NonDeep_baselines.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "33dee702",
   "metadata": {},
   "source": [
    "Use scikit-learn implementation of Isolation Forest, One Class SVM and Local Outlier Factor for comparison. This notebook applies the 3 non-deep AD methods with and without PCA dimensionality reduction beforehand. A PCA dimensionality reduction preserving 95% of the data variance was used for shallow baselines in both DeepSVDD and DeepMSVDD papers."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "3f7e2ba8",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<style>.container { width:100% !important; }</style>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from IPython.core.display import display, HTML, Markdown\n",
    "display(HTML(\"<style>.container { width:100% !important; }</style>\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "24cb3cb4",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch.utils.data as data_utils\n",
    "import torch\n",
    "import random\n",
    "import matplotlib.pyplot as plt\n",
    "from seaborn import boxplot\n",
    "from sklearn.decomposition import PCA\n",
    "from sklearn.ensemble import IsolationForest\n",
    "from sklearn.neighbors import LocalOutlierFactor\n",
    "from sklearn.svm import OneClassSVM\n",
    "from sklearn.metrics import roc_auc_score\n",
    "import numpy as np\n",
    "from tqdm import tqdm\n",
    "from utils import *"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "c21ce5b7",
   "metadata": {},
   "outputs": [],
   "source": [
    "# dataset = 'MNIST'\n",
    "dataset = 'FashionMNIST'\n",
    "# dataset = 'CIFAR10'\n",
    "normal_cls = [0,1,2]\n",
    "nbr_seeds = 3 # how many different random init to try"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "876a8ff8",
   "metadata": {},
   "outputs": [],
   "source": [
    "pca = PCA(n_components=0.95,svd_solver='full')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1b7f1c18",
   "metadata": {},
   "source": [
    "# IF, OC-SVM and LOF without PCA dimensionality reduction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "15525a1b",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "  0%|          | 0/3 [00:00<?, ?it/s]/home/blupon/anaconda3/envs/recent/lib/python3.9/site-packages/torchvision/datasets/mnist.py:498: UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at  /opt/conda/conda-bld/pytorch_1631630797748/work/torch/csrc/utils/tensor_numpy.cpp:180.)\n",
      "  return torch.from_numpy(parsed.astype(m[2], copy=False)).view(*s)\n",
      "100%|██████████| 3/3 [00:18<00:00,  6.33s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "MEAN VALID AUC +/- STD: 0.8286018371582031 +/- 0.008531280793249607\n",
      "MEAN TEST AUC +/- STD: 0.8212025761604309 +/- 0.004212884698063135\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAZAAAAElCAYAAADKuLQKAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjMuNCwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8QVMy6AAAACXBIWXMAAAsTAAALEwEAmpwYAAAwU0lEQVR4nO3deZxcVZn/8c83AUxCgIgkENKEoAQY9McawW0YFmEA2VUW0QHHGVRAxHFjBMXRcQYZQc2oIAgjIIsoMKIiqyJuDEnY97Rha5KQhJA9kO35/XFOh0qnuvpWddfS3d/369Wvrqp7zz1PFaGevufe8xxFBGZmZtUa0uwAzMysf3ICMTOzmjiBmJlZTZxAzMysJk4gZmZWEycQMzOriROIGSDpbkn/VEO7kLR9PWKqF0nPSlou6apmx9ITSb+V9KqkPzY7FlufE4g1RP7SeknSxiWv/ZOku5sY1mB2eER8pPOJpAmSfidpmaQnJb236IEkHSvpz7nt3dUEIekkSdMkLZLUIel8SRt0bo+I/YFPVHNMaxwnEGukDYBP9/YgSvxvt29dCzwAvAk4G/i5pNEF284HvgOcV0O/I4AzgS2AvYEDgM/VcBxrAv9PaI30X8DnJI0qt1HSuyRNkbQw/35Xyba7JX1D0p+AZcCb8/DRqZKmS1os6euS3iLpL/kv2uslbZTbv1HSryTNlfRKftxWJGhJQyV9SdJfcz/TJG1TZr/3SXog9/2CpK+WbBsm6SeSXpa0IL+/LfO2kyXNyMd+RtKJJe3+UdITOebbJG2bX5ekb0uakz+vhyW9rcj7KRP3DsAewLkRsTwibgAeAd5fpH1E3BkR1wMzq+07Ii6KiD9ExIqIeBG4Gnh3tcex5nACsUaaCtxNmb8wJW0O/BqYTPor+ELg15LeVLLbR4BTgE2A5/JrBwN7Au8AvgBcApwIbAO8DTgh7zcE+B9gW2A8sBz4XsG4/yUf51BgU+AfSUmsq6XAPwCjgPcBn5R0VN52ErBZjutNpGGZ5XlIbzJwSERsArwLeDB/JkcBXwKOAUYDfyCdKQAcBOwD7JD7Ow54Obf7kKSHC743gLcCMyJicclrD+XXG20f4LEm9Gs1cAKxRvsK8KkywyPvA6ZHxFURsSoirgWeBA4v2efHEfFY3r4yv/bNiFgUEY8BjwK3R8SMiFgI/AbYHSAiXo6IGyJiWf6i/AbwdwVj/ifgnIh4KpKHIuLlrjtFxN0R8UhErImIh0lf9p19rCQlju0jYnVETIuIRXnbGuBtkoZHxKz8XgA+DvxnRDwREauA/wB2y2chK0mJdCdAeZ9ZOY5rImKXgu8NYCSwsMtrC/PxG0bSR4FJwLca2a/VzgnEGioiHgV+BZzVZdPWvH5W0ek5YFzJ8xfKHPKlksfLyzwfCSBphKQfSnpO0iLgHmCUpKEFwt4G+GtPO0naO1+InitpIeksY4u8+SrgNuA6STPzxeINI2Ip6ezhE8AsSb+WtFNusy3w3TzktYB0rUHAuIj4LekM6vvAS5IukbRpgfdSzhLSmVWpTYHFZfati3y2dR7pTGxeo/q13nECsWY4F/hn1k0OM0lfmKXGAy+WPO9N6ejPAjsCe0fEpqShEkhfyD15AXhLgf2uAW4GtomIzYCLO48fESsj4t8iYmfSMNVhpOEuIuK2iDgQGEs667q0pN+PR8Sokp/hEfHn3G5yROxJGmraAfh8gRjLeYx0Tan0jGNXGjSUJOlg0ns+PCIeaUSf1jecQKzhIqId+ClwRsnLtwA75PH7DSQdB+xMOlvpC5uQzkgW5Ost51bR9kfA1yVNzBevd+lybaa0j/kR8aqkvYAPdW6QtJ+k/5fPeBaRhqBWS9pS0hH5WshrpLOB1bnZxcC/SnprPsZmkj6YH789n/FsSLr28mpJu6pExNOk6y7n5ov9RwO7ADfkvibkGxYmlGufbzIYRrrLbkg+xoYl25+VdHI3bfcnXTh/f0TcV0v81jxOINYsXwPWzgnJ1xQOI50pvEy6IH5YHw5nfAcYDswD7gVuraLthcD1wO2kL//L8rG6OhX4mqTFpGs915ds2wr4eW7/BPB74Cek/wc/SzoDm0+6ZnIqQETcBHyTNOy1iHSN55B8vE1Jf7W/Qhrqe5l87UDSiZKqPXs4nnT94RXSUNIHImJu3rZN7uPFbtp+hJScLwL+Nj++NMeyEenaz73dtP0y6eaCWyQtyT+/qTJ2axJ5QSmzwUXSU6Thspsi4qQC+58DzI2IH9bQ13uA0yLihB53Lt/+DtIddvdFxAG1HMPqxwnEzMxq4iEsMzOriROImZnVxAnEzMxq4gRiZmY1cQKxHuX7+Lst753nRsyQ9HiZbW+VdHsuBrhAqRDhoflW087bNpdLWlPyfEk3/Rwp6UGlYoXzJN3VOTdB0lclrSw9Rp693dk2JC3tsv0LXdouzj9PS/qepLEV3vPJklZ3Od4SSVtLGpk/s9J5IJtIel7SB/LzH0takdvMl3RH5wz0Kt/Li5IuVMmMepWsbSJp37z/jV3i3zW/fncVn1F0zkPJr23QOT9E0m9K2qwseW9LJF3czWfYJulqpQKTSyXdJ+mwLvuEpEdUUn1Z0r9L+nF+3DlHZYP8fLhScc1/6HKccyX9Sa7i3Kf8YVpf2AcYQ5rN/PYu234J3AFsmfc5A1gUEVdHxMiIGEma2zCz83l+bR1KizZdSZozsRmwHfADUh2pTj8tPUZEjOpymF27bD+/S9tNgM2Bo0nzNqZVSiLAX7ocb2REzIyIJaSij9/V6zW/zgemRsTPS9qfn99rGzAH+HG174U0b+Q4UoHH7swF3qV1Jz+eBDxdZt9Kn9F80jyX9cq/RMQhJf/tru58b/lnvfU8lCZz/hFYQZpJvwXwbeCaziRbYmvSPJUeRcRy4GPAhXq92vHfkApifiwi1lRqb9VxArG+cBLwC9Js8rXzCiRtQfqivzSX614REX+KiFpWl9sNeCYi7soFDRfn4ojP98Ub6JRLjjxG+lKeS0pYtRzndnJ1YUn7AscCp3Wz7zJSGZSqy7HnWf1/In0+3VkB/C/5SzgngGNJX/TVuDUf68PVxlnGZ0iz7j8WEbMjlZG/llTk8gJJpSVmzgf+TSULTVUSEfeQKh18Lx/nUlJRyif7IG4r4QRivSJpBPAB0pfR1cDxymtwkGZHtwM/kXRU51+ENbof2ElpDYz9JK13ltKXImI1KSn+bS8O8xlgX9IM9M91VsvtKr+XE0kLOlUlD3v9LelzruRKcu0t4O9Jda6qXb8jSDPHz1VJqZIaHQjcUOaM4HpSDbQdSl67kTSD/+Qqjv9F4O2kcizDSGvRWB9zArHeOoZUw+l2Ut2qDUil2Yk0S3U/4FngAlK12XskTay2k4iYQfoyHkf6kpmXryOUJpJjlSvX5p/fdTnM/V22/30P3c4kDWl15x1djrdOxd6IeIX0RT2C9CXY1efytY12UtXgk6t8L0tJZVHuJg3ndSsXYNxc0o6kRHJlN7tW/Iwi4mbSmVnV68d3sQVQLqHOKtm+tltS4vqKpDcUOXgeRjyNNBz5sfwHgfUxJxDrrZOA6yOt0fEa6Yty7TBWRHRExOkR8RZStd2ldP/lVVFE3BsRx0bEaNJf3fuQll/tdH2XyrX7dTnEHl2239ZDl+NI4/7dubfL8dap2Cvpw8AE4E5STauuvpXbbRURR0REaQLq8b2Qks5xpKVgN6ZnVwGnk5L6Td3sU+QzOof0uQ8r0Gd35pHKqXQ1tmT7WhFxC/A86dpSUY91+W19zAnEaqa0JOz+wIclzZY0mzScdWi+/rGOiHiBtH5FTUuvdjnWFFKy6vWxysl36xxOWgWwlvZjSBeF/5m0MNSxkvap3Ko6+VrQ9cBfSMUbe3IVqVDjLfm6S6393kE6azq11mOQkur7y9wVdSypjH25C/ydiWtEL/q1PuQEYkVtqFSmu/NnA1IV1qdJ62zsln92ADqAE5TWIf83SdtLGpKTyj/SfWXWbkl6j6R/zl/MnWP/R9RyrB762TDftXMt6U6sC2s81PeA/42I3+VrH18ALi06BFOl84BTJG1VaaeIeIZ019bZlfYr6GzSe6rVt0kVhS+TtFX+N3VCPu7no0yRvoi4m7RWe7kCkG/o8u/T320N4A/ZirqFVKa78+erpP+Rf5Dvoln7Q1rH4iTSHTsTSH9tdpYjf43qLoZ2WkBKGI8ozRO5lTQMU3qb6XFaf17GmJLtD3XZ9p2ubXM/N5NuANgzIipdaH5nmf7errS63nsoWeApIn5ESqxFzhSKvJe1Ii3C9HsKLCgVEX/s4T1V+oxKj/MnoOb1OyKV738PaRjscdLn/S/ARyLipxWankP561JLWPff5/61xmbFuRqvmZnVxGcgZmZWEycQMzOriROImZnVxAnEzMxqUqi2zECxxRZbxIQJE5odhplZvzJt2rR5eQLvOgZVApkwYQJTp05tdhhmZv2KpOfKve4hLDMzq4kTiJmZ1cQJxMzMauIEYmZmNXECMTOzmjiBmJlZTZxAzMysJoNqHkh/MXnyZNrbyy9x3dHRAUBbW1vZ7dtvvz1nnHFG3WIzM+vkBNLPLF++vNkhmJkBTiAtqdIZROe2yZMnNyocM7OyfA3EzMxq4jOQLmq9/uBrD2Y22DiBVMHXH8zMXucE0oWvP5iZFeMEYmZV8TCvdWrqRXRJB0t6SlK7pLPKbN9J0l8kvSbpc9W0NbPGW758uYd6B5GmnYFIGgp8HzgQ6ACmSLo5Ih4v2W0+cAZwVA1tbQDwpMrW42Fe69TMM5C9gPaImBERK4DrgCNLd4iIORExBVhZbVsb+PzXrllzNfMayDjghZLnHcDefd1W0inAKQDjx4+vPkprKv+1a9a6mplAVOa16Ou2EXEJcAnApEmTih6/ISoNz3Rn+vTpQOUv1u54SMfM+lIzE0gHsE3J8zZgZgPatoz29naefvR+xo9cXbjNRivTqOOrz06pqq/nlwytan/w3TZmVlmPCUTS+cC/A8uBW4FdgTMj4ie97HsKMFHSdsCLwPHAhxrQtqWMH7macyYtqXs//z51ZJ8ez9cezKzIGchBEfEFSUeT/vL/IPA7oFcJJCJWSToduA0YClweEY9J+kTefrGkrYCpwKbAGklnAjtHxKJybXsTj63P1x/MrJIiCWTD/PtQ4NqImC+VuwRRvYi4Bbily2sXlzyeTRqeKtTWrF48nGe2viIJ5JeSniQNYZ0qaTTwan3DMus/PJxng1WPCSQizpL0TWBRRKyWtBTPuegTHR0dLF08tM+vT5Tz3OKhbJz/UrbqeTjPbH09TiSU9EFgVU4e55CufWxd98jMzKylFRnC+nJE/EzSe4C/B74FXETxSX/Wjba2Nl5dNathd2EN66bkh5lZLYqUMumcpPA+4KKI+AWwUf1CMjOz/qBIAnlR0g+BY4FbJL2hYDszMxvAiiSCY0nzLQ6OiAXA5sDn6xmUmZm1vh4TSEQsi4gbgYWSxpPmhTxZ98jMzKylFbkL6whJ04FngN/n37+pd2BmZtbaityF9XXgHcCdEbG7pP2AE+ob1uDx/JLq5oG8tCzl/C1HrKm6nx2qamFmVlmRBLIyIl6WNETSkIj4XZ5Y2G/VUkYdai+l3l05i+23377qGFbkGIZNmFhVux1q7K9RXNrerP8pkkAWSBoJ3ANcLWkOsKq+YdVXe3s7DzzyOGtGbF5VO61Iy4lM++vswm2GLJvf7bZavsAG6qzn9vZ2nnzwQbaqok3n+OuCBx+sqq/i//XMrJIiCeRIUu2rzwAnApsBX6tnUI2wZsTmvLrzYXXvZ9jjv6p7HwPFVsDHyq4V1rcuK7xumZlVUqQW1tKSp1fUMRYzM+tHuk0gkhaTlokV6y4XKyAiYtM6x2ZmZi2s2wQSEZs0MhAzM+tfiswDeYekTUqej5TkQopmZoNckVImFwGl5WKX5dfMzGwQK5JAFBFrr4FExBqK3b1lZmYDWJEEMkPSGZI2zD+fBmbUOzAzM2ttRc4kPgFMBs4h3Y11F3BKPYOywaejo4PFNGaOxixgiZf3Neu1IvNA5gDHNyCWhuno6GDIsoUNmeQ3ZNnLdHT064n7Ngi1Srkfa22+lmEtoa2tjQXz5jVsJvooL+9bUXt7O089+gTbbFJNcRnYcFUaFV/23CuF27yw2MVl+qtBmUDa2tp46bUNGlbKpK2tuv8JzVrBNptsxWf3+mjd+7ngvv+pex9WH91eRJd0TL07l3SwpKcktUs6q8x2SZqctz8saY+Sbc9KekTSg5Km1jtWMzNbV6UzkHOAG+vVsaShwPeBA4EOYIqkmyPi8ZLdDgEm5p+9SfNPSicx7hcR8+oVow0uHvc3q04zh7D2AtojYgaApOtIlX9LE8iRwJV5Hsq9kkZJGhsRsxofrg107e3tPPbIE4waMaaqdmtWpOs2L/715cJtFiybU1UfZq2oUgLZSdLDZV7vLKa4Sy/7Hge8UPK8g3XPLrrbZxzpTswAbpcUwA8j4pJynUg6hXzb8fjx43sZ8sDjv7rXNWrEGPbbqf43Hf7uyevq3odZvVVKIM8Ah9ex73K323SdBFBpn3dHxExJY4A7JD0ZEfest3NKLJcATJo0yQtBdNHe3s4Djz0Ao6psmFfUfeDFB4q3WVBlH2bW0iolkBUR8Vwd++4Atil53gbMLLpPRHT+niPpJtKQ2HoJxAoYBWv2rW6N9VoMubty4YPZVDeRsHPA6E1VxjGb6vOlma2vUgL5U537ngJMlLQd8CJpsuKHuuxzM3B6vj6yN7AwImZJ2hgYEhGL8+ODGACrJHaqNKzU09BRqw8RdaeW9drn5s9i1MTq1ocfVWN/ZrauSgnkNknbdp6FSPoK8H7gOeDTEfFMbzqOiFWSTgduA4YCl0fEY5I+kbdfDNwCHAq0k6oAd96UviVwk6TO93BNRNzam3j6i+HDhzc7hLrw+vBm/U+lBPIN4B0Akg4DPgycAOwOXAz8fW87j4hbSEmi9LWLSx4HcFqZdjOAXXvbf6vqj2cQZjb4VEogERHL8uNjgMsiYhowTdKp9Q/NzJqlo6ODpYsXN2SW+AuLZ7Nxx9K692N9r9JVTeXVB4cAB5Cq8HYaVt+wzMys1VU6A/kO8CCwCHgiIqYCSNqdNA/DbEDp6Ohg4bLFDZmjsWDZHKJjed37qVVbWxvLVr/SsFpYI9reWPd+rO91m0Ai4nJJtwFjgIdKNs3m9YvZZmY2SHWbQEoLFwK75TueSj1fl4gaZMiy+VWvB6JXFwEQwzatqh9o3Wq8HR0dsLDnORp9YgF0ROsu5NTW1oZee7lhM9HHtVU7g8VaTaVb7jvyomVtZZYO6K+323dVaQjrggrbAti/j2NpmFrnAEyfvhiAiW+pJiFs5TkHZoPQ8uWtO0TZVyoNYe3XyEAaqdbMPxDnHbS1tTFXcxs2E71tnBdysoGj0nfJQPy+6KrSENY+FdpFRPyhDvGYmVk/UWkI6/NlXgvSBL420uxxMzMbpCoNYa1TiVfSe4CzSbfwnl7nuMzMrMX1uKCUpAOAL5POPv4jIu6oe1RmZtbyKl0DeR/pjGMhcHZE1Ls6r5mZ9SOVzkB+SVqP42Xgi13ngUTEEXWMy8zMWlylBDJgb+M1686CZXOqLmWy5NVXABg5rHg5jgXL5jCu6qWwzFpLpYvov29kIGbNVvsE0/kAjHtL8YQwjje1/ATTFxbPrroa75xl6bMYM2LzqvrZEdfC6o8qXQN5uFLDiNil78Mxax5PMH1drclt5fR5AIzYtnhC2JE3tnwytfIqDWGtId15dQ3pesjAn5dvZoCTqRXTbQW9iNiNtALhSFIS+QbwVuDFzmVuzcxs8Ko4DyQingTOBc6VdBxwJfBN4L8aEJuZlai18isMnOqv1loqJhBJ44DjgaOBV4DPADc1IC4zq8JgqPxqrafSRfTfA5sA1wMnA/Pzpo0kbR4R87tra/3MghrWA1mSf4+srh/GVdeNvW6wV3611lPpDGRb0kX0jwOnlLyu/Pqb6xiXNUjtt65OB2DiuInFG42rvT8zaz2V5oFMaGAc1iStfrdNpXH/ziTW3XvwuL9ZfRUat5A0TtK7JO3T+dMXnUs6WNJTktolnVVmuyRNztsfLl1mt6e2NvANHz6c4cOHNzsMs0GrSDXebwLHAY8Dq/PLAdzTm44lDQW+DxxIqrk1RdLNEfF4yW6HABPzz97ARcDeBdvaAOAzCLPW1WMCAY4CdoyI1/q4772A9oiYASDpOuBIUqLqdCRwZUQEcK+kUZLGAhMKtDUzszoqMoQ1A9iwDn2PA14oed7B+vfodLdPkbYASDpF0lRJU+fOndvroM3MLClyBrIMeFDSXcDas5CI6O3Ygsq8FgX3KdI2vRhxCXAJwKRJk8ruY2Zm1SuSQG7OP32tA9im5HkbMLPgPhsVaGtmZnXUYwKJiCskbQTskF96KiJW9kHfU4CJkrYDXiTNeP9Ql31uBk7P1zj2BhZGxCxJcwu0NTOzOipyF9a+wBXAs6Sho20knRQRvboLKyJWSToduA0YClweEY9J+kTefjFwC3Ao0E4aSvtopba9icfMzKpTZAjrAuCgiHgKQNIOwLXAnr3tPCJuISWJ0tcuLnkcwGlF25qZWeMUuQtrw87kARART1Ofu7LMzKwfKXIGMlXSZcBV+fmJwLT6hWRmZv1BkQTySdIw0hmkayD3AD+oZ1BmZtb6ityF9RpwYf4xMzMDKq8Hcn1EHCvpEcpM0ouIXeoamZmZtbRKZyCfzr8Pa0QgZmbWv1RaD2RW/v1c48IxM7P+osfbeCUdI2m6pIWSFklaLGlRI4IzM7PWVeQurPOBwyPiiXoHY9aqal0Z0asi2kBWJIG85ORh1j2vimiDVaW7sI7JD6dK+inwv6xbzv3G+oZm1jp8FmG2vkpnIIeXPF4GHFTyPAAnEDOzQazSXVgfbWQgZmbWvxQp534+8O/AcuBWYFfgzIj4SZ1jawpfLDUzK6ZINd6DImIRaUJhB2lhqc/XNaoWNXz4cF8wNTPLityF1Vm6/VDg2oiYL5Vbknxg8FmEmVkxRRLILyU9SRrCOlXSaODV+oZlZta9SkPNHR0dALS1tZXd7uHmvtPjEFZEnAW8E5iU10JfChxZ78DMzGqxfPlyli9f3uwwBoUiZyAA44ADJQ0ree3KOsRjZtajSmcQndsmT57cqHAGrSJ3YZ0L7AvsTFqD/BDgjziBmJkNakXuwvoAcAAwO88N2RV4Q12jMjOzlldkCGt5RKyRtErSpsAc4M11jsvMWpTnSlmnIglkqqRRwKXANGAJcF89gzKz/snzpAaXisUUI+LGiDhV0hsj4mJJtwKbRsTDvelU0ubAT4EJwLPAsRHxSpn9Dga+CwwFfhQR5+XXvwr8MzA37/qliLilNzGZtYpKf+F3p9Jf/pXUclbgswjrVOkM5BxeL5h4F7BHRDzbR/2eBdwVEedJOis//2LpDpKGAt8HDiTNgJ8i6eaIeDzv8u2I+FYfxWPWMtrb23n0oYfYZKOiN0nCqlWrAXjuiccKt1m8YlXVsZmVqvQvVN087gtHku7sArgCuJsuCQTYC2iPiBkAkq7L7R7HbIDbZKMN2GvLN9a1j/teWu+k36wqlRLIcEm7k+7UGpYfr00kEXF/L/rdsmTN9VmSxpTZZxzwQsnzDmDvkuenS/oHYCrw2XJDYACSTgFOARg/fnwvQh58fLHUzCqplEBmARfmx7NLHkNaD2T/SgeWdCewVZlNZxeMrdxZT+TfFwFfz8+/DlwA/GO5g0TEJcAlAJMmTYpy+1j1fLHUrDU0s6xLpfVA9qv5qKn9e7vbJuklSWPz2cdY0q3BXXUA25Q8bwNm5mO/VHKsS4Ff9SZWK89nEWb9W71LuhS/Ste3bgZOAs7Lv39RZp8pwERJ2wEvAscDHwLoTD55v6OBR+sesZlZC2pmWZdmJZDzgOslfQx4HvgggKStSbfrHhoRqySdDtxGuo338ojovMXkfEm7kYawngU+3uD4zcwGvaYkkIh4mVQepevrM0nrjnQ+v4VUf6vrfh+pa4BmZlkt83KgsXNzmqVQApE0Dti2dP+IuKdeQZmZtYr29nYef/wBthhd7T046T6gOXOL37A6b27/WqyvSDXebwLHkeZfrM4vB+AEYmaDwhajg2OOWVH3fm68caO699GXipyBHAXsGBGv1TkWMzPrR4qUc5/B6+uim5mZAZWLKf43aahqGfCgpLuAtWchEdE/rvKY9TMdHR0sXrGq7qVGFq9YtXaimVktKg1hTc2/p5HmbZiZma1VaSb6FQCSNgZejYjV+flQvCKhWd20tbWxevHChhRT7K7ERato9dL2g12Ri+h3Ae8lLSQFMBy4HXhXvYIyM4Nc2v7RRxk5cmThNitXrgTg2WefLdxmyZIlPe9k6ymSQIZFxNpPNyKWSBpRx5jMzNYaOXIke+yxR137uP/+3hQXH7yK3IW1VNLa/3qS9gTqW6HLzMxaXpEzkDOBn0mamZ+PJU0sNDOzQazHBBIRUyTtBOxImpv/ZESsrHtkZmbW0ooWU9wR2BkYBuwuiYi4sn5hmZlZqytSC+tc0vrlO5Mq4x4C/BFwAjGzuuro6GDx4sV1v8i9ePFiT6qsQZGL6B8glV6fHREfBXbF80DMzAa9IkNYyyNijaRVkjYlLT/75jrHZWZGW1sbq1atashtvK0+qbIVFUkgUyWNAi4llTVZAtxXz6DMBrtqa2EtW5VWWhixwdCq+jDrjSJ3YZ2aH14s6VZg04h4uL5hmQ1e22+/fdVtOst3bDtxYt37MutU5CK6gBOBN0fE1ySNl7RXRPgsxKwOaqnH1Nlm8uTJfR2OWbeKXET/AfBO4IT8fDHw/bpFZGZm/UKRayB7R8Qekh4AiIhXJPWvdRfNzKzPFUkgK3MJ9wCQNBpYU9eozMyyJUuWVDUPZNmyZQCMGFG85mularwdHR0sWqSGrFc+b65Y8Vr5+SitWNq+SAKZDNwEjJH0DdK8kHOqisbMrAa9uaFgwoQJde+rkdrb23ng8SdYPXrLwm2G5KsUU+fOL9xm6NyXCu9b5C6sqyVNI00mFHBURDxRuIcyJG0O/BSYADwLHBsR692zKOly4DBgTkS8rdr2Zta/tcINBW1tbcyZO4djjlnRJ8er5MYbN2LM6O7no6wevSVL3/+Rusaw8Q1XFd63yEV0IuLJiPh+RHyvt8kjOwu4KyImkhasOqub/X4MHNyL9mZmVieFEkgdHAlckR9fARxVbqeIuAcod+5VqL2ZmdVPsxLIlhExCyD/HtPg9mZm1ksVr4Hku69ui4j3VntgSXcCW5XZdHa1x+oNSacApwCMHz++kV2bmQ1oFRNIRKyWtEzSZhGxsJoDV0o6kl6SNDYiZkkaSyrQWI3C7SPiEuASgEmTJkWV/ZiZWTeK3Mb7KvCIpDuApZ0vRkT1t0e87mbgJOC8/PsXDW5vZma9VCSB/Dr/9KXzgOslfQx4HvgggKStgR9FxKH5+bWkxay2kNQBnBsRl3XX3szMGqfIPJArcumSHfJLT/V2TfSIeJk0r6Tr6zOBQ0uen9B1n0rtzcyscYpU492XdKvss6SJhNtIOinfYmtmZoNUkSGsC4CDIuIpAEk7ANcCe9YzMDMze11HRwdDFy2uaqZ4LYbOfYmO15YV2rfIPJANO5MHQEQ8DWxYY2xmZjZAFF3S9jKgM+2dSFra1szMGqStrY3Zc+c3pBZW2+jNC+1bJIF8EjgNOIN0DeQe0iJTZmY2iHWbQCTdFREHAF+LiC8CFzYuLDMza3WVzkDGSvo74AhJ15HOPtaKiOIrvJiZ2YBTKYF8hVQmvY31zz4C2L9eQZmZWevrNoFExM+Bn0v6ckR8vYExmZlVVGl5156WcS26XKv1rMhMdCcPM+s3hg8f3ufHnDe3+jXRFy5Io/6bjSpew3XeXDFmdFXdNFWRu7DMzFpKI88gal0rfeGCdCY0ZvTEwm3GjG79tdlLOYGYmVVQa7Lq67XZIc0Sr2rN8gWvALBm1Bur6oO+mgci6S1AR0S8luti7QJcGRELCkdkZma9UsuZyfQFLwMwsWBCAGD05oX7KnIGcgMwSdL2wGWktTiuoaRqrpmZ1VctZ0L1OAsqVaQW1pqIWAUcDXwnIj4DjK1LNGZm1m8USSArJZ1AWvnvV/k1F1M0MxvkigxhfRT4BPCNiHhG0nbAT+oblpl15bkP1mqKzAN5XNIXgfH5+TOkJWXNrEXUY+6DWU+K3IV1OPAtYCNgO0m7kQosHlHn2MyshM8grNUUuQbyVWAvYAFARDwIbFe3iMzMrF8okkBWRcTCLq8Vn5tvZmYDUpGL6I9K+hAwVNJE0sJSf65vWGZm1uqKnIF8Cngr8BppAuFC4Mw6xmRmZv1AxTMQSUOBmyPivcDZjQnJzMz6g4pnIBGxGlgmabO+7FTS5pLukDQ9/y5b6UvS5ZLmSHq0y+tflfSipAfzj8uqmJk1WJEhrFeBRyRdJmly508v+z0LuCsiJgJ35efl/Bg4uJtt346I3fLPLb2Mx8zMqlTkIvqv809fOhLYNz++Argb+GLXnSLiHkkT+rhvMzPrA0Vmol8haSNgh/zSUxGxspf9bhkRs/LxZ0kaU8MxTpf0D8BU4LMR8Uq5nSSdApwCMH78+FrjNTOzLnocwsprgEwHvg/8AHha0j4F2t0p6dEyP0f2NmjgIuAtwG7ALOCC7naMiEsiYlJETBo9uh+tFWlm1uKKDGFdABwUEU8BSNoBuBbYs1KjfOdWWZJekjQ2n32MBeZUETMR8VLJsS7l9SrBZmbWIEUuom/YmTwAIuJpel/O/WZSeXjy719U0zgnnU5HA492t6+ZmdVHkQQyNd+BtW/+uRSY1st+zwMOlDQdODA/R9LWktbeUSXpWuAvwI6SOiR9LG86X9Ijkh4G9gM+08t4zMysSkWGsD4JnEYqYSLgHtK1kJpFxMvAAWVen0nJUrkRcUI37T/Sm/7NzKz3iiSQDYDvRsSFsHZ2+hvqGpWZmbW8IkNYdwGlq9UMB+6sTzhmZtZfFEkgwyJiSeeT/HhE/UIyM7P+oEgCWSppj84nkvYEltcvJDMz6w+KXAM5E/iZpJn5+VjguLpFZGZm/UKRUiZTJO0E7Ei6C+vJPihlYmZm/Vy3CUTS24EXImJ2RKzMw1jvB56T9NWImN+wKM3MrKzJkyfT3t5edtv06dMBOOOMM8pu33777bvdVkSlayA/BFYA5NpX5wFXklYkvKTmHs3MrCGGDx/O8OHDe96xRpWGsIaWnGUcB1wSETcAN0h6sG4RmZlZYb05g+itSmcgQyV1JpgDgN+WbCty8d3MzAawSongWuD3kuaRbtv9A4Ck7UnDWGZmNoh1m0Ai4huS7iLdtnt7RETeNAT4VCOCMzOz1lVxKCoi7i3z2tP1C8fMzPqLIjPRzczM1uMEYmZmNXECMTOzmjiBmJlZTZxAzMysJk4gZmZWE88oNzOrUa2FDHtbxLBVOIGYmdVBPYsYtgonEDOzGg2Es4je8DUQMzOrSVMSiKTNJd0haXr+/cYy+2wj6XeSnpD0mKRPV9PezMzqq1lnIGcBd0XEROCu/LyrVcBnI+JvgHcAp0nauYr2ZmZWR81KIEcCV+THVwBHdd0hImZFxP358WLgCWBc0fZmZlZfzUogW0bELEiJAhhTaWdJE4Ddgf+rtr2kUyRNlTR17ty5fRG7mZlRx7uwJN0JbFVm09lVHmckcANwZkQsqjaOiLiEvIb7pEmToofdzcysoLolkIh4b3fbJL0kaWxEzJI0FpjTzX4bkpLH1RFxY8mmQu3NzKx+mjWEdTNwUn58EvCLrjtIEnAZ8EREXFhtezMzqy+9vlJtAzuV3gRcD4wHngc+GBHzJW0N/CgiDpX0HtI67I8Aa3LTL0XELd21L9DvXOC5Xoa/BTCvl8forVaIAVojjlaIAVojjlaIAVojjlaIAVojjr6IYduIGN31xaYkkP5M0tSImDTYY2iVOFohhlaJoxViaJU4WiGGVomjnjF4JrqZmdXECcTMzGriBFK9S5odAK0RA7RGHK0QA7RGHK0QA7RGHK0QA7RGHHWLwddAzMysJj4DMTOzmjiBmJlZTZxACpJ0sKSnJLVLakr1X0mXS5oj6dFm9J9j6LbMfoPjGCbpPkkP5Tj+rRlx5FiGSnpA0q+aGMOzkh6R9KCkqU2KYZSkn0t6Mv/7eGcTYtgxfwadP4skndmEOD6T/10+KulaScOaEMOnc/+P1esz8DWQAiQNBZ4GDgQ6gCnACRHxeIPj2AdYAlwZEW9rZN8lMYwFxkbE/ZI2AaYBRzXhsxCwcUQsySVv/gh8OiLubWQcOZZ/ASYBm0bEYY3uP8fwLDApIpo2aU3SFcAfIuJHkjYCRkTEgibGMxR4Edg7Ino7gbiafseR/j3uHBHLJV0P3BIRP25gDG8DrgP2AlYAtwKfjIjpfdmPz0CK2Qtoj4gZEbGC9B/myEYHERH3AD3OuK9zDJXK7DcyjoiIJfnphvmn4X8NSWoD3gf8qNF9txJJmwL7kMoPERErmpk8sgOAvzYyeZTYABguaQNgBDCzwf3/DXBvRCyLiFXA74Gj+7oTJ5BixgEvlDzvoAlfmq2mTJn9Rvc/VNKDpGKad0REM+L4DvAFXi+30ywB3C5pmqRTmtD/m4G5wP/k4bwfSdq4CXGUOh64ttGdRsSLwLdIZZZmAQsj4vYGh/EosI+kN0kaARwKbNPXnTiBFKMyrw3qsb/eltnvCxGxOiJ2A9qAvfJpe8NIOgyYExHTGtlvN94dEXsAh5BW79ynwf1vAOwBXBQRuwNLaeJKoXkI7QjgZ03o+42kEYrtgK2BjSV9uJExRMQTwDeBO0jDVw+RVnntU04gxXSwbvZuo/GnpC2jQpn9pshDJXcDBze463cDR+TrD9cB+0v6SYNjACAiZubfc4CbSMOujdQBdJScBf6clFCa5RDg/oh4qQl9vxd4JiLmRsRK4EbgXY0OIiIui4g9ImIf0tB3n17/ACeQoqYAEyVtl/+yOZ5UUn7Q6aHMfiPjGC1pVH48nPQ/7ZONjCEi/jUi2iJiAunfxG8joqF/aQJI2jjf0EAeNjqINITRMBExG3hB0o75pQOAht5Y0cUJNGH4KnseeIekEfn/lwNI1wobStKY/Hs8cAx1+DzqtqDUQBIRqySdDtwGDAUuj4jHGh2HpGuBfYEtJHUA50bEZQ0O493AR4BH8vUHyGX2GxzHWOCKfKfNEOD6iGjabbRNtiVwU/quYgPgmoi4tQlxfAq4Ov+RNQP4aBNiII/5Hwh8vBn9R8T/Sfo5cD9p2OgBmlPS5Ia89MVK4LSIeKWvO/BtvGZmVhMPYZmZWU2cQMzMrCZOIGZmVhMnEDMzq4kTiJmZ1cQJxAYMSau7VGKdUGX7ZyVtUeb1I2qtwCxpgqSQ9PWS17aQtFLS9/Lzr0pa1nnffn5tSdfHkoZImpwrrD4iaUqem/R/+f0+L2luNe9f0r7NrCJs/ZvngdhAsjyXNulTEXEzvZs4OgM4DPhyfv5BoOs8onnAZ4EvVjjOcaTSGLtExJpcyHFpROwNIOlkUkXe03sRq1lhPgOxAUvSSEl3Sbo//8V+ZH59Y0m/VlpL5FFJx5U0+1TJ/jvl/U8uOVvYNh/z4fx7fH79x/ns4M+SZkj6QMkxlwNPSJqUnx8HXN8l3MuB4yRtXuEtjQVmRcQagIjoqGZymKS35/geUlpLZZMu2/fK2x/Iv3fMr7817/9gft8Te/gMbZBwArGBZHjJ8M1NwKvA0bnI4H7ABbm0xMHAzIjYNa+rUjpre17e/yLgc2X6+B5pPZZdgKuBySXbxgLvIZ1tnNel3XXA8fmsYTXr11JbQkoilRbouh44PL+/CyTtXmHfdeTZ4T8lrZmyK6n0y/Iuuz0J7JOLIX4F+I/8+ieA7+azu0mkuleVPkMbJJxAbCBZHhG75Z+jSVWU/0PSw8CdpBL8WwKPAO+V9E1JfxsRC0uO0VkcchowoUwf7wSuyY+vIiWMTv8bEWvy4lpbdml3K6m8xgmkL/JyJgMnKa2tsZ6I6AB2BP6VVD7+LkkHdHOsrnYknb1MycdalNeJKLUZ8DOlFS+/Dbw1v/4X4EuSvghsGxHLqfwZ2iDhBGID2YnAaGDP/NfzS8CwiHga2JP0Jfifkr5S0ua1/Hs1xa4RltYCeq3k8TpLAOSFyKaRrnPcUPZAqarwNcCp3XYW8VpE/CYiPk86QziqQIyd8fRUt+jrwO/yGcXhwLDc5zWk0ujLgdsk7d/DZ2iDhC+i20C2GWm9jpWS9gO2BZC0NTA/In6S73A6uYpj/plUefcqUoL6YxVtLwB+HxEv58KH5VxIqv683v+bkvYAZkfETElDgF2Ahwv2/SSwtaS3R8SUfP2j6xDWZqQlYKHkM5H0ZmBGREzOj3eR9CS1f4Y2QDiB2EB2NfBLSVOBB3m93Pv/A/5L0hpSpdJPVnHMM4DLJX2etAJf4YqzuYJzxSrOETEvX7/5TJnNY4BLJb0hP7+PdE2mSN8r8oXu/1Yqf7+cdB2k1PmkCsf/Avy25PXjgA9LWgnMBr4GvJ3aP0MbIFyN18zMauJrIGZmVhMnEDMzq4kTiJmZ1cQJxMzMauIEYmZmNXECMTOzmjiBmJlZTf4/FkJRjcBEAT4AAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "valid_AUCs = torch.zeros((nbr_seeds,))\n",
    "test_AUCs = torch.zeros((nbr_seeds,))\n",
    "\n",
    "for seed_idx in tqdm(range(nbr_seeds)):\n",
    "    \n",
    "    torch.manual_seed(seed_idx)\n",
    "    random.seed(seed_idx)\n",
    "    np.random.seed(seed_idx)\n",
    "    \n",
    "    # 128 is batch_size, useless parameter here but we re-use the deep models import function for consistency\n",
    "    _, _, _, x_train, y_train, x_valid, y_valid, x_test, y_test = get_dataloaders_MNIST_FashionMNIST(128, normal_cls, dataset, seed=seed_idx)\n",
    "    \n",
    "    x_train = np.reshape(x_train, (x_train.shape[0], -1)) / 255\n",
    "    x_valid = np.reshape(x_valid, (x_valid.shape[0], -1)) / 255\n",
    "    x_test = np.reshape(x_test, (x_test.shape[0], -1)) / 255\n",
    "    \n",
    "    y_valid_AD = convert_labels(y_valid, normal_cls)\n",
    "    y_test_AD = convert_labels(y_test, normal_cls)\n",
    "    \n",
    "    clf = IsolationForest(random_state=seed_idx).fit(x_train)\n",
    "    y_score_val = clf.decision_function(x_valid)\n",
    "    y_score_test = clf.decision_function(x_test)\n",
    "    \n",
    "    valid_AUCs[seed_idx] = roc_auc_score(y_valid_AD, y_score_val*-1)\n",
    "    test_AUCs[seed_idx] = roc_auc_score(y_test_AD, y_score_test*-1)\n",
    "\n",
    "print(\"MEAN VALID AUC +/- STD: {} +/- {}\".format(torch.mean(valid_AUCs),torch.std(valid_AUCs)))\n",
    "print(\"MEAN TEST AUC +/- STD: {} +/- {}\".format(torch.mean(test_AUCs),torch.std(test_AUCs)))\n",
    "\n",
    "plt.figure()\n",
    "boxplot(y=y_score_test, x=y_test, showfliers=False)\n",
    "plt.xlabel('{} class'.format(dataset))\n",
    "plt.ylabel('Scores for each {} class'.format(dataset))\n",
    "plt.title('Normal classes: {}\\nLAST SEED EXPERIMENT ONLY'.format(normal_cls))\n",
    "\n",
    "with open('last_results.txt', \"a\") as file:\n",
    "    file.write(\"IF-{}-Norm{}-{}seeds - MEAN TEST AUC +/- STD: {} +/- {}\\n\".format(dataset, normal_cls, nbr_seeds, torch.mean(test_AUCs),torch.std(test_AUCs)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "da7d71a9",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 3/3 [01:31<00:00, 30.36s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "MEAN VALID AUC +/- STD: 0.7892044186592102 +/- 0.003036376554518938\n",
      "MEAN TEST AUC +/- STD: 0.7815338969230652 +/- 0.005380204878747463\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "valid_AUCs = torch.zeros((nbr_seeds,))\n",
    "test_AUCs = torch.zeros((nbr_seeds,))\n",
    "\n",
    "for seed_idx in tqdm(range(nbr_seeds)):\n",
    "    \n",
    "    torch.manual_seed(seed_idx)\n",
    "    random.seed(seed_idx)\n",
    "    np.random.seed(seed_idx)\n",
    "    \n",
    "    # 128 is batch_size, useless parameter here but we re-use the deep models import function for consistency\n",
    "    _, _, _, x_train, y_train, x_valid, y_valid, x_test, y_test = get_dataloaders_MNIST_FashionMNIST(128, normal_cls, dataset, seed=seed_idx)\n",
    "    \n",
    "    x_train = np.reshape(x_train, (x_train.shape[0], -1)) / 255\n",
    "    x_valid = np.reshape(x_valid, (x_valid.shape[0], -1)) / 255\n",
    "    x_test = np.reshape(x_test, (x_test.shape[0], -1)) / 255\n",
    "    \n",
    "    y_valid_AD = convert_labels(y_valid, normal_cls)\n",
    "    y_test_AD = convert_labels(y_test, normal_cls)\n",
    "    \n",
    "    clf = OneClassSVM(nu=0.1).fit(x_train)\n",
    "    y_score_val = clf.decision_function(x_valid)\n",
    "    y_score_test = clf.decision_function(x_test)\n",
    "    \n",
    "    valid_AUCs[seed_idx] = roc_auc_score(y_valid_AD, y_score_val*-1)\n",
    "    test_AUCs[seed_idx] = roc_auc_score(y_test_AD, y_score_test*-1)\n",
    "\n",
    "print(\"MEAN VALID AUC +/- STD: {} +/- {}\".format(torch.mean(valid_AUCs),torch.std(valid_AUCs)))\n",
    "print(\"MEAN TEST AUC +/- STD: {} +/- {}\".format(torch.mean(test_AUCs),torch.std(test_AUCs)))\n",
    "\n",
    "plt.figure()\n",
    "boxplot(y=y_score_test, x=y_test, showfliers=False)\n",
    "plt.xlabel('{} class'.format(dataset))\n",
    "plt.ylabel('Scores for each {} class'.format(dataset))\n",
    "plt.title('Normal classes: {}\\nLAST SEED EXPERIMENT ONLY'.format(normal_cls))\n",
    "\n",
    "with open('last_results.txt', \"a\") as file:\n",
    "    file.write(\"OCSVM-{}-Norm{}-{}seeds - MEAN TEST AUC +/- STD: {} +/- {}\\n\".format(dataset, normal_cls, nbr_seeds, torch.mean(test_AUCs),torch.std(test_AUCs)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "b2c191bf",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 3/3 [00:29<00:00,  9.86s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "MEAN VALID AUC +/- STD: 0.865561306476593 +/- 0.0013079264899715781\n",
      "MEAN TEST AUC +/- STD: 0.8641157150268555 +/- 0.0077929929830133915\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "valid_AUCs = torch.zeros((nbr_seeds,))\n",
    "test_AUCs = torch.zeros((nbr_seeds,))\n",
    "\n",
    "for seed_idx in tqdm(range(nbr_seeds)):\n",
    "    \n",
    "    torch.manual_seed(seed_idx)\n",
    "    random.seed(seed_idx)\n",
    "    np.random.seed(seed_idx)\n",
    "    \n",
    "    # 128 is batch_size, useless parameter here but we re-use the deep models import function for consistency\n",
    "    _, _, _, x_train, y_train, x_valid, y_valid, x_test, y_test = get_dataloaders_MNIST_FashionMNIST(128, normal_cls, dataset, seed=seed_idx)\n",
    "    \n",
    "    x_train = np.reshape(x_train, (x_train.shape[0], -1)) / 255\n",
    "    x_valid = np.reshape(x_valid, (x_valid.shape[0], -1)) / 255\n",
    "    x_test = np.reshape(x_test, (x_test.shape[0], -1)) / 255\n",
    "    \n",
    "    y_valid_AD = convert_labels(y_valid, normal_cls)\n",
    "    y_test_AD = convert_labels(y_test, normal_cls)\n",
    "    \n",
    "    clf = LocalOutlierFactor(n_neighbors=64, contamination=0.1, novelty=True).fit(x_train)\n",
    "    y_score_val = clf.decision_function(x_valid)\n",
    "    y_score_test = clf.decision_function(x_test)\n",
    "    \n",
    "    valid_AUCs[seed_idx] = roc_auc_score(y_valid_AD, y_score_val*-1)\n",
    "    test_AUCs[seed_idx] = roc_auc_score(y_test_AD, y_score_test*-1)\n",
    "\n",
    "print(\"MEAN VALID AUC +/- STD: {} +/- {}\".format(torch.mean(valid_AUCs),torch.std(valid_AUCs)))\n",
    "print(\"MEAN TEST AUC +/- STD: {} +/- {}\".format(torch.mean(test_AUCs),torch.std(test_AUCs)))\n",
    "\n",
    "plt.figure()\n",
    "boxplot(y=y_score_test, x=y_test, showfliers=False)\n",
    "plt.xlabel('{} class'.format(dataset))\n",
    "plt.ylabel('Scores for each {} class'.format(dataset))\n",
    "plt.title('Normal classes: {}\\nLAST SEED EXPERIMENT ONLY'.format(normal_cls))\n",
    "\n",
    "with open('last_results.txt', \"a\") as file:\n",
    "    file.write(\"LOF-{}-Norm{}-{}seeds - MEAN TEST AUC +/- STD: {} +/- {}\\n\".format(dataset, normal_cls, nbr_seeds, torch.mean(test_AUCs),torch.std(test_AUCs)))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a2543dd9",
   "metadata": {},
   "source": [
    "# IF, OC-SVM and LOF with PCA dimensionality reduction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "b0281885",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 3/3 [00:06<00:00,  2.00s/it]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "MEAN VALID AUC +/- STD: 0.8462562561035156 +/- 0.006732582114636898\n",
      "MEAN TEST AUC +/- STD: 0.8434601426124573 +/- 0.01108610164374113\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "valid_AUCs = torch.zeros((nbr_seeds,))\n",
    "test_AUCs = torch.zeros((nbr_seeds,))\n",
    "\n",
    "for seed_idx in tqdm(range(nbr_seeds)):\n",
    "    \n",
    "    torch.manual_seed(seed_idx)\n",
    "    random.seed(seed_idx)\n",
    "    np.random.seed(seed_idx)\n",
    "    \n",
    "    # 128 is batch_size, useless parameter here but we re-use the deep models import function for consistency\n",
    "    _, _, _, x_train, y_train, x_valid, y_valid, x_test, y_test = get_dataloaders_MNIST_FashionMNIST(128, normal_cls, dataset, seed=seed_idx)\n",
    "    \n",
    "    x_train = np.reshape(x_train, (x_train.shape[0], -1)) / 255\n",
    "    x_valid = np.reshape(x_valid, (x_valid.shape[0], -1)) / 255\n",
    "    x_test = np.reshape(x_test, (x_test.shape[0], -1)) / 255\n",
    "    \n",
    "    x_train = pca.fit_transform(x_train)\n",
    "    x_valid = pca.transform(x_valid)\n",
    "    x_test = pca.transform(x_test)\n",
    "    \n",
    "    y_valid_AD = convert_labels(y_valid, normal_cls)\n",
    "    y_test_AD = convert_labels(y_test, normal_cls)\n",
    "    \n",
    "    clf = IsolationForest(random_state=seed_idx).fit(x_train)\n",
    "    y_score_val = clf.decision_function(x_valid)\n",
    "    y_score_test = clf.decision_function(x_test)\n",
    "    \n",
    "    valid_AUCs[seed_idx] = roc_auc_score(y_valid_AD, y_score_val*-1)\n",
    "    test_AUCs[seed_idx] = roc_auc_score(y_test_AD, y_score_test*-1)\n",
    "\n",
    "print(\"MEAN VALID AUC +/- STD: {} +/- {}\".format(torch.mean(valid_AUCs),torch.std(valid_AUCs)))\n",
    "print(\"MEAN TEST AUC +/- STD: {} +/- {}\".format(torch.mean(test_AUCs),torch.std(test_AUCs)))\n",
    "\n",
    "plt.figure()\n",
    "boxplot(y=y_score_test, x=y_test, showfliers=False)\n",
    "plt.xlabel('{} class'.format(dataset))\n",
    "plt.ylabel('Scores for each {} class'.format(dataset))\n",
    "plt.title('Normal classes: {}\\nLAST SEED EXPERIMENT ONLY'.format(normal_cls))\n",
    "\n",
    "with open('last_results.txt', \"a\") as file:\n",
    "    file.write(\"IF-PCA-{}-Norm{}-{}seeds - MEAN TEST AUC +/- STD: {} +/- {}\\n\".format(dataset, normal_cls, nbr_seeds, torch.mean(test_AUCs),torch.std(test_AUCs)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "6f4cd23b",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 3/3 [00:24<00:00,  8.15s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "MEAN VALID AUC +/- STD: 0.782134473323822 +/- 0.004652039147913456\n",
      "MEAN TEST AUC +/- STD: 0.7723626494407654 +/- 0.005159242078661919\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "valid_AUCs = torch.zeros((nbr_seeds,))\n",
    "test_AUCs = torch.zeros((nbr_seeds,))\n",
    "\n",
    "for seed_idx in tqdm(range(nbr_seeds)):\n",
    "    \n",
    "    torch.manual_seed(seed_idx)\n",
    "    random.seed(seed_idx)\n",
    "    np.random.seed(seed_idx)\n",
    "    \n",
    "    # 128 is batch_size, useless parameter here but we re-use the deep models import function for consistency\n",
    "    _, _, _, x_train, y_train, x_valid, y_valid, x_test, y_test = get_dataloaders_MNIST_FashionMNIST(128, normal_cls, dataset, seed=seed_idx)\n",
    "    \n",
    "    x_train = np.reshape(x_train, (x_train.shape[0], -1)) / 255\n",
    "    x_valid = np.reshape(x_valid, (x_valid.shape[0], -1)) / 255\n",
    "    x_test = np.reshape(x_test, (x_test.shape[0], -1)) / 255\n",
    "    \n",
    "    x_train = pca.fit_transform(x_train)\n",
    "    x_valid = pca.transform(x_valid)\n",
    "    x_test = pca.transform(x_test)\n",
    "    \n",
    "    y_valid_AD = convert_labels(y_valid, normal_cls)\n",
    "    y_test_AD = convert_labels(y_test, normal_cls)\n",
    "    \n",
    "    clf = OneClassSVM(nu=0.1).fit(x_train)\n",
    "    y_score_val = clf.decision_function(x_valid)\n",
    "    y_score_test = clf.decision_function(x_test)\n",
    "    \n",
    "    valid_AUCs[seed_idx] = roc_auc_score(y_valid_AD, y_score_val*-1)\n",
    "    test_AUCs[seed_idx] = roc_auc_score(y_test_AD, y_score_test*-1)\n",
    "\n",
    "print(\"MEAN VALID AUC +/- STD: {} +/- {}\".format(torch.mean(valid_AUCs),torch.std(valid_AUCs)))\n",
    "print(\"MEAN TEST AUC +/- STD: {} +/- {}\".format(torch.mean(test_AUCs),torch.std(test_AUCs)))\n",
    "\n",
    "plt.figure()\n",
    "boxplot(y=y_score_test, x=y_test, showfliers=False)\n",
    "plt.xlabel('{} class'.format(dataset))\n",
    "plt.ylabel('Scores for each {} class'.format(dataset))\n",
    "plt.title('Normal classes: {}\\nLAST SEED EXPERIMENT ONLY'.format(normal_cls))\n",
    "\n",
    "with open('last_results.txt', \"a\") as file:\n",
    "    file.write(\"OCSVM-PCA-{}-Norm{}-{}seeds - MEAN TEST AUC +/- STD: {} +/- {}\\n\".format(dataset, normal_cls, nbr_seeds, torch.mean(test_AUCs),torch.std(test_AUCs)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "f968f566",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 3/3 [00:21<00:00,  7.27s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "MEAN VALID AUC +/- STD: 0.8596253395080566 +/- 0.0031973335426300764\n",
      "MEAN TEST AUC +/- STD: 0.8559381365776062 +/- 0.00891717430204153\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "valid_AUCs = torch.zeros((nbr_seeds,))\n",
    "test_AUCs = torch.zeros((nbr_seeds,))\n",
    "\n",
    "for seed_idx in tqdm(range(nbr_seeds)):\n",
    "    \n",
    "    torch.manual_seed(seed_idx)\n",
    "    random.seed(seed_idx)\n",
    "    np.random.seed(seed_idx)\n",
    "    \n",
    "    # 128 is batch_size, useless parameter here but we re-use the deep models import function for consistency\n",
    "    _, _, _, x_train, y_train, x_valid, y_valid, x_test, y_test = get_dataloaders_MNIST_FashionMNIST(128, normal_cls, dataset, seed=seed_idx)\n",
    "    \n",
    "    x_train = np.reshape(x_train, (x_train.shape[0], -1)) / 255\n",
    "    x_valid = np.reshape(x_valid, (x_valid.shape[0], -1)) / 255\n",
    "    x_test = np.reshape(x_test, (x_test.shape[0], -1)) / 255\n",
    "    \n",
    "    x_train = pca.fit_transform(x_train)\n",
    "    x_valid = pca.transform(x_valid)\n",
    "    x_test = pca.transform(x_test)\n",
    "    \n",
    "    y_valid_AD = convert_labels(y_valid, normal_cls)\n",
    "    y_test_AD = convert_labels(y_test, normal_cls)\n",
    "    \n",
    "    clf = LocalOutlierFactor(n_neighbors=64, contamination=0.1, novelty=True).fit(x_train)\n",
    "    y_score_val = clf.decision_function(x_valid)\n",
    "    y_score_test = clf.decision_function(x_test)\n",
    "    \n",
    "    valid_AUCs[seed_idx] = roc_auc_score(y_valid_AD, y_score_val*-1)\n",
    "    test_AUCs[seed_idx] = roc_auc_score(y_test_AD, y_score_test*-1)\n",
    "\n",
    "print(\"MEAN VALID AUC +/- STD: {} +/- {}\".format(torch.mean(valid_AUCs),torch.std(valid_AUCs)))\n",
    "print(\"MEAN TEST AUC +/- STD: {} +/- {}\".format(torch.mean(test_AUCs),torch.std(test_AUCs)))\n",
    "\n",
    "plt.figure()\n",
    "boxplot(y=y_score_test, x=y_test, showfliers=False)\n",
    "plt.xlabel('{} class'.format(dataset))\n",
    "plt.ylabel('Scores for each {} class'.format(dataset))\n",
    "plt.title('Normal classes: {}\\nLAST SEED EXPERIMENT ONLY'.format(normal_cls))\n",
    "\n",
    "with open('last_results.txt', \"a\") as file:\n",
    "    file.write(\"LOF-PCA-{}-Norm{}-{}seeds - MEAN TEST AUC +/- STD: {} +/- {}\\n\".format(dataset, normal_cls, nbr_seeds, torch.mean(test_AUCs),torch.std(test_AUCs)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "individual-judge",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}