{"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":"none","dataSources":[{"sourceType":"competition","sourceId":25563,"databundleVersionId":2094376},{"sourceType":"datasetVersion","sourceId":15619837,"datasetId":9996899,"databundleVersionId":16554068}],"dockerImageVersionId":31328,"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\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,"execution":{"iopub.status.busy":"2026-04-15T14:48:45.493021Z","iopub.execute_input":"2026-04-15T14:48:45.494063Z","iopub.status.idle":"2026-04-15T14:48:45.794878Z","shell.execute_reply.started":"2026-04-15T14:48:45.494029Z","shell.execute_reply":"2026-04-15T14:48:45.794043Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 1: Imports","metadata":{}},{"cell_type":"code","source":"import os\nos.environ[\"TF_CPP_MIN_LOG_LEVEL\"] = \"2\"\n\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nfrom sklearn.preprocessing import MultiLabelBinarizer\nfrom sklearn.model_selection import train_test_split\n\ntf.get_logger().setLevel('ERROR')\n\nprint(\"TensorFlow:\", tf.__version__)\nprint(\"GPU available:\", len(tf.config.list_physical_devices('GPU')) > 0)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:48:45.796064Z","iopub.execute_input":"2026-04-15T14:48:45.796517Z","iopub.status.idle":"2026-04-15T14:49:30.829049Z","shell.execute_reply.started":"2026-04-15T14:48:45.796486Z","shell.execute_reply":"2026-04-15T14:49:30.828001Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 2: Set Paths","metadata":{}},{"cell_type":"code","source":"# Competition data\nBASE_DIR   = \"/kaggle/input/competitions/plant-pathology-2021-fgvc8\"\nTRAIN_DIR  = os.path.join(BASE_DIR, \"train_images\")\nTEST_DIR   = os.path.join(BASE_DIR, \"test_images\")\n\n# Offline weights — uploaded as a Kaggle Dataset (see setup instructions above)\n# Change the filename below if yours is named differently\nWEIGHTS_PATH = \"/kaggle/input/datasets/violentfish/efficientnetv2b0/efficientnetv2-b0_notop.h5\"\nprint(\"Train images:\", TRAIN_DIR)\nprint(\"Test  images:\", TEST_DIR)\nprint(\"Weights file exists:\", os.path.exists(WEIGHTS_PATH))\nprint(\"Sample train files:\", os.listdir(TRAIN_DIR)[:3])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:30.829922Z","iopub.execute_input":"2026-04-15T14:49:30.830334Z","iopub.status.idle":"2026-04-15T14:49:31.070244Z","shell.execute_reply.started":"2026-04-15T14:49:30.830314Z","shell.execute_reply":"2026-04-15T14:49:31.069088Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 3: Load and Explore the Data\n\nThis is a **multi-label** problem — one leaf can have multiple diseases at the same time.  \nLabels are stored as space-separated strings, e.g. `\"scab frog_eye_leaf_spot\"`.","metadata":{}},{"cell_type":"code","source":"train_df  = pd.read_csv(os.path.join(BASE_DIR, \"train.csv\"))\nsample_df = pd.read_csv(os.path.join(BASE_DIR, \"sample_submission.csv\"))\n\nprint(\"Train shape:\", train_df.shape)\nprint(train_df.head())\n\nALL_LABELS  = ['complex', 'frog_eye_leaf_spot', 'healthy', 'powdery_mildew', 'rust', 'scab']\nNUM_CLASSES = len(ALL_LABELS)\nprint(\"\\nClasses:\", ALL_LABELS)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:31.072830Z","iopub.execute_input":"2026-04-15T14:49:31.073206Z","iopub.status.idle":"2026-04-15T14:49:31.128918Z","shell.execute_reply.started":"2026-04-15T14:49:31.073169Z","shell.execute_reply":"2026-04-15T14:49:31.127753Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Show label distribution\nlabel_counts = train_df['labels'].value_counts().head(10)\n\nplt.figure(figsize=(10, 4))\nlabel_counts.plot(kind='bar')\nplt.title(\"Top 10 Label Combinations\")\nplt.xticks(rotation=30, ha='right')\nplt.tight_layout()\nplt.show()\n\nprint(\"\\nUnique label combinations:\", train_df['labels'].nunique())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:31.129905Z","iopub.execute_input":"2026-04-15T14:49:31.130125Z","iopub.status.idle":"2026-04-15T14:49:31.465917Z","shell.execute_reply.started":"2026-04-15T14:49:31.130103Z","shell.execute_reply":"2026-04-15T14:49:31.464389Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 4: Show Sample Images","metadata":{}},{"cell_type":"code","source":"plt.figure(figsize=(16, 4))\nfor i in range(8):\n    row   = train_df.iloc[i]\n    path  = os.path.join(TRAIN_DIR, row['image'])\n    image = tf.keras.utils.load_img(path, target_size=(128, 128))\n    image = tf.keras.utils.img_to_array(image).astype(\"uint8\")\n    plt.subplot(1, 8, i + 1)\n    plt.imshow(image)\n    plt.title(row['labels'], fontsize=6)\n    plt.axis(\"off\")\nplt.suptitle(\"Sample Training Images\", fontsize=13)\nplt.tight_layout()\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:31.467150Z","iopub.execute_input":"2026-04-15T14:49:31.467502Z","iopub.status.idle":"2026-04-15T14:49:32.613710Z","shell.execute_reply.started":"2026-04-15T14:49:31.467472Z","shell.execute_reply":"2026-04-15T14:49:32.612767Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 5: Encode Labels\n\nWe convert space-separated label strings into **binary vectors** of length 6.  \nExample: `\"scab frog_eye_leaf_spot\"` → `[0, 1, 0, 0, 0, 1]`  \nThis is called **multi-hot encoding** — multiple positions can be 1 at the same time.","metadata":{}},{"cell_type":"code","source":"train_df['label_list'] = train_df['labels'].apply(lambda x: x.split())\n\nmlb   = MultiLabelBinarizer(classes=ALL_LABELS)\ny_all = mlb.fit_transform(train_df['label_list']).astype('float32')\n\nprint(\"Label matrix shape:\", y_all.shape)   # (18632, 6)\nprint(\"Classes:\", mlb.classes_)\nprint(\"\\nExample — label:\", train_df['labels'].iloc[0], \"→ encoded:\", y_all[0])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:32.614811Z","iopub.execute_input":"2026-04-15T14:49:32.615161Z","iopub.status.idle":"2026-04-15T14:49:32.643264Z","shell.execute_reply.started":"2026-04-15T14:49:32.615130Z","shell.execute_reply":"2026-04-15T14:49:32.641903Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 6: Build the tf.data Pipeline","metadata":{}},{"cell_type":"code","source":"IMG_SIZE   = 224\nBATCH_SIZE = 32\nAUTOTUNE   = tf.data.AUTOTUNE\n\nX_train, X_val, y_train, y_val = train_test_split(\n    train_df['image'].values, y_all,\n    test_size=0.2, random_state=42\n)\nprint(f\"Train: {len(X_train)} | Val: {len(X_val)}\")\n\n\ndef load_image(path, label):\n    full_path = tf.strings.join([TRAIN_DIR + \"/\", path])\n    image = tf.io.read_file(full_path)\n    image = tf.image.decode_jpeg(image, channels=3)\n    image = tf.image.resize(image, (IMG_SIZE, IMG_SIZE))\n    image = tf.cast(image, tf.float32) / 255.0\n    return image, label\n\ndef load_test_image(path):\n    full_path = tf.strings.join([TEST_DIR + \"/\", path])\n    image = tf.io.read_file(full_path)\n    image = tf.image.decode_jpeg(image, channels=3)\n    image = tf.image.resize(image, (IMG_SIZE, IMG_SIZE))\n    image = tf.cast(image, tf.float32) / 255.0\n    return image\n\n\ntrain_ds = (\n    tf.data.Dataset.from_tensor_slices((X_train, y_train))\n    .shuffle(2000)\n    .map(load_image, num_parallel_calls=AUTOTUNE)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTOTUNE)\n)\n\nval_ds = (\n    tf.data.Dataset.from_tensor_slices((X_val, y_val))\n    .map(load_image, num_parallel_calls=AUTOTUNE)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTOTUNE)\n)\n\ntest_paths = sample_df['image'].values\ntest_ds = (\n    tf.data.Dataset.from_tensor_slices(test_paths)\n    .map(load_test_image, num_parallel_calls=AUTOTUNE)\n    .batch(BATCH_SIZE)\n    .prefetch(AUTOTUNE)\n)\n\nprint(\"Pipelines ready!\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:32.644830Z","iopub.execute_input":"2026-04-15T14:49:32.645103Z","iopub.status.idle":"2026-04-15T14:49:32.863316Z","shell.execute_reply.started":"2026-04-15T14:49:32.645081Z","shell.execute_reply":"2026-04-15T14:49:32.862437Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 7: Define Augmentation Blocks\n\nCombined geometric + color augmentation from Lectures 11a and 11b.","metadata":{}},{"cell_type":"code","source":"# Geometric augmentation\ngeometric_augmentation = tf.keras.Sequential([\n    tf.keras.layers.RandomFlip(\"horizontal\"),\n    tf.keras.layers.RandomRotation(factor=0.05),\n    tf.keras.layers.RandomZoom(height_factor=(-0.1, 0.1), width_factor=(-0.1, 0.1)),\n    tf.keras.layers.RandomTranslation(height_factor=0.1, width_factor=0.1),\n], name=\"geometric_augmentation\")\n\n\n# Color augmentation custom layers\nclass RandomBrightnessLayer(tf.keras.layers.Layer):\n    def __init__(self, max_delta=0.2, **kwargs):\n        super().__init__(**kwargs)\n        self.max_delta = max_delta\n    def call(self, x, training=False):\n        if training:\n            x = tf.image.random_brightness(x, self.max_delta)\n            x = tf.clip_by_value(x, 0.0, 1.0)\n        return x\n\nclass RandomContrastLayer(tf.keras.layers.Layer):\n    def __init__(self, lower=0.7, upper=1.3, **kwargs):\n        super().__init__(**kwargs)\n        self.lower, self.upper = lower, upper\n    def call(self, x, training=False):\n        if training:\n            x = tf.image.random_contrast(x, self.lower, self.upper)\n            x = tf.clip_by_value(x, 0.0, 1.0)\n        return x\n\nclass RandomSaturationLayer(tf.keras.layers.Layer):\n    def __init__(self, lower=0.6, upper=1.4, **kwargs):\n        super().__init__(**kwargs)\n        self.lower, self.upper = lower, upper\n    def call(self, x, training=False):\n        if training:\n            x = tf.image.random_saturation(x, self.lower, self.upper)\n            x = tf.clip_by_value(x, 0.0, 1.0)\n        return x\n\nclass RandomHueLayer(tf.keras.layers.Layer):\n    def __init__(self, max_delta=0.1, **kwargs):\n        super().__init__(**kwargs)\n        self.max_delta = max_delta\n    def call(self, x, training=False):\n        if training:\n            x = tf.image.random_hue(x, self.max_delta)\n            x = tf.clip_by_value(x, 0.0, 1.0)\n        return x\n\ncolor_augmentation = tf.keras.Sequential([\n    RandomBrightnessLayer(max_delta=0.2),\n    RandomContrastLayer(lower=0.7, upper=1.3),\n    RandomSaturationLayer(lower=0.6, upper=1.4),\n    RandomHueLayer(max_delta=0.1),\n], name=\"color_augmentation\")\n\nprint(\"Augmentation blocks ready!\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:32.864712Z","iopub.execute_input":"2026-04-15T14:49:32.864952Z","iopub.status.idle":"2026-04-15T14:49:32.898877Z","shell.execute_reply.started":"2026-04-15T14:49:32.864932Z","shell.execute_reply":"2026-04-15T14:49:32.897513Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 8: Build the Model\n\n**Key difference from previous notebooks:**  \nThis is **multi-label** classification:\n- Output activation: `sigmoid` — each of the 6 classes is predicted independently  \n- Loss: `binary_crossentropy` — loss is computed separately per class  \n\n**Offline weights:**  \nWe use `weights=None` and load the `.h5` file manually.  \nThis is required because **internet is OFF** in this competition.","metadata":{}},{"cell_type":"code","source":"inputs = tf.keras.Input(shape=(IMG_SIZE, IMG_SIZE, 3), name=\"input_image\")\n\n# 1) Geometric augmentation\nx = geometric_augmentation(inputs)\n\n# 2) Color augmentation (expects [0,1])\nx = color_augmentation(x)\n\n# 3) Scale back to [0,255] for EfficientNet preprocessing\nx = x * 255.0\nx = tf.keras.applications.efficientnet_v2.preprocess_input(x)\n\n# 4) Backbone — weights=None because internet is OFF\n#    We load the weights manually from our uploaded dataset file\nbase_model = tf.keras.applications.EfficientNetV2B0(\n    include_top=False,\n    weights=None          # do NOT download — we load from file below\n)\nbase_model.load_weights(WEIGHTS_PATH, by_name=True, skip_mismatch=True)\nprint(\"Weights loaded from:\", WEIGHTS_PATH)\n\nbase_model.trainable = False   # freeze backbone for first training phase\nx = base_model(x, training=False)\n\n# 5) Classification head — sigmoid for multi-label\nx = tf.keras.layers.GlobalAveragePooling2D()(x)\nx = tf.keras.layers.Dropout(0.3)(x)\noutputs = tf.keras.layers.Dense(\n    NUM_CLASSES, activation=\"sigmoid\", name=\"predictions\"\n)(x)\n\nmodel = tf.keras.Model(inputs, outputs, name=\"plant_pathology_model\")\nmodel.summary()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:32.900758Z","iopub.execute_input":"2026-04-15T14:49:32.901009Z","iopub.status.idle":"2026-04-15T14:49:35.355230Z","shell.execute_reply.started":"2026-04-15T14:49:32.900988Z","shell.execute_reply":"2026-04-15T14:49:35.353815Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 9: Compile and Train (Frozen Backbone)","metadata":{}},{"cell_type":"code","source":"model.compile(\n    optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3),\n    loss=\"binary_crossentropy\",\n    metrics=[\"accuracy\"]\n)\n\nprint(\"Training frozen backbone (1 epochs)...\")\nhistory = model.fit(\n    train_ds,\n    validation_data=val_ds,\n    epochs=1\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T14:49:35.356179Z","iopub.execute_input":"2026-04-15T14:49:35.356450Z","iopub.status.idle":"2026-04-15T15:01:43.600454Z","shell.execute_reply.started":"2026-04-15T14:49:35.356427Z","shell.execute_reply":"2026-04-15T15:01:43.599136Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 10: Fine-Tune the Backbone (improves score)\n\nUnfreeze the backbone and retrain with a smaller learning rate.  \nThis usually gives a significant improvement in F1 score.","metadata":{}},{"cell_type":"code","source":"base_model.trainable = True\n\nmodel.compile(\n    optimizer=tf.keras.optimizers.Adam(learning_rate=1e-4),  # lower LR for fine-tuning\n    loss=\"binary_crossentropy\",\n    metrics=[\"accuracy\"]\n)\n\nprint(\"Fine-tuning full model (1 epochs)...\")\nhistory_ft = model.fit(\n    train_ds,\n    validation_data=val_ds,\n    epochs=1\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T15:02:34.831765Z","iopub.execute_input":"2026-04-15T15:02:34.832079Z","iopub.status.idle":"2026-04-15T15:35:26.124147Z","shell.execute_reply.started":"2026-04-15T15:02:34.832056Z","shell.execute_reply":"2026-04-15T15:35:26.121512Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 11: Make Predictions on Test Set\n\n**Threshold = 0.5** — if probability > 0.5 we predict that class.  \nIf no class passes the threshold, we pick the highest probability class.","metadata":{}},{"cell_type":"code","source":"preds = model.predict(test_ds, verbose=1)   # shape: (N, 6)\n\nTHRESHOLD = 0.5\n\ndef probs_to_label(prob_row, threshold=THRESHOLD):\n    selected = [ALL_LABELS[i] for i, p in enumerate(prob_row) if p >= threshold]\n    if len(selected) == 0:\n        selected = [ALL_LABELS[np.argmax(prob_row)]]\n    return \" \".join(selected)\n\npredicted_labels = [probs_to_label(row) for row in preds]\n\nprint(\"Sample predictions:\")\nfor i in range(3):\n    print(f\"  {test_paths[i]}  →  {predicted_labels[i]}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T15:35:52.854232Z","iopub.execute_input":"2026-04-15T15:35:52.854834Z","iopub.status.idle":"2026-04-15T15:35:55.002313Z","shell.execute_reply.started":"2026-04-15T15:35:52.854804Z","shell.execute_reply":"2026-04-15T15:35:55.000788Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Step 12: Save submission.csv","metadata":{}},{"cell_type":"code","source":"submission = pd.DataFrame({\n    'image':  test_paths,\n    'labels': predicted_labels\n})\n\nsubmission.to_csv('submission.csv', index=False)\n\nprint(\"submission.csv saved!\")\nprint(submission.head())\nprint(\"Shape:\", submission.shape)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-04-15T15:35:59.526644Z","iopub.execute_input":"2026-04-15T15:35:59.527909Z","iopub.status.idle":"2026-04-15T15:35:59.565672Z","shell.execute_reply.started":"2026-04-15T15:35:59.527871Z","shell.execute_reply":"2026-04-15T15:35:59.564158Z"}},"outputs":[],"execution_count":null}]}