siamese_model.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Siamese network implementation with the strips dataset\n",
    "\n",
    "With denoised and averaged images"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Dataset structure\n",
    "\n",
    "The images themselves are stored already cut as strips, with the characteristic extracted by denoising them and getting the residual, as described in Lukas et al. We have an indexing file to retrieve them (info data) in the following structure:\n",
    "\n",
    "* Info data:\n",
    "    * Student 0:\n",
    "        * Image 0:\n",
    "            * Strip 0,0\n",
    "            * Strip 0,256\n",
    "            * Strip 0,512\n",
    "            * ...\n",
    "        * Image 1:\n",
    "            * Strip 0,0\n",
    "            * ...\n",
    "        * ...\n",
    "    * Student 1:\n",
    "        * ...\n",
    "\n",
    "Strip naming convention:\n",
    "\n",
    "* Eurecom_NNN_picXG_III_XXXX_YYYY.PNG -> Strip from:\n",
    "    * position (XXXX,YYYY) (top, left), from: \n",
    "    * image picXG_III, from:\n",
    "    * Student with id NNN\n",
    "\n",
    "Then, the function get_batch generates a random sample from the dataset. It allows to draw either from the training or validation split of the dataset (defined at the level of the students, 0.8 of the students are in train and the rest in validation).\n",
    "The generated batch has the following shape:\n",
    "\n",
    "* (label, (batch_size, examples, height, width, channels), (batch_size, examples, height, width, channels))  Since images are grayscale, in fact channels should be 1\n",
    "\n",
    "More clearly:\n",
    "\n",
    "* batch:\n",
    "    * label 0:\n",
    "        * strip_x-y from Image j from Student i  <---------->  strip_w-t from Image k from Student l\n",
    "        * strip_x-y from Image j2 from Student i  <---------->  strip_w-t from Image k2 from Student l\n",
    "        * strip_x-y from Image j3 from Student i  <---------->  strip_w-t from Image k3 from Student l\n",
    "        * strip_x-y from Image j4 from Student i  <---------->  strip_w-t from Image k4 from Student l\n",
    "    * label 0:\n",
    "        * ...\n",
    "        \n",
    "    * label 1:\n",
    "        * strip_x-y from Image j from Student i  <---------->  strip_w-t from Image k from Student i\n",
    "        * strip_x-y from Image j2 from Student i  <---------->  strip_w-t from Image k2 from Student i\n",
    "        * strip_x-y from Image j3 from Student i  <---------->  strip_w-t from Image k3 from Student i\n",
    "        * strip_x-y from Image j4 from Student i  <---------->  strip_w-t from Image k4 from Student i\n",
    "    * label 1:\n",
    "        * ...\n",
    "        \n",
    "Note that ${i \\neq l, j \\neq j_i}$ etc. \n",
    "\n",
    "            \n",
    "            "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training pipeline\n",
    "\n",
    "During training, a batch is generated at random from the training slice, ensuring the previously described invariant. The next steps are as follows:\n",
    "\n",
    "* Images are normalized\n",
    "* We define the left and right input, according to the tuple in the pair (for label == 0, all strips belong to the same camera and position, while for label == 1, each input belongs to a different camera.)\n",
    "* All images from each input are averaged element-wise (separetely, of course)\n",
    "* The result is fed, one at a time to a CNN model defined at sequential_block. Note that the weights are the same for each input\n",
    "* The CNN outputs a feature vector of size 2048\n",
    "* The output from each of the inputs is compared by calculating the absolute distance between the two vectors\n",
    "* This result is fed into a final Dense layer, for classification, with sigmoid activation. \n",
    "    * A result close to 0 indicates that the strips from each input belong to different devices.\n",
    "    * A result close to 1 indicates that the strips from each input where taken by the same device.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install opencv-python"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import numpy as np\n",
    "import random as rng\n",
    "import cv2\n",
    "import json"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "info_data = []\n",
    "with open('/home/data/strips_socrates/dataset_info.json') as json_file: \n",
    "    info_data = json.load(json_file) "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "STRIP_SIZE = 256\n",
    "SUB_STRIP_SIZE = 128\n",
    "CHANNELS = 3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_batch(info_data, batch_size, examples, with_id=False, dataset = 'train'):\n",
    "    pairs = [np.zeros((batch_size, examples, 128, 128, CHANNELS)) for i in range(2)]\n",
    "    labels = np.zeros((batch_size, ))\n",
    "    labels[batch_size//2:] = 1\n",
    "    split_lhs = 0\n",
    "    split_rhs = len(info_data)-1\n",
    "    if dataset == 'train':\n",
    "        split_rhs = int((len(info_data)-1)*0.55)\n",
    "    elif dataset == 'validation':\n",
    "        split_lhs = int((len(info_data)-1)*0.55)+1\n",
    "        split_rhs = int((len(info_data)-1)*0.62)\n",
    "    elif dataset == 'test':\n",
    "        split_lhs = int((len(info_data)-1)*0.62)+1\n",
    "        split_rhs = int((len(info_data)-1))\n",
    "    num_stds = split_rhs-split_lhs+1\n",
    "    students = [rng.randint(split_lhs, split_rhs) for _ in range(batch_size)]\n",
    "    strips_loc = [rng.randint(0, len(info_data[i][0][1])-1) for i in students]\n",
    "    id_pairs = []\n",
    "    for i in range(batch_size):\n",
    "        std = students[i]\n",
    "        avl_images = [j for j in range(len(info_data[std]))]\n",
    "        exl = rng.sample(avl_images, examples)\n",
    "        avl_images = [j for j in avl_images if j not in exl]\n",
    "        srl = strips_loc[i]\n",
    "        strips1 = [ info_data[std][e][1][srl] for e in exl ]\n",
    "        strips2 = []\n",
    "        if i >= batch_size // 2:\n",
    "            exl2 = rng.sample(avl_images, examples)\n",
    "            strips2 = [info_data[std][e][1][srl] for e in exl2]\n",
    "        else:\n",
    "            std2 = split_lhs + (((std-split_lhs) + rng.randint(1, num_stds-1)) % num_stds)\n",
    "            srl2 = rng.randint(0, len(info_data[std2][0][1])-1)\n",
    "            avl_images2 = [j for j in range(len(info_data[std2]))]\n",
    "            exl2 = rng.sample(avl_images2, examples)\n",
    "            strips2 = [ info_data[std2][e][1][srl2] for e in exl2]\n",
    "        for k in range(examples):\n",
    "            rch = rng.randint(0, 127)\n",
    "            rcw = rng.randint(0, 127)\n",
    "            im1 = cv2.imread(strips1[k], cv2.IMREAD_UNCHANGED)\n",
    "            if im1 is None:\n",
    "                print(f\"Error reading {strips1[k]}, dataset:{dataset}\")\n",
    "            im2 = cv2.imread(strips2[k], cv2.IMREAD_UNCHANGED)\n",
    "            if im2 is None:\n",
    "                print(f\"Error reading {strips2[k]}, dataset:{dataset}\")\n",
    "            pairs[0][i,k,...] = im1[rch:rch+128,rcw:rcw+128,:]/255\n",
    "            pairs[1][i,k,...] = im2[rch:rch+128,rcw:rcw+128,:]/255\n",
    "        id_pairs.append((strips1, strips2))\n",
    "    if with_id:\n",
    "        return pairs, labels, id_pairs\n",
    "    else:\n",
    "        return pairs, labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "p,l,ids = get_batch(info_data, 10, 4, True, dataset=\"validation\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "p[0][0][0].shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.imshow(p[0][0][0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "for r,s in zip(l, ids):\n",
    "    print(r)\n",
    "    for a in s:\n",
    "        for b in a:\n",
    "            print(b)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def generate(info_data, examples, batch_size, dataset='train'):\n",
    "    while True:\n",
    "        pairs, labels = get_batch(info_data, batch_size, examples, False, dataset=dataset)\n",
    "        yield(pairs, labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "from tensorflow import keras\n",
    "from tensorflow.keras import layers\n",
    "from tensorflow.keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def sequential_block(input_shape = (128,128, 3), base_filters=64):\n",
    "    model = keras.Sequential()\n",
    "    model.add(layers.Conv2D(base_filters, (3,3), input_shape=input_shape, activation=\"relu\"))\n",
    "    #model.add(layers.SpatialDropout2D(0.1))\n",
    "    model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=1))\n",
    "    model.add(layers.BatchNormalization(renorm=True))\n",
    "    model.add(layers.Conv2D(base_filters*2, (3,3) ,activation=\"relu\"))\n",
    "    model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))\n",
    "    model.add(layers.BatchNormalization(renorm=True))\n",
    "    model.add(layers.Conv2D(base_filters*4, (3,3), activation=\"relu\"))\n",
    "    model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))\n",
    "    model.add(layers.BatchNormalization(renorm=True))\n",
    "    model.add(layers.Conv2D(base_filters*4, (3,3), activation=\"relu\"))\n",
    "    model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))\n",
    "    model.add(layers.BatchNormalization(renorm=True))\n",
    "    model.add(layers.Conv2D(base_filters*4, (3,3), activation=\"relu\"))\n",
    "    model.add(layers.MaxPooling2D(pool_size=(2, 2), strides=2))\n",
    "    model.add(layers.BatchNormalization(renorm=True))\n",
    "    #model.add(layers.GlobalAveragePooling2D())\n",
    "    model.add(layers.Flatten())\n",
    "    model.add(layers.Dropout(0.3))\n",
    "    model.add(layers.Dense(base_filters*32, activation='relu'))\n",
    "    \n",
    "    \n",
    "    \n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "seq = sequential_block(base_filters=32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "seq.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_siamese_model(input_shape = (4, 128,128,3), base_filters=16):\n",
    "\n",
    "    left_input = layers.Input(input_shape)\n",
    "    right_input = layers.Input(input_shape)\n",
    "    \n",
    "    # Convolutional Neural Network\n",
    "    subshape = input_shape[1:4]\n",
    "    model = sequential_block(subshape, base_filters)\n",
    "    \n",
    "    #average noise patterns for image examples\n",
    "    avged_input_l = layers.Average()([left_input[:,k,...] for k in range(input_shape[0])])\n",
    "    avged_input_r = layers.Average()([right_input[:,k,...] for k in range(input_shape[0])])\n",
    "\n",
    "    encoded_l = model(avged_input_l)\n",
    "    encoded_r = model(avged_input_r)\n",
    "    \n",
    "    # Add a customized layer to compute the absolute difference between the encodings\n",
    "    L1_layer = layers.Lambda(lambda tensors:K.abs(tensors[0] - tensors[1]))\n",
    "    L1_distance = L1_layer([encoded_l, encoded_r])\n",
    "    \n",
    "    # Add a dense layer with a sigmoid unit to generate the similarity score\n",
    "    prediction = layers.Dense(1,activation='sigmoid')(L1_distance)\n",
    "    \n",
    "    # Connect the inputs with the outputs\n",
    "    siamese_net = keras.Model(inputs=[left_input,right_input],outputs=prediction)\n",
    "    \n",
    "    # return the model\n",
    "    return siamese_net"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model\"\n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "Layer (type)                                                      Output Shape                                Param #                 Connected to                                                      \n",
      "========================================================================================================================================================================================================\n",
      "input_1 (InputLayer)                                              [(None, 4, 128, 128, 3)]                    0                                                                                         \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "input_2 (InputLayer)                                              [(None, 4, 128, 128, 3)]                    0                                                                                         \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem (SlicingOpLambda)                        (None, 128, 128, 3)                         0                       input_1[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_1 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_1[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_2 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_1[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_3 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_1[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_4 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_2[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_5 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_2[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_6 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_2[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "tf.__operators__.getitem_7 (SlicingOpLambda)                      (None, 128, 128, 3)                         0                       input_2[0][0]                                                     \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "average (Average)                                                 (None, 128, 128, 3)                         0                       tf.__operators__.getitem[0][0]                                    \n",
      "                                                                                                                                      tf.__operators__.getitem_1[0][0]                                  \n",
      "                                                                                                                                      tf.__operators__.getitem_2[0][0]                                  \n",
      "                                                                                                                                      tf.__operators__.getitem_3[0][0]                                  \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "average_1 (Average)                                               (None, 128, 128, 3)                         0                       tf.__operators__.getitem_4[0][0]                                  \n",
      "                                                                                                                                      tf.__operators__.getitem_5[0][0]                                  \n",
      "                                                                                                                                      tf.__operators__.getitem_6[0][0]                                  \n",
      "                                                                                                                                      tf.__operators__.getitem_7[0][0]                                  \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "sequential (Sequential)                                           (None, 1024)                                3669600                 average[0][0]                                                     \n",
      "                                                                                                                                      average_1[0][0]                                                   \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "lambda (Lambda)                                                   (None, 1024)                                0                       sequential[0][0]                                                  \n",
      "                                                                                                                                      sequential[1][0]                                                  \n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n",
      "dense_1 (Dense)                                                   (None, 1)                                   1025                    lambda[0][0]                                                      \n",
      "========================================================================================================================================================================================================\n",
      "Total params: 3,670,625\n",
      "Trainable params: 3,668,225\n",
      "Non-trainable params: 2,400\n",
      "________________________________________________________________________________________________________________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = get_siamese_model(base_filters=32)\n",
    "model.summary(200)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "initial_learning_rate = 1e-5\n",
    "end_learning_rate = 1e-7\n",
    "decay_steps = 12800 * 5\n",
    "lr_schedule = keras.optimizers.schedules.PolynomialDecay(\n",
    "    initial_learning_rate, decay_steps, end_learning_rate, power=1\n",
    ")\n",
    "model.compile(\n",
    "        loss=\"binary_crossentropy\",\n",
    "        optimizer=keras.optimizers.Adam(learning_rate=lr_schedule),\n",
    "        metrics=[keras.metrics.Precision(name='precision'),\n",
    "             keras.metrics.Recall(name='recall'), \"accuracy\"] \n",
    "    )\n",
    "run_name = \"rand_crop05\"\n",
    "tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=\"./logs/\"+run_name)\n",
    "\n",
    "checkpoint_filepath = \"checkPts/siamese_\"+run_name+\".h5\"\n",
    "model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(\n",
    "    filepath=checkpoint_filepath,\n",
    "    save_best_only=True)\n",
    "\n",
    "early_stopping_callback = keras.callbacks.EarlyStopping(monitor='val_loss', patience=40, verbose=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 16\n",
    "epochs = 500\n",
    "examples = 4\n",
    "steps_epoch = 128"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "val_set = get_batch(info_data, 300, 4, False, 'validation')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.fit(generate(info_data, examples, batch_size, dataset='train'), \n",
    "         epochs = epochs,\n",
    "          initial_epoch=9,\n",
    "         steps_per_epoch= steps_epoch,\n",
    "          callbacks=[tensorboard_callback,\n",
    "                     model_checkpoint_callback,\n",
    "                     early_stopping_callback],\n",
    "          validation_batch_size=16,\n",
    "          validation_data=val_set,\n",
    "         shuffle = True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save(\"last-hack02.h5\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.load_weights(\"checkPts/siamese_rand_crop05.h5\") #best model so far"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Validation:\n",
    "Simple, just feed pairs from the validation split and see the performance."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "val_set = get_batch(info_data, 1, 4, False, 'whole')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "preds = model.predict(val_set[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "corr = [y == (p>0) for y, p in zip (val_set, preds)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "round(preds[0][0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "cs = np.zeros((2,2))\n",
    "for y,p in zip (val_set[1], preds):\n",
    "    cs[int(y), round(p[0])] += 1 \n",
    "acc = (cs[0,0] + cs [1 , 1]) / np.sum(cs)\n",
    "rec = cs[1,1] / np.sum(cs[1])\n",
    "print(cs, acc, rec)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Extended validation: \n",
    "Take two images, and compare strip by strip, then do a majority voting."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_extended(model, info_data, std1, std2, ims1, ims2, max_test_per_pair):\n",
    "    examples = len(ims1)\n",
    "    pair = [np.zeros((1, examples, 128, 128, CHANNELS)) for i in range(2)]\n",
    "    nstrips1 = len(info_data[std1][0][1])\n",
    "    nstrips2 = len(info_data[std2][0][1])\n",
    "        \n",
    "    all_strips1 = [[info_data[std1][im][1][i] for im in ims1] for i in range(nstrips1)]\n",
    "    all_strips2 = [[info_data[std2][im][1][i] for im in ims2] for i in range(nstrips2)]\n",
    "    \n",
    "    predicts = []\n",
    "    mt = [0]*max_test_per_pair\n",
    "    for strips1, strips2, _ in zip(all_strips1, all_strips2, mt):    \n",
    "        for k in range(examples):\n",
    "            s1 = strips1[k]\n",
    "            s2 = strips2[k]\n",
    "            im1 = cv2.imread(s1)[64:256-64, 64:256-64,:]\n",
    "            im2 = cv2.imread(s2)[64:256-64, 64:256-64,:]\n",
    "            if im1 is None:\n",
    "                print(f\"Error reading {strips1[k]}, dataset:{dataset}\")\n",
    "            if im2 is None:\n",
    "                print(f\"Error reading {strips2[k]}, dataset:{dataset}\")\n",
    "            pair[0][0, k,:,:,:] = im1/255\n",
    "            pair[1][0, k,:,:,:] = im2/255\n",
    "        predicts.append(model.predict(pair))\n",
    "    return predicts"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "def test_extended(model, info_data, max_pairs, max_test_per_pair,  examples, dataset, thr):\n",
    "    split_lhs = 0\n",
    "    split_rhs = len(info_data)-1\n",
    "    if dataset == 'train':\n",
    "        split_rhs = int((len(info_data)-1)*0.52)\n",
    "    elif dataset == 'validation':\n",
    "        split_lhs = int((len(info_data)-1)*0.52)+1\n",
    "        split_rhs = int((len(info_data)-1)*0.61)\n",
    "    elif dataset == 'test':\n",
    "        split_lhs = int((len(info_data)-1)*0.61)+1\n",
    "        split_rhs = int((len(info_data)-1))\n",
    "    num_stds = split_rhs-split_lhs+1\n",
    "    students = [rng.randint(split_lhs, split_rhs) for _ in range(max_pairs)]\n",
    "    cs = np.zeros((2,2))\n",
    "    all_preds = [(0,0)]*max_pairs\n",
    "    progress = int (max_pairs * 0.1)\n",
    "    for i in range(max_pairs):\n",
    "        if i % progress == 0:\n",
    "            print(f'Progress: {i} of {max_pairs}')\n",
    "        std = students[i]\n",
    "        avl_images = [j for j in range(len(info_data[std]))]\n",
    "        exl = rng.sample(avl_images, examples)\n",
    "        avl_images = [j for j in avl_images if j not in exl]\n",
    "        if i >= max_pairs // 2:\n",
    "            exl2 = rng.sample(avl_images, examples)\n",
    "            preds = predict_extended(model, info_data, std, std, exl, exl2, max_test_per_pair)\n",
    "            score = sum(preds)/len(preds)\n",
    "            all_preds[i] = (1,score)\n",
    "            cs[1][int(score>thr)] += 1\n",
    "        else:\n",
    "            std2 = split_lhs + (((std-split_lhs) + rng.randint(1, num_stds-1)) % num_stds)\n",
    "            avl_images2 = [j for j in range(len(info_data[std2]))]\n",
    "            exl2 = rng.sample(avl_images2, examples)\n",
    "            preds = predict_extended(model, info_data, std, std2, exl, exl2, max_test_per_pair)\n",
    "            score = sum(preds)/len(preds)\n",
    "            all_preds[i] = (0,score)\n",
    "            cs[0][int(score>thr)] += 1\n",
    "    return all_preds, cs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Final testing, with test (unseen until now) set\n",
    "Note that I'm using 'whole' to have no split, but this is because the training and validation sets have been deleted from the system to make space for the test set. Thus all images from the following test belong to devices unseen before this point. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Progress: 0 of 1000\n",
      "Progress: 100 of 1000\n",
      "Progress: 200 of 1000\n",
      "Progress: 300 of 1000\n",
      "Progress: 400 of 1000\n",
      "Progress: 500 of 1000\n",
      "Progress: 600 of 1000\n",
      "Progress: 700 of 1000\n",
      "Progress: 800 of 1000\n",
      "Progress: 900 of 1000\n"
     ]
    }
   ],
   "source": [
    "preds, cs = test_extended(model, info_data, 1000, 40, 4, 'whole', 0.3 )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_true = [p[0] for p in preds]\n",
    "y_pred = [p[1].flatten() for p in preds]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.metrics import roc_auc_score, roc_curve\n",
    "from scipy.optimize import brentq\n",
    "from scipy.interpolate import interp1d\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [],
   "source": [
    "def conf_mat(y_true, y_pred, thr):\n",
    "    cm = np.zeros((2,2))\n",
    "    for yt, yp in zip(y_true,y_pred):\n",
    "        cm[yt][int(yp>thr)] +=1\n",
    "    return cm"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[429.,  71.],\n",
       "       [ 65., 435.]])"
      ]
     },
     "execution_count": 57,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "conf_mat(y_true,y_pred, 0.42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "roc_auc =roc_auc_score(y_true, y_pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [],
   "source": [
    "fpr, tpr, thrs = roc_curve(y_true, y_pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [],
   "source": [
    "eer = brentq(lambda x : 1. - x - interp1d(fpr, tpr)(x), 0., 1.)\n",
    "thresh = interp1d(fpr, thrs)(eer)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.1360000000005653"
      ]
     },
     "execution_count": 51,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "eer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.944748"
      ]
     },
     "execution_count": 52,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "roc_auc"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x720 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(10,10))\n",
    "lw = 2\n",
    "plt.plot(fpr, tpr, color='darkorange',\n",
    "         lw=lw, label='ROC curve (area = %0.2f)' % roc_auc)\n",
    "plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')\n",
    "plt.xlim([0.0, 1.0])\n",
    "plt.ylim([0.0, 1.05])\n",
    "plt.xlabel('False Positive Rate', fontsize=16)\n",
    "plt.ylabel('True Positive Rate', fontsize=18)\n",
    "plt.title('ROC (Test set 1000 random pairs)', fontsize=25)\n",
    "plt.legend(loc=\"lower right\", fontsize=20)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.6.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}