face_reg.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "import face_recognition\n",
    "from PIL import Image, ImageDraw\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# import libaries\n",
    "\n",
    "from PIL import Image\n",
    "from tqdm.notebook import tqdm\n",
    "import numpy as np\n",
    "import random\n",
    "import sys\n",
    "import pandas as pd\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from torch.utils.data import Dataset, DataLoader, SubsetRandomSampler\n",
    "from torchinfo import summary\n",
    "from torchvision import transforms\n",
    "from torch import optim\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, ReLU, ReLU6, Sigmoid, Dropout2d, Dropout, AvgPool2d, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module, Parameter\n",
    "from statistics import mean\n",
    "\n",
    "import datetime\n",
    "import os\n",
    "\n",
    "from sklearn.model_selection import KFold, train_test_split\n",
    "from sklearn.metrics import confusion_matrix\n",
    "from mtcnn import MTCNN\n",
    "import cv2\n",
    " \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Python version: 3.9.0 (tags/v3.9.0:9cf6752, Oct  5 2020, 15:34:40) [MSC v.1927 64 bit (AMD64)]\n",
      "CUDA Available: True\n",
      "GPU Name: NVIDIA GeForce RTX 2070\n",
      "GPU Properties:\n",
      " _CudaDeviceProperties(name='NVIDIA GeForce RTX 2070', major=7, minor=5, total_memory=8191MB, multi_processor_count=36)\n",
      "Tue Apr  4 11:40:57 2023       \n",
      "+-----------------------------------------------------------------------------+\n",
      "| NVIDIA-SMI 526.98       Driver Version: 526.98       CUDA Version: 12.0     |\n",
      "|-------------------------------+----------------------+----------------------+\n",
      "| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |\n",
      "| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |\n",
      "|                               |                      |               MIG M. |\n",
      "|===============================+======================+======================|\n",
      "|   0  NVIDIA GeForce ... WDDM  | 00000000:01:00.0  On |                  N/A |\n",
      "|  0%   45C    P8    25W / 175W |   1527MiB /  8192MiB |     16%      Default |\n",
      "|                               |                      |                  N/A |\n",
      "+-------------------------------+----------------------+----------------------+\n",
      "                                                                               \n",
      "+-----------------------------------------------------------------------------+\n",
      "| Processes:                                                                  |\n",
      "|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |\n",
      "|        ID   ID                                                   Usage      |\n",
      "|=============================================================================|\n",
      "|    0   N/A  N/A      4032    C+G   D:\\Downloads\\AnyDesk.exe        N/A      |\n",
      "|    0   N/A  N/A      7088    C+G   ...8bbwe\\WindowsTerminal.exe    N/A      |\n",
      "|    0   N/A  N/A     10588    C+G   ...e\\PhoneExperienceHost.exe    N/A      |\n",
      "|    0   N/A  N/A     12664    C+G   ...lPanel\\SystemSettings.exe    N/A      |\n",
      "|    0   N/A  N/A     13072    C+G   ...perience\\NVIDIA Share.exe    N/A      |\n",
      "|    0   N/A  N/A     14060    C+G   C:\\Windows\\explorer.exe         N/A      |\n",
      "|    0   N/A  N/A     15416    C+G   ...perience\\NVIDIA Share.exe    N/A      |\n",
      "|    0   N/A  N/A     16316    C+G   ...5n1h2txyewy\\SearchApp.exe    N/A      |\n",
      "|    0   N/A  N/A     16720      C   ...ython\\Python39\\python.exe    N/A      |\n",
      "|    0   N/A  N/A     17968    C+G   ...8wekyb3d8bbwe\\Cortana.exe    N/A      |\n",
      "|    0   N/A  N/A     18384    C+G   ...ge\\Application\\msedge.exe    N/A      |\n",
      "|    0   N/A  N/A     21532    C+G   ...2txyewy\\TextInputHost.exe    N/A      |\n",
      "|    0   N/A  N/A     24708    C+G   ...8wekyb3d8bbwe\\GameBar.exe    N/A      |\n",
      "+-----------------------------------------------------------------------------+\n"
     ]
    }
   ],
   "source": [
    "print('Python version:', sys.version)\n",
    "print('CUDA Available:', torch.cuda.is_available())\n",
    "\n",
    "if torch.cuda.is_available():\n",
    "    print('GPU Name:', torch.cuda.get_device_name())\n",
    "    print('GPU Properties:\\n', torch.cuda.get_device_properties('cuda'))\n",
    "    device = \"cuda\"\n",
    "    torch.cuda.set_per_process_memory_fraction(0.95, 0)\n",
    "    torch.cuda.empty_cache()\n",
    "else:\n",
    "    print(\"Cuda is not available, please use cpu instead\")\n",
    "    device = \"cpu\"\n",
    "!nvidia-smi"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "seed = 41\n",
    "img_size = 112"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "x\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "(True, <PIL.Image.Image image mode=RGB size=74x74>)"
      ]
     },
     "execution_count": 40,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "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": [
    "def face_reg(path: str):\n",
    "    try:\n",
    "        image = face_recognition.load_image_file(path)\n",
    "        face_locations = face_recognition.face_locations(image)\n",
    "\n",
    "        # print(\"I found {} face(s) in this photograph.\".format(len(face_locations)))\n",
    "\n",
    "        for face_location in face_locations:\n",
    "\n",
    "            # Print the location of each face in this image\n",
    "            top, right, bottom, left = face_location\n",
    "            if top - bottom < 224:\n",
    "                print('x')\n",
    "\n",
    "            # You can access the actual face itself like this:\n",
    "            face_image = image[top:bottom, left:right]\n",
    "            pil_image = Image.fromarray(face_image)\n",
    "            plt.imshow(pil_image)\n",
    "            # print(pil_image)\n",
    "        return (True, pil_image)\n",
    "    except:\n",
    "        print(False, None)\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "raw_ds_df = pd.read_csv('./csv_files/yt_ds.csv')\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Unnamed: 0</th>\n",
       "      <th>sub_identity</th>\n",
       "      <th>identity</th>\n",
       "      <th>path</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>1</td>\n",
       "      <td>0</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\0\\0.556.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>2</td>\n",
       "      <td>0</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\0\\0.557.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3</td>\n",
       "      <td>0</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\0\\0.558.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>4</td>\n",
       "      <td>0</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\0\\0.559.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>390403</th>\n",
       "      <td>390403</td>\n",
       "      <td>1</td>\n",
       "      <td>Zydrunas_Ilgauskas</td>\n",
       "      <td>./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1491.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>390404</th>\n",
       "      <td>390404</td>\n",
       "      <td>1</td>\n",
       "      <td>Zydrunas_Ilgauskas</td>\n",
       "      <td>./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1492.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>390405</th>\n",
       "      <td>390405</td>\n",
       "      <td>1</td>\n",
       "      <td>Zydrunas_Ilgauskas</td>\n",
       "      <td>./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1493.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>390406</th>\n",
       "      <td>390406</td>\n",
       "      <td>1</td>\n",
       "      <td>Zydrunas_Ilgauskas</td>\n",
       "      <td>./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1494.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>390407</th>\n",
       "      <td>390407</td>\n",
       "      <td>1</td>\n",
       "      <td>Zydrunas_Ilgauskas</td>\n",
       "      <td>./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1495.jpg</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>390408 rows × 4 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "        Unnamed: 0  sub_identity            identity  \\\n",
       "0                0             0       Aaron_Eckhart   \n",
       "1                1             0       Aaron_Eckhart   \n",
       "2                2             0       Aaron_Eckhart   \n",
       "3                3             0       Aaron_Eckhart   \n",
       "4                4             0       Aaron_Eckhart   \n",
       "...            ...           ...                 ...   \n",
       "390403      390403             1  Zydrunas_Ilgauskas   \n",
       "390404      390404             1  Zydrunas_Ilgauskas   \n",
       "390405      390405             1  Zydrunas_Ilgauskas   \n",
       "390406      390406             1  Zydrunas_Ilgauskas   \n",
       "390407      390407             1  Zydrunas_Ilgauskas   \n",
       "\n",
       "                                                     path  \n",
       "0             ./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg  \n",
       "1             ./frame_images_DB/Aaron_Eckhart\\0\\0.556.jpg  \n",
       "2             ./frame_images_DB/Aaron_Eckhart\\0\\0.557.jpg  \n",
       "3             ./frame_images_DB/Aaron_Eckhart\\0\\0.558.jpg  \n",
       "4             ./frame_images_DB/Aaron_Eckhart\\0\\0.559.jpg  \n",
       "...                                                   ...  \n",
       "390403  ./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1491.jpg  \n",
       "390404  ./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1492.jpg  \n",
       "390405  ./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1493.jpg  \n",
       "390406  ./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1494.jpg  \n",
       "390407  ./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1495.jpg  \n",
       "\n",
       "[390408 rows x 4 columns]"
      ]
     },
     "execution_count": 34,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "raw_ds_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "isFace = list()\n",
    "for index, row in raw_ds_df.iterrows():\n",
    "    face_recognition"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>identity</th>\n",
       "      <th>sub_identity</th>\n",
       "      <th>path</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/AJ_Cook\\0\\0.123.jpg, ./fram...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/AJ_Cook\\2\\2.131.jpg, ./fram...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\2\\2.1718.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3420</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/Zoran_Djindjic\\2\\2.1973.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3421</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\1\\1.1977.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3422</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\2\\2.1051.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3423</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>4</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\4\\4.155.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3424</th>\n",
       "      <td>Zydrunas_Ilgauskas</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1296...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>3425 rows × 3 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                identity  sub_identity  \\\n",
       "0                AJ_Cook             0   \n",
       "1                AJ_Cook             2   \n",
       "2          Aaron_Eckhart             0   \n",
       "3          Aaron_Eckhart             1   \n",
       "4          Aaron_Eckhart             2   \n",
       "...                  ...           ...   \n",
       "3420      Zoran_Djindjic             2   \n",
       "3421      Zulfiqar_Ahmed             1   \n",
       "3422      Zulfiqar_Ahmed             2   \n",
       "3423      Zulfiqar_Ahmed             4   \n",
       "3424  Zydrunas_Ilgauskas             1   \n",
       "\n",
       "                                                   path  \n",
       "0     [./frame_images_DB/AJ_Cook\\0\\0.123.jpg, ./fram...  \n",
       "1     [./frame_images_DB/AJ_Cook\\2\\2.131.jpg, ./fram...  \n",
       "2     [./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg, ...  \n",
       "3     [./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg, ...  \n",
       "4     [./frame_images_DB/Aaron_Eckhart\\2\\2.1718.jpg,...  \n",
       "...                                                 ...  \n",
       "3420  [./frame_images_DB/Zoran_Djindjic\\2\\2.1973.jpg...  \n",
       "3421  [./frame_images_DB/Zulfiqar_Ahmed\\1\\1.1977.jpg...  \n",
       "3422  [./frame_images_DB/Zulfiqar_Ahmed\\2\\2.1051.jpg...  \n",
       "3423  [./frame_images_DB/Zulfiqar_Ahmed\\4\\4.155.jpg,...  \n",
       "3424  [./frame_images_DB/Zydrunas_Ilgauskas\\1\\1.1296...  \n",
       "\n",
       "[3425 rows x 3 columns]"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# ds_df = pd.read_csv('./csv_files/vgg_face.csv')\n",
    "raw_ds_df = pd.read_csv('./csv_files/yt_ds.csv')\n",
    "# ds_df = pd.read_csv('./digiface_csv_files/digi_all.csv')\n",
    "# ds_df = pd.read_csv('./csv_files/lfw_ds.csv')\n",
    "ds_df = raw_ds_df.groupby(['identity', 'sub_identity'])['path'].apply(list).reset_index()\n",
    "ds_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "name_more = list()\n",
    "# more than 2 videos \n",
    "for name in list(ds_df['identity'].unique()):\n",
    "    if ds_df[ds_df['identity'] == name].shape[0] >= 2: \n",
    "        name_more.append(name)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "ds_df[ds_df['identity'].isin(name_more)]\n",
    "\n",
    "anc_df = pd.DataFrame()\n",
    "# append all identity \n",
    "for identity in name_more: \n",
    "    temp_anc = ds_df[ds_df['identity'] == identity].reset_index().T.pop(0)\n",
    "    anc_df = pd.concat([anc_df, pd.DataFrame(temp_anc).T], axis = 0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "anc_df= anc_df.set_index('index')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>identity</th>\n",
       "      <th>sub_identity</th>\n",
       "      <th>path</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>index</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/AJ_Cook\\0\\0.123.jpg, ./fram...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>6</th>\n",
       "      <td>Aaron_Sorkin</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Aaron_Sorkin\\0\\0.100.jpg, ....</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>8</th>\n",
       "      <td>Aaron_Tippin</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Aaron_Tippin\\0\\0.406.jpg, ....</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>10</th>\n",
       "      <td>Abba_Eban</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Abba_Eban\\0\\0.100.jpg, ./fr...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3410</th>\n",
       "      <td>Zhang_Ziyi</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Zhang_Ziyi\\1\\1.1000.jpg, ./...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3413</th>\n",
       "      <td>Zhong_Nanshan</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Zhong_Nanshan\\0\\0.1000.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3416</th>\n",
       "      <td>Zhu_Rongji</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Zhu_Rongji\\0\\0.561.jpg, ./f...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3418</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Zoran_Djindjic\\0\\0.1693.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3421</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\1\\1.1977.jpg...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>1004 rows × 3 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "             identity sub_identity  \\\n",
       "index                                \n",
       "0             AJ_Cook            0   \n",
       "2       Aaron_Eckhart            0   \n",
       "6        Aaron_Sorkin            0   \n",
       "8        Aaron_Tippin            0   \n",
       "10          Abba_Eban            0   \n",
       "...               ...          ...   \n",
       "3410       Zhang_Ziyi            1   \n",
       "3413    Zhong_Nanshan            0   \n",
       "3416       Zhu_Rongji            0   \n",
       "3418   Zoran_Djindjic            0   \n",
       "3421   Zulfiqar_Ahmed            1   \n",
       "\n",
       "                                                    path  \n",
       "index                                                     \n",
       "0      [./frame_images_DB/AJ_Cook\\0\\0.123.jpg, ./fram...  \n",
       "2      [./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg, ...  \n",
       "6      [./frame_images_DB/Aaron_Sorkin\\0\\0.100.jpg, ....  \n",
       "8      [./frame_images_DB/Aaron_Tippin\\0\\0.406.jpg, ....  \n",
       "10     [./frame_images_DB/Abba_Eban\\0\\0.100.jpg, ./fr...  \n",
       "...                                                  ...  \n",
       "3410   [./frame_images_DB/Zhang_Ziyi\\1\\1.1000.jpg, ./...  \n",
       "3413   [./frame_images_DB/Zhong_Nanshan\\0\\0.1000.jpg,...  \n",
       "3416   [./frame_images_DB/Zhu_Rongji\\0\\0.561.jpg, ./f...  \n",
       "3418   [./frame_images_DB/Zoran_Djindjic\\0\\0.1693.jpg...  \n",
       "3421   [./frame_images_DB/Zulfiqar_Ahmed\\1\\1.1977.jpg...  \n",
       "\n",
       "[1004 rows x 3 columns]"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "anc_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>identity</th>\n",
       "      <th>sub_identity</th>\n",
       "      <th>path</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/AJ_Cook\\0\\0.123.jpg, ./fram...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/AJ_Cook\\2\\2.131.jpg, ./fram...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>0</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\2\\2.1718.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3419</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Zoran_Djindjic\\1\\1.104.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3420</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/Zoran_Djindjic\\2\\2.1973.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3421</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>1</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\1\\1.1977.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3422</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>2</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\2\\2.1051.jpg...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3423</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>4</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\4\\4.155.jpg,...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>2834 rows × 3 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "            identity  sub_identity  \\\n",
       "0            AJ_Cook             0   \n",
       "1            AJ_Cook             2   \n",
       "2      Aaron_Eckhart             0   \n",
       "3      Aaron_Eckhart             1   \n",
       "4      Aaron_Eckhart             2   \n",
       "...              ...           ...   \n",
       "3419  Zoran_Djindjic             1   \n",
       "3420  Zoran_Djindjic             2   \n",
       "3421  Zulfiqar_Ahmed             1   \n",
       "3422  Zulfiqar_Ahmed             2   \n",
       "3423  Zulfiqar_Ahmed             4   \n",
       "\n",
       "                                                   path  \n",
       "0     [./frame_images_DB/AJ_Cook\\0\\0.123.jpg, ./fram...  \n",
       "1     [./frame_images_DB/AJ_Cook\\2\\2.131.jpg, ./fram...  \n",
       "2     [./frame_images_DB/Aaron_Eckhart\\0\\0.555.jpg, ...  \n",
       "3     [./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg, ...  \n",
       "4     [./frame_images_DB/Aaron_Eckhart\\2\\2.1718.jpg,...  \n",
       "...                                                 ...  \n",
       "3419  [./frame_images_DB/Zoran_Djindjic\\1\\1.104.jpg,...  \n",
       "3420  [./frame_images_DB/Zoran_Djindjic\\2\\2.1973.jpg...  \n",
       "3421  [./frame_images_DB/Zulfiqar_Ahmed\\1\\1.1977.jpg...  \n",
       "3422  [./frame_images_DB/Zulfiqar_Ahmed\\2\\2.1051.jpg...  \n",
       "3423  [./frame_images_DB/Zulfiqar_Ahmed\\4\\4.155.jpg,...  \n",
       "\n",
       "[2834 rows x 3 columns]"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# remove less than 2 identity\n",
    "test_df = ds_df[ds_df['identity'].isin(list(anc_df['identity']))]\n",
    "test_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>identity</th>\n",
       "      <th>sub_identity</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>7</th>\n",
       "      <td>Aaron_Sorkin</td>\n",
       "      <td>3</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>9</th>\n",
       "      <td>Aaron_Tippin</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3417</th>\n",
       "      <td>Zhu_Rongji</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3419</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3420</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3422</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>2</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3423</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>4</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>1830 rows × 2 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "            identity sub_identity\n",
       "1            AJ_Cook            2\n",
       "3      Aaron_Eckhart            1\n",
       "4      Aaron_Eckhart            2\n",
       "7       Aaron_Sorkin            3\n",
       "9       Aaron_Tippin            1\n",
       "...              ...          ...\n",
       "3417      Zhu_Rongji            1\n",
       "3419  Zoran_Djindjic            1\n",
       "3420  Zoran_Djindjic            2\n",
       "3422  Zulfiqar_Ahmed            2\n",
       "3423  Zulfiqar_Ahmed            4\n",
       "\n",
       "[1830 rows x 2 columns]"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# remove the anchor identity and sub_identity\n",
    "test_df = pd.concat([test_df[['identity','sub_identity']],anc_df[['identity','sub_identity']]]).drop_duplicates(keep=False)\n",
    "test_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Unnamed: 0</th>\n",
       "      <th>sub_identity</th>\n",
       "      <th>identity</th>\n",
       "      <th>path</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>85</td>\n",
       "      <td>1</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>86</td>\n",
       "      <td>1</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\1\\1.184.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>87</td>\n",
       "      <td>1</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\1\\1.185.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>88</td>\n",
       "      <td>1</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\1\\1.186.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>89</td>\n",
       "      <td>1</td>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>./frame_images_DB/Aaron_Eckhart\\1\\1.187.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>206484</th>\n",
       "      <td>390203</td>\n",
       "      <td>4</td>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>./frame_images_DB/Zulfiqar_Ahmed\\4\\4.263.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>206485</th>\n",
       "      <td>390204</td>\n",
       "      <td>4</td>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>./frame_images_DB/Zulfiqar_Ahmed\\4\\4.264.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>206486</th>\n",
       "      <td>390205</td>\n",
       "      <td>4</td>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>./frame_images_DB/Zulfiqar_Ahmed\\4\\4.265.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>206487</th>\n",
       "      <td>390206</td>\n",
       "      <td>4</td>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>./frame_images_DB/Zulfiqar_Ahmed\\4\\4.266.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>206488</th>\n",
       "      <td>390207</td>\n",
       "      <td>4</td>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>./frame_images_DB/Zulfiqar_Ahmed\\4\\4.267.jpg</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>206489 rows × 4 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "        Unnamed: 0 sub_identity        identity  \\\n",
       "0               85            1   Aaron_Eckhart   \n",
       "1               86            1   Aaron_Eckhart   \n",
       "2               87            1   Aaron_Eckhart   \n",
       "3               88            1   Aaron_Eckhart   \n",
       "4               89            1   Aaron_Eckhart   \n",
       "...            ...          ...             ...   \n",
       "206484      390203            4  Zulfiqar_Ahmed   \n",
       "206485      390204            4  Zulfiqar_Ahmed   \n",
       "206486      390205            4  Zulfiqar_Ahmed   \n",
       "206487      390206            4  Zulfiqar_Ahmed   \n",
       "206488      390207            4  Zulfiqar_Ahmed   \n",
       "\n",
       "                                                path  \n",
       "0        ./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg  \n",
       "1        ./frame_images_DB/Aaron_Eckhart\\1\\1.184.jpg  \n",
       "2        ./frame_images_DB/Aaron_Eckhart\\1\\1.185.jpg  \n",
       "3        ./frame_images_DB/Aaron_Eckhart\\1\\1.186.jpg  \n",
       "4        ./frame_images_DB/Aaron_Eckhart\\1\\1.187.jpg  \n",
       "...                                              ...  \n",
       "206484  ./frame_images_DB/Zulfiqar_Ahmed\\4\\4.263.jpg  \n",
       "206485  ./frame_images_DB/Zulfiqar_Ahmed\\4\\4.264.jpg  \n",
       "206486  ./frame_images_DB/Zulfiqar_Ahmed\\4\\4.265.jpg  \n",
       "206487  ./frame_images_DB/Zulfiqar_Ahmed\\4\\4.266.jpg  \n",
       "206488  ./frame_images_DB/Zulfiqar_Ahmed\\4\\4.267.jpg  \n",
       "\n",
       "[206489 rows x 4 columns]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# get ungroup identity\n",
    "temp_ds_df = raw_ds_df.copy()\n",
    "test_df = temp_ds_df.merge(test_df, on = ['identity','sub_identity'])\n",
    "display(test_df)\n",
    "# test_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>identity</th>\n",
       "      <th>path</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>AJ_Cook</td>\n",
       "      <td>[./frame_images_DB/AJ_Cook\\2\\2.131.jpg, ./fram...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Aaron_Eckhart</td>\n",
       "      <td>[./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Aaron_Sorkin</td>\n",
       "      <td>[./frame_images_DB/Aaron_Sorkin\\3\\3.425.jpg, ....</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Aaron_Tippin</td>\n",
       "      <td>[./frame_images_DB/Aaron_Tippin\\1\\1.1595.jpg, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Abba_Eban</td>\n",
       "      <td>[./frame_images_DB/Abba_Eban\\5\\5.10.jpg, ./fra...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>999</th>\n",
       "      <td>Zhang_Ziyi</td>\n",
       "      <td>[./frame_images_DB/Zhang_Ziyi\\2\\2.116.jpg, ./f...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1000</th>\n",
       "      <td>Zhong_Nanshan</td>\n",
       "      <td>[./frame_images_DB/Zhong_Nanshan\\1\\1.1675.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1001</th>\n",
       "      <td>Zhu_Rongji</td>\n",
       "      <td>[./frame_images_DB/Zhu_Rongji\\1\\1.431.jpg, ./f...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1002</th>\n",
       "      <td>Zoran_Djindjic</td>\n",
       "      <td>[./frame_images_DB/Zoran_Djindjic\\1\\1.104.jpg,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1003</th>\n",
       "      <td>Zulfiqar_Ahmed</td>\n",
       "      <td>[./frame_images_DB/Zulfiqar_Ahmed\\2\\2.1051.jpg...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>1004 rows × 2 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "            identity                                               path\n",
       "0            AJ_Cook  [./frame_images_DB/AJ_Cook\\2\\2.131.jpg, ./fram...\n",
       "1      Aaron_Eckhart  [./frame_images_DB/Aaron_Eckhart\\1\\1.183.jpg, ...\n",
       "2       Aaron_Sorkin  [./frame_images_DB/Aaron_Sorkin\\3\\3.425.jpg, ....\n",
       "3       Aaron_Tippin  [./frame_images_DB/Aaron_Tippin\\1\\1.1595.jpg, ...\n",
       "4          Abba_Eban  [./frame_images_DB/Abba_Eban\\5\\5.10.jpg, ./fra...\n",
       "...              ...                                                ...\n",
       "999       Zhang_Ziyi  [./frame_images_DB/Zhang_Ziyi\\2\\2.116.jpg, ./f...\n",
       "1000   Zhong_Nanshan  [./frame_images_DB/Zhong_Nanshan\\1\\1.1675.jpg,...\n",
       "1001      Zhu_Rongji  [./frame_images_DB/Zhu_Rongji\\1\\1.431.jpg, ./f...\n",
       "1002  Zoran_Djindjic  [./frame_images_DB/Zoran_Djindjic\\1\\1.104.jpg,...\n",
       "1003  Zulfiqar_Ahmed  [./frame_images_DB/Zulfiqar_Ahmed\\2\\2.1051.jpg...\n",
       "\n",
       "[1004 rows x 2 columns]"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_df = test_df.groupby(['identity'])['path'].apply(list).reset_index()\n",
    "test_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True    1004\n",
       "dtype: int64"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "check = list()\n",
    "for inx, each in enumerate(list(test_df['identity'])):\n",
    "    if list(anc_df['identity'])[inx] == each: \n",
    "        check.append(True)\n",
    "    else: \n",
    "        check.append(False)\n",
    "pd.Series(check).value_counts()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define custom dataset\n",
    "class FaceDataset(Dataset):\n",
    "    def __init__(self, df,df2, train, num_sample = None, transform = None, num_img_pool = 10, ):\n",
    "        # set random seed for FaceDataset\n",
    "        np.random.seed(seed)\n",
    "        random.seed(seed)\n",
    "        # create constructors\n",
    "        self.unique_img_name = None\n",
    "        self.train = train\n",
    "        self.data = dict()\n",
    "        self.images = list()\n",
    "        self.images2 = list()\n",
    "        self.identities = list()\n",
    "        self.identities2 = list()\n",
    "        # label to indices\n",
    "        # self.label_to_indices = dict()\n",
    "        self.labels = list()\n",
    "        self.labels2 = list()\n",
    "        self.img_name = list()\n",
    "        self.img_name2 = list()\n",
    "        # read csv file\n",
    "        self.df = df\n",
    "        self.df2 = df2\n",
    "        # set the transformation\n",
    "        self.transform = transform\n",
    "        # drop last n row from dataframe\n",
    "        self.df = self.df.head(num_sample)\n",
    "        #get the length of entire dataset\n",
    "        self.len_ = len(self.df)\n",
    "        if num_sample is None or num_sample > self.len_:\n",
    "            num_sample = self.len_\n",
    "        # load imgs\n",
    "        self.load_imgs(self.df, self.df2, num_imgs = num_img_pool, max = num_sample)\n",
    "\n",
    "    def __len__(self):\n",
    "        return self.len_\n",
    "\n",
    "    def __getitem__(self, idx):\n",
    "        if self.train:\n",
    "            anchor_img = self.images[idx]\n",
    "            anchor_label = self.labels[idx]\n",
    "            pos_idx = np.random.choice(np.arange(len(self.images))[self.labels == anchor_label])\n",
    "            neg_idx = np.random.choice(np.arange(len(self.images))[self.labels != anchor_label])\n",
    "\n",
    "            pos_img = self.images[pos_idx]\n",
    "            neg_img = self.images[neg_idx]\n",
    "\n",
    "            pos_label = self.labels[pos_idx]\n",
    "            neg_label = self.labels[neg_idx]\n",
    "\n",
    "            if self.transform is None:\n",
    "                img_to_tensor = transforms.ToTensor()\n",
    "                anchor_img = img_to_tensor(anchor_img)\n",
    "                pos_img = img_to_tensor(pos_img)\n",
    "                neg_img = img_to_tensor(neg_img)\n",
    "            else:\n",
    "                anchor_img = self.transform(anchor_img)\n",
    "                pos_img = self.transform(pos_img)\n",
    "                neg_img = self.transform(neg_img)\n",
    "            return anchor_img, pos_img, neg_img\n",
    "\n",
    "        else:\n",
    "            label = False\n",
    "            anchor_img = self.images[idx]\n",
    "            anchor_label = self.labels[idx]\n",
    "            if idx % 2 == 0:\n",
    "                test_idx = np.random.choice(np.arange(len(self.images))[self.labels == anchor_label])\n",
    "                label = True\n",
    "\n",
    "            else:\n",
    "                test_idx = np.random.choice(np.arange(len(self.images))[self.labels != anchor_label])\n",
    "\n",
    "            neg_idx = np.random.choice(np.arange(len(self.images))[self.labels != anchor_label])\n",
    "            test_img = self.images2[test_idx]\n",
    "            neg_img = self.images2[neg_idx]\n",
    "\n",
    "            if self.transform is None:\n",
    "                img_to_tensor = transforms.ToTensor()\n",
    "                anchor_img = img_to_tensor(anchor_img)\n",
    "                test_img = img_to_tensor(test_img)\n",
    "                neg_img = img_to_tensor(neg_img)\n",
    "\n",
    "            else:\n",
    "                anchor_img = self.transform(anchor_img)\n",
    "                test_img = self.transform(test_img)\n",
    "                neg_img = self.transform(neg_img)\n",
    "\n",
    "            return anchor_img, test_img, neg_img ,label, self.identities[idx], self.img_name[idx], self.identities[test_idx], self.img_name[test_idx]\n",
    "    \n",
    "    def face_regs(self, path: str):\n",
    "        try: \n",
    "            image = face_recognition.load_image_file(path)\n",
    "            face_locations = face_recognition.face_locations(image)\n",
    "\n",
    "            # print(\"I found {} face(s) in this photograph.\".format(len(face_locations)))\n",
    "\n",
    "            for face_location in face_locations:\n",
    "\n",
    "                # Print the location of each face in this image\n",
    "                top, right, bottom, left = face_location\n",
    "            #     print(\"A face is located at pixel location Top: {}, Left: {}, Bottom: {}, Right: {}\".format(top, left, bottom, right))\n",
    "\n",
    "                # You can access the actual face itself like this:\n",
    "                face_image = image[top:bottom, left:right]\n",
    "                pil_image = Image.fromarray(face_image).convert('RGB')\n",
    "    #             plt.imshow(pil_image)\n",
    "#                 print(pil_image)\n",
    "            return (True, pil_image)\n",
    "        except: \n",
    "            return (False, None) \n",
    "\n",
    "    # load imgs from pandas to memory and define the maximum number of images\n",
    "    def load_imgs(self, df, df2, num_imgs, max):\n",
    "        # iterate thought each row\n",
    "        count_row = 0 \n",
    "        for i, row in tqdm(df.iterrows(), total = max):\n",
    "            # get identity of each row\n",
    "            row_identity = row['identity']\n",
    "            row_identity2 = df2.iloc[count_row]\n",
    "            \n",
    "            # append each identity to numberical value\n",
    "            # self.label_to_indices[int(row_identity)] = i\n",
    "            count_img = 0\n",
    "            # loop imgs in each identity\n",
    "            for index, img_name in enumerate(row['path']):\n",
    "                if count_img > num_imgs:\n",
    "                    break\n",
    "                # concatenate the directoru and image name\n",
    "                path_to_image = img_name\n",
    "                path_to_image2 = np.random.choice(row_identity2['path'])\n",
    "                (isImg, img) = self.face_regs(path_to_image)\n",
    "                (isImg2, img2) = self.face_regs(path_to_image2)\n",
    "                if not (isImg & isImg2):\n",
    "                    continue\n",
    "                # open image and convert to RGB\n",
    "#                 img = Image.open(path_to_image).convert('RGB')\n",
    "                self.labels.append(i)\n",
    "                self.images.append(img)\n",
    "                self.img_name.append(path_to_image)\n",
    "                self.identities.append(str(row_identity))\n",
    "                \n",
    "                self.labels2.append(i)\n",
    "                self.images2.append(img2)\n",
    "                self.img_name2.append(path_to_image2)\n",
    "                self.identities2.append(str(row_identity2))\n",
    "                \n",
    "                count_img += 1  # print('Added img '+ str(row_identity))\n",
    "            count_row +=1 \n",
    "                \n",
    "        self.labels = np.array(self.labels)\n",
    "    def return_val(self):\n",
    "        return self.labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "False None\n"
     ]
    }
   ],
   "source": [
    "face_reg(\"./frame_images_DB\\\\Aaron_Tippin\\\\1\\\\1.1211.jpg\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'./weights/mobile_network_120_updated20230329 020936.pth'"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_transform = transforms.Compose([\n",
    "    transforms.Resize((256, 256)),\n",
    "    transforms.CenterCrop(img_size),\n",
    "    transforms.ToTensor(),\n",
    "    transforms.Normalize(mean = [0.5, 0.5, 0.5],\n",
    "                        std = [0.5, 0.5, 0.5])\n",
    "    ])\n",
    "\n",
    "# define batch size for testing\n",
    "test_batch_size = 64\n",
    "# create dataloader for validation set\n",
    "val_triplet_dataset = FaceDataset(df = anc_df,df2 = test_df, train = False, num_sample = 100000, transform = test_transform)\n",
    "val_triplet_dataloader = DataLoader(val_triplet_dataset, batch_size = test_batch_size, shuffle = True,\n",
    "                                     pin_memory = True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "class Flatten(Module):\n",
    "    def forward(self, input):\n",
    "        return input.view(input.size(0), -1)\n",
    "\n",
    "\n",
    "def l2_norm(input, axis = 1):\n",
    "    norm = torch.norm(input, 2, axis, True)\n",
    "    output = torch.div(input, norm)\n",
    "    return output\n",
    "\n",
    "\n",
    "class h_sigmoid(Module):\n",
    "    def __init__(self, inplace = True):\n",
    "        super(h_sigmoid, self).__init__()\n",
    "        self.relu = ReLU6(inplace = inplace)\n",
    "\n",
    "    def forward(self, x):\n",
    "        return self.relu(x + 3) / 6\n",
    "\n",
    "\n",
    "class h_swish(Module):\n",
    "    def __init__(self, inplace = True):\n",
    "        super(h_swish, self).__init__()\n",
    "        self.sigmoid = h_sigmoid(inplace = inplace)\n",
    "\n",
    "    def forward(self, x):\n",
    "        return x * self.sigmoid(x)\n",
    "\n",
    "\n",
    "class SELayer(Module):\n",
    "    def __init__(self, channel, reduction = 4):\n",
    "        super(SELayer, self).__init__()\n",
    "        self.avg_pool = AdaptiveAvgPool2d(1)\n",
    "        self.fc = Sequential(Linear(channel, channel // reduction), ReLU(inplace = True),\n",
    "                             Linear(channel // reduction, channel), h_sigmoid())\n",
    "\n",
    "    def forward(self, x):\n",
    "        b, c, _, _ = x.size()\n",
    "        y = self.avg_pool(x).view(b, c)\n",
    "        y = self.fc(y).view(b, c, 1, 1)\n",
    "        return x * y\n",
    "\n",
    "\n",
    "class PermutationBlock(Module):\n",
    "    def __init__(self, groups):\n",
    "        super(PermutationBlock, self).__init__()\n",
    "        self.groups = groups\n",
    "\n",
    "    def forward(self, input):\n",
    "        n, c, h, w = input.size()\n",
    "        G = self.groups\n",
    "        output = input.view(n, G, c // G, h, w).permute(0, 2, 1, 3, 4).contiguous().view(n, c, h, w)\n",
    "        return output\n",
    "\n",
    "\n",
    "class Conv_block(Module):\n",
    "    def __init__(self, in_c, out_c, kernel = (1, 1), stride = (1, 1), padding = (0, 0), groups = 1):\n",
    "        super(Conv_block, self).__init__()\n",
    "        self.conv = Conv2d(in_c, out_channels = out_c, kernel_size = kernel, groups = groups, stride = stride,\n",
    "                           padding = padding, bias = False)\n",
    "        self.bn = BatchNorm2d(out_c)\n",
    "        self.prelu = PReLU(out_c)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.conv(x)\n",
    "        x = self.bn(x)\n",
    "        x = self.prelu(x)\n",
    "        return x\n",
    "\n",
    "\n",
    "class Linear_block(Module):\n",
    "    def __init__(self, in_c, out_c, kernel = (1, 1), stride = (1, 1), padding = (0, 0), groups = 1):\n",
    "        super(Linear_block, self).__init__()\n",
    "        self.conv = Conv2d(in_c, out_channels = out_c, kernel_size = kernel, groups = groups, stride = stride,\n",
    "                           padding = padding, bias = False)\n",
    "        self.bn = BatchNorm2d(out_c)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.conv(x)\n",
    "        x = self.bn(x)\n",
    "        return x\n",
    "\n",
    "\n",
    "class Depth_Wise(Module):\n",
    "    def __init__(self, in_c, out_c, residual = False, kernel = (3, 3), stride = (2, 2), padding = (1, 1), groups = 1):\n",
    "        super(Depth_Wise, self).__init__()\n",
    "        self.conv = Conv_block(in_c, out_c = groups, kernel = (1, 1), padding = (0, 0), stride = (1, 1))\n",
    "        self.conv_dw = Conv_block(groups, groups, groups = groups, kernel = kernel, padding = padding, stride = stride)\n",
    "        self.project = Linear_block(groups, out_c, kernel = (1, 1), padding = (0, 0), stride = (1, 1))\n",
    "        self.residual = residual\n",
    "\n",
    "    def forward(self, x):\n",
    "        if self.residual:\n",
    "            short_cut = x\n",
    "        x = self.conv(x)\n",
    "        x = self.conv_dw(x)\n",
    "        x = self.project(x)\n",
    "        if self.residual:\n",
    "            output = short_cut + x\n",
    "        else:\n",
    "            output = x\n",
    "        return output\n",
    "\n",
    "\n",
    "class Residual(Module):\n",
    "    def __init__(self, c, num_block, groups, kernel = (3, 3), stride = (1, 1), padding = (1, 1)):\n",
    "        super(Residual, self).__init__()\n",
    "        modules = []\n",
    "        for _ in range(num_block):\n",
    "            modules.append(\n",
    "                Depth_Wise(c, c, residual = True, kernel = kernel, padding = padding, stride = stride, groups = groups))\n",
    "        self.model = Sequential(*modules)\n",
    "\n",
    "    def forward(self, x):\n",
    "        return self.model(x)\n",
    "\n",
    "\n",
    "class LinearScheduler(nn.Module):\n",
    "    def __init__(self, dropblock, start_value, stop_value, nr_steps):\n",
    "        super(LinearScheduler, self).__init__()\n",
    "        self.dropblock = dropblock\n",
    "        self.i = 0\n",
    "        self.drop_values = np.linspace(start = start_value, stop = stop_value, num = nr_steps)\n",
    "\n",
    "    def forward(self, x):\n",
    "        return self.dropblock(x)\n",
    "\n",
    "    def step(self):\n",
    "        if self.i < len(self.drop_values):\n",
    "            self.dropblock.drop_prob = self.drop_values[self.i]\n",
    "\n",
    "        self.i += 1\n",
    "\n",
    "\n",
    "class DropBlock2D(nn.Module):\n",
    "\n",
    "    def __init__(self, drop_prob, block_size):\n",
    "        super(DropBlock2D, self).__init__()\n",
    "\n",
    "        self.drop_prob = drop_prob\n",
    "        self.block_size = block_size\n",
    "\n",
    "    def forward(self, x):\n",
    "        # shape: (bsize, channels, height, width)\n",
    "\n",
    "        assert x.dim() == 4, \"Expected input with 4 dimensions (bsize, channels, height, width)\"\n",
    "\n",
    "        if not self.training or self.drop_prob == 0.:\n",
    "            return x\n",
    "        else:\n",
    "            # get gamma value\n",
    "            gamma = self._compute_gamma(x)\n",
    "\n",
    "            # sample mask\n",
    "            mask = (torch.rand(x.shape[0], *x.shape[2:]) < gamma).float()\n",
    "\n",
    "            # place mask on input device\n",
    "            mask = mask.to(x.device)\n",
    "\n",
    "            # compute block mask\n",
    "            block_mask = self._compute_block_mask(mask)\n",
    "\n",
    "            # apply block mask\n",
    "            out = x * block_mask[:, None, :, :]\n",
    "\n",
    "            # scale output\n",
    "            out = out * block_mask.numel() / block_mask.sum()\n",
    "\n",
    "            return out\n",
    "\n",
    "    def _compute_block_mask(self, mask):\n",
    "        block_mask = F.max_pool2d(input = mask[:, None, :, :], kernel_size = (\n",
    "            self.block_size, self.block_size), stride = (1, 1), padding = self.block_size // 2)\n",
    "\n",
    "        if self.block_size % 2 == 0:\n",
    "            block_mask = block_mask[:, :, :-1, :-1]\n",
    "\n",
    "        block_mask = 1 - block_mask.squeeze(1)\n",
    "\n",
    "        return block_mask\n",
    "\n",
    "    def _compute_gamma(self, x):\n",
    "        return self.drop_prob / (self.block_size ** 2)\n",
    "\n",
    "class MobileFaceNet(Module):\n",
    "    def __init__(self, embedding_size = 512):\n",
    "        super(MobileFaceNet, self).__init__()\n",
    "        self.conv1 = Conv_block(3, 64, kernel = (3, 3), stride = (2, 2), padding = (1, 1))\n",
    "        self.conv2_dw = Conv_block(64, 64, kernel = (3, 3), stride = (1, 1), padding = (1, 1), groups = 64)\n",
    "        self.conv_23 = Depth_Wise(64, 64, kernel = (3, 3), stride = (2, 2), padding = (1, 1), groups = 128)\n",
    "        self.conv_3 = Residual(64, num_block = 4, groups = 128, kernel = (3, 3), stride = (1, 1), padding = (1, 1))\n",
    "        self.conv_34 = Depth_Wise(64, 128, kernel = (3, 3), stride = (2, 2), padding = (1, 1), groups = 256)\n",
    "        self.conv_4 = Residual(128, num_block = 6, groups = 256, kernel = (3, 3), stride = (1, 1), padding = (1, 1))\n",
    "        self.conv_45 = Depth_Wise(128, 128, kernel = (3, 3), stride = (2, 2), padding = (1, 1), groups = 512)\n",
    "        self.conv_5 = Residual(128, num_block = 2, groups = 256, kernel = (3, 3), stride = (1, 1), padding = (1, 1))\n",
    "        self.conv_6_sep = Conv_block(128, 512, kernel = (1, 1), stride = (1, 1), padding = (0, 0))\n",
    "        self.conv_6_dw = Linear_block(512, 512, groups = 512, kernel = (7, 7), stride = (1, 1), padding = (0, 0))\n",
    "        self.conv_6_flatten = Flatten()\n",
    "        self.linear = Linear(512, embedding_size, bias = False)\n",
    "        self.bn = BatchNorm1d(embedding_size)\n",
    "        self.dropblock = DropBlock2D(block_size = 3, drop_prob = 0.25)\n",
    "        self.dropout = Dropout(0.25)\n",
    "\n",
    "    def forward_once(self, x):\n",
    "        out = self.conv1(x)\n",
    "\n",
    "        out = self.conv2_dw(out)\n",
    "\n",
    "        out = self.conv_23(out)\n",
    "\n",
    "        out = self.conv_3(out)\n",
    "\n",
    "        out = self.conv_34(out)\n",
    "        \n",
    "        out = self.dropblock(out)\n",
    "\n",
    "        out = self.conv_4(out)\n",
    "        \n",
    "        out = self.dropout(out)\n",
    "\n",
    "        out = self.conv_45(out)\n",
    "        \n",
    "        out = self.dropblock(out)\n",
    "        \n",
    "        out = self.conv_5(out)\n",
    "\n",
    "        out = self.conv_6_sep(out)\n",
    "\n",
    "        out = self.dropblock(out)\n",
    "\n",
    "        out = self.conv_6_dw(out)\n",
    "\n",
    "        out = self.dropblock(out)\n",
    "\n",
    "        out = self.conv_6_flatten(out)\n",
    "\n",
    "        out = self.linear(out)\n",
    "\n",
    "        out = self.bn(out)\n",
    "\n",
    "        return l2_norm(out)\n",
    "\n",
    "    def forward(self, img):\n",
    "        anchor = self.forward_once(img)\n",
    "        return anchor\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'MobileFaceNet' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[1;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[1;32m~\\AppData\\Local\\Temp/ipykernel_13860/3249437949.py\u001b[0m in \u001b[0;36m<module>\u001b[1;34m\u001b[0m\n\u001b[1;32m----> 1\u001b[1;33m \u001b[0mtriplet_model\u001b[0m \u001b[1;33m=\u001b[0m \u001b[0mMobileFaceNet\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m\u001b[0;32m      2\u001b[0m \u001b[0mprint\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0msummary\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mtriplet_model\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0minput_size\u001b[0m \u001b[1;33m=\u001b[0m \u001b[1;33m[\u001b[0m\u001b[1;33m(\u001b[0m\u001b[1;36m32\u001b[0m\u001b[1;33m,\u001b[0m \u001b[1;36m3\u001b[0m\u001b[1;33m,\u001b[0m \u001b[1;36m112\u001b[0m\u001b[1;33m,\u001b[0m \u001b[1;36m112\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m]\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
      "\u001b[1;31mNameError\u001b[0m: name 'MobileFaceNet' is not defined"
     ]
    }
   ],
   "source": [
    "triplet_model = MobileFaceNet()\n",
    "print(summary(triplet_model, input_size = [(32, 3, 112, 112)]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "class TripletLoss(nn.Module):\n",
    "    def __init__(self, margin):\n",
    "        super(TripletLoss, self).__init__()\n",
    "        self.margin = margin\n",
    "\n",
    "    def forward(self, anchor, positive, negative, size_average = True):\n",
    "        # calculate Euclidean's distance\n",
    "        distance_positive = (anchor - positive).pow(2).sum(1)\n",
    "        distance_negative = (anchor - negative).pow(2).sum(1)\n",
    "\n",
    "        losses = F.relu(distance_positive - distance_negative + self.margin)\n",
    "\n",
    "        return torch.mean(losses)\n",
    "    \n",
    "# def double_triplet_loss(nn.Module): \n",
    "    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "def calculate_accuracy(threshold, dist, actual_issame):\n",
    "    # If distance is less than threshold, then prediction is set to True\n",
    "    predict_issame = np.less(dist, threshold)\n",
    "\n",
    "    true_positives = np.sum(np.logical_and(predict_issame, actual_issame))\n",
    "    false_positives = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))\n",
    "    true_negatives = np.sum(np.logical_and(np.logical_not(predict_issame), np.logical_not(actual_issame)))\n",
    "    false_negatives = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))\n",
    "\n",
    "    # For dealing with Divide By Zero exception\n",
    "    true_positive_rate = 0 if (true_positives + false_negatives == 0) else float(true_positives) / float(\n",
    "        true_positives + false_negatives)\n",
    "\n",
    "    false_positive_rate = 0 if (false_positives + true_negatives == 0) else float(false_positives) / float(\n",
    "        false_positives + true_negatives)\n",
    "\n",
    "    precision = 0 if (true_positives + false_positives) == 0 else float(true_positives) / float(\n",
    "        true_positives + false_positives)\n",
    "\n",
    "    recall = 0 if (true_positives + false_negatives) == 0 else float(true_positives) / float(\n",
    "        true_positives + false_negatives)\n",
    "\n",
    "    accuracy = float(true_positives + true_negatives) / dist.size\n",
    "\n",
    "    #     return true_positive_rate, false_positive_rate, precision, recall, accuracy\n",
    "    return [true_positive_rate, false_positive_rate, accuracy]\n",
    "\n",
    "# save the model weight\n",
    "def save_weight(custom_name: str, model, csv_file: str):\n",
    "    # check csv file exist \n",
    "    if not os.path.exists(csv_file):\n",
    "        # create new csv_file\n",
    "        pd.DataFrame(columns = ['date_added', 'path']).to_csv(csv_file)\n",
    "    # read csv file    \n",
    "    weight_df = pd.read_csv(csv_file)\n",
    "\n",
    "    datetime_now = str(datetime.datetime.now())\n",
    "    save_path = './weights/' + custom_name + datetime_now[:-7].replace('-', '').replace(':', '') + '.pth'\n",
    "\n",
    "    # create dataframe\n",
    "    temp_df = pd.DataFrame([[datetime_now, save_path]], columns = ['date_added', 'path'])\n",
    "    weight_df = pd.concat([weight_df, temp_df], axis = 0, ignore_index = True).drop(['Unnamed: 0'], axis = 1)\n",
    "    weight_df = weight_df[['date_added', 'path']]\n",
    "    # save csv \n",
    "    weight_df.to_csv(csv_file)\n",
    "    # save weight \n",
    "    save_weight_path = save_path\n",
    "    torch.save(model.state_dict(), save_weight_path)\n",
    "    print('Save Success! @' + datetime_now + ' in ' + save_path)\n",
    "    return save_path\n",
    "\n",
    "\n",
    "# get the latest weight\n",
    "def get_latest_weight(csv_path: str) -> str:\n",
    "    # read csv file\n",
    "    weight_df = pd.read_csv(csv_path)\n",
    "    # convert object to datetime type\n",
    "    weight_df['date_added'] = pd.to_datetime(weight_df['date_added'])\n",
    "    # sorting date\n",
    "    weight_df = weight_df.sort_values(by = 'date_added', ascending = False)\n",
    "    # get latest item\n",
    "    latest_record = weight_df.head(1)\n",
    "    print('It was saved at ' + str(latest_record['date_added'].values[0]))\n",
    "    return weight_df['path'].head(1).values[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train_model(model, train_loader, train_optimizer, train_criterion, weight_path = None):\n",
    "    train_loss = 0\n",
    "    train_accuracy = list()\n",
    "    label_s = list()\n",
    "    # applied weight\n",
    "    if weight_path != None:\n",
    "        model.load_state_dict(weight_path)\n",
    "    # set to train mode\n",
    "    model.train()\n",
    "    # loop though loader\n",
    "    with torch.set_grad_enabled(True):\n",
    "        for anchor_img, pos_img, neg_img in tqdm(train_loader):\n",
    "            # for calculate accuracy\n",
    "            threshold_list = np.full(anchor_img.shape[0], model_config['margin'])\n",
    "            is_same = np.full(anchor_img.shape[0], 1)\n",
    "\n",
    "            # throw img to CUDA\n",
    "            anchor_img, pos_img, neg_img = anchor_img.to(device), pos_img.to(device), neg_img.to(device)\n",
    "            # prevent grad vansishing\n",
    "            train_optimizer.zero_grad()\n",
    "            # training model\n",
    "            # anchor_img, pos_img, neg_img = model(anchor_img, pos_img,neg_img)\n",
    "            anchor_out = model(anchor_img)\n",
    "            pos_out = model(pos_img)\n",
    "            neg_out = model(neg_img)\n",
    "            # calculate distances \n",
    "            pos_dists = (anchor_out - pos_out).pow(2).sum(1)\n",
    "            neg_dists = (anchor_out - neg_out).pow(2).sum(1)\n",
    "            \n",
    "            # triplet selections\n",
    "            if False: \n",
    "                first_condition = (neg_dists - pos_dists < model_config['threshold']).cpu().numpy().flatten()\n",
    "                second_condition = (pos_dists < neg_dists).cpu().numpy().flatten()\n",
    "                all = (np.logical_and(first_condition, second_condition))\n",
    "                valid_triplets = np.where(all == 1)\n",
    "            else: \n",
    "                all = (neg_dists - pos_dists < model_config['threshold']).cpu().numpy().flatten()\n",
    "                valid_triplets = np.where(all == 1)\n",
    "                \n",
    "            # calculate loss\n",
    "            loss = train_criterion(anchor_out, pos_out, neg_out)\n",
    "            # backpropragation\n",
    "            loss.backward()\n",
    "            # compute gradient\n",
    "            train_optimizer.step()\n",
    "            # compute train loss\n",
    "            train_loss += loss.item() * anchor_img.size(0)\n",
    "    # test_acc, distances, labels =test_model(model, train_loader)\n",
    "\n",
    "    return train_loss\n",
    "\n",
    "\n",
    "def test_model(model, test_loader,train_criterion, weight_path: str = None):\n",
    "    \n",
    "    test_accuracy = list()\n",
    "    label_s = list()\n",
    "    distance_s = list()\n",
    "    distances_hist = list()\n",
    "    \n",
    "    test_imgs_name = list()\n",
    "    anch_imgs_name = list()\n",
    "    \n",
    "    test_identities = list()\n",
    "    anch_identities = list()\n",
    "    \n",
    "    test_loss = 0\n",
    "    if weight_path is not None:\n",
    "        model.load_state_dict(torch.load(weight_path))\n",
    "    # set to train mode\n",
    "    model.eval()\n",
    "    # enable grad\n",
    "    with torch.no_grad():\n",
    "        # loop though loader\n",
    "        #anchor_img, test_img, neg_img ,label, self.identities[idx], self.img_name[idx], self.identities[test_idx], self.img_name[test_idx], \n",
    "\n",
    "        for index, (anchor_img, test_img, neg_img, label, anch_iden, anch_name, test_iden, test_name) in enumerate(test_loader):\n",
    "            \n",
    "            # for calculate accuracy\n",
    "            threshold_list = np.full(anchor_img.shape[0], model_config['margin'])\n",
    "            is_same = np.full(anchor_img.shape[0], 1)\n",
    "            # throw img to CUDA\n",
    "            anchor_img, test_img, neg_img = anchor_img.to(device), test_img.to(device), neg_img.to(device)\n",
    "            # compute distance from model\n",
    "            anchor_out = model(anchor_img)\n",
    "            test_out = model(test_img)\n",
    "            neg_out = model(neg_img)\n",
    "            # calculate loss\n",
    "            loss = train_criterion(anchor_out,test_out,neg_out)\n",
    "            # calculate distance between anchor and test images\n",
    "            distance_val = (anchor_out - test_out).pow(2).sum(1)\n",
    "            # append label coordinate with each testing pair\n",
    "            label_s.append(label.cpu().data.numpy())\n",
    "            distance_s.append(distance_val.cpu().data.numpy())\n",
    "            test_loss += loss.item() * anchor_img.size(0)\n",
    "            \n",
    "            test_imgs_name.append(test_name)\n",
    "            anch_imgs_name.append(anch_name)\n",
    "            test_identities.append(test_iden)\n",
    "            anch_identities.append(anch_iden)\n",
    "            \n",
    "        labels_hist = [sub_label for each_label in label_s for sub_label in each_label]\n",
    "        distance_hist = [sub_distance for each_distance in distance_s for sub_distance in each_distance]\n",
    "        \n",
    "        test_imgs_name = [sub for each in test_imgs_name for sub in each]\n",
    "        anch_imgs_name = [sub for each in anch_imgs_name for sub in each]\n",
    "        test_identities = [sub for each in test_identities for sub in each]\n",
    "        anch_identities = [sub for each in anch_identities for sub in each]\n",
    "        \n",
    "        test_accuracy.append(calculate_accuracy(np.full(np.size(distance_hist),\n",
    "                                                        model_config['margin']), np.array(distance_hist), labels_hist)[\n",
    "                                 2])\n",
    "    return mean(test_accuracy), distance_hist, labels_hist, test_loss, anch_imgs_name, test_imgs_name, test_identities, anch_identities\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_config = {\n",
    "    'margin': 0.9,\n",
    "    'lr': 1e-3,\n",
    "    'patience': 5,\n",
    "    'factor': 0.1,\n",
    "    'min_lr': 1e-10,\n",
    "    'threshold': 1e-3,\n",
    "    'weight_decay': 1e-3\n",
    "    }\n",
    "criterion = TripletLoss(margin = model_config['margin'])\n",
    "optimizer = optim.Adam(triplet_model.parameters(), lr = model_config['lr'], weight_decay = model_config['weight_decay'])\n",
    "scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience = model_config['patience'],\n",
    "                                                 factor = model_config['factor'], min_lr = model_config['min_lr'],\n",
    "                                                 threshold = model_config['threshold'], verbose = True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "It was saved at 2023-03-31T05:43:36.786046000\n",
      "Accuracy: 0.5139442231075697\n"
     ]
    }
   ],
   "source": [
    "# get the latest weight\n",
    "latest_weight_path = get_latest_weight('./csv_files/weight_records.csv')\n",
    "\n",
    "acc, distances, labels,loss, anch_imgs_name, test_imgs_name, test_identities, anch_identities = test_model(MobileFaceNet().to(device), val_triplet_dataloader, criterion, './weights/mobile_network_120_updated20230323 035808.pth')\n",
    "print(f'Accuracy: {acc}')\n",
    "predict_issame = np.less(distances, (np.full(np.size(np.array(distances)), 0.98)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "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": "3.9.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}