VAE WADI.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.2.0\n",
      "0.10.0\n"
     ]
    }
   ],
   "source": [
    "# Imports\n",
    "from __future__ import absolute_import\n",
    "from __future__ import division\n",
    "from __future__ import print_function\n",
    "\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import os\n",
    "\n",
    "import tensorflow as tf\n",
    "from tensorflow.python import tf2\n",
    "if not tf2.enabled():\n",
    "    import tensorflow.compat.v2 as tf\n",
    "    tf.enable_v2_behavior()\n",
    "    assert tf2.enabled()\n",
    "import tensorflow_probability as tfp\n",
    "\n",
    "tfk = tf.keras\n",
    "tfkl = tf.keras.layers\n",
    "tfpl = tfp.layers\n",
    "tfd = tfp.distributions\n",
    "\n",
    "from tensorflow.keras.utils import plot_model\n",
    "from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from sklearn.model_selection import train_test_split \n",
    "from sklearn.linear_model import LogisticRegression\n",
    "from sklearn.metrics import classification_report, accuracy_score, roc_auc_score, roc_curve\n",
    "from sklearn.preprocessing import MinMaxScaler\n",
    "from sklearn.manifold import TSNE\n",
    "from sklearn import preprocessing\n",
    "\n",
    "%matplotlib inline\n",
    "np.random.seed(0)\n",
    "tf.random.set_seed(0)\n",
    "print(tf.__version__)\n",
    "print(tfp.__version__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Utility Functions\n",
    "## Plots\n",
    "def tsne_plot(x1, y1, name=\"graph.png\"):\n",
    "    tsne = TSNE(n_components=2, random_state=0)\n",
    "    X_t = tsne.fit_transform(x1)\n",
    "    plt.scatter(X_t[np.where(y1 == 0), 0], X_t[np.where(y1 == 0), 1], marker='o', color='g', linewidth='1', alpha=0.8, label='Normal', s=2)\n",
    "    plt.scatter(X_t[np.where(y1 == 1), 0], X_t[np.where(y1 == 1), 1], marker='o', color='r', linewidth='1', alpha=0.8, label='Attack', s=2)\n",
    "    plt.legend(loc='best');\n",
    "    plt.savefig(name);\n",
    "    plt.show();\n",
    "    \n",
    "# Plot Keras training history\n",
    "def plot_loss(hist):\n",
    "    plt.plot(hist.history['loss'])\n",
    "    plt.plot(hist.history['val_loss'])\n",
    "    plt.title('model loss')\n",
    "    plt.ylabel('loss')\n",
    "    plt.xlabel('epoch')\n",
    "    plt.legend(['train', 'test'], loc='upper left')\n",
    "    plt.yscale('log',basey=10)\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "df = pd.read_csv(\"C:/Users/Ruben/Datasets/WADI/WADI_attackdataLABLE.csv\", low_memory=False)\n",
    "\n",
    "df.columns = [\n",
    "    'ROW','DATE','TIME','1_AIT_001_PV','1_AIT_002_PV','1_AIT_003_PV','1_AIT_004_PV','1_AIT_005_PV','1_FIT_001_PV','1_LS_001_AL',\n",
    "    '1_LS_002_AL','1_LT_001_PV','1_MV_001_STATUS','1_MV_002_STATUS','1_MV_003_STATUS','1_MV_004_STATUS','1_P_001_STATUS',\n",
    "    '1_P_002_STATUS','1_P_003_STATUS','1_P_004_STATUS','1_P_005_STATUS','1_P_006_STATUS','2_DPIT_001_PV','2_FIC_101_CO',\n",
    "    '2_FIC_101_PV','2_FIC_101_SP','2_FIC_201_CO','2_FIC_201_PV','2_FIC_201_SP','2_FIC_301_CO','2_FIC_301_PV','2_FIC_301_SP',\n",
    "    '2_FIC_401_CO','2_FIC_401_PV','2_FIC_401_SP','2_FIC_501_CO','2_FIC_501_PV','2_FIC_501_SP','2_FIC_601_CO','2_FIC_601_PV',\n",
    "    '2_FIC_601_SP','2_FIT_001_PV','2_FIT_002_PV','2_FIT_003_PV','2_FQ_101_PV','2_FQ_201_PV','2_FQ_301_PV','2_FQ_401_PV',\n",
    "    '2_FQ_501_PV','2_FQ_601_PV','2_LS_001_AL','2_LS_002_AL','2_LS_101_AH','2_LS_101_AL','2_LS_201_AH','2_LS_201_AL',\n",
    "    '2_LS_301_AH','2_LS_301_AL','2_LS_401_AH','2_LS_401_AL','2_LS_501_AH','2_LS_501_AL','2_LS_601_AH','2_LS_601_AL',\n",
    "    '2_LT_001_PV','2_LT_002_PV','2_MCV_007_CO','2_MCV_101_CO','2_MCV_201_CO','2_MCV_301_CO','2_MCV_401_CO','2_MCV_501_CO',\n",
    "    '2_MCV_601_CO','2_MV_001_STATUS','2_MV_002_STATUS','2_MV_003_STATUS','2_MV_004_STATUS','2_MV_005_STATUS','2_MV_006_STATUS',\n",
    "    '2_MV_009_STATUS','2_MV_101_STATUS','2_MV_201_STATUS','2_MV_301_STATUS','2_MV_401_STATUS','2_MV_501_STATUS',\n",
    "    '2_MV_601_STATUS','2_P_001_STATUS','2_P_002_STATUS','2_P_003_SPEED','2_P_003_STATUS','2_P_004_SPEED','2_P_004_STATUS',\n",
    "    '2_PIC_003_CO','2_PIC_003_PV','2_PIC_003_SP','2_PIT_001_PV','2_PIT_002_PV','2_PIT_003_PV','2_SV_101_STATUS',\n",
    "    '2_SV_201_STATUS','2_SV_301_STATUS','2_SV_401_STATUS','2_SV_501_STATUS','2_SV_601_STATUS','2A_AIT_001_PV','2A_AIT_002_PV',\n",
    "    '2A_AIT_003_PV','2A_AIT_004_PV','2B_AIT_001_PV','2B_AIT_002_PV','2B_AIT_003_PV','2B_AIT_004_PV','3_AIT_001_PV',\n",
    "    '3_AIT_002_PV','3_AIT_003_PV','3_AIT_004_PV','3_AIT_005_PV','3_FIT_001_PV','3_LS_001_AL','3_LT_001_PV','3_MV_001_STATUS',\n",
    "    '3_MV_002_STATUS','3_MV_003_STATUS','3_P_001_STATUS','3_P_002_STATUS','3_P_003_STATUS','3_P_004_STATUS','LEAK_DIFF_PRESSURE',\n",
    "    'PLANT_START_STOP_LOG','TOTAL_CONS_REQUIRED_FLOW','OUTCOME'\n",
    "]\n",
    "\n",
    "df = df.iloc[1:] # Row only contains labels\n",
    "df = df.drop([172802,172803]) # Rows contain no information, only nan values\n",
    "\n",
    "# 'ROW', 'DATE' and 'TIME' are irrelevant for the thesis.\n",
    "# '2_LS_001_AL', '2_LS_002_AL', '2_P_001_STATUS', '2_P_002_STATUS' contain only nan values\n",
    "# Other dropped columns contain either only 0s, only 1s or only 2s and are therefore irrelevant\n",
    "\n",
    "df = df.drop(['ROW', 'DATE', 'TIME', '1_LS_001_AL', '1_LS_002_AL', '1_P_002_STATUS', '1_P_004_STATUS', '2_LS_001_AL', \n",
    "              '2_LS_002_AL', '2_MV_001_STATUS', '2_MV_002_STATUS', '2_MV_004_STATUS', '2_MV_005_STATUS', '2_MV_009_STATUS',\n",
    "              '2_P_001_STATUS', '2_P_002_STATUS', '2_P_004_STATUS', '2_SV_101_STATUS', '2_SV_201_STATUS', '2_SV_301_STATUS',\n",
    "              '2_SV_401_STATUS', '2_SV_501_STATUS', '2_SV_601_STATUS', '3_AIT_001_PV', '3_LS_001_AL', '3_MV_001_STATUS',\n",
    "              '3_MV_002_STATUS', '3_MV_003_STATUS', '3_P_001_STATUS', '3_P_002_STATUS', '3_P_003_STATUS', '3_P_004_STATUS',\n",
    "              'PLANT_START_STOP_LOG'], axis = 1)\n",
    "\n",
    "# set OUTCOME data type to int and change normal values to 0 / attack values to -1 \n",
    "df['OUTCOME'] = df['OUTCOME'].astype(int)\n",
    "df['OUTCOME'].replace(to_replace = [1], value = 0, inplace = True)\n",
    "df['OUTCOME'].replace(to_replace = [-1], value = 1, inplace = True)\n",
    "\n",
    "# data types need to be numeric to be encoded to z-scores --> convert column object data types to numerics\n",
    "cols = df.columns[df.columns != 'OUTCOME']\n",
    "df[cols] = df[cols].apply(pd.to_numeric, errors='coerce')\n",
    "\n",
    "# Encoding the feature vectors to z-scores\n",
    "cols = list(df.columns[df.columns != 'OUTCOME'])\n",
    "for col in cols:\n",
    "    df[col] = ((df[col] - df[col].mean())/df[col].std(ddof=0))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# define a method to scale data, looping thru the columns, and passing a scaler\n",
    "from sklearn.preprocessing import MinMaxScaler\n",
    "scaler = MinMaxScaler(feature_range=(-1,1))\n",
    "cols = list(df.columns[df.columns != 'OUTCOME'])\n",
    "for col in cols:\n",
    "    df[col] = scaler.fit_transform(df[col].values.reshape(-1, 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "df, data_test = train_test_split(df, test_size=0.25)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(9977, 98)\n"
     ]
    }
   ],
   "source": [
    "attack = df[df['OUTCOME'] == 1]\n",
    "print(attack.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "normal = df[df['OUTCOME'] == 0].sample(9900)\n",
    "attack = df[df['OUTCOME'] == 1].sample(9900)\n",
    "\n",
    "df1 = normal.append(attack).sample(frac=1).reset_index(drop=True)\n",
    "\n",
    "X = df1.drop(['OUTCOME'], axis = 1).values\n",
    "Y = df1[\"OUTCOME\"].values\n",
    "\n",
    "# tsne_plot(X, Y, \"TSNE_WADI.png\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(19800, 97)\n",
      "97\n",
      "(97,)\n",
      "(19800,)\n"
     ]
    }
   ],
   "source": [
    "print(X.shape)\n",
    "print(X.shape[1])\n",
    "print(X[0].shape)\n",
    "print(Y.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:From C:\\Users\\Ruben\\miniconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\ops\\linalg\\linear_operator_lower_triangular.py:158: calling LinearOperator.__init__ (from tensorflow.python.ops.linalg.linear_operator) with graph_parents is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "Do not pass `graph_parents`.  They will  no longer be used.\n",
      "Model: \"encoder\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "dense (Dense)                (None, 80)                7840      \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 70)                5670      \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 60)                4260      \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 55)                3355      \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 14)                784       \n",
      "_________________________________________________________________\n",
      "multivariate_normal_tri_l (M ((None, 4), (None, 4))    0         \n",
      "=================================================================\n",
      "Total params: 21,909\n",
      "Trainable params: 21,909\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"decoder\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "dense_5 (Dense)              (None, 55)                275       \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 60)                3360      \n",
      "_________________________________________________________________\n",
      "dense_7 (Dense)              (None, 70)                4270      \n",
      "_________________________________________________________________\n",
      "dense_8 (Dense)              (None, 80)                5680      \n",
      "_________________________________________________________________\n",
      "dense_9 (Dense)              (None, 194)               15714     \n",
      "_________________________________________________________________\n",
      "independent_normal (Independ ((None, 97), (None, 97))  0         \n",
      "=================================================================\n",
      "Total params: 29,299\n",
      "Trainable params: 29,299\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"vae_mlp\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "encoder_input (InputLayer)   [(None, 97)]              0         \n",
      "_________________________________________________________________\n",
      "dense (Dense)                (None, 80)                7840      \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 70)                5670      \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 60)                4260      \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 55)                3355      \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 14)                784       \n",
      "_________________________________________________________________\n",
      "multivariate_normal_tri_l (M ((None, 4), (None, 4))    0         \n",
      "_________________________________________________________________\n",
      "decoder (Sequential)         (None, 97)                29299     \n",
      "=================================================================\n",
      "Total params: 51,208\n",
      "Trainable params: 51,208\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "original_dim = X.shape[1]\n",
    "input_shape = X[0].shape\n",
    "latent_dim = 4\n",
    "batch_size = 64\n",
    "max_epochs = 80\n",
    "\n",
    "prior = tfd.MultivariateNormalDiag(loc=tf.zeros([latent_dim]), scale_identity_multiplier=1.0)\n",
    "\n",
    "encoder = tfk.Sequential([\n",
    "    tfkl.InputLayer(input_shape=input_shape, name='encoder_input'),\n",
    "    tfkl.Dense(80, activation='relu'),\n",
    "    tfkl.Dense(70, activation='relu'),\n",
    "    tfkl.Dense(60, activation='relu'),\n",
    "    tfkl.Dense(55, activation='relu'),\n",
    "    tfkl.Dense(tfpl.MultivariateNormalTriL.params_size(latent_dim), activation=None),\n",
    "    tfpl.MultivariateNormalTriL(latent_dim, activity_regularizer=tfpl.KLDivergenceRegularizer(prior)),], name='encoder')\n",
    "\n",
    "\n",
    "encoder.summary()\n",
    "plot_model(encoder, to_file='vae_mlp_encoder.png', show_shapes=True)\n",
    "\n",
    "decoder = tfk.Sequential([\n",
    "    tfkl.InputLayer(input_shape=[latent_dim]),\n",
    "    tfkl.Dense(55, activation='relu'),\n",
    "    tfkl.Dense(60, activation='relu'),\n",
    "    tfkl.Dense(70, activation='relu'),\n",
    "    tfkl.Dense(80, activation='relu'),\n",
    "    tfkl.Dense(tfpl.IndependentNormal.params_size(original_dim), activation=None), \n",
    "    tfpl.IndependentNormal(original_dim),], name='decoder')\n",
    "\n",
    "decoder.summary()\n",
    "plot_model(decoder, to_file='vae_mlp_decoder.png', show_shapes=True)\n",
    "\n",
    "vae = tfk.Model(inputs=encoder.inputs, outputs=decoder(encoder.outputs[0]), name='vae_mlp')\n",
    "\n",
    "negloglik = lambda x, rv_x: -rv_x.log_prob(x)\n",
    "\n",
    "vae.compile(optimizer=tf.keras.optimizers.Nadam(), loss=negloglik, metrics=['accuracy'])\n",
    "\n",
    "vae.summary()\n",
    "\n",
    "plot_model(vae, to_file='vae_mlp.png', show_shapes=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "x = df.drop([\"OUTCOME\"], axis=1)\n",
    "y = df[\"OUTCOME\"].values\n",
    "\n",
    "x_norm, x_fraud = x.values[y == 0], x.values[y == 1]\n",
    "\n",
    "x_norm_sample = x_norm[np.random.randint(x_norm.shape[0], size=100000), :]\n",
    "x_norm_train_sample, x_norm_val_sample = train_test_split(x_norm_sample, test_size=0.2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(162824, 97)\n",
      "(9977, 97)\n",
      "(100000, 97)\n",
      "(80000, 97)\n",
      "(20000, 97)\n"
     ]
    }
   ],
   "source": [
    "print(x_norm.shape)\n",
    "print(x_fraud.shape)\n",
    "print(x_norm_sample.shape)\n",
    "print(x_norm_train_sample.shape)\n",
    "print(x_norm_val_sample.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 64\n",
    "max_epochs = 50"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf_train = tf.data.Dataset.from_tensor_slices(\n",
    "    (x_norm_train_sample, x_norm_train_sample)).batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE).shuffle(int(10e4))\n",
    "\n",
    "tf_val = tf.data.Dataset.from_tensor_slices(\n",
    "    (x_norm_val_sample, x_norm_val_sample)).batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE).shuffle(int(10e4))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "checkpointer = ModelCheckpoint(filepath='bestmodel.h5', verbose=0, save_best_only=True)\n",
    "earlystopper = EarlyStopping(monitor='val_loss', mode='min', min_delta=0.005, patience=20, verbose=0, restore_best_weights=True)\n",
    "\n",
    "hist = vae.fit(x_norm_train_sample, x_norm_train_sample, epochs=max_epochs, shuffle=True, verbose=1,\n",
    "               validation_data=(x_norm_val_sample, x_norm_val_sample), callbacks=[checkpointer, earlystopper])\n",
    "\n",
    "plot_loss(hist)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "reconstruct_samples_n = 100\n",
    "\n",
    "def reconstruction_log_prob(eval_samples, reconstruct_samples_n):\n",
    "    encoder_out = encoder(eval_samples)\n",
    "    encoder_samples = encoder_out.sample(reconstruct_samples_n)\n",
    "    return np.mean(decoder(encoder_samples).log_prob(eval_samples), axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "latent_x_mean = encoder(X).mean()\n",
    "plt.scatter(latent_x_mean[:, 0], latent_x_mean[:, 1], c=Y, cmap='RdYlGn_r', s=2)\n",
    "plt.title('latent means')\n",
    "plt.ylabel('mean[1]')\n",
    "plt.xlabel('mean[0]')\n",
    "plt.show()\n",
    "# tsne_plot(latent_x_mean, data_sample[\"Class\"].values, \"raw.png\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "latent_x_std = encoder(X).stddev()\n",
    "plt.scatter(latent_x_std[:, 0], latent_x_std[:, 1], c=Y, cmap='RdYlGn_r', s=2)\n",
    "plt.title('latent standard deviations')\n",
    "plt.ylabel('stddev[1]')\n",
    "plt.xlabel('stddev[0]')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "latent_x = encoder(X).sample()\n",
    "plt.scatter(latent_x[:, 0], latent_x[:, 1], c=Y, cmap='RdYlGn_r', s=2)\n",
    "plt.title('latent vector samples')\n",
    "plt.ylabel('z[1]')\n",
    "plt.xlabel('z[0]')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_log_prob = reconstruction_log_prob(X, reconstruct_samples_n)\n",
    "ax = plt.hist([x_log_prob[Y==0], x_log_prob[Y==1]], 97)\n",
    "plt.title('reconstruction log probability')\n",
    "plt.ylabel('frequency')\n",
    "plt.xlabel(\"log p(x|x')\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "fpr, tpr, thresh = roc_curve(Y, -x_log_prob)\n",
    "auc = roc_auc_score(Y, -x_log_prob)\n",
    "\n",
    "plt.plot(fpr,tpr,label=\"linear in-sample, auc=\"+str(auc))\n",
    "plt.title('VAE roc curve - training')\n",
    "plt.ylabel('True Positive Rate')\n",
    "plt.xlabel(\"False Positive Rate\")\n",
    "plt.legend(loc='best')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_test_log_prob = reconstruction_log_prob(data_test.drop(['OUTCOME'], axis = 1).values, reconstruct_samples_n)\n",
    "test_y = data_test[\"OUTCOME\"].values\n",
    "\n",
    "fpr, tpr, thresh = roc_curve(test_y, -x_test_log_prob)\n",
    "auc = roc_auc_score(test_y, -x_test_log_prob)\n",
    "\n",
    "plt.plot(fpr,tpr,label=\"linear in-sample, auc=\"+str(auc))\n",
    "plt.title('VAE roc curve - test')\n",
    "plt.ylabel('True Positive Rate')\n",
    "plt.xlabel(\"False Positive Rate\")\n",
    "plt.legend(loc='best')\n",
    "plt.show()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.7 (tensorflow)",
   "language": "python",
   "name": "tensorflow"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}