{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":126118,"databundleVersionId":14933142,"sourceType":"competition"}],"dockerImageVersionId":31260,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"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","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!pip install -q rasterio timm albumentations\n!pip install -q scikit-learn matplotlib pandas numpy\n\nprint(\"āœ… All libraries installed!\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T21:04:38.342185Z","iopub.execute_input":"2026-02-15T21:04:38.342456Z","iopub.status.idle":"2026-02-15T21:04:48.371253Z","shell.execute_reply.started":"2026-02-15T21:04:38.342426Z","shell.execute_reply":"2026-02-15T21:04:48.370284Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport matplotlib.gridspec as gridspec\nfrom pathlib import Path\nimport warnings\nimport time\nwarnings.filterwarnings('ignore')\n\ntry:\n import rasterio\nexcept ImportError:\n os.system('pip install rasterio')\n import rasterio\n\n# ===== PATHS =====\nBASE_PATH = '/kaggle/input/beyond-visible-spectrum-ai-for-agriculture-2026p2'\nTRAIN_DIR = os.path.join(BASE_PATH, 'ICPR02', 'kaggle')\nEVAL_DIR = os.path.join(BASE_PATH, 'ICPR02', 'kaggle', 'evaluation')\nARCHIVE_DIR = os.path.join(BASE_PATH, 'archive', 'share')\nCLASS_NAMES = ['Aphid', 'Blast', 'RPH', 'Rust']\n\nprint(\"šŸ“‚ Checking all paths:\")\nfor name, path in [('TRAIN_DIR', TRAIN_DIR), ('EVAL_DIR', EVAL_DIR), ('ARCHIVE_DIR', ARCHIVE_DIR)]:\n exists = os.path.exists(path)\n count = len(os.listdir(path)) if exists else 0\n print(f\" {'āœ…' if exists else 'āŒ'} {name}: {path} ({count} items)\")\n\nprint(f\"\\nšŸ“ Training class folders:\")\nif os.path.exists(TRAIN_DIR):\n for item in sorted(os.listdir(TRAIN_DIR)):\n full = os.path.join(TRAIN_DIR, item)\n if os.path.isdir(full):\n print(f\" šŸ“ {item}/ ā†’ {len(os.listdir(full))} samples\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T22:19:47.374267Z","iopub.execute_input":"2026-02-15T22:19:47.374585Z","iopub.status.idle":"2026-02-15T22:19:47.391111Z","shell.execute_reply.started":"2026-02-15T22:19:47.374557Z","shell.execute_reply":"2026-02-15T22:19:47.390478Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def print_tree(path, prefix=\"\", max_depth=3, depth=0, max_items=8):\n if depth >= max_depth: return\n try: items = sorted(os.listdir(path))\n except: return\n for i, item in enumerate(items[:max_items]):\n full = os.path.join(path, item)\n last = (i == min(len(items), max_items) - 1)\n conn = \"ā””ā”€ā”€ \" if last else \"ā”œā”€ā”€ \"\n if os.path.isdir(full):\n print(f\"{prefix}{conn}šŸ“ {item}/ ({len(os.listdir(full))} items)\")\n print_tree(full, prefix + (\" \" if last else \"ā”‚ \"), max_depth, depth+1, max_items)\n else:\n print(f\"{prefix}{conn}šŸ“„ {item} ({os.path.getsize(full):,} bytes)\")\n if len(items) > max_items:\n print(f\"{prefix}ā””ā”€ā”€ ... +{len(items)-max_items} more\")\n\nprint(\"=\" * 60)\nprint(\"šŸ“‚ FULL DATA STRUCTURE\")\nprint(\"=\" * 60)\nprint_tree(BASE_PATH, max_depth=4)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T21:07:04.906662Z","iopub.execute_input":"2026-02-15T21:07:04.907296Z","iopub.status.idle":"2026-02-15T21:07:05.263619Z","shell.execute_reply.started":"2026-02-15T21:07:04.907268Z","shell.execute_reply":"2026-02-15T21:07:05.263015Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"BAND_NAMES = ['B1','B2','B3','B4','B5','B6','B7','B8','B8A','B9','B11','B12']\nBAND_INFO = {\n 'B1':('Coastal/Aerosol','60m','#9400D3'), 'B2':('Blue','10m','#0000FF'),\n 'B3':('Green','10m','#00FF00'), 'B4':('Red','10m','#FF0000'),\n 'B5':('Red Edge 1','20m','#FF4500'), 'B6':('Red Edge 2','20m','#FF6347'),\n 'B7':('Red Edge 3','20m','#FF7F50'), 'B8':('NIR','10m','#8B0000'),\n 'B8A':('NIR Narrow','20m','#A52A2A'), 'B9':('Water Vapour','60m','#00CED1'),\n 'B11':('SWIR 1','20m','#FFD700'), 'B12':('SWIR 2','20m','#FFA500'),\n}\n\ndef load_bands(folder_path):\n bands = {}\n for bn in BAND_NAMES:\n tif = os.path.join(folder_path, f\"{bn}.tif\")\n if os.path.exists(tif):\n with rasterio.open(tif) as src:\n bands[bn] = src.read(1).astype(np.float32)\n return bands\n\ndef normalize_band(band, pct=(2,98)):\n valid = band[band > 0]\n if len(valid) == 0: return np.zeros_like(band, dtype=np.float32)\n lo, hi = np.percentile(valid, pct)\n if hi == lo: return np.zeros_like(band, dtype=np.float32)\n return np.clip((band - lo) / (hi - lo), 0, 1)\n\ndef load_any_image(path):\n \"\"\"Load image from path (folder with .tif bands, single .tif, or image file).\"\"\"\n if os.path.isdir(path):\n bands = load_bands(path)\n if all(b in bands for b in ['B4','B3','B2']):\n return np.stack([normalize_band(bands['B4']),\n normalize_band(bands['B3']),\n normalize_band(bands['B2'])], axis=-1), 'rgb', bands\n elif bands:\n first = list(bands.values())[0]\n return normalize_band(first), 'gray', bands\n else:\n fpath = path if os.path.isfile(path) else None\n if fpath and fpath.endswith(('.tif','.tiff')):\n with rasterio.open(fpath) as src:\n data = src.read().astype(np.float32)\n if data.shape[0] >= 3:\n return np.stack([normalize_band(data[2]),\n normalize_band(data[1]),\n normalize_band(data[0])], axis=-1), 'rgb', None\n else:\n return normalize_band(data[0]), 'gray', None\n elif fpath and fpath.endswith(('.png','.jpg','.jpeg')):\n from PIL import Image\n return np.array(Image.open(fpath)), 'rgb', None\n return None, None, None\n\n# --- Class Distribution ---\nprint(\"šŸ·ļø LABELED DATASET\")\nprint(\"=\" * 60)\nclass_counts = {}\nfor cls in CLASS_NAMES:\n d = os.path.join(TRAIN_DIR, cls)\n class_counts[cls] = len(os.listdir(d)) if os.path.exists(d) else 0\n\nfig, axes = plt.subplots(1, 2, figsize=(14, 5))\ncolors = ['#FF6B6B', '#4ECDC4', '#45B7D1', '#96CEB4']\nnames, counts = list(class_counts.keys()), list(class_counts.values())\n\nbars = axes[0].bar(names, counts, color=colors, edgecolor='black', lw=0.5)\naxes[0].set_title('Class Distribution', fontsize=13, fontweight='bold')\naxes[0].set_ylabel('Samples')\nfor b, c in zip(bars, counts):\n axes[0].text(b.get_x()+b.get_width()/2, b.get_height()+1, str(c), ha='center', fontweight='bold')\n\naxes[1].pie(counts, labels=names, autopct='%1.1f%%', colors=colors, startangle=90, explode=[0.05]*4)\naxes[1].set_title('Class Distribution (%)', fontsize=13, fontweight='bold')\nplt.suptitle('ļæ½ Crop Disease Dataset', fontsize=15, fontweight='bold')\nplt.tight_layout(); plt.savefig('/kaggle/working/class_dist.png', dpi=150, bbox_inches='tight'); plt.show()\n\n# --- Sample Images Per Class ---\nn_show = 4\nfig, axes = plt.subplots(len(CLASS_NAMES), n_show, figsize=(4*n_show, 4*len(CLASS_NAMES)))\n\nfor row, cls in enumerate(CLASS_NAMES):\n cls_dir = os.path.join(TRAIN_DIR, cls)\n samples = sorted(os.listdir(cls_dir))[:n_show] if os.path.exists(cls_dir) else []\n axes[row, 0].set_ylabel(cls, fontsize=14, fontweight='bold', rotation=0, labelpad=60)\n\n for col in range(n_show):\n ax = axes[row, col]\n if col < len(samples):\n img, mode, _ = load_any_image(os.path.join(cls_dir, samples[col]))\n if img is not None:\n ax.imshow(img, cmap='gray' if mode == 'gray' else None)\n ax.set_title(samples[col], fontsize=7)\n ax.axis('off')\n\nplt.suptitle('šŸ·ļø Sample Images by Disease Class', fontsize=14, fontweight='bold', y=1.02)\nplt.tight_layout(); plt.savefig('/kaggle/working/labeled_samples.png', dpi=150, bbox_inches='tight'); plt.show()\n\n# --- Band Details for One Sample ---\nprint(\"\\nšŸ“Š Inspecting one sample in detail:\")\nfor cls in CLASS_NAMES:\n cls_dir = os.path.join(TRAIN_DIR, cls)\n if os.path.exists(cls_dir):\n sample = sorted(os.listdir(cls_dir))[0]\n sp = os.path.join(cls_dir, sample)\n img, mode, bands = load_any_image(sp)\n print(f\"\\n Class: {cls} | Sample: {sample}\")\n if os.path.isdir(sp):\n print(f\" Type: Folder with files: {os.listdir(sp)[:5]}\")\n else:\n print(f\" Type: File ({os.path.getsize(sp):,} bytes)\")\n if bands:\n for bn, bd in bands.items():\n info = BAND_INFO.get(bn, ('?','?','#000'))\n print(f\" {bn:5s} {info[0]:18s} {info[1]:5s} shape={bd.shape} \"\n f\"min={bd.min():.1f} max={bd.max():.1f} mean={bd.mean():.1f}\")\n break\n\n# --- Evaluation Data Peek ---\nprint(\"\\nšŸ“‹ EVALUATION (Test) DATA:\")\nif os.path.exists(EVAL_DIR):\n items = sorted(os.listdir(EVAL_DIR))\n print(f\" {len(items)} items\")\n print(f\" First 5: {items[:5]}\")\n first = os.path.join(EVAL_DIR, items[0])\n if os.path.isdir(first):\n print(f\" Inside '{items[0]}': {os.listdir(first)[:5]}\")\nelse:\n print(\" āŒ Not found\")\n\nprint(\"\\nāœ… Data exploration complete!\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T21:08:32.532336Z","iopub.execute_input":"2026-02-15T21:08:32.532980Z","iopub.status.idle":"2026-02-15T21:08:40.216693Z","shell.execute_reply.started":"2026-02-15T21:08:32.532952Z","shell.execute_reply":"2026-02-15T21:08:40.216094Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\nfrom sklearn.model_selection import StratifiedKFold\nfrom sklearn.metrics import classification_report, accuracy_score, f1_score\nimport timm\nimport rasterio\nimport numpy as np\nimport pandas as pd\nimport os, time\n\nDEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nprint(f\"šŸ–„ļø Device: {DEVICE}\")\nif torch.cuda.is_available():\n print(f\" GPU: {torch.cuda.get_device_name(0)}\")\n print(f\" VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB\")\n\n# ===== PATHS =====\nBASE_PATH = '/kaggle/input/beyond-visible-spectrum-ai-for-agriculture-2026p2'\nTRAIN_DIR = os.path.join(BASE_PATH, 'ICPR02', 'kaggle')\nEVAL_DIR = os.path.join(BASE_PATH, 'ICPR02', 'kaggle', 'evaluation')\n\nBAND_NAMES = ['B1','B2','B3','B4','B5','B6','B7','B8','B8A','B9','B11','B12']\nNUM_BANDS = 12\nIMG_SIZE = 64 # Start small for stable training (increase later)\nCLASS_NAMES = ['Aphid', 'Blast', 'RPH', 'Rust']\nCLASS_MAP = {n: i for i, n in enumerate(CLASS_NAMES)}\nINV_MAP = {i: n for n, i in CLASS_MAP.items()}\n\n\ndef build_labeled_data():\n \"\"\"Build (file_paths, labels) from class folders.\"\"\"\n paths, labels = [], []\n for cls in CLASS_NAMES:\n cls_dir = os.path.join(TRAIN_DIR, cls)\n if not os.path.exists(cls_dir):\n print(f\" āš ļø Missing: {cls_dir}\")\n continue\n for item in sorted(os.listdir(cls_dir)):\n paths.append(os.path.join(cls_dir, item))\n labels.append(CLASS_MAP[cls])\n labels = np.array(labels)\n print(f\"šŸ“Š Labeled data: {len(paths)} samples\")\n for cls in CLASS_NAMES:\n print(f\" {cls}: {(labels == CLASS_MAP[cls]).sum()}\")\n return paths, labels\n\n\ndef build_test_data():\n \"\"\"Build file paths for evaluation folder.\"\"\"\n paths, ids = [], []\n if not os.path.exists(EVAL_DIR):\n print(f\"āš ļø Eval dir missing: {EVAL_DIR}\")\n return paths, ids\n for item in sorted(os.listdir(EVAL_DIR)):\n paths.append(os.path.join(EVAL_DIR, item))\n ids.append(item)\n print(f\"šŸ“Š Test data: {len(paths)} samples\")\n return paths, ids\n\n\nclass SentinelDataset(Dataset):\n \"\"\"Loads Sentinel-2 images (folder-of-bands or single-file).\"\"\"\n def __init__(self, file_paths, labels=None, img_size=IMG_SIZE, augment=False):\n self.paths = file_paths\n self.labels = labels\n self.img_size = img_size\n self.augment = augment\n\n def _load(self, path):\n bands = []\n if os.path.isdir(path):\n for bn in BAND_NAMES:\n tif = os.path.join(path, f\"{bn}.tif\")\n if os.path.exists(tif):\n with rasterio.open(tif) as src:\n bands.append(src.read(1).astype(np.float32))\n else:\n bands.append(None)\n else:\n fpath = path if os.path.isfile(path) else path + '.tif'\n if os.path.exists(fpath):\n with rasterio.open(fpath) as src:\n data = src.read().astype(np.float32)\n for i in range(data.shape[0]):\n bands.append(data[i])\n while len(bands) < NUM_BANDS:\n bands.append(None)\n\n result = []\n for b in bands[:NUM_BANDS]:\n if b is not None:\n t = torch.from_numpy(b).unsqueeze(0).unsqueeze(0)\n t = F.interpolate(t, size=(self.img_size, self.img_size),\n mode='bilinear', align_corners=False)\n result.append(t.squeeze())\n else:\n result.append(torch.zeros(self.img_size, self.img_size))\n\n img = torch.stack(result, dim=0) # (NUM_BANDS, H, W)\n\n # Normalize each band using percentile clipping (more robust than min-max)\n for i in range(img.shape[0]):\n b = img[i]\n valid = b[b > 0]\n if len(valid) > 0:\n lo = torch.quantile(valid, 0.02)\n hi = torch.quantile(valid, 0.98)\n if hi > lo:\n img[i] = torch.clamp((b - lo) / (hi - lo), 0.0, 1.0)\n else:\n img[i] = torch.zeros_like(b)\n else:\n img[i] = torch.zeros_like(b)\n return img\n\n def __len__(self):\n return len(self.paths)\n\n def __getitem__(self, idx):\n img = self._load(self.paths[idx])\n if self.augment:\n if torch.rand(1) > 0.5: img = torch.flip(img, [2])\n if torch.rand(1) > 0.5: img = torch.flip(img, [1])\n k = torch.randint(0, 4, (1,)).item()\n img = torch.rot90(img, k, [1, 2])\n if self.labels is not None:\n return img, self.labels[idx]\n return img\n\n\n# --- SANITY CHECK: Load one sample and print stats ---\nprint(\"āœ… Utilities ready!\")\ntrain_paths, train_labels = build_labeled_data()\ntest_paths, test_ids = build_test_data()\n\nprint(\"\\nšŸ” Sanity check ā€” loading first sample:\")\n_ds = SentinelDataset(train_paths[:1])\n_img = _ds[0]\nprint(f\" Shape: {_img.shape}\")\nprint(f\" Dtype: {_img.dtype}\")\nfor i, bn in enumerate(BAND_NAMES):\n b = _img[i]\n print(f\" {bn:5s}: min={b.min():.3f} max={b.max():.3f} mean={b.mean():.3f} nonzero={(b>0).sum().item()}\")\ndel _ds, _img\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T22:17:23.953436Z","iopub.execute_input":"2026-02-15T22:17:23.954173Z","iopub.status.idle":"2026-02-15T22:17:24.059018Z","shell.execute_reply.started":"2026-02-15T22:17:23.954144Z","shell.execute_reply":"2026-02-15T22:17:24.058274Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class CropDiseaseModel(nn.Module):\n \"\"\"\n Pre-trained ImageNet backbone adapted for 12-channel satellite input.\n timm's in_chans parameter handles all architectures automatically.\n \"\"\"\n def __init__(self, backbone='resnet50', num_classes=4, pretrained=True):\n super().__init__()\n self.model = timm.create_model(\n backbone,\n pretrained=pretrained,\n num_classes=num_classes,\n in_chans=NUM_BANDS\n )\n print(f\"āœ… {backbone}: {NUM_BANDS}-ch input, {num_classes} classes (pretrained={pretrained})\")\n\n def forward(self, x):\n return self.model(x)\n\n\n# Quick sanity test\n_m = CropDiseaseModel('resnet50', pretrained=False)\n_d = torch.randn(2, NUM_BANDS, IMG_SIZE, IMG_SIZE)\n_o = _m(_d)\nprint(f\"āœ… Model test: {_d.shape} ā†’ {_o.shape}\")\ndel _m, _d, _o","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T22:18:42.232604Z","iopub.execute_input":"2026-02-15T22:18:42.232958Z","iopub.status.idle":"2026-02-15T22:18:42.579548Z","shell.execute_reply.started":"2026-02-15T22:18:42.232926Z","shell.execute_reply":"2026-02-15T22:18:42.578746Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"BACKBONES = {\n 'resnet50': {'img_size': 64, 'bs': 32, 'lr': 3e-4},\n 'efficientnet_b2': {'img_size': 64, 'bs': 32, 'lr': 3e-4},\n 'convnext_tiny': {'img_size': 64, 'bs': 16, 'lr': 1e-4},\n 'swin_tiny_patch4_window7_224': {'img_size': 224, 'bs': 8, 'lr': 1e-4},\n}\n\nCOMPARE_EPOCHS = 20\nCOMPARE_FOLDS = 3\n\nbackbone_results = {}\n\nfor bb_name, bb_cfg in BACKBONES.items():\n print(f\"\\n{'šŸ”·'*25}\")\n print(f\"šŸ—ļø BACKBONE: {bb_name}\")\n print(f\" img_size={bb_cfg['img_size']} | bs={bb_cfg['bs']} | lr={bb_cfg['lr']}\")\n print(f\"{'šŸ”·'*25}\")\n\n skf = StratifiedKFold(n_splits=COMPARE_FOLDS, shuffle=True, random_state=42)\n fold_accs, fold_f1s = [], []\n start_time = time.time()\n\n for fold, (tr_idx, va_idx) in enumerate(skf.split(train_paths, train_labels)):\n print(f\"\\n šŸ“‚ Fold {fold+1}/{COMPARE_FOLDS}\")\n\n tr_paths_f = [train_paths[i] for i in tr_idx]\n va_paths_f = [train_paths[i] for i in va_idx]\n tr_labels_f = train_labels[tr_idx]\n va_labels_f = train_labels[va_idx]\n\n tr_loader = DataLoader(\n SentinelDataset(tr_paths_f, tr_labels_f, img_size=bb_cfg['img_size'], augment=True),\n batch_size=bb_cfg['bs'], shuffle=True, num_workers=2, pin_memory=True)\n va_loader = DataLoader(\n SentinelDataset(va_paths_f, va_labels_f, img_size=bb_cfg['img_size']),\n batch_size=bb_cfg['bs'], num_workers=2, pin_memory=True)\n\n try:\n model = CropDiseaseModel(bb_name, num_classes=4, pretrained=True).to(DEVICE)\n except Exception as e:\n print(f\" āŒ Failed: {e}\")\n break\n\n # Class weights\n counts = np.bincount(tr_labels_f, minlength=4).astype(np.float32)\n weights = torch.tensor(counts.sum() / (4 * counts + 1e-6)).to(DEVICE)\n criterion = nn.CrossEntropyLoss(weight=weights, label_smoothing=0.1)\n\n # Single optimizer ā€” same LR for all params (simpler, more stable)\n optimizer = torch.optim.AdamW(model.parameters(), lr=bb_cfg['lr'], weight_decay=1e-4)\n\n # Warmup + Cosine schedule\n warmup_epochs = 3\n def lr_lambda(epoch):\n if epoch < warmup_epochs:\n return (epoch + 1) / warmup_epochs\n return 0.5 * (1 + np.cos(np.pi * (epoch - warmup_epochs) / (COMPARE_EPOCHS - warmup_epochs)))\n scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)\n\n best_acc, best_f1 = 0, 0\n for ep in range(COMPARE_EPOCHS):\n model.train()\n ep_loss = 0\n for imgs, lbls in tr_loader:\n imgs, lbls = imgs.to(DEVICE), lbls.to(DEVICE)\n loss = criterion(model(imgs), lbls)\n optimizer.zero_grad()\n loss.backward()\n torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)\n optimizer.step()\n ep_loss += loss.item()\n scheduler.step()\n\n model.eval()\n vp, vt = [], []\n with torch.no_grad():\n for imgs, lbls in va_loader:\n preds = model(imgs.to(DEVICE)).argmax(1).cpu().numpy()\n vp.extend(preds); vt.extend(lbls.numpy())\n acc = accuracy_score(vt, vp)\n f1 = f1_score(vt, vp, average='macro')\n\n if acc > best_acc:\n best_acc = acc\n best_f1 = f1\n\n if (ep+1) % 5 == 0:\n lr_now = optimizer.param_groups[0]['lr']\n print(f\" Ep {ep+1:2d}/{COMPARE_EPOCHS} | Loss: {ep_loss/len(tr_loader):.4f} | \"\n f\"Acc: {acc:.4f} | F1: {f1:.4f} | LR: {lr_now:.1e}\")\n\n fold_accs.append(best_acc)\n fold_f1s.append(best_f1)\n print(f\" āœ… Fold {fold+1} ā†’ Acc: {best_acc:.4f} | F1: {best_f1:.4f}\")\n del model; torch.cuda.empty_cache()\n\n elapsed = time.time() - start_time\n if fold_accs:\n backbone_results[bb_name] = {\n 'mean_acc': np.mean(fold_accs),\n 'std_acc': np.std(fold_accs),\n 'mean_f1': np.mean(fold_f1s),\n 'time_min': elapsed / 60,\n 'cfg': bb_cfg,\n }\n\n# --- Comparison Table ---\nprint(f\"\\n\\n{'='*70}\")\nprint(f\"šŸ“Š BACKBONE COMPARISON ({COMPARE_FOLDS}-Fold, {COMPARE_EPOCHS} epochs)\")\nprint(f\"{'='*70}\")\nprint(f\" {'Backbone':<35} {'Acc':>10} {'F1':>10} {'Time':>10}\")\nprint(\"-\" * 70)\nbest_bb, best_score = None, 0\nfor name, res in sorted(backbone_results.items(), key=lambda x: -x[1]['mean_f1']):\n print(f\" {name:<33} {res['mean_acc']:.4f}Ā±{res['std_acc']:.3f}\"\n f\" {res['mean_f1']:.4f} {res['time_min']:.1f} min\")\n if res['mean_f1'] > best_score:\n best_score = res['mean_f1']\n best_bb = name\n\nprint(f\"\\nšŸ† BEST: {best_bb} (F1: {best_score:.4f})\")\n\n# Bar chart\nimport matplotlib.pyplot as plt\nfig, axes = plt.subplots(1, 2, figsize=(14, 5))\nnames = list(backbone_results.keys())\naccs = [backbone_results[n]['mean_acc'] for n in names]\nf1s = [backbone_results[n]['mean_f1'] for n in names]\ncolors = ['#FF6B6B', '#4ECDC4', '#45B7D1', '#96CEB4']\n\nfor ax, vals, title in [(axes[0], accs, 'Accuracy'), (axes[1], f1s, 'F1 Score')]:\n bars = ax.bar(range(len(names)), vals, color=colors[:len(names)], edgecolor='black', lw=0.5)\n ax.set_xticks(range(len(names)))\n ax.set_xticklabels(names, rotation=25, ha='right', fontsize=8)\n ax.set_ylabel(title); ax.set_title(f'{title} by Backbone', fontweight='bold')\n for i, v in enumerate(vals): ax.text(i, v+0.005, f'{v:.3f}', ha='center', fontsize=9)\nplt.suptitle('šŸ—ļø Backbone Comparison', fontsize=14, fontweight='bold')\nplt.tight_layout()\nplt.savefig('/kaggle/working/backbone_comparison.png', dpi=150, bbox_inches='tight')\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-15T22:19:59.766671Z","iopub.execute_input":"2026-02-15T22:19:59.767341Z","iopub.status.idle":"2026-02-16T00:21:40.192217Z","shell.execute_reply.started":"2026-02-15T22:19:59.767313Z","shell.execute_reply":"2026-02-16T00:21:40.191496Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"BEST_BB = best_bb\nBEST_CFG = backbone_results[BEST_BB]['cfg']\nFT_EPOCHS = 40\nN_FOLDS = 5\n\nprint(f\"šŸš€ Full training: {BEST_BB}\")\nprint(f\" img_size={BEST_CFG['img_size']} | bs={BEST_CFG['bs']} | lr={BEST_CFG['lr']}\")\n\nskf = StratifiedKFold(n_splits=N_FOLDS, shuffle=True, random_state=42)\nfold_scores = []\nall_val_preds, all_val_true = [], []\n\nfor fold, (tr_idx, va_idx) in enumerate(skf.split(train_paths, train_labels)):\n print(f\"\\n{'='*50}\\nšŸ“‚ FOLD {fold+1}/{N_FOLDS}\\n{'='*50}\")\n\n tr_paths_f = [train_paths[i] for i in tr_idx]\n va_paths_f = [train_paths[i] for i in va_idx]\n tr_labels_f = train_labels[tr_idx]\n va_labels_f = train_labels[va_idx]\n\n tr_loader = DataLoader(\n SentinelDataset(tr_paths_f, tr_labels_f, img_size=BEST_CFG['img_size'], augment=True),\n batch_size=BEST_CFG['bs'], shuffle=True, num_workers=2, pin_memory=True)\n va_loader = DataLoader(\n SentinelDataset(va_paths_f, va_labels_f, img_size=BEST_CFG['img_size']),\n batch_size=BEST_CFG['bs'], num_workers=2, pin_memory=True)\n\n model = CropDiseaseModel(BEST_BB, num_classes=4, pretrained=True).to(DEVICE)\n\n counts = np.bincount(tr_labels_f, minlength=4).astype(np.float32)\n weights = torch.tensor(counts.sum() / (4 * counts + 1e-6)).to(DEVICE)\n criterion = nn.CrossEntropyLoss(weight=weights, label_smoothing=0.1)\n\n optimizer = torch.optim.AdamW(model.parameters(), lr=BEST_CFG['lr'], weight_decay=1e-4)\n warmup_epochs = 5\n def lr_lambda(epoch):\n if epoch < warmup_epochs:\n return (epoch + 1) / warmup_epochs\n return 0.5 * (1 + np.cos(np.pi * (epoch - warmup_epochs) / (FT_EPOCHS - warmup_epochs)))\n scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)\n\n best_acc = 0\n for ep in range(FT_EPOCHS):\n model.train()\n train_loss = 0\n for imgs, lbls in tr_loader:\n imgs, lbls = imgs.to(DEVICE), lbls.to(DEVICE)\n loss = criterion(model(imgs), lbls)\n optimizer.zero_grad()\n loss.backward()\n torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)\n optimizer.step()\n train_loss += loss.item()\n scheduler.step()\n\n model.eval()\n vp, vt = [], []\n with torch.no_grad():\n for imgs, lbls in va_loader:\n preds = model(imgs.to(DEVICE)).argmax(1).cpu().numpy()\n vp.extend(preds); vt.extend(lbls.numpy())\n\n acc = accuracy_score(vt, vp)\n f1 = f1_score(vt, vp, average='macro')\n\n if acc > best_acc:\n best_acc = acc\n torch.save(model.state_dict(), f'/kaggle/working/best_fold{fold}.pth')\n\n if (ep+1) % 5 == 0 or ep == 0:\n print(f\" Ep {ep+1:2d}/{FT_EPOCHS} | Loss: {train_loss/len(tr_loader):.4f} | \"\n f\"Acc: {acc:.4f} | F1: {f1:.4f} | Best: {best_acc:.4f}\")\n\n fold_scores.append(best_acc)\n all_val_preds.extend(vp)\n all_val_true.extend(vt)\n print(f\"\\n āœ… Fold {fold+1} Best: {best_acc:.4f}\")\n print(classification_report(vt, vp, target_names=CLASS_NAMES))\n del model; torch.cuda.empty_cache()\n\nprint(f\"\\n{'='*50}\\nšŸ“Š FINAL CV RESULTS ({BEST_BB})\\n{'='*50}\")\nfor i, s in enumerate(fold_scores):\n print(f\" Fold {i+1}: {s:.4f}\")\nprint(f\" Mean: {np.mean(fold_scores):.4f} Ā± {np.std(fold_scores):.4f}\")\nprint(f\" Overall F1: {f1_score(all_val_true, all_val_preds, average='macro'):.4f}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-16T00:34:40.783801Z","iopub.execute_input":"2026-02-16T00:34:40.784503Z","iopub.status.idle":"2026-02-16T04:13:51.546953Z","shell.execute_reply.started":"2026-02-16T00:34:40.784474Z","shell.execute_reply":"2026-02-16T04:13:51.546049Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay\n\ncm = confusion_matrix(all_val_true, all_val_preds)\nfig, ax = plt.subplots(figsize=(8, 6))\ndisp = ConfusionMatrixDisplay(cm, display_labels=CLASS_NAMES)\ndisp.plot(ax=ax, cmap='Blues', values_format='d')\nax.set_title(f'Confusion Matrix ā€” {BEST_BB}', fontsize=14, fontweight='bold')\nplt.tight_layout()\nplt.savefig('/kaggle/working/confusion_matrix.png', dpi=150, bbox_inches='tight')\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-16T04:39:53.738385Z","iopub.execute_input":"2026-02-16T04:39:53.739178Z","iopub.status.idle":"2026-02-16T04:39:54.155067Z","shell.execute_reply.started":"2026-02-16T04:39:53.739140Z","shell.execute_reply":"2026-02-16T04:39:54.154352Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(\"šŸŽÆ GENERATING SUBMISSION\")\n\nif len(test_paths) > 0:\n test_ds = SentinelDataset(test_paths, img_size=BEST_CFG['img_size'])\n all_preds = []\n\n for fold in range(N_FOLDS):\n mp = f'/kaggle/working/best_fold{fold}.pth'\n if not os.path.exists(mp):\n continue\n\n model = CropDiseaseModel(BEST_BB, num_classes=4, pretrained=False).to(DEVICE)\n model.load_state_dict(torch.load(mp, map_location=DEVICE))\n model.eval()\n\n fold_preds = []\n with torch.no_grad():\n loader = DataLoader(test_ds, batch_size=BEST_CFG['bs'], num_workers=2)\n for imgs in loader:\n if isinstance(imgs, (list, tuple)):\n imgs = imgs[0]\n pred = F.softmax(model(imgs.to(DEVICE)), dim=1).cpu().numpy()\n fold_preds.append(pred)\n\n all_preds.append(np.concatenate(fold_preds, axis=0))\n del model; torch.cuda.empty_cache()\n print(f\" āœ… Fold {fold+1} done\")\n\n if all_preds:\n ensemble = np.mean(all_preds, axis=0)\n final_classes = ensemble.argmax(axis=1)\n\n submission = pd.DataFrame({\n 'Id': test_ids,\n 'Category': [INV_MAP[c] for c in final_classes]\n })\n submission.to_csv('/kaggle/working/submission.csv', index=False)\n\n print(f\"\\nāœ… submission.csv saved! ({len(submission)} predictions)\")\n print(f\"\\nDistribution:\\n{submission['Category'].value_counts().to_string()}\")\n print(f\"\\nFirst 10:\\n{submission.head(10).to_string(index=False)}\")\n\n max_conf = ensemble.max(axis=1)\n print(f\"\\nšŸ“Š Confidence: mean={max_conf.mean():.3f} | \"\n f\"min={max_conf.min():.3f} | max={max_conf.max():.3f}\")\nelse:\n print(\"āš ļø No test data found.\")\n\nprint(\"\\nšŸŽ‰ Done! Submit 'submission.csv' on the competition page.\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-02-16T04:40:39.329974Z","iopub.execute_input":"2026-02-16T04:40:39.330712Z","iopub.status.idle":"2026-02-16T04:41:01.280939Z","shell.execute_reply.started":"2026-02-16T04:40:39.330686Z","shell.execute_reply":"2026-02-16T04:41:01.280162Z"}},"outputs":[],"execution_count":null}]}