basil_bird_or_plane.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "pip install virtualenv"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!virtualenv myenv\n",
    "!source myenv/bin/activate\n",
    "!pip install tensorflow"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import ssl\n",
    "ssl._create_default_https_context = ssl._create_unverified_context"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## CIFAR-10 Dataset\n",
    "\n",
    "The CIFAR-10 dataset offers a great starting point for exploring anomaly detection techniques. It is a well-structured, labeled image dataset that can be adapted for various anomaly detection scenarios. Let's delve into how we can utilize CIFAR-10 for this purpose:\n",
    "\n",
    "* CIFAR-10 consists of 60,000 32x32 color images in 10 classes (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck), with 6,000 images per class.\n",
    "* The dataset is divided into 50,000 training images and 10,000 test images.\n",
    "\n",
    "We can transform CIFAR-10, originally designed for image classification, into an anomaly detection problem using these strategies:\n",
    "\n",
    "1. **Single-Class Anomaly Detection:**\n",
    "   * Choose one class as \"normal\" and treat all other classes as \"anomalies.\"\n",
    "   * Train a model on the \"normal\" class to learn its distribution.\n",
    "   * During testing, the model identifies instances that deviate significantly from the learned \"normal\" distribution as anomalies.\n",
    "\n",
    "2. **Multi-Class Anomaly Detection:**\n",
    "   * Select a few classes as \"normal\" and the rest as \"anomalies.\"\n",
    "   * Train a model on the \"normal\" classes.\n",
    "   * Identify instances from the \"anomalous\" classes as deviations from the learned \"normal\" pattern.\n",
    "\n",
    "## **Techniques for Anomaly Detection with CIFAR-10:**\n",
    "\n",
    "* **Autoencoders:** Train an autoencoder on the \"normal\" class images. Anomalies will have higher reconstruction errors due to their deviation from the learned representation.\n",
    "* **One-Class SVM:** Train a One-Class SVM on the \"normal\" class features. Instances outside the learned boundary are considered anomalies.\n",
    "* **Isolation Forest:** Build an Isolation Forest on the \"normal\" class features. Anomalies are isolated with fewer splits compared to \"normal\" instances.\n",
    "\n",
    "## **Implementation Strategy**\n",
    "\n",
    "1. **Load CIFAR-10 Dataset:** Utilize libraries like TensorFlow or PyTorch to load the CIFAR-10 dataset.\n",
    "2. **Preprocess Data:** Normalize pixel values and potentially apply data augmentation techniques.\n",
    "3. **Select Normal and Anomalous Classes:** Choose the classes based on your anomaly detection scenario.\n",
    "4. **Train Anomaly Detection Model:** Implement your chosen technique and train it on the \"normal\" class data.\n",
    "5. **Evaluate on Test Data:** Assess the model's performance in identifying anomalies from the \"anomalous\" classes using metrics like precision, recall, and F1-score.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "from tensorflow.keras.models import Model\n",
    "from tensorflow.keras.layers import Input, Dense, Flatten, Reshape\n",
    "from tensorflow.keras.preprocessing.image import ImageDataGenerator\n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Load the CIFAR-10 dataset\n",
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Normalize pixel values to be between 0 and 1\n",
    "x_train, x_test = x_train / 255.0, x_test / 255.0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Filter data for 'bird' class (label 2) as normal and 'airplane' class (label 0) as anomaly\n",
    "x_train_normal = x_train[y_train.flatten() == 2]\n",
    "x_test_normal = x_test[y_test.flatten() == 2]\n",
    "x_test_anomaly = x_test[y_test.flatten() == 0]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "x_train_normal shape: (5000, 32, 32, 3)\n",
      "x_test_normal shape: (1000, 32, 32, 3)\n",
      "x_test_anomaly shape: (1000, 32, 32, 3)\n"
     ]
    }
   ],
   "source": [
    "# Print shapes of filtered data\n",
    "print(\"x_train_normal shape:\", x_train_normal.shape)\n",
    "print(\"x_test_normal shape:\", x_test_normal.shape)\n",
    "print(\"x_test_anomaly shape:\", x_test_anomaly.shape)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Train Anomaly Detection Model\n",
    "\n",
    "1. **Import Necessary Layers:** We import `Model`, `Input`, `Dense`, `Flatten`, and `Reshape` from `tensorflow.keras.models` and `tensorflow.keras.layers`.\n",
    "\n",
    "2. **Define Input Shape:** We get the shape of the input images from `x_train_normal`.\n",
    "\n",
    "3. **Encoder:**\n",
    "   * `Input`: Defines the input layer with the specified shape.\n",
    "   * `Flatten`: Flattens the input image into a single-dimensional vector.\n",
    "   * `Dense`: Two dense layers with ReLU activation for dimensionality reduction. The last layer (`encoded`) represents the compressed representation of the input.\n",
    "\n",
    "4. **Decoder:**\n",
    "   * `Dense`: Two dense layers with ReLU activation to reconstruct the original dimensions.\n",
    "   * `Dense`: Final dense layer with sigmoid activation to output pixel values between 0 and 1.\n",
    "   * `Reshape`: Reshapes the output back to the original image dimensions.\n",
    "\n",
    "5. **Autoencoder Model:** We create the autoencoder model using the `Model` class, taking `input_img` as input and `decoded` as output.\n",
    "\n",
    "6. **Compile:** We compile the model with the Adam optimizer and mean squared error (MSE) loss. MSE measures the difference between the original and reconstructed images.\n",
    "\n",
    "7. **Train:** We train the autoencoder on the `x_train_normal` data for 50 epochs with a batch size of 256. We also use the `x_test_normal` data for validation during training.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m1s\u001b[0m 20ms/step - loss: 0.0539 - val_loss: 0.0493\n",
      "Epoch 2/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0464 - val_loss: 0.0402\n",
      "Epoch 3/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0387 - val_loss: 0.0356\n",
      "Epoch 4/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0337 - val_loss: 0.0323\n",
      "Epoch 5/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0331 - val_loss: 0.0301\n",
      "Epoch 6/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0298 - val_loss: 0.0291\n",
      "Epoch 7/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0277 - val_loss: 0.0263\n",
      "Epoch 8/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0255 - val_loss: 0.0255\n",
      "Epoch 9/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0240 - val_loss: 0.0248\n",
      "Epoch 10/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0242 - val_loss: 0.0243\n",
      "Epoch 11/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0239 - val_loss: 0.0241\n",
      "Epoch 12/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0225 - val_loss: 0.0229\n",
      "Epoch 13/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0221 - val_loss: 0.0225\n",
      "Epoch 14/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0214 - val_loss: 0.0219\n",
      "Epoch 15/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0209 - val_loss: 0.0214\n",
      "Epoch 16/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0204 - val_loss: 0.0214\n",
      "Epoch 17/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0204 - val_loss: 0.0206\n",
      "Epoch 18/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0197 - val_loss: 0.0203\n",
      "Epoch 19/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0193 - val_loss: 0.0203\n",
      "Epoch 20/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0192 - val_loss: 0.0207\n",
      "Epoch 21/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0193 - val_loss: 0.0198\n",
      "Epoch 22/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0189 - val_loss: 0.0195\n",
      "Epoch 23/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0186 - val_loss: 0.0196\n",
      "Epoch 24/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 20ms/step - loss: 0.0184 - val_loss: 0.0198\n",
      "Epoch 25/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0186 - val_loss: 0.0194\n",
      "Epoch 26/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0183 - val_loss: 0.0191\n",
      "Epoch 27/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0185 - val_loss: 0.0190\n",
      "Epoch 28/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0183 - val_loss: 0.0188\n",
      "Epoch 29/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0182 - val_loss: 0.0185\n",
      "Epoch 30/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0175 - val_loss: 0.0183\n",
      "Epoch 31/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0177 - val_loss: 0.0183\n",
      "Epoch 32/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0174 - val_loss: 0.0183\n",
      "Epoch 33/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0175 - val_loss: 0.0182\n",
      "Epoch 34/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0173 - val_loss: 0.0184\n",
      "Epoch 35/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0177 - val_loss: 0.0185\n",
      "Epoch 36/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0176 - val_loss: 0.0182\n",
      "Epoch 37/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0173 - val_loss: 0.0183\n",
      "Epoch 38/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0172 - val_loss: 0.0180\n",
      "Epoch 39/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0172 - val_loss: 0.0180\n",
      "Epoch 40/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0171 - val_loss: 0.0184\n",
      "Epoch 41/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0170 - val_loss: 0.0179\n",
      "Epoch 42/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0169 - val_loss: 0.0184\n",
      "Epoch 43/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0170 - val_loss: 0.0177\n",
      "Epoch 44/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0164 - val_loss: 0.0172\n",
      "Epoch 45/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0163 - val_loss: 0.0189\n",
      "Epoch 46/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 23ms/step - loss: 0.0170 - val_loss: 0.0170\n",
      "Epoch 47/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0158 - val_loss: 0.0169\n",
      "Epoch 48/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0159 - val_loss: 0.0168\n",
      "Epoch 49/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0157 - val_loss: 0.0171\n",
      "Epoch 50/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 20ms/step - loss: 0.0161 - val_loss: 0.0169\n",
      "Epoch 51/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 21ms/step - loss: 0.0160 - val_loss: 0.0167\n",
      "Epoch 52/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 23ms/step - loss: 0.0155 - val_loss: 0.0166\n",
      "Epoch 53/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0156 - val_loss: 0.0169\n",
      "Epoch 54/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0156 - val_loss: 0.0166\n",
      "Epoch 55/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 20ms/step - loss: 0.0157 - val_loss: 0.0166\n",
      "Epoch 56/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 22ms/step - loss: 0.0156 - val_loss: 0.0178\n",
      "Epoch 57/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0157 - val_loss: 0.0164\n",
      "Epoch 58/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0153 - val_loss: 0.0164\n",
      "Epoch 59/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0153 - val_loss: 0.0163\n",
      "Epoch 60/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0153 - val_loss: 0.0166\n",
      "Epoch 61/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0151 - val_loss: 0.0161\n",
      "Epoch 62/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 20ms/step - loss: 0.0147 - val_loss: 0.0159\n",
      "Epoch 63/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 20ms/step - loss: 0.0149 - val_loss: 0.0165\n",
      "Epoch 64/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0149 - val_loss: 0.0158\n",
      "Epoch 65/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0146 - val_loss: 0.0158\n",
      "Epoch 66/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0145 - val_loss: 0.0158\n",
      "Epoch 67/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0146 - val_loss: 0.0158\n",
      "Epoch 68/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 21ms/step - loss: 0.0145 - val_loss: 0.0157\n",
      "Epoch 69/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0145 - val_loss: 0.0159\n",
      "Epoch 70/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0145 - val_loss: 0.0159\n",
      "Epoch 71/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0144 - val_loss: 0.0158\n",
      "Epoch 72/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0144 - val_loss: 0.0156\n",
      "Epoch 73/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0144 - val_loss: 0.0158\n",
      "Epoch 74/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0144 - val_loss: 0.0159\n",
      "Epoch 75/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0143 - val_loss: 0.0157\n",
      "Epoch 76/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0143 - val_loss: 0.0156\n",
      "Epoch 77/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0141 - val_loss: 0.0156\n",
      "Epoch 78/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0140 - val_loss: 0.0156\n",
      "Epoch 79/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0143 - val_loss: 0.0158\n",
      "Epoch 80/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0141 - val_loss: 0.0159\n",
      "Epoch 81/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0144 - val_loss: 0.0155\n",
      "Epoch 82/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0139 - val_loss: 0.0154\n",
      "Epoch 83/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0139 - val_loss: 0.0153\n",
      "Epoch 84/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0137 - val_loss: 0.0153\n",
      "Epoch 85/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0138 - val_loss: 0.0158\n",
      "Epoch 86/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0139 - val_loss: 0.0154\n",
      "Epoch 87/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0138 - val_loss: 0.0153\n",
      "Epoch 88/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0139 - val_loss: 0.0152\n",
      "Epoch 89/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0136 - val_loss: 0.0153\n",
      "Epoch 90/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0139 - val_loss: 0.0156\n",
      "Epoch 91/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0144 - val_loss: 0.0155\n",
      "Epoch 92/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 17ms/step - loss: 0.0137 - val_loss: 0.0152\n",
      "Epoch 93/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0133 - val_loss: 0.0151\n",
      "Epoch 94/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0132 - val_loss: 0.0155\n",
      "Epoch 95/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0136 - val_loss: 0.0151\n",
      "Epoch 96/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0135 - val_loss: 0.0150\n",
      "Epoch 97/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0133 - val_loss: 0.0150\n",
      "Epoch 98/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0132 - val_loss: 0.0149\n",
      "Epoch 99/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 19ms/step - loss: 0.0134 - val_loss: 0.0150\n",
      "Epoch 100/100\n",
      "\u001b[1m20/20\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 18ms/step - loss: 0.0131 - val_loss: 0.0149\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.src.callbacks.history.History at 0x40cbc0e90>"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from tensorflow.keras.models import Model\n",
    "from tensorflow.keras.layers import Input, Dense, Flatten, Reshape\n",
    "\n",
    "# Define the input shape\n",
    "input_shape = x_train_normal.shape[1:]\n",
    "\n",
    "# Encoder\n",
    "input_img = Input(shape=input_shape)\n",
    "x = Flatten()(input_img)\n",
    "x = Dense(512, activation='relu')(x)  # Increased number of neurons\n",
    "x = Dense(256, activation='relu')(x)\n",
    "x = Dense(128, activation='relu')(x)\n",
    "encoded = Dense(32, activation='relu')(x)  # Another layer added\n",
    "\n",
    "# Decoder\n",
    "x = Dense(128, activation='relu')(encoded)  # Another layer added\n",
    "x = Dense(256, activation='relu')(x)\n",
    "x = Dense(512, activation='relu')(x)  # Increased number of neurons\n",
    "x = Dense(np.prod(input_shape), activation='sigmoid')(x)\n",
    "decoded = Reshape(input_shape)(x)\n",
    "\n",
    "# Autoencoder model\n",
    "autoencoder = Model(input_img, decoded)\n",
    "autoencoder.compile(optimizer='adam', loss='mse')\n",
    "\n",
    "# # Data augmentation (optional, you can experiment with this)\n",
    "# datagen = ImageDataGenerator(\n",
    "#     rotation_range=10,\n",
    "#     width_shift_range=0.1,\n",
    "#     height_shift_range=0.1,\n",
    "#     horizontal_flip=True,\n",
    "#     brightness_range=[0.8, 1.2],\n",
    "# )\n",
    "# datagen.fit(x_train_normal)\n",
    "\n",
    "# Train the autoencoder\n",
    "autoencoder.fit(x_train_normal, x_train_normal, batch_size=256,\n",
    "                epochs=100,\n",
    "                shuffle=True,\n",
    "                validation_data=(x_test_normal, x_test_normal))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Evaluate on Test Data\n",
    "\n",
    "Now that we've trained the autoencoder, we can:\n",
    "\n",
    "1. **Evaluate on Test Data:** Use the trained autoencoder to reconstruct both normal ('airplane') and anomalous images from the test set.\n",
    "2. **Calculate Reconstruction Errors:** Compute the reconstruction errors (MSE) for each image.\n",
    "3. **Set Threshold:** Determine a threshold value for the reconstruction error. Images with errors above this threshold will be classified as anomalies.\n",
    "4. **Assess Performance:** Evaluate the model's anomaly detection performance using metrics like precision, recall, and F1-score."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\u001b[1m63/63\u001b[0m \u001b[32m━━━━━━━━━━━━━━━━━━━━\u001b[0m\u001b[37m\u001b[0m \u001b[1m0s\u001b[0m 2ms/step\n",
      "Accuracy: 0.552\n",
      "Precision: 0.5368271954674221\n",
      "Recall: 0.758\n",
      "F1-Score: 0.6285240464344942\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "\n",
    "# Get all test data (normal and anomalous)\n",
    "x_test_all = np.concatenate((x_test_normal, x_test_anomaly), axis=0)\n",
    "\n",
    "# Reconstruct all test images\n",
    "decoded_imgs = autoencoder.predict(x_test_all)\n",
    "\n",
    "# Calculate reconstruction errors (MSE)\n",
    "mse = np.mean(np.power(x_test_all - decoded_imgs, 2), axis=(1, 2, 3))\n",
    "\n",
    "# Set threshold\n",
    "threshold = 0.01\n",
    "\n",
    "# Classify as anomaly if reconstruction error is above threshold\n",
    "y_pred = (mse > threshold).astype(int)\n",
    "\n",
    "# Get true labels (1 for airplanes (anomalies), 0 for birds)\n",
    "y_true = np.concatenate((np.zeros(x_test_normal.shape[0]), np.ones(x_test_anomaly.shape[0])), axis=0)\n",
    "\n",
    "# Evaluate\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "precision = precision_score(y_true, y_pred)\n",
    "recall = recall_score(y_true, y_pred)\n",
    "f1 = f1_score(y_true, y_pred)\n",
    "\n",
    "print(\"Accuracy:\", accuracy)\n",
    "print(\"Precision:\", precision)\n",
    "print(\"Recall:\", recall)\n",
    "print(\"F1-Score:\", f1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1000x400 with 10 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Visualize a few samples of birds and airplanes\n",
    "fig, axes = plt.subplots(2, 5, figsize=(10, 4))\n",
    "for i in range(5):\n",
    "    axes[0, i].imshow(x_test_normal[i])\n",
    "    axes[0, i].axis('off')\n",
    "    axes[0, i].set_title(\"Bird (Normal)\")\n",
    "\n",
    "    axes[1, i].imshow(x_test_anomaly[i])\n",
    "    axes[1, i].axis('off')\n",
    "    axes[1, i].set_title(\"Airplane (Anomaly)\")\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\"birds\" = the normal class \n",
    "\"airplanes\" = the anomalies\n",
    "\n",
    "**Accuracy: 0.4995** \n",
    "\n",
    "This indicates that the model correctly classified approximately 49.95% of the images in the test set as either normal (birds) or anomalous (airplanes). This accuracy is close to random guessing (50%), suggesting that the model might be struggling to differentiate between the two classes effectively.\n",
    "\n",
    "**Precision: 0.4997**\n",
    "\n",
    "This signifies that out of all the images the model predicted as anomalies (airplanes), around 49.97% were actually airplanes. This low precision suggests a high rate of false positives, meaning the model is incorrectly flagging many birds as airplanes.\n",
    "\n",
    "**Recall: 0.999** \n",
    "\n",
    "This shows that the model was able to correctly identify 99.9% of all the actual anomalies (airplanes) in the test set. This is a very high recall, indicating the model is highly sensitive to detecting airplanes.\n",
    "\n",
    "**F1-Score: 0.666**\n",
    "\n",
    "The F1-score, being the harmonic mean of precision and recall, provides a balanced measure of the model's performance. In this case, the relatively low F1-score (compared to a perfect score of 1) reflects the imbalance between the high recall and the low precision.\n",
    "\n",
    "**Overall Assessment:**\n",
    "\n",
    "The results suggest that the model is very good at detecting airplanes (anomalies) as evidenced by the high recall. However, it struggles to distinguish them from birds (normal class), leading to a high number of false positives and consequently a low precision. The overall accuracy is also poor due to this imbalance.\n",
    "\n",
    "**Possible Reasons and Improvements:**\n",
    "\n",
    "* **Model Complexity:** The autoencoder architecture might be too simple to capture the subtle differences between birds and airplanes. Consider increasing the complexity by adding more layers or neurons.\n",
    "* **Data Imbalance:** The CIFAR-10 dataset has an equal number of images per class. This might lead to the model being biased towards the majority class (birds in this case). You could try techniques like oversampling the minority class (airplanes) or undersampling the majority class during training.\n",
    "* **Hyperparameter Tuning:** Experiment with different hyperparameters like learning rate, batch size, and the number of epochs to see if you can improve the model's performance.\n",
    "* **Threshold Adjustment:** The current threshold might be too low, leading to many false positives. Try increasing the threshold to reduce false alarms, but be mindful of the impact on recall.\n",
    "* **Feature Engineering:** Explore additional features or preprocessing techniques that might help the model better distinguish between birds and airplanes.\n",
    "\n",
    "Anomaly detection can be challenging, especially when the normal and anomalous classes share visual similarities. It often requires careful experimentation and fine-tuning to achieve satisfactory results."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 94,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "from sklearn.svm import OneClassSVM\n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "\n",
    "# Load the CIFAR-10 dataset\n",
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\n",
    "\n",
    "# Normalize pixel values to be between 0 and 1\n",
    "x_train, x_test = x_train / 255.0, x_test / 255.0\n",
    "\n",
    "# Filter data for 'bird' class (label 2) as normal and 'airplane' class (label 0) as anomaly\n",
    "x_train_normal = x_train[y_train.flatten() == 2]\n",
    "x_test_normal = x_test[y_test.flatten() == 2]\n",
    "x_test_anomaly = x_test[y_test.flatten() == 4]\n",
    "\n",
    "# Flatten the images for One-Class SVM\n",
    "x_train_normal_flat = x_train_normal.reshape(x_train_normal.shape[0], -1)\n",
    "x_test_all_flat = np.concatenate((x_test_normal, x_test_anomaly), axis=0).reshape(-1, 32*32*3)\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 95,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.4585\n",
      "Precision: 0.4546448087431694\n",
      "Recall: 0.416\n",
      "F1-Score: 0.43446475195822454\n"
     ]
    }
   ],
   "source": [
    "# Initialize and train the One-Class SVM\n",
    "ocsvm = OneClassSVM(kernel=\"rbf\")  # You might need to tune nu and gamma\n",
    "ocsvm.fit(x_train_normal_flat)\n",
    "\n",
    "# Predict on the test data\n",
    "y_pred = ocsvm.predict(x_test_all_flat)\n",
    "\n",
    "# Convert predictions to binary labels (1 for anomaly, -1 for normal)\n",
    "y_pred = np.where(y_pred == 1, 0, 1) \n",
    "\n",
    "# Get true labels (1 for airplanes (anomalies), 0 for birds)\n",
    "y_true = np.concatenate((np.zeros(x_test_normal.shape[0]), np.ones(x_test_anomaly.shape[0])), axis=0)\n",
    "\n",
    "# Evaluate\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "precision = precision_score(y_true, y_pred)\n",
    "recall = recall_score(y_true, y_pred)\n",
    "f1 = f1_score(y_true, y_pred)\n",
    "\n",
    "print(\"Accuracy:\", accuracy)\n",
    "print(\"Precision:\", precision)\n",
    "print(\"Recall:\", recall)\n",
    "print(\"F1-Score:\", f1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Isolation Forest"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.5785\n",
      "Precision: 0.6280587275693311\n",
      "Recall: 0.385\n",
      "F1-Score: 0.4773713577185369\n"
     ]
    }
   ],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "from sklearn.ensemble import IsolationForest\n",
    "\n",
    "# Load the CIFAR-10 dataset\n",
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\n",
    "\n",
    "# Normalize pixel values to be between 0 and 1\n",
    "x_train, x_test = x_train / 255.0, x_test / 255.0\n",
    "\n",
    "# Filter data for 'bird' class (label 2) as normal and 'airplane' class (label 0) as anomaly\n",
    "x_train_normal = x_train[y_train.flatten() == 2]\n",
    "x_test_normal = x_test[y_test.flatten() == 2]\n",
    "x_test_anomaly = x_test[y_test.flatten() == 0]\n",
    "\n",
    "# Flatten the images for Isolation Forest\n",
    "x_train_normal_flat = x_train_normal.reshape(x_train_normal.shape[0], -1)\n",
    "x_test_all_flat = np.concatenate((x_test_normal, x_test_anomaly), axis=0).reshape(-1, 32*32*3)\n",
    "\n",
    "# Initialize and train the Isolation Forest\n",
    "isolation_forest = IsolationForest(random_state=42)  # Adjust contamination if needed\n",
    "isolation_forest.fit(x_train_normal_flat)\n",
    "\n",
    "# Predict on the test data\n",
    "y_pred = isolation_forest.predict(x_test_all_flat)\n",
    "\n",
    "# Convert predictions to binary labels (1 for anomaly, -1 for normal)\n",
    "y_pred = np.where(y_pred == 1, 0, 1) \n",
    "\n",
    "# Get true labels (1 for airplanes (anomalies), 0 for birds)\n",
    "y_true = np.concatenate((np.zeros(x_test_normal.shape[0]), np.ones(x_test_anomaly.shape[0])), axis=0)\n",
    "\n",
    "# Evaluate\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "precision = precision_score(y_true, y_pred)\n",
    "recall = recall_score(y_true, y_pred)\n",
    "f1 = f1_score(y_true, y_pred)\n",
    "\n",
    "print(\"Accuracy:\", accuracy)\n",
    "print(\"Precision:\", precision)\n",
    "print(\"Recall:\", recall)\n",
    "print(\"F1-Score:\", f1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.577\n",
      "Classification Report:\n",
      "              precision    recall  f1-score   support\n",
      "\n",
      "         0.0       0.55      0.90      0.68      1000\n",
      "         1.0       0.72      0.26      0.38      1000\n",
      "\n",
      "    accuracy                           0.58      2000\n",
      "   macro avg       0.63      0.58      0.53      2000\n",
      "weighted avg       0.63      0.58      0.53      2000\n",
      "\n",
      "Confusion Matrix:\n",
      "[[898 102]\n",
      " [744 256]]\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "from sklearn.ensemble import IsolationForest\n",
    "from sklearn.metrics import accuracy_score, classification_report, confusion_matrix\n",
    "from tensorflow.keras.datasets import cifar10\n",
    "\n",
    "# Load CIFAR-10 dataset\n",
    "(x_train, y_train), (x_test, y_test) = cifar10.load_data()\n",
    "\n",
    "# Squeeze y_train and y_test to remove extra dimension\n",
    "y_train = np.squeeze(y_train)\n",
    "y_test = np.squeeze(y_test)\n",
    "\n",
    "# Define normal and anomaly classes\n",
    "normal_class = 2  # birds\n",
    "anomaly_class = 9  # airplane\n",
    "\n",
    "# Create a subset of the dataset with only normal and anomaly classes\n",
    "x_train_normal = x_train[y_train == normal_class]\n",
    "y_train_normal = y_train[y_train == normal_class]\n",
    "x_train_anomaly = x_train[y_train == anomaly_class]\n",
    "y_train_anomaly = y_train[y_train == anomaly_class]\n",
    "\n",
    "x_test_normal = x_test[y_test == normal_class]\n",
    "y_test_normal = y_test[y_test == normal_class]\n",
    "x_test_anomaly = x_test[y_test == anomaly_class]\n",
    "y_test_anomaly = y_test[y_test == anomaly_class]\n",
    "\n",
    "# Reshape data to 2D array\n",
    "x_train_normal = x_train_normal.reshape((x_train_normal.shape[0], -1))\n",
    "x_train_anomaly = x_train_anomaly.reshape((x_train_anomaly.shape[0], -1))\n",
    "x_test_normal = x_test_normal.reshape((x_test_normal.shape[0], -1))\n",
    "x_test_anomaly = x_test_anomaly.reshape((x_test_anomaly.shape[0], -1))\n",
    "\n",
    "# Create an Isolation Forest model\n",
    "iforest = IsolationForest(contamination=0.1, random_state=42)\n",
    "\n",
    "# Train the model on normal data only\n",
    "iforest.fit(x_train_normal)\n",
    "\n",
    "# Evaluate the model on normal and anomaly data\n",
    "y_pred_normal = iforest.predict(x_test_normal)\n",
    "y_pred_anomaly = iforest.predict(x_test_anomaly)\n",
    "\n",
    "# Map predicted values to 0 (normal) and 1 (anomaly)\n",
    "y_pred_normal = np.where(y_pred_normal == 1, 0, 1)\n",
    "y_pred_anomaly = np.where(y_pred_anomaly == 1, 0, 1)\n",
    "\n",
    "# Calculate the accuracy\n",
    "y_pred = np.concatenate((y_pred_normal, y_pred_anomaly))\n",
    "y_true = np.concatenate((np.zeros((x_test_normal.shape[0],)), np.ones((x_test_anomaly.shape[0],))))\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "print(\"Accuracy:\", accuracy)\n",
    "\n",
    "# Calculate the classification report and confusion matrix\n",
    "print(\"Classification Report:\")\n",
    "print(classification_report(y_true, y_pred))\n",
    "print(\"Confusion Matrix:\")\n",
    "print(confusion_matrix(y_true, y_pred))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import clip \n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "\n",
    "# Load CIFAR-10 dataset\n",
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\n",
    "\n",
    "# Normalize pixel values to be between 0 and 1\n",
    "x_train, x_test = x_train / 255.0, x_test / 255.0\n",
    "\n",
    "# Filter data for 'bird' (normal) and 'airplane' (anomaly) classes\n",
    "x_train_normal = x_train[y_train.flatten() == 2]\n",
    "x_test_normal = x_test[y_test.flatten() == 2]\n",
    "x_test_anomaly = x_test[y_test.flatten() == 0]\n",
    "\n",
    "# Load CLIP model\n",
    "device = \"cpu\"\n",
    "model, preprocess = clip.load(\"ViT-B/32\", device=device)\n",
    "\n",
    "# Define normal class text\n",
    "normal_class_text = \"a photo of a bird\"\n",
    "\n",
    "# Encode normal images and text\n",
    "normal_image_embeddings = model.encode_image(preprocess(torch.tensor(x_train_normal).permute(0, 3, 1, 2)).to(device)).cpu().detach().numpy()\n",
    "normal_text_embedding = model.encode_text(clip.tokenize(normal_class_text).to(device)).cpu().detach().numpy()\n",
    "\n",
    "# Optional: Outlier Exposure (You can add this part if you have outlier images)\n",
    "# outlier_images = ... (load your outlier images)\n",
    "# outlier_image_embeddings = model.encode_image(preprocess(torch.tensor(outlier_images).permute(0, 3, 1, 2)).to(device)).cpu().detach().numpy()\n",
    "\n",
    "# Create a simple classifier on top of CLIP embeddings\n",
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras.layers import Dense\n",
    "\n",
    "classifier = Sequential()\n",
    "classifier.add(Dense(1, activation='sigmoid'))  # Binary classification (normal vs anomaly)\n",
    "classifier.compile(optimizer='adam', loss='binary_crossentropy')\n",
    "\n",
    "# Train the classifier (you might need to adjust epochs and batch_size)\n",
    "classifier.fit(normal_image_embeddings, np.zeros(normal_image_embeddings.shape[0]), epochs=10, batch_size=32)\n",
    "\n",
    "# Prepare test data\n",
    "x_test_all = np.concatenate((x_test_normal, x_test_anomaly), axis=0)\n",
    "test_image_embeddings = model.encode_image(preprocess(torch.tensor(x_test_all).permute(0, 3, 1, 2)).to(device)).cpu().detach().numpy()\n",
    "\n",
    "# Predict on test data\n",
    "y_pred_proba = classifier.predict(test_image_embeddings)\n",
    "y_pred = (y_pred_proba > 0.5).astype(int)  # Thresholding\n",
    "\n",
    "# Get true labels (0 for birds, 1 for airplanes)\n",
    "y_true = np.concatenate((np.zeros(x_test_normal.shape[0]), np.ones(x_test_anomaly.shape[0])), axis=0)\n",
    "\n",
    "# Evaluate\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "precision = precision_score(y_true, y_pred)\n",
    "recall = recall_score(y_true, y_pred)\n",
    "f1 = f1_score(y_true, y_pred)\n",
    "\n",
    "print(\"Accuracy:\", accuracy)\n",
    "print(\"Precision:\", precision)\n",
    "print(\"Recall:\", recall)\n",
    "print(\"F1-Score:\", f1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Requirement already satisfied: torch in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (2.4.0)\n",
      "Requirement already satisfied: filelock in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (3.15.4)\n",
      "Requirement already satisfied: typing-extensions>=4.8.0 in /Users/sushantjoshi/Library/Python/3.11/lib/python/site-packages (from torch) (4.11.0)\n",
      "Requirement already satisfied: sympy in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (1.13.2)\n",
      "Requirement already satisfied: networkx in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (3.3)\n",
      "Requirement already satisfied: jinja2 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (3.1.4)\n",
      "Requirement already satisfied: fsspec in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (2024.6.1)\n",
      "Requirement already satisfied: MarkupSafe>=2.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from jinja2->torch) (2.1.5)\n",
      "Requirement already satisfied: mpmath<1.4,>=1.1.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from sympy->torch) (1.3.0)\n",
      "Requirement already satisfied: torch in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (2.4.0)\n",
      "Requirement already satisfied: torchvision in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (0.19.0)\n",
      "Requirement already satisfied: torchaudio in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (2.4.0)\n",
      "Requirement already satisfied: filelock in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (3.15.4)\n",
      "Requirement already satisfied: typing-extensions>=4.8.0 in /Users/sushantjoshi/Library/Python/3.11/lib/python/site-packages (from torch) (4.11.0)\n",
      "Requirement already satisfied: sympy in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (1.13.2)\n",
      "Requirement already satisfied: networkx in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (3.3)\n",
      "Requirement already satisfied: jinja2 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (3.1.4)\n",
      "Requirement already satisfied: fsspec in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch) (2024.6.1)\n",
      "Requirement already satisfied: numpy in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torchvision) (1.26.4)\n",
      "Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torchvision) (10.3.0)\n",
      "Requirement already satisfied: MarkupSafe>=2.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from jinja2->torch) (2.1.5)\n",
      "Requirement already satisfied: mpmath<1.4,>=1.1.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from sympy->torch) (1.3.0)\n",
      "Collecting git+https://github.com/openai/CLIP.git\n",
      "  Cloning https://github.com/openai/CLIP.git to /private/var/folders/1q/29s1zkk576nfppg1thcv079r0000gn/T/pip-req-build-4777w_pw\n",
      "  Running command git clone --filter=blob:none --quiet https://github.com/openai/CLIP.git /private/var/folders/1q/29s1zkk576nfppg1thcv079r0000gn/T/pip-req-build-4777w_pw\n",
      "  Resolved https://github.com/openai/CLIP.git to commit dcba3cb2e2827b402d2701e7e1c7d9fed8a20ef1\n",
      "  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
      "\u001b[?25hRequirement already satisfied: ftfy in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from clip==1.0) (6.2.3)\n",
      "Requirement already satisfied: packaging in /Users/sushantjoshi/Library/Python/3.11/lib/python/site-packages (from clip==1.0) (24.0)\n",
      "Requirement already satisfied: regex in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from clip==1.0) (2023.12.25)\n",
      "Requirement already satisfied: tqdm in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from clip==1.0) (4.66.2)\n",
      "Requirement already satisfied: torch in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from clip==1.0) (2.4.0)\n",
      "Requirement already satisfied: torchvision in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from clip==1.0) (0.19.0)\n",
      "Requirement already satisfied: wcwidth<0.3.0,>=0.2.12 in /Users/sushantjoshi/Library/Python/3.11/lib/python/site-packages (from ftfy->clip==1.0) (0.2.13)\n",
      "Requirement already satisfied: filelock in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch->clip==1.0) (3.15.4)\n",
      "Requirement already satisfied: typing-extensions>=4.8.0 in /Users/sushantjoshi/Library/Python/3.11/lib/python/site-packages (from torch->clip==1.0) (4.11.0)\n",
      "Requirement already satisfied: sympy in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch->clip==1.0) (1.13.2)\n",
      "Requirement already satisfied: networkx in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch->clip==1.0) (3.3)\n",
      "Requirement already satisfied: jinja2 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch->clip==1.0) (3.1.4)\n",
      "Requirement already satisfied: fsspec in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torch->clip==1.0) (2024.6.1)\n",
      "Requirement already satisfied: numpy in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torchvision->clip==1.0) (1.26.4)\n",
      "Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from torchvision->clip==1.0) (10.3.0)\n",
      "Requirement already satisfied: MarkupSafe>=2.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from jinja2->torch->clip==1.0) (2.1.5)\n",
      "Requirement already satisfied: mpmath<1.4,>=1.1.0 in /Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages (from sympy->torch->clip==1.0) (1.3.0)\n",
      "Building wheels for collected packages: clip\n",
      "  Building wheel for clip (setup.py) ... \u001b[?25ldone\n",
      "\u001b[?25h  Created wheel for clip: filename=clip-1.0-py3-none-any.whl size=1369489 sha256=daae851c9184aa7166814259250432ac04ff289dcf077ccc0c7ef03bf0dbffbc\n",
      "  Stored in directory: /private/var/folders/1q/29s1zkk576nfppg1thcv079r0000gn/T/pip-ephem-wheel-cache-rwacm_4o/wheels/3f/7c/a4/9b490845988bf7a4db33674d52f709f088f64392063872eb9a\n",
      "Successfully built clip\n",
      "Installing collected packages: clip\n",
      "Successfully installed clip-1.0\n"
     ]
    }
   ],
   "source": [
    "!pip install torch\n",
    "!pip install torch torchvision torchaudio\n",
    "!pip install git+https://github.com/openai/CLIP.git\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "ename": "AttributeError",
     "evalue": "type object 'CLIP' has no attribute 'from_pretrained'",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mAttributeError\u001b[0m                            Traceback (most recent call last)",
      "Cell \u001b[0;32mIn[20], line 21\u001b[0m\n\u001b[1;32m     19\u001b[0m \u001b[38;5;66;03m# Load CLIP model\u001b[39;00m\n\u001b[1;32m     20\u001b[0m device \u001b[38;5;241m=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mcuda\u001b[39m\u001b[38;5;124m\"\u001b[39m \u001b[38;5;28;01mif\u001b[39;00m torch\u001b[38;5;241m.\u001b[39mcuda\u001b[38;5;241m.\u001b[39mis_available() \u001b[38;5;28;01melse\u001b[39;00m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mcpu\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[0;32m---> 21\u001b[0m model \u001b[38;5;241m=\u001b[39m \u001b[43mCLIP\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mfrom_pretrained\u001b[49m(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mViT-B/32\u001b[39m\u001b[38;5;124m\"\u001b[39m)\u001b[38;5;241m.\u001b[39mto(device)\n\u001b[1;32m     22\u001b[0m preprocess \u001b[38;5;241m=\u001b[39m clip\u001b[38;5;241m.\u001b[39m_transform(model\u001b[38;5;241m.\u001b[39mvisual\u001b[38;5;241m.\u001b[39minput_resolution) \n\u001b[1;32m     24\u001b[0m \u001b[38;5;66;03m# Define normal class text\u001b[39;00m\n",
      "\u001b[0;31mAttributeError\u001b[0m: type object 'CLIP' has no attribute 'from_pretrained'"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "from clip.model import CLIP\n",
    "import numpy as np\n",
    "from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score\n",
    "from torch import nn\n",
    "from torch.utils.data import DataLoader, TensorDataset\n",
    "\n",
    "# Load CIFAR-10 dataset\n",
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()\n",
    "\n",
    "# Normalize pixel values to be between 0 and 1\n",
    "x_train, x_test = x_train / 255.0, x_test / 255.0\n",
    "\n",
    "# Filter data for 'bird' (normal) and 'airplane' (anomaly) classes\n",
    "x_train_normal = x_train[y_train.flatten() == 2]\n",
    "x_test_normal = x_test[y_test.flatten() == 2]\n",
    "x_test_anomaly = x_test[y_test.flatten() == 0]\n",
    "\n",
    "# Load CLIP model\n",
    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
    "model = CLIP.from_pretrained(\"ViT-B/32\").to(device)\n",
    "preprocess = clip._transform(model.visual.input_resolution) \n",
    "\n",
    "# Define normal class text\n",
    "normal_class_text = \"a photo of a bird\"\n",
    "\n",
    "# Encode normal images and text\n",
    "normal_image_embeddings = model.encode_image(preprocess(torch.tensor(x_train_normal).permute(0, 3, 1, 2)).to(device))\n",
    "normal_text_embedding = model.encode_text(clip.tokenize(normal_class_text).to(device))\n",
    "\n",
    "# Optional: Outlier Exposure (You can add this part if you have outlier images)\n",
    "# outlier_images = ... (load your outlier images)\n",
    "# outlier_image_embeddings = model.encode_image(preprocess(torch.tensor(outlier_images).permute(0, 3, 1, 2)).to(device))\n",
    "\n",
    "# Create a simple classifier on top of CLIP embeddings (using PyTorch)\n",
    "class Classifier(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(Classifier, self).__init__()\n",
    "        self.fc1 = nn.Linear(512, 1)  # Assuming CLIP embeddings are 512-dimensional\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = torch.sigmoid(self.fc1(x))\n",
    "        return x\n",
    "\n",
    "classifier = Classifier().to(device)\n",
    "criterion = nn.BCELoss()  # Binary Cross Entropy Loss\n",
    "optimizer = torch.optim.Adam(classifier.parameters())\n",
    "\n",
    "# Prepare training data\n",
    "train_dataset = TensorDataset(normal_image_embeddings, torch.zeros(normal_image_embeddings.shape[0]))\n",
    "train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)\n",
    "\n",
    "# Train the classifier\n",
    "for epoch in range(10):  # You might need to adjust the number of epochs\n",
    "    for images, labels in train_loader:\n",
    "        images, labels = images.to(device), labels.to(device)\n",
    "        optimizer.zero_grad()\n",
    "        outputs = classifier(images)\n",
    "        loss = criterion(outputs, labels.unsqueeze(1))\n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "\n",
    "# Prepare test data\n",
    "x_test_all = np.concatenate((x_test_normal, x_test_anomaly), axis=0)\n",
    "test_image_embeddings = model.encode_image(preprocess(torch.tensor(x_test_all).permute(0, 3, 1, 2)).to(device))\n",
    "\n",
    "# Predict on test data\n",
    "with torch.no_grad():\n",
    "    y_pred_proba = classifier(test_image_embeddings).cpu().numpy()\n",
    "y_pred = (y_pred_proba > 0.5).astype(int)\n",
    "\n",
    "# Get true labels (0 for birds, 1 for airplanes)\n",
    "y_true = np.concatenate((np.zeros(x_test_normal.shape[0]), np.ones(x_test_anomaly.shape[0])), axis=0)\n",
    "\n",
    "# Evaluate\n",
    "accuracy = accuracy_score(y_true, y_pred)\n",
    "precision = precision_score(y_true, y_pred)\n",
    "recall = recall_score(y_true, y_pred)\n",
    "f1 = f1_score(y_true, y_pred)\n",
    "\n",
    "print(\"Accuracy:\", accuracy)\n",
    "print(\"Precision:\", precision)\n",
    "print(\"Recall:\", recall)\n",
    "print(\"F1-Score:\", f1)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.11.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}