{"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":"gpu","dataSources":[{"sourceId":4521,"databundleVersionId":326986,"sourceType":"competition"}],"dockerImageVersionId":31193,"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 timm --quiet\n\nimport os\nimport zipfile\nfrom pathlib import Path\n\nimport torch\nimport torch.nn as nn\nfrom torch.optim import Adam\nfrom torch.utils.data import Dataset, DataLoader\n\nfrom torchvision import transforms\nimport torchvision.models as models\nimport timm\n\nimport pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom sklearn.metrics import confusion_matrix\nfrom PIL import Image\n\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nprint(\"PyTorch:\", torch.__version__)\nprint(\"CUDA available:\", torch.cuda.is_available())\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:34:21.074936Z","iopub.execute_input":"2025-11-30T04:34:21.075431Z","iopub.status.idle":"2025-11-30T04:35:42.721257Z","shell.execute_reply.started":"2025-11-30T04:34:21.075405Z","shell.execute_reply":"2025-11-30T04:35:42.720557Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"DATA_DIR = Path(\"/kaggle/input/noaa-right-whale-recognition\")\nWORK_DIR = Path(\"/kaggle/working\")\nIMG_DIR = WORK_DIR/\"imgs\"\n\nif not IMG_DIR.exists():\n    print(\"Extracting imgs.zip ...\")\n    with zipfile.ZipFile(DATA_DIR/\"imgs.zip\", \"r\") as z:\n        z.extractall(WORK_DIR)\n    print(\"Done.\")\nelse:\n    print(\"Images already extracted.\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:37:38.278882Z","iopub.execute_input":"2025-11-30T04:37:38.279532Z","iopub.status.idle":"2025-11-30T04:39:11.006467Z","shell.execute_reply.started":"2025-11-30T04:37:38.279501Z","shell.execute_reply":"2025-11-30T04:39:11.005744Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"df = pd.read_csv(DATA_DIR/\"train.csv\")\n\n# Remove corrupted image\ndf = df[df[\"Image\"] != \"w_7489.jpg\"]\n\n# Count samples per whale\ncounts = df[\"whaleID\"].value_counts()\n\n# Single-image whales → MUST stay in training\nsingle_whales = counts[counts == 1].index\ndf[\"is_single\"] = df[\"whaleID\"].isin(single_whales)\n\n# Split\ntrain_df = df[df[\"is_single\"] == True]\nother_df = df[df[\"is_single\"] == False]\n\nvalid_df = other_df.sample(frac=0.18, random_state=42)\ntrain_df = pd.concat([train_df, other_df.drop(valid_df.index)])\n\ntrain_df = train_df.reset_index(drop=True)\nvalid_df = valid_df.reset_index(drop=True)\n\nprint(\"Train size:\", len(train_df))\nprint(\"Valid size:\", len(valid_df))\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:39:45.249877Z","iopub.execute_input":"2025-11-30T04:39:45.250686Z","iopub.status.idle":"2025-11-30T04:39:45.268268Z","shell.execute_reply.started":"2025-11-30T04:39:45.250661Z","shell.execute_reply":"2025-11-30T04:39:45.267461Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class WhaleDataset(Dataset):\n    def __init__(self, df, img_dir, transform=None):\n        self.df = df.reset_index(drop=True)\n        self.img_dir = img_dir\n        self.transform = transform\n\n        # label mapping\n        self.labels = sorted(self.df[\"whaleID\"].unique())\n        self.lab2idx = {label:i for i,label in enumerate(self.labels)}\n\n    def __len__(self):\n        return len(self.df)\n\n    def __getitem__(self, idx):\n        row = self.df.iloc[idx]\n        img_path = self.img_dir/row[\"Image\"]\n        img = Image.open(img_path).convert(\"RGB\")\n\n        if self.transform:\n            img = self.transform(img)\n\n        label = self.lab2idx[row[\"whaleID\"]]\n        return img, label\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:40:03.559902Z","iopub.execute_input":"2025-11-30T04:40:03.560176Z","iopub.status.idle":"2025-11-30T04:40:03.565675Z","shell.execute_reply.started":"2025-11-30T04:40:03.560158Z","shell.execute_reply":"2025-11-30T04:40:03.564946Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"transform = transforms.Compose([\n    transforms.Resize((224,224)),   # you can reduce to 160×160 for faster speed\n    transforms.ToTensor(),\n])\n\ntrain_ds = WhaleDataset(train_df, IMG_DIR, transform)\nvalid_ds = WhaleDataset(valid_df, IMG_DIR, transform)\n\ntrain_loader = DataLoader(train_ds, batch_size=32, shuffle=True, num_workers=2)\nvalid_loader = DataLoader(valid_ds, batch_size=32, shuffle=False, num_workers=2)\n\nNUM_CLASSES = len(train_ds.labels)\nprint(\"Total whale classes:\", NUM_CLASSES)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:40:26.124668Z","iopub.execute_input":"2025-11-30T04:40:26.125383Z","iopub.status.idle":"2025-11-30T04:40:26.133158Z","shell.execute_reply.started":"2025-11-30T04:40:26.125358Z","shell.execute_reply":"2025-11-30T04:40:26.132547Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def train_model(model, name, epochs=5):\n    device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    model.to(device)\n\n    optimizer = Adam(model.parameters(), lr=1e-4)\n    criterion = nn.CrossEntropyLoss()\n\n    history = {\"train_loss\": [], \"val_loss\": [], \"val_acc\": []}\n\n    for epoch in range(epochs):\n        model.train()\n        total_loss = 0\n\n        for imgs, labels in train_loader:\n            imgs, labels = imgs.to(device), labels.to(device)\n\n            optimizer.zero_grad()\n\n            # AMP for speed\n            with torch.cuda.amp.autocast():\n                preds = model(imgs)\n                loss = criterion(preds, labels)\n\n            loss.backward()\n            optimizer.step()\n            total_loss += loss.item()\n\n        avg_train_loss = total_loss / len(train_loader)\n        history[\"train_loss\"].append(avg_train_loss)\n\n        # validation\n        model.eval()\n        val_loss, correct, total = 0, 0, 0\n\n        with torch.no_grad():\n            for imgs, labels in valid_loader:\n                imgs, labels = imgs.to(device), labels.to(device)\n                preds = model(imgs)\n                loss = criterion(preds, labels)\n                val_loss += loss.item()\n\n                _, predicted = preds.max(1)\n                correct += (predicted == labels).sum().item()\n                total += labels.size(0)\n\n        avg_val_loss = val_loss / len(valid_loader)\n        acc = correct / total\n\n        history[\"val_loss\"].append(avg_val_loss)\n        history[\"val_acc\"].append(acc)\n\n        print(f\"{name} | Epoch {epoch+1}/{epochs} | Train={avg_train_loss:.4f}, Val={avg_val_loss:.4f}, Acc={acc:.4f}\")\n\n    # Save model\n    torch.save(model.state_dict(), f\"/kaggle/working/{name}.pth\")\n    return model, history\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:40:47.716445Z","iopub.execute_input":"2025-11-30T04:40:47.717006Z","iopub.status.idle":"2025-11-30T04:40:47.725119Z","shell.execute_reply.started":"2025-11-30T04:40:47.716981Z","shell.execute_reply":"2025-11-30T04:40:47.724334Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"res34 = models.resnet34(weights=\"IMAGENET1K_V1\")\nres34.fc = nn.Linear(res34.fc.in_features, NUM_CLASSES)\n\nm1, h1 = train_model(res34, \"resnet34\", epochs=5)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T04:41:05.214614Z","iopub.execute_input":"2025-11-30T04:41:05.215409Z","iopub.status.idle":"2025-11-30T05:02:08.678898Z","shell.execute_reply.started":"2025-11-30T04:41:05.215384Z","shell.execute_reply":"2025-11-30T05:02:08.677956Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"res50 = models.resnet50(weights=\"IMAGENET1K_V2\")\nres50.fc = nn.Linear(res50.fc.in_features, NUM_CLASSES)\n\nm2, h2 = train_model(res50, \"resnet50\", epochs=5)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T05:02:50.177085Z","iopub.execute_input":"2025-11-30T05:02:50.177785Z","iopub.status.idle":"2025-11-30T05:17:58.325083Z","shell.execute_reply.started":"2025-11-30T05:02:50.177753Z","shell.execute_reply":"2025-11-30T05:17:58.324174Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"effnet = timm.create_model(\"efficientnet_b0\", pretrained=True, num_classes=NUM_CLASSES)\n\nm3, h3 = train_model(effnet, \"efficientnet_b0\", epochs=5)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T05:36:31.319699Z","iopub.execute_input":"2025-11-30T05:36:31.320358Z","iopub.status.idle":"2025-11-30T05:52:18.568228Z","shell.execute_reply.started":"2025-11-30T05:36:31.32032Z","shell.execute_reply":"2025-11-30T05:52:18.567279Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def plot_history(history, name):\n    plt.figure(figsize=(10,4))\n\n    plt.subplot(1,2,1)\n    plt.plot(history[\"train_loss\"], label=\"Train\")\n    plt.plot(history[\"val_loss\"], label=\"Validation\")\n    plt.title(f\"{name} Loss\"); plt.legend()\n\n    plt.subplot(1,2,2)\n    plt.plot(history[\"val_acc\"], label=\"Val Accuracy\")\n    plt.title(f\"{name} Accuracy\"); plt.legend()\n\n    plt.savefig(f\"/kaggle/working/{name}_curves.png\")\n    plt.show()\n\nplot_history(h1, \"resnet34\")\nplot_history(h2, \"resnet50\")\nplot_history(h3, \"efficientnet_b0\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T05:52:29.33494Z","iopub.execute_input":"2025-11-30T05:52:29.335259Z","iopub.status.idle":"2025-11-30T05:52:30.570958Z","shell.execute_reply.started":"2025-11-30T05:52:29.335229Z","shell.execute_reply":"2025-11-30T05:52:30.570187Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def plot_confusion(model, name):\n    model.eval()\n    device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n\n    preds, actual = [], []\n\n    with torch.no_grad():\n        for imgs, labels in valid_loader:\n            imgs = imgs.to(device)\n            outputs = model(imgs)\n            _, predicted = outputs.max(1)\n            preds += predicted.cpu().tolist()\n            actual += labels.tolist()\n\n    cm = confusion_matrix(actual, preds)\n    plt.figure(figsize=(6,6))\n    plt.imshow(cm, cmap=\"Blues\")\n    plt.title(f\"{name} Confusion Matrix\")\n    plt.colorbar()\n    plt.savefig(f\"/kaggle/working/{name}_cm.png\")\n    plt.show()\n\nplot_confusion(m1, \"resnet34\")\nplot_confusion(m2, \"resnet50\")\nplot_confusion(m3, \"efficientnet_b0\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T05:53:01.94328Z","iopub.execute_input":"2025-11-30T05:53:01.943585Z","iopub.status.idle":"2025-11-30T05:54:43.91984Z","shell.execute_reply.started":"2025-11-30T05:53:01.943564Z","shell.execute_reply":"2025-11-30T05:54:43.919025Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"accs = {\n    \"resnet34\": h1[\"val_acc\"][-1],\n    \"resnet50\": h2[\"val_acc\"][-1],\n    \"efficientnet_b0\": h3[\"val_acc\"][-1],\n}\n\nplt.figure(figsize=(6,4))\nplt.bar(accs.keys(), accs.values())\nplt.title(\"Model Accuracy Comparison\")\nplt.ylabel(\"Accuracy\")\nplt.savefig(\"/kaggle/working/model_comparison.png\")\nplt.show()\n\nprint(accs)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T05:55:07.20055Z","iopub.execute_input":"2025-11-30T05:55:07.200898Z","iopub.status.idle":"2025-11-30T05:55:07.379749Z","shell.execute_reply.started":"2025-11-30T05:55:07.200871Z","shell.execute_reply":"2025-11-30T05:55:07.378988Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sample = pd.read_csv(DATA_DIR/\"sample_submission.csv\")\n\ntest_tf = transforms.Compose([\n    transforms.Resize((224,224)),\n    transforms.ToTensor(),\n])\n\ndef create_submission(model):\n    device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n    model.eval()\n\n    all_probs = []\n\n    with torch.no_grad():\n        for img_name in sample[\"Image\"]:\n            img = Image.open(IMG_DIR/img_name).convert(\"RGB\")\n            img = test_tf(img).unsqueeze(0).to(device)\n            out = model(img)\n            probs = torch.softmax(out, dim=1)[0].cpu().numpy()\n            all_probs.append(probs)\n\n    df = pd.DataFrame(all_probs, columns=train_ds.labels)\n    df[\"Image\"] = sample[\"Image\"]\n\n    df.to_csv(\"/kaggle/working/submission.csv\", index=False)\n    print(\"submission.csv saved!\")\n\n# choose best model (your choice)\ncreate_submission(m3)   # EfficientNet recommended\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T05:55:34.473381Z","iopub.execute_input":"2025-11-30T05:55:34.473653Z","iopub.status.idle":"2025-11-30T06:07:27.529675Z","shell.execute_reply.started":"2025-11-30T05:55:34.473634Z","shell.execute_reply":"2025-11-30T06:07:27.528779Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# CELL: get embeddings from model (replace last fc with identity)\nimport torch\nimport torch.nn as nn\nfrom sklearn.neighbors import KNeighborsClassifier\n\ndef get_embedding_model(model, global_pool='avg'):\n    # For torchvision ResNet: replace fc with identity\n    emb_model = model\n    # ensure fc exists\n    if hasattr(emb_model, 'fc'):\n        in_feat = emb_model.fc.in_features\n        emb_model.fc = nn.Identity()\n    elif hasattr(emb_model, 'classifier'):\n        emb_model.classifier = nn.Identity()\n    return emb_model\n\ndef compute_embeddings(model, loader, device=None):\n    device = device or (\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    model.to(device).eval()\n    embeddings, labels = [], []\n    with torch.no_grad():\n        for imgs, lbl in loader:\n            imgs = imgs.to(device)\n            out = model(imgs)\n            out = out.cpu().numpy()\n            embeddings.append(out)\n            labels.extend(lbl.numpy())\n    embeddings = np.vstack(embeddings)\n    labels = np.array(labels)\n    return embeddings, labels\n\n# Example: convert res34 into embedding extractor\nemb_model = get_embedding_model(res34)\n# compute embeddings for training set and validation set (use smaller loaders or batch size)\ntrain_emb_loader = DataLoader(train_ds, batch_size=64, shuffle=False, num_workers=2)\nvalid_emb_loader = DataLoader(valid_ds, batch_size=64, shuffle=False, num_workers=2)\n\ntrain_embs, train_lbls = compute_embeddings(emb_model, train_emb_loader)\nvalid_embs, valid_lbls = compute_embeddings(emb_model, valid_emb_loader)\n\n# fit k-NN on training embeddings\nknn = KNeighborsClassifier(n_neighbors=3, metric='cosine')\nknn.fit(train_embs, train_lbls)\n\n# predict on validation embeddings\npreds = knn.predict(valid_embs)\n\nfrom sklearn.metrics import accuracy_score, confusion_matrix\nacc = accuracy_score(valid_lbls, preds)\nprint(\"k-NN on embeddings accuracy:\", acc)\n\ncm = confusion_matrix(valid_lbls, preds)\nplt.figure(figsize=(6,6))\nplt.imshow(cm, cmap='Blues')\nplt.title(\"Confusion Matrix (k-NN on embeddings)\")\nplt.colorbar()\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T15:56:55.885425Z","iopub.execute_input":"2025-11-30T15:56:55.885586Z","iopub.status.idle":"2025-11-30T15:57:00.865652Z","shell.execute_reply.started":"2025-11-30T15:56:55.885569Z","shell.execute_reply":"2025-11-30T15:57:00.864557Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nfrom sklearn.neighbors import KNeighborsClassifier\nfrom torch.utils.data import DataLoader\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# --- Embedding model extractor ---\ndef get_embedding_model(model, global_pool='avg'):\n    # For torchvision ResNet: replace fc with identity\n    emb_model = model\n    if hasattr(emb_model, 'fc'):\n        emb_model.fc = nn.Identity()\n    elif hasattr(emb_model, 'classifier'):\n        emb_model.classifier = nn.Identity()\n    return emb_model\n\n# --- Compute embeddings ---\ndef compute_embeddings(model, loader, device=None):\n    device = device or (\"cuda\" if torch.cuda.is_available() else \"cpu\")\n    model.to(device).eval()\n    embeddings, labels = [], []\n\n    with torch.no_grad():\n        for imgs, lbl in loader:\n            imgs = imgs.to(device)\n            out = model(imgs)\n            embeddings.append(out.cpu().numpy())\n            labels.extend(lbl.numpy())\n\n    embeddings = np.vstack(embeddings)\n    labels = np.array(labels)\n    return embeddings, labels\n\n# --------------------------\n# 👉 FIX: Use your actual model variable\n# --------------------------\n\nemb_model = get_embedding_model(resnet34)   # <-- FIXED HERE\n\n# Make dataloaders for embeddings\ntrain_emb_loader = DataLoader(train_ds, batch_size=64, shuffle=False)\nvalid_emb_loader = DataLoader(valid_ds, batch_size=64, shuffle=False)\n\n# Compute embeddings\ntrain_embs, train_lbls = compute_embeddings(emb_model, train_emb_loader)\nvalid_embs, valid_lbls = compute_embeddings(emb_model, valid_emb_loader)\n\n# Fit k-NN\nknn = KNeighborsClassifier(n_neighbors=3, metric='cosine')\nknn.fit(train_embs, train_lbls)\n\n# Predict\npreds = knn.predict(valid_embs)\n\n# Metrics\nfrom sklearn.metrics import accuracy_score, confusion_matrix\nacc = accuracy_score(valid_lbls, preds)\nprint(\"k-NN on embeddings accuracy:\", acc)\n\n# Confusion matrix\ncm = confusion_matrix(valid_lbls, preds)\nplt.figure(figsize=(6,6))\nplt.imshow(cm, cmap='Blues')\nplt.title(\"Confusion Matrix (k-NN on embeddings)\")\nplt.colorbar()\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T15:58:44.697197Z","iopub.execute_input":"2025-11-30T15:58:44.697999Z","iopub.status.idle":"2025-11-30T15:58:44.715257Z","shell.execute_reply.started":"2025-11-30T15:58:44.697973Z","shell.execute_reply":"2025-11-30T15:58:44.714342Z"}},"outputs":[],"execution_count":null}]}