{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.12.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"none","dataSources":[],"dockerImageVersionId":28755,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# ============================================================\n# PETALS TO THE METAL - STRONG FLOWER CLASSIFIER\n# EfficientNetB2 + Data Augmentation + Fine Tuning\n# ============================================================\n\nimport os\nimport glob\nimport numpy as np\nimport pandas as pd\nimport tensorflow as tf\n\nfrom sklearn.metrics import f1_score\n\nprint(\"TensorFlow:\", tf.__version__)\nprint(\"GPUs:\", tf.config.list_physical_devices(\"GPU\"))\n\n\n# ============================================================\n# 1. SETTINGS\n# ============================================================\n\nDATA_PATH = \"/kaggle/input/competitions/tpu-getting-started\"\n\nIMG_SIZE = 260\nBATCH_SIZE = 32\nNUM_CLASSES = 104\n\nEPOCHS_HEAD = 5\nEPOCHS_FINE = 5\n\nAUTOTUNE = tf.data.AUTOTUNE\n\n\n# ============================================================\n# 2. FIND DATA\n# ============================================================\n\nTRAIN_FILES = sorted(\n glob.glob(\n DATA_PATH +\n \"/tfrecords-jpeg-224x224/train/*.tfrec\"\n )\n)\n\nVAL_FILES = sorted(\n glob.glob(\n DATA_PATH +\n \"/tfrecords-jpeg-224x224/val/*.tfrec\"\n )\n)\n\nTEST_FILES = sorted(\n glob.glob(\n DATA_PATH +\n \"/tfrecords-jpeg-224x224/test/*.tfrec\"\n )\n)\n\nprint(\"\\nTrain files:\", len(TRAIN_FILES))\nprint(\"Validation files:\", len(VAL_FILES))\nprint(\"Test files:\", len(TEST_FILES))\n\n\n# ============================================================\n# 3. READ TRAIN / VALIDATION TFRECORDS\n# ============================================================\n\ndef read_labeled(example):\n\n features = {\n \"image\": tf.io.FixedLenFeature([], tf.string),\n \"class\": tf.io.FixedLenFeature([], tf.int64)\n }\n\n example = tf.io.parse_single_example(\n example,\n features\n )\n\n image = tf.image.decode_jpeg(\n example[\"image\"],\n channels=3\n )\n\n image = tf.image.resize(\n image,\n [IMG_SIZE, IMG_SIZE]\n )\n\n image = tf.cast(image, tf.float32)\n\n label = tf.cast(\n example[\"class\"],\n tf.int32\n )\n\n return image, label\n\n\ntrain_ds = tf.data.TFRecordDataset(\n TRAIN_FILES,\n num_parallel_reads=AUTOTUNE\n)\n\ntrain_ds = train_ds.map(\n read_labeled,\n num_parallel_calls=AUTOTUNE\n)\n\ntrain_ds = train_ds.shuffle(\n 4000,\n reshuffle_each_iteration=True\n)\n\ntrain_ds = train_ds.batch(\n BATCH_SIZE,\n drop_remainder=True\n)\n\ntrain_ds = train_ds.prefetch(AUTOTUNE)\n\n\nval_ds = tf.data.TFRecordDataset(\n VAL_FILES,\n num_parallel_reads=AUTOTUNE\n)\n\nval_ds = val_ds.map(\n read_labeled,\n num_parallel_calls=AUTOTUNE\n)\n\nval_ds = val_ds.batch(\n BATCH_SIZE\n)\n\nval_ds = val_ds.prefetch(AUTOTUNE)\n\n\n# ============================================================\n# 4. CALCULATE CLASS WEIGHTS\n# ============================================================\n\nprint(\"\\nCalculating class weights...\")\n\nclass_counts = np.zeros(\n NUM_CLASSES,\n dtype=np.int64\n)\n\nfor _, labels in train_ds.unbatch():\n\n label = int(labels.numpy())\n\n class_counts[label] += 1\n\n\n# Avoid division by zero\nclass_counts = np.maximum(\n class_counts,\n 1\n)\n\ntotal_samples = np.sum(\n class_counts\n)\n\nclass_weights = (\n total_samples /\n (NUM_CLASSES * class_counts)\n)\n\n# Don't allow extreme weights\nclass_weights = np.clip(\n class_weights,\n 0.5,\n 3.0\n)\n\nclass_weight_dict = {\n i: float(class_weights[i])\n for i in range(NUM_CLASSES)\n}\n\nprint(\"Class weights calculated.\")\n\n\n# ============================================================\n# 5. DATA AUGMENTATION\n# ============================================================\n\naugmentation = tf.keras.Sequential([\n\n tf.keras.layers.RandomFlip(\n \"horizontal\"\n ),\n\n tf.keras.layers.RandomRotation(\n 0.12\n ),\n\n tf.keras.layers.RandomZoom(\n 0.15\n ),\n\n tf.keras.layers.RandomContrast(\n 0.10\n )\n])\n\n\n# ============================================================\n# 6. BUILD EFFICIENTNETB2\n# ============================================================\n\nprint(\"\\nLoading EfficientNetB2...\")\n\nbase_model = tf.keras.applications.EfficientNetB2(\n\n include_top=False,\n\n weights=\"imagenet\",\n\n input_shape=(\n IMG_SIZE,\n IMG_SIZE,\n 3\n )\n)\n\n# First train only the new classification head\nbase_model.trainable = False\n\n\ninputs = tf.keras.Input(\n shape=(\n IMG_SIZE,\n IMG_SIZE,\n 3\n )\n)\n\nx = augmentation(inputs)\n\nx = base_model(\n x,\n training=False\n)\n\nx = tf.keras.layers.GlobalAveragePooling2D()(x)\n\nx = tf.keras.layers.BatchNormalization()(x)\n\nx = tf.keras.layers.Dropout(\n 0.35\n)(x)\n\nx = tf.keras.layers.Dense(\n 256,\n activation=\"relu\"\n)(x)\n\nx = tf.keras.layers.Dropout(\n 0.25\n)(x)\n\noutputs = tf.keras.layers.Dense(\n NUM_CLASSES,\n activation=\"softmax\"\n)(x)\n\nmodel = tf.keras.Model(\n inputs,\n outputs\n)\n\n\n# ============================================================\n# 7. COMPILE\n# ============================================================\n\nmodel.compile(\n\n optimizer=tf.keras.optimizers.Adam(\n learning_rate=0.001\n ),\n\n loss=\"sparse_categorical_crossentropy\",\n\n metrics=[\n \"accuracy\"\n ]\n)\n\nprint(\"\\nModel ready!\")\nmodel.summary()\n\n\n# ============================================================\n# 8. MACRO F1 CALLBACK\n# ============================================================\n\nclass MacroF1Callback(\n tf.keras.callbacks.Callback\n):\n\n def __init__(self, validation_data):\n\n super().__init__()\n\n self.validation_data = validation_data\n\n self.best_f1 = 0.0\n\n\n def on_epoch_end(\n self,\n epoch,\n logs=None\n ):\n\n y_true = []\n y_pred = []\n\n for images, labels in self.validation_data:\n\n predictions = self.model.predict(\n images,\n verbose=0\n )\n\n predictions = np.argmax(\n predictions,\n axis=1\n )\n\n y_true.extend(\n labels.numpy()\n )\n\n y_pred.extend(\n predictions\n )\n\n\n score = f1_score(\n y_true,\n y_pred,\n average=\"macro\"\n )\n\n print(\n \"\\nšŸŒø Validation Macro F1:\",\n round(score, 5)\n )\n\n\n if score > self.best_f1:\n\n self.best_f1 = score\n\n self.model.save(\n \"/kaggle/working/best_flower_model.keras\"\n )\n\n print(\n \"ā­ New best model saved!\"\n )\n\n\nf1_callback = MacroF1Callback(\n val_ds\n)\n\n\n# ============================================================\n# 9. TRAIN CLASSIFICATION HEAD\n# ============================================================\n\nprint(\"\\n======================================\")\nprint(\"PHASE 1: TRAINING CLASSIFICATION HEAD\")\nprint(\"======================================\")\n\nhistory1 = model.fit(\n\n train_ds,\n\n validation_data=val_ds,\n\n epochs=EPOCHS_HEAD,\n\n class_weight=class_weight_dict,\n\n callbacks=[\n f1_callback,\n\n tf.keras.callbacks.ReduceLROnPlateau(\n monitor=\"val_loss\",\n factor=0.5,\n patience=1,\n min_lr=1e-6\n )\n ]\n)\n\n\n# ============================================================\n# 10. FINE-TUNE EFFICIENTNET\n# ============================================================\n\nprint(\"\\n======================================\")\nprint(\"PHASE 2: FINE-TUNING\")\nprint(\"======================================\")\n\n\nbase_model.trainable = True\n\n\n# Freeze early layers\n# Train only the later layers\n\nfor layer in base_model.layers[:-50]:\n\n layer.trainable = False\n\n\nmodel.compile(\n\n optimizer=tf.keras.optimizers.Adam(\n learning_rate=0.00005\n ),\n\n loss=\"sparse_categorical_crossentropy\",\n\n metrics=[\n \"accuracy\"\n ]\n)\n\n\nhistory2 = model.fit(\n\n train_ds,\n\n validation_data=val_ds,\n\n epochs=EPOCHS_FINE,\n\n class_weight=class_weight_dict,\n\n callbacks=[\n\n f1_callback,\n\n tf.keras.callbacks.ReduceLROnPlateau(\n monitor=\"val_loss\",\n factor=0.5,\n patience=1,\n min_lr=1e-7\n )\n ]\n)\n\n\n# ============================================================\n# 11. LOAD BEST MODEL\n# ============================================================\n\nprint(\"\\nLoading best model...\")\n\nmodel = tf.keras.models.load_model(\n \"/kaggle/working/best_flower_model.keras\"\n)\n\n\n# ============================================================\n# 12. FINAL VALIDATION F1\n# ============================================================\n\nprint(\"\\nCalculating final validation score...\")\n\ny_true = []\ny_pred = []\n\nfor images, labels in val_ds:\n\n predictions = model.predict(\n images,\n verbose=0\n )\n\n predictions = np.argmax(\n predictions,\n axis=1\n )\n\n y_true.extend(\n labels.numpy()\n )\n\n y_pred.extend(\n predictions\n )\n\n\nfinal_f1 = f1_score(\n y_true,\n y_pred,\n average=\"macro\"\n)\n\nprint(\n \"\\n======================================\"\n)\n\nprint(\n \"FINAL VALIDATION MACRO F1:\",\n round(final_f1, 5)\n)\n\nprint(\n \"======================================\"\n)\n\n\n# ============================================================\n# 13. READ TEST DATA\n# ============================================================\n\ndef read_test(example):\n\n features = {\n\n \"image\": tf.io.FixedLenFeature(\n [],\n tf.string\n ),\n\n \"id\": tf.io.FixedLenFeature(\n [],\n tf.string\n )\n }\n\n example = tf.io.parse_single_example(\n example,\n features\n )\n\n image = tf.image.decode_jpeg(\n example[\"image\"],\n channels=3\n )\n\n image = tf.image.resize(\n image,\n [IMG_SIZE, IMG_SIZE]\n )\n\n image = tf.cast(\n image,\n tf.float32\n )\n\n return image, example[\"id\"]\n\n\ntest_ds = tf.data.TFRecordDataset(\n TEST_FILES,\n num_parallel_reads=AUTOTUNE\n)\n\ntest_ds = test_ds.map(\n read_test,\n num_parallel_calls=AUTOTUNE\n)\n\ntest_ds = test_ds.batch(\n BATCH_SIZE\n)\n\ntest_ds = test_ds.prefetch(\n AUTOTUNE\n)\n\n\n# ============================================================\n# 14. TEST PREDICTIONS\n# ============================================================\n\nprint(\"\\nCreating test predictions...\")\n\nids = []\npredicted_labels = []\n\n\nfor images, image_ids in test_ds:\n\n probabilities = model.predict(\n images,\n verbose=0\n )\n\n labels = np.argmax(\n probabilities,\n axis=1\n )\n\n predicted_labels.extend(\n labels\n )\n\n ids.extend([\n\n x.decode(\"utf-8\")\n\n for x in image_ids.numpy()\n\n ])\n\n\n# ============================================================\n# 15. CREATE SUBMISSION\n# ============================================================\n\nsubmission = pd.DataFrame({\n\n \"id\": ids,\n\n \"label\": predicted_labels\n\n})\n\n\nsubmission.to_csv(\n\n \"/kaggle/working/submission.csv\",\n\n index=False\n\n)\n\n\nprint(\"\\n======================================\")\nprint(\"šŸŽ‰ SUBMISSION CREATED\")\nprint(\"======================================\")\n\nprint(\n \"Predictions:\",\n len(submission)\n)\n\nprint(\n \"\\nValidation Macro F1:\",\n round(final_f1, 5)\n)\n\nprint(\n \"\\nFirst 10 predictions:\"\n)\n\nprint(\n submission.head(10)\n)\n\nprint(\n \"\\nFile:\"\n)\n\nprint(\n \"/kaggle/working/submission.csv\"\n)\n\nprint(\n \"\\nREADY FOR KAGGLE SUBMISSION šŸš€\"\n)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2026-08-27T15:05:44.470698Z","iopub.execute_input":"2026-08-27T15:05:44.471417Z","iopub.status.idle":"2026-08-27T15:29:34.869717Z","shell.execute_reply.started":"2026-08-27T15:05:44.471391Z","shell.execute_reply":"2026-08-27T15:29:34.868903Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}