{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "import mlflow\n",
    "\n",
    "mlflow_server_uri = \"https://champi.heuzef.com\"\n",
    "\n",
    "mlflow.set_tracking_uri(mlflow_server_uri)\n",
    "mlflow.set_experiment(\"champi\") # Le nom du projet\n",
    "\n",
    "with mlflow.start_run():\n",
    "    for epoch, metrics in epochs_data.items():\n",
    "        mlflow.log_metric(\"accuracy\", metrics[\"accuracy\"], step=epoch)\n",
    "        mlflow.log_metric(\"loss\", metrics[\"loss\"], step=epoch)\n",
    "        mlflow.log_metric(\"val_accuracy\", metrics[\"val_accuracy\"], step=epoch)\n",
    "        mlflow.log_metric(\"val_loss\", metrics[\"val_loss\"], step=epoch)\n",
    "        mlflow.log_metric(\"learning_rate\", metrics[\"learning_rate\"], step=epoch)\n",
    "    mlflow.log_param(\"model\", \"resnet50_callbacks_5layers_unfreeze\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "epochs_data = {\n",
    "    1: {\n",
    "        \"accuracy\": 0.0847,\n",
    "        \"loss\": 5.8520,\n",
    "        \"val_accuracy\": 0.4417,\n",
    "        \"val_loss\": 3.9198,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    2: {\n",
    "        \"accuracy\": 0.1360,\n",
    "        \"loss\": 4.2380,\n",
    "        \"val_accuracy\": 0.5399,\n",
    "        \"val_loss\": 3.6218,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    3: {\n",
    "        \"accuracy\": 0.2402,\n",
    "        \"loss\": 3.9099,\n",
    "        \"val_accuracy\": 0.6810,\n",
    "        \"val_loss\": 3.2485,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    4: {\n",
    "        \"accuracy\": 0.3660,\n",
    "        \"loss\": 3.5690,\n",
    "        \"val_accuracy\": 0.7362,\n",
    "        \"val_loss\": 2.7518,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    5: {\n",
    "        \"accuracy\": 0.4996,\n",
    "        \"loss\": 3.1509,\n",
    "        \"val_accuracy\": 0.7730,\n",
    "        \"val_loss\": 2.3880,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    6: {\n",
    "        \"accuracy\": 0.6404,\n",
    "        \"loss\": 2.7446,\n",
    "        \"val_accuracy\": 0.7914,\n",
    "        \"val_loss\": 2.2776,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    7: {\n",
    "        \"accuracy\": 0.7238,\n",
    "        \"loss\": 2.4671,\n",
    "        \"val_accuracy\": 0.8160,\n",
    "        \"val_loss\": 2.2215,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    8: {\n",
    "        \"accuracy\": 0.7835,\n",
    "        \"loss\": 2.2709,\n",
    "        \"val_accuracy\": 0.8221,\n",
    "        \"val_loss\": 2.1896,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    9: {\n",
    "        \"accuracy\": 0.8369,\n",
    "        \"loss\": 2.0898,\n",
    "        \"val_accuracy\": 0.8344,\n",
    "        \"val_loss\": 2.1191,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    10: {\n",
    "        \"accuracy\": 0.8759,\n",
    "        \"loss\": 1.9587,\n",
    "        \"val_accuracy\": 0.8589,\n",
    "        \"val_loss\": 2.1327,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    11: {\n",
    "        \"accuracy\": 0.9061,\n",
    "        \"loss\": 1.8641,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 2.0924,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    12: {\n",
    "        \"accuracy\": 0.9285,\n",
    "        \"loss\": 1.7836,\n",
    "        \"val_accuracy\": 0.8405,\n",
    "        \"val_loss\": 2.1150,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    13: {\n",
    "        \"accuracy\": 0.9438,\n",
    "        \"loss\": 1.7205,\n",
    "        \"val_accuracy\": 0.8466,\n",
    "        \"val_loss\": 2.0380,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    14: {\n",
    "        \"accuracy\": 0.9596,\n",
    "        \"loss\": 1.6534,\n",
    "        \"val_accuracy\": 0.8405,\n",
    "        \"val_loss\": 2.0600,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    15: {\n",
    "        \"accuracy\": 0.9639,\n",
    "        \"loss\": 1.6170,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 2.0319,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    16: {\n",
    "        \"accuracy\": 0.9740,\n",
    "        \"loss\": 1.5687,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 2.0340,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    17: {\n",
    "        \"accuracy\": 0.9745,\n",
    "        \"loss\": 1.5440,\n",
    "        \"val_accuracy\": 0.8405,\n",
    "        \"val_loss\": 2.0015,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    18: {\n",
    "        \"accuracy\": 0.9821,\n",
    "        \"loss\": 1.5040,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.9666,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    19: {\n",
    "        \"accuracy\": 0.9877,\n",
    "        \"loss\": 1.4689,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.9311,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    20: {\n",
    "        \"accuracy\": 0.9911,\n",
    "        \"loss\": 1.4414,\n",
    "        \"val_accuracy\": 0.8773,\n",
    "        \"val_loss\": 1.9213,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    21: {\n",
    "        \"accuracy\": 0.9876,\n",
    "        \"loss\": 1.4215,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 1.9028,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    22: {\n",
    "        \"accuracy\": 0.9938,\n",
    "        \"loss\": 1.3837,\n",
    "        \"val_accuracy\": 0.8773,\n",
    "        \"val_loss\": 1.8655,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    23: {\n",
    "        \"accuracy\": 0.9908,\n",
    "        \"loss\": 1.3609,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 1.9050,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    24: {\n",
    "        \"accuracy\": 0.9949,\n",
    "        \"loss\": 1.3233,\n",
    "        \"val_accuracy\": 0.8589,\n",
    "        \"val_loss\": 1.8978,\n",
    "        \"learning_rate\": 1.0000e-05\n",
    "    },\n",
    "    25: {\n",
    "        \"accuracy\": 0.9954,\n",
    "        \"loss\": 1.2955,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.8539,\n",
    "        \"learning_rate\": 2.0000e-06\n",
    "    },\n",
    "    26: {\n",
    "        \"accuracy\": 0.9962,\n",
    "        \"loss\": 1.2885,\n",
    "        \"val_accuracy\": 0.8712,\n",
    "        \"val_loss\": 1.8314,\n",
    "        \"learning_rate\": 2.0000e-06\n",
    "    },\n",
    "    27: {\n",
    "        \"accuracy\": 0.9958,\n",
    "        \"loss\": 1.2817,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.8324,\n",
    "        \"learning_rate\": 2.0000e-06\n",
    "    },\n",
    "    28: {\n",
    "        \"accuracy\": 0.9976,\n",
    "        \"loss\": 1.2710,\n",
    "        \"val_accuracy\": 0.8589,\n",
    "        \"val_loss\": 1.8316,\n",
    "        \"learning_rate\": 2.0000e-06\n",
    "    },\n",
    "    29: {\n",
    "        \"accuracy\": 0.9963,\n",
    "        \"loss\": 1.2652,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.8263,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    30: {\n",
    "        \"accuracy\": 0.9975,\n",
    "        \"loss\": 1.2591,\n",
    "        \"val_accuracy\": 0.8712,\n",
    "        \"val_loss\": 1.8105,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    31: {\n",
    "        \"accuracy\": 0.9979,\n",
    "        \"loss\": 1.2559,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.8199,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    32: {\n",
    "        \"accuracy\": 0.9978,\n",
    "        \"loss\": 1.2485,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.8312,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    33: {\n",
    "        \"accuracy\": 0.9972,\n",
    "        \"loss\": 1.2443,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 1.8201,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    34: {\n",
    "        \"accuracy\": 0.9980,\n",
    "        \"loss\": 1.2402,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.8054,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    35: {\n",
    "        \"accuracy\": 0.9987,\n",
    "        \"loss\": 1.2324,\n",
    "        \"val_accuracy\": 0.8528,\n",
    "        \"val_loss\": 1.7969,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    36: {\n",
    "        \"accuracy\": 0.9985,\n",
    "        \"loss\": 1.2261,\n",
    "        \"val_accuracy\": 0.8589,\n",
    "        \"val_loss\": 1.7965,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    37: {\n",
    "        \"accuracy\": 0.9986,\n",
    "        \"loss\": 1.2198,\n",
    "        \"val_accuracy\": 0.8589,\n",
    "        \"val_loss\": 1.7878,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    38: {\n",
    "        \"accuracy\": 0.9990,\n",
    "        \"loss\": 1.2131,\n",
    "        \"val_accuracy\": 0.8650,\n",
    "        \"val_loss\": 1.7906,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    },\n",
    "    39: {\n",
    "        \"accuracy\": 0.9977,\n",
    "        \"loss\": 1.2106,\n",
    "        \"val_accuracy\": 0.8589,\n",
    "        \"val_loss\": 1.7810,\n",
    "        \"learning_rate\": 1.0000e-06\n",
    "    }\n",
    "}\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "import tensorflow as tf\n",
    "from tensorflow.keras.applications import ResNet50\n",
    "from tensorflow.keras.layers import Dense, GlobalAveragePooling2D, Dropout\n",
    "from tensorflow.keras.models import Model\n",
    "from tensorflow.keras.optimizers import Adam\n",
    "from tensorflow.keras.regularizers import l2\n",
    "from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau\n",
    "\n",
    "# Define paths to directories\n",
    "train_dir = 'C:/Users/vsavelev/GITHUB/DS_projet/LAYER2/MO/train'\n",
    "validation_dir = 'C:/Users/vsavelev/GITHUB/DS_projet/LAYER2/MO/validation'\n",
    "\n",
    "# Create Dataset from directories\n",
    "def create_dataset(directory, batch_size):\n",
    "    dataset = tf.keras.preprocessing.image_dataset_from_directory(\n",
    "        directory,\n",
    "        image_size=(224, 224),\n",
    "        batch_size=batch_size,\n",
    "        label_mode='categorical',  # Labels are one-hot encoded\n",
    "        shuffle=True\n",
    "    )\n",
    "    return dataset\n",
    "\n",
    "# Load data from directories\n",
    "batch_size = 32\n",
    "train_dataset = create_dataset(train_dir, batch_size)\n",
    "validation_dataset = create_dataset(validation_dir, batch_size)\n",
    "\n",
    "# Build ResNet50 Model\n",
    "base_model = ResNet50(weights='imagenet', include_top=False, input_shape=(224, 224, 3))\n",
    "x = base_model.output\n",
    "x = GlobalAveragePooling2D()(x)\n",
    "x = Dropout(0.7)(x)\n",
    "x = Dense(1024, activation='relu', kernel_regularizer=l2(0.001))(x)\n",
    "x = Dropout(0.7)(x)\n",
    "predictions = Dense(len(train_dataset.class_names), activation='softmax', kernel_regularizer=l2(0.01))(x)\n",
    "\n",
    "model = Model(inputs=base_model.input, outputs=predictions)\n",
    "\n",
    "# Unfreeze the last few layers\n",
    "for layer in base_model.layers[-5:]:\n",
    "    layer.trainable = True\n",
    "\n",
    "# Compile the model\n",
    "model.compile(optimizer=Adam(learning_rate=1e-5),\n",
    "              loss='categorical_crossentropy',\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# Callbacks for early stopping and learning rate reduction\n",
    "early_stopping = EarlyStopping(\n",
    "    monitor='val_loss',  \n",
    "    patience=3,          \n",
    "    restore_best_weights=True  \n",
    ")\n",
    "\n",
    "reduce_lr = ReduceLROnPlateau(\n",
    "    monitor='val_loss',   \n",
    "    factor=0.2,           \n",
    "    patience=2,           \n",
    "    min_lr=1e-6           \n",
    ")\n",
    "\n",
    "# Custom callback to save metrics per epoch to a JSON file\n",
    "class MetricsLogger(tf.keras.callbacks.Callback):\n",
    "    def __init__(self, log_file='training_metrics.json'):\n",
    "        super().__init__()\n",
    "        self.log_file = log_file\n",
    "        self.logs = []\n",
    "\n",
    "    def on_epoch_end(self, epoch, logs=None):\n",
    "        self.logs.append({\n",
    "            'epoch': epoch,\n",
    "            'loss': logs['loss'],\n",
    "            'accuracy': logs['accuracy'],\n",
    "            'val_loss': logs['val_loss'],\n",
    "            'val_accuracy': logs['val_accuracy']\n",
    "        })\n",
    "        with open(self.log_file, 'w') as f:\n",
    "            json.dump(self.logs, f, indent=4)\n",
    "\n",
    "# Train the model and log metrics\n",
    "metrics_logger = MetricsLogger(log_file='training_metrics.json')\n",
    "\n",
    "history = model.fit(\n",
    "    train_dataset,\n",
    "    epochs=50,\n",
    "    validation_data=validation_dataset,\n",
    "    callbacks=[reduce_lr, early_stopping, metrics_logger]\n",
    ")\n",
    "\n",
    "# Save the model\n",
    "model.save('resnet50_model_callbacks.h5')\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import mlflow\n",
    "import mlflow.keras\n",
    "import json\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Function to load metrics from JSON\n",
    "def load_metrics_from_json(log_file='training_metrics.json'):\n",
    "    with open(log_file, 'r') as f:\n",
    "        return json.load(f)\n",
    "\n",
    "# Function to plot training history\n",
    "def plot_training_history(metrics):\n",
    "    epochs = [m['epoch'] for m in metrics]\n",
    "    train_accuracy = [m['accuracy'] for m in metrics]\n",
    "    val_accuracy = [m['val_accuracy'] for m in metrics]\n",
    "    train_loss = [m['loss'] for m in metrics]\n",
    "    val_loss = [m['val_loss'] for m in metrics]\n",
    "    \n",
    "    plt.figure(figsize=(12, 8))\n",
    "    \n",
    "    # Plot training & validation accuracy values\n",
    "    plt.subplot(1, 2, 1)\n",
    "    plt.plot(epochs, train_accuracy)\n",
    "    plt.plot(epochs, val_accuracy)\n",
    "    plt.title('Model accuracy')\n",
    "    plt.ylabel('Accuracy')\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.legend(['Train', 'Validation'], loc='upper left')\n",
    "\n",
    "    # Plot training & validation loss values\n",
    "    plt.subplot(1, 2, 2)\n",
    "    plt.plot(epochs, train_loss)\n",
    "    plt.plot(epochs, val_loss)\n",
    "    plt.title('Model loss')\n",
    "    plt.ylabel('Loss')\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.legend(['Train', 'Validation'], loc='upper left')\n",
    "\n",
    "    plt.savefig(\"training_history.png\")\n",
    "    plt.show()\n",
    "\n",
    "# Load metrics and log to MLflow\n",
    "with mlflow.start_run():\n",
    "    # Load metrics from JSON\n",
    "    metrics = load_metrics_from_json('training_metrics.json')\n",
    "\n",
    "    # Log metrics for each epoch\n",
    "    for m in metrics:\n",
    "        mlflow.log_metric('train_loss', m['loss'], step=m['epoch'])\n",
    "        mlflow.log_metric('train_accuracy', m['accuracy'], step=m['epoch'])\n",
    "        mlflow.log_metric('val_loss', m['val_loss'], step=m['epoch'])\n",
    "        mlflow.log_metric('val_accuracy', m['val_accuracy'], step=m['epoch'])\n",
    "\n",
    "    # Log hyperparameters\n",
    "    mlflow.log_param(\"learning_rate\", 1e-5)\n",
    "    mlflow.log_param(\"batch_size\", 32)\n",
    "    mlflow.log_param(\"dropout_rate\", 0.7)\n",
    "    mlflow.log_param(\"l2_regularization\", 0.01)\n",
    "    \n",
    "    # Plot and log training history\n",
    "    plot_training_history(metrics)\n",
    "    mlflow.log_artifact(\"training_history.png\")\n",
    "\n",
    "    # Log the trained model\n",
    "    mlflow.keras.log_model(tf.keras.models.load_model('resnet50_model_callbacks.h5'), \"resnet50_model\")\n",
    "\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "base",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}