute_reply.started":"2025-11-29T19:42:16.376882Z","shell.execute_reply":"2025-11-29T19:42:16.391437Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Hyperparameter search using YOLOv8 ```tune()```","metadata":{}},{"cell_type":"code","source":"# import os\n# os.environ[\"WANDB_DISABLED\"] = \"true\"\n# os.environ[\"WANDB_MODE\"] = \"disabled\"\n\n# from ultralytics import settings, YOLO\n# settings.update({\"wandb\": False})\n\n# DATASET = \"cotton_weed.yaml\"\n\n# search_space = {\n# \"lr0\": (1e-5, 1e-1),\n# \"lrf\": (0.01, 1.0),\n# \"momentum\": (0.6, 0.98),\n# \"weight_decay\": (0.0, 0.001),\n# \"warmup_epochs\": (0.0, 5.0),\n# \"warmup_momentum\": (0.6, 0.98),\n# \"warmup_bias_lr\": (0.01, 0.5),\n# \"box\": (1.0, 12.0),\n# \"cls\": (0.1, 3.0),\n# \"dfl\": (0.5, 3.0),\n# \"dropout\": (0.0, 0.2),\n# \"hsv_h\": (0.0, 0.1),\n# \"hsv_s\": (0.0, 0.9),\n# \"hsv_v\": (0.0, 0.9),\n# \"degrees\": (0.0, 15.0),\n# \"translate\": (0.0, 0.3),\n# \"scale\": (0.0, 0.9),\n# \"shear\": (0.0, 10.0),\n# \"perspective\": (0.0, 0.001),\n# \"flipud\": (0.0, 0.2),\n# \"fliplr\": (0.0, 0.7),\n# \"bgr\": (0.0, 0.5),\n# \"mosaic\": (0.0, 1.0),\n# \"mixup\": (0.0, 0.7),\n# }\n\n# from pathlib import Path\n\n# def run_tune_and_get_yaml(model_path, optimizer_name, epochs=10, iterations=12, batch=32):\n# model = YOLO(model_path)\n# print(f\"\\n=== Tuning {optimizer_name}: {iterations} iters x {epochs} epochs, batch={batch} ===\\n\")\n# model.tune(\n# data=DATASET,\n# epochs=epochs,\n# iterations=iterations,\n# optimizer=optimizer_name,\n# batch=batch,\n# imgsz=640,\n# space=search_space,\n# plots=False,\n# save=False,\n# val=True,\n# project=\"cotton_weed\",\n# name=f\"tune_{optimizer_name.lower()}\",\n# )\n\n# # Pick the latest best_hyperparameters.yaml\n# candidates = list(Path(\"runs\").rglob(\"best_hyperparameters.yaml\"))\n# assert candidates, \"No best_hyperparameters.yaml found; check runs/\"\n# best_yaml = max(candidates, key=lambda p: p.stat().st_mtime)\n# print(\"Best hyperparameters yaml:\", best_yaml)\n# return str(best_yaml)\n\n# # Example: tune AdamW\n# best_yaml_adam = run_tune_and_get_yaml(\"yolov8n.pt\", optimizer_name=\"AdamW\",\n# epochs=40, iterations=10, batch=64)\n# print(\"AdamW best hyperparameters saved at:\", best_yaml_adam)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:14:53.637923Z","iopub.execute_input":"2025-11-29T19:14:53.638309Z","iopub.status.idle":"2025-11-29T19:14:53.643936Z","shell.execute_reply.started":"2025-11-29T19:14:53.638279Z","shell.execute_reply":"2025-11-29T19:14:53.643012Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Training diverse YOLOv8n models for the ensemble","metadata":{}},{"cell_type":"code","source":"# # # MODEL A\n# !yolo task=detect mode=train \\\n# model=yolov8n.pt \\\n# data=/kaggle/working/cotton_weed.yaml \\\n# epochs=200 \\\n# patience=50 \\\n# imgsz=640 \\\n# batch=64 \\\n# workers=4 \\\n# optimizer=AdamW \\\n# lr0=0.005 \\\n# lrf=0.011 \\\n# momentum=0.723 \\\n# weight_decay=0.0004 \\\n# warmup_epochs=3 \\\n# warmup_momentum=0.66 \\\n# warmup_bias_lr=0.096 \\\n# box=6.345 \\\n# cls=0.652 \\\n# dfl=1.512 \\\n# dropout=0.0 \\\n# mosaic=0.721 \\\n# close_mosaic=20 \\\n# hsv_h=0.015 \\\n# hsv_s=0.8 \\\n# hsv_v=0.3 \\\n# degrees=0.0 \\\n# translate=0.094 \\\n# scale=0.421 \\\n# shear=0.0 \\\n# perspective=0.0 \\\n# flipud=0.0 \\\n# fliplr=0.466 \\\n# bgr=0.0 \\\n# mixup=0.0 \\\n# seed=42\n\n\n# # # MODEL B ************************************\n# !yolo task=detect mode=train \\\n# model=yolov8n.pt \\\n# data=/kaggle/working/cotton_weed.yaml \\\n# epochs=200 patience=50 \\\n# imgsz=640 batch=64 workers=4 \\\n# optimizer=SGD \\\n# lr0=0.01 lrf=0.01 momentum=0.937 weight_decay=0.0005 \\\n# warmup_epochs=3 warmup_momentum=0.8 warmup_bias_lr=0.1 \\\n# box=7.5 cls=0.5 dfl=1.5 \\\n# dropout=0.0 \\\n# mosaic=1.0 close_mosaic=10 \\\n# hsv_h=0.02 hsv_s=0.8 hsv_v=0.45 \\\n# degrees=10.0 translate=0.10 scale=0.50 shear=2.0 perspective=0.0 \\\n# flipud=0.0 fliplr=0.5 bgr=0.0 mixup=0.0 \\\n# seed=0\n\n# # # MODEL C ***********************************************************************************************************\n# !yolo task=detect mode=train \\\n# model=yolov8n.pt \\\n# data=/kaggle/working/cotton_weed.yaml \\\n# epochs=150 patience=40 \\\n# imgsz=640 batch=64 workers=4 \\\n# optimizer=AdamW \\\n# lr0=0.003 lrf=0.01 momentum=0.8 weight_decay=0.0006 \\\n# warmup_epochs=2 warmup_momentum=0.7 warmup_bias_lr=0.08 \\\n# box=5.5 cls=1.0 dfl=1.4 \\\n# dropout=0.0 \\\n# mosaic=0.3 close_mosaic=5 \\\n# hsv_h=0.01 hsv_s=0.6 hsv_v=0.25 \\\n# degrees=0.0 translate=0.05 scale=0.25 shear=0.0 perspective=0.0 \\\n# flipud=0.0 fliplr=0.3 bgr=0.0 mixup=0.0 \\\n# seed=123\n\n# # MODEL D\n# !yolo task=detect mode=train \\\n# model=yolov8n.pt \\\n# data=/kaggle/working/cotton_weed.yaml \\\n# device=0,1 \\\n# epochs=600 patience=80 \\\n# imgsz=640 batch=128 workers=4 \\\n# optimizer=AdamW \\\n# lr0=0.004 lrf=0.01 momentum=0.80 weight_decay=0.0006 \\\n# warmup_epochs=5 warmup_momentum=0.70 warmup_bias_lr=0.10 \\\n# box=5.8 cls=1.2 dfl=1.5 \\\n# label_smoothing=0.05 \\\n# cos_lr=True \\\n# dropout=0.0 \\\n# mosaic=0.9 close_mosaic=30 \\\n# hsv_h=0.03 hsv_s=0.90 hsv_v=0.50 \\\n# degrees=10.0 translate=0.15 scale=0.60 shear=4.0 perspective=0.0005 \\\n# flipud=0.10 fliplr=0.60 bgr=0.0 \\\n# mixup=0.20 copy_paste=0.30 \\\n# seed=7\n\n\n# MODEL E\n# !yolo task=detect mode=train \\\n# model=yolov8n.pt \\\n# data=/kaggle/working/cotton_weed.yaml \\\n# device=0,1 \\\n# epochs=900 patience=120 \\\n# imgsz=640 batch=128 workers=4 \\\n# optimizer=SGD \\\n# lr0=0.009 lrf=0.01 momentum=0.94 weight_decay=0.0005 \\\n# warmup_epochs=3 warmup_momentum=0.85 warmup_bias_lr=0.10 \\\n# box=7.8 cls=0.45 dfl=1.6 \\\n# label_smoothing=0.02 \\\n# cos_lr=True \\\n# dropout=0.0 \\\n# mosaic=0.7 close_mosaic=40 \\\n# hsv_h=0.02 hsv_s=0.80 hsv_v=0.40 \\\n# degrees=5.0 translate=0.12 scale=0.70 shear=3.0 perspective=0.0005 \\\n# flipud=0.05 fliplr=0.50 bgr=0.0 \\\n# mixup=0.10 copy_paste=0.20 \\\n# seed=2025","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Prediction And Genrating Solution ","metadata":{}},{"cell_type":"code","source":"MODEL_PATHS = [\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_A.pt\",\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_B.pt\",\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_C.pt\",\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_D1.pt\",\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_D2.pt\",\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_D3.pt\",\n r\"/kaggle/input/3cl-cotton-weed-det-yolov8n-models/pytorch/default/1/yolov8n_model_E.pt\",\n]\n\n# Test dataset root (images/ + labels/)\nTEST_ROOT = r\"/kaggle/working/CottonWeedDetection/test\"\nTEST_IMAGES_DIR = os.path.join(TEST_ROOT, \"images\")\n\n# YOLO inference settings for raw predictions\nINFER_IMGSZ = 640\nINFER_CONF = 0.01 # very low, keep almost everything\nINFER_IOU = 0.9 # NMS IoU (just to remove exact dups per model)\n\n# WBF settings\nWBF_IOU_THR = 0.45\nWBF_SKIP_BOX_THR = 0.01\nWBF_WEIGHTS = [1, 1, 0.75, 0.5, 0.5, 1, 1] \n\n# Number of classes\nNUM_CLASSES = 3\n\nOUT_SUBMISSION_CSV = \"submission.csv\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T20:00:03.979484Z","iopub.execute_input":"2025-11-29T20:00:03.980317Z","iopub.status.idle":"2025-11-29T20:00:03.986115Z","shell.execute_reply.started":"2025-11-29T20:00:03.980285Z","shell.execute_reply":"2025-11-29T20:00:03.985249Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 1. Utils Functions","metadata":{}},{"cell_type":"code","source":"# =========================\n# 1. LOAD MODELS\n# =========================\n\ndef load_models(model_paths):\n \"\"\"Load YOLO models from given paths.\"\"\"\n models = []\n for p in model_paths:\n print(f\"Loading model: {p}\")\n models.append(YOLO(p))\n return models","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:44:39.294540Z","iopub.execute_input":"2025-11-29T19:44:39.295741Z","iopub.status.idle":"2025-11-29T19:44:39.301455Z","shell.execute_reply.started":"2025-11-29T19:44:39.295704Z","shell.execute_reply":"2025-11-29T19:44:39.300321Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =========================\n# 2. RAW PREDICTIONS (ALL MODELS, LOW CONF/IOU)\n# =========================\n\ndef get_raw_predictions(models, images_dir, imgsz=640, conf=0.001, iou=0.7):\n \"\"\"\n Run all models on all images in images_dir with low conf/IoU and\n return a DataFrame of raw predictions (normalized xyxy).\n Columns: image_id, model_idx, class_id, score, x1,y1,x2,y2\n \"\"\"\n image_paths = sorted(\n list(Path(images_dir).glob(\"*.jpg\")) + list(Path(images_dir).glob(\"*.png\"))\n )\n rows = []\n\n for img_path in tqdm(image_paths, desc=\"Running models on test images\"):\n image_id = img_path.stem\n for m_idx, model in enumerate(models):\n results = model.predict(\n source=str(img_path),\n conf=conf,\n iou=iou,\n imgsz=imgsz,\n verbose=False\n )[0]\n\n if results.boxes is None or results.boxes.shape[0] == 0:\n continue\n\n # Normalized xyxy coordinates in [0, 1]\n xyxyn = results.boxes.xyxyn.cpu().numpy()\n scores = results.boxes.conf.cpu().numpy()\n classes = results.boxes.cls.cpu().numpy().astype(int)\n\n for b, s, c in zip(xyxyn, scores, classes):\n rows.append(\n {\n \"image_id\": image_id,\n \"model_idx\": m_idx,\n \"class_id\": int(c),\n \"score\": float(s),\n \"x1\": float(b[0]),\n \"y1\": float(b[1]),\n \"x2\": float(b[2]),\n \"y2\": float(b[3]),\n }\n )\n\n df = pd.DataFrame(rows)\n return df","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:44:52.177809Z","iopub.execute_input":"2025-11-29T19:44:52.178197Z","iopub.status.idle":"2025-11-29T19:44:52.192021Z","shell.execute_reply.started":"2025-11-29T19:44:52.178157Z","shell.execute_reply":"2025-11-29T19:44:52.191023Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =========================\n# 3. APPLY WBF ON DATAFRAME\n# =========================\n\ndef apply_wbf_to_dataframe(df_preds, num_models, weights=None,\n iou_thr=0.55, skip_box_thr=0.01):\n \"\"\"\n Apply Weighted Boxes Fusion per image over predictions from multiple models.\n\n df_preds: DataFrame with columns [image_id, model_idx, class_id, score, x1,y1,x2,y2]\n Returns fused_df: DataFrame with columns [image_id, class_id, score, x1,y1,x2,y2]\n \"\"\"\n fused_rows = []\n\n for image_id, df_img in tqdm(df_preds.groupby(\"image_id\"), desc=\"Applying WBF\"):\n boxes_list = []\n scores_list = []\n labels_list = []\n\n # Prepare list entries for each model 0..num_models-1\n for m_idx in range(num_models):\n df_m = df_img[df_img[\"model_idx\"] == m_idx]\n if df_m.empty:\n boxes_list.append([])\n scores_list.append([])\n labels_list.append([])\n continue\n\n boxes = df_m[[\"x1\", \"y1\", \"x2\", \"y2\"]].values.tolist()\n scores = df_m[\"score\"].values.tolist()\n labels = df_m[\"class_id\"].values.tolist()\n\n boxes_list.append(boxes)\n scores_list.append(scores)\n labels_list.append(labels)\n\n if all(len(b) == 0 for b in boxes_list):\n # No boxes from any model\n continue\n\n fused_boxes, fused_scores, fused_labels = weighted_boxes_fusion(\n boxes_list,\n scores_list,\n labels_list,\n weights=weights,\n iou_thr=iou_thr,\n skip_box_thr=skip_box_thr,\n conf_type='box_and_model_avg'\n )\n\n for b, s, c in zip(fused_boxes, fused_scores, fused_labels):\n fused_rows.append(\n {\n \"image_id\": image_id,\n \"class_id\": int(c),\n \"score\": float(s),\n \"x1\": float(b[0]),\n \"y1\": float(b[1]),\n \"x2\": float(b[2]),\n \"y2\": float(b[3]),\n }\n )\n\n fused_df = pd.DataFrame(fused_rows)\n return fused_df","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:45:05.215380Z","iopub.execute_input":"2025-11-29T19:45:05.215724Z","iopub.status.idle":"2025-11-29T19:45:05.229452Z","shell.execute_reply.started":"2025-11-29T19:45:05.215697Z","shell.execute_reply":"2025-11-29T19:45:05.228405Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =========================\n# 5. BUILD SUBMISSION CSV\n# =========================\n\ndef build_submission_csv(df_preds, images_dir, out_csv_path):\n \"\"\"\n Convert final predictions (normalized xyxy) into Kaggle submission format:\n image_id,prediction_string\n \"\"\"\n image_paths = sorted(\n list(Path(images_dir).glob(\"*.jpg\")) + list(Path(images_dir).glob(\"*.png\"))\n )\n img_ids = [p.stem for p in image_paths]\n\n grouped = {k: v for k, v in df_preds.groupby(\"image_id\")} if not df_preds.empty else {}\n\n rows = []\n for img_id in img_ids:\n if img_id not in grouped:\n rows.append({\"image_id\": img_id, \"prediction_string\": \"no box\"})\n continue\n\n df_img = grouped[img_id]\n parts = []\n for row in df_img.itertuples(index=False):\n x1, y1, x2, y2 = row.x1, row.y1, row.x2, row.y2\n w = max(0.0, x2 - x1)\n h = max(0.0, y2 - y1)\n xc = x1 + w / 2.0\n yc = y1 + h / 2.0\n\n xc = min(max(xc, 0.0), 1.0)\n yc = min(max(yc, 0.0), 1.0)\n w = min(max(w, 0.0), 1.0)\n h = min(max(h, 0.0), 1.0)\n\n parts.extend([\n str(int(row.class_id)),\n f\"{row.score:.6f}\",\n f\"{xc:.6f}\",\n f\"{yc:.6f}\",\n f\"{w:.6f}\",\n f\"{h:.6f}\",\n ])\n\n pred_str = \" \".join(parts) if parts else \"no box\"\n rows.append({\"image_id\": img_id, \"prediction_string\": pred_str})\n\n sub_df = pd.DataFrame(rows)\n sub_df.to_csv(str(out_csv_path), index=False)\n print(f\"Saved submission CSV to: {out_csv_path}\")\n return sub_df","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:45:29.461919Z","iopub.execute_input":"2025-11-29T19:45:29.462255Z","iopub.status.idle":"2025-11-29T19:45:29.474749Z","shell.execute_reply.started":"2025-11-29T19:45:29.462233Z","shell.execute_reply":"2025-11-29T19:45:29.473587Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### 2. Prediction","metadata":{}},{"cell_type":"code","source":"# 1) load models\nmodels = load_models(MODEL_PATHS)\n\n# 2) raw predictions from all models on test set\nraw_df = get_raw_predictions(\n models,\n TEST_IMAGES_DIR,\n imgsz=INFER_IMGSZ,\n conf=INFER_CONF,\n iou=INFER_IOU,\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:47:07.273786Z","iopub.execute_input":"2025-11-29T19:47:07.274990Z","iopub.status.idle":"2025-11-29T19:53:10.458639Z","shell.execute_reply.started":"2025-11-29T19:47:07.274955Z","shell.execute_reply":"2025-11-29T19:53:10.457704Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# 3) WBF fusion\nfused_df = apply_wbf_to_dataframe(\n raw_df,\n num_models=len(models),\n weights=WBF_WEIGHTS,\n iou_thr=WBF_IOU_THR,\n skip_box_thr=WBF_SKIP_BOX_THR,\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T19:58:31.053898Z","iopub.execute_input":"2025-11-29T19:58:31.054552Z","iopub.status.idle":"2025-11-29T19:58:33.669053Z","shell.execute_reply.started":"2025-11-29T19:58:31.054520Z","shell.execute_reply":"2025-11-29T19:58:33.668260Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# 5) build submission CSV (for Kaggle)\n_ = build_submission_csv(fused_df, TEST_IMAGES_DIR, OUT_SUBMISSION_CSV)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T20:00:08.054955Z","iopub.execute_input":"2025-11-29T20:00:08.055299Z","iopub.status.idle":"2025-11-29T20:00:08.170307Z","shell.execute_reply.started":"2025-11-29T20:00:08.055271Z","shell.execute_reply":"2025-11-29T20:00:08.168982Z"}},"outputs":[],"execution_count":null}]}