01 Model Training.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# `00` Loading Required Libraries:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "import mediapipe as mp\n",
    "import cv2\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import pandas as pd\n",
    "import pickle"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# `01` Loading Data:"
   ]
  },
  {
   "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>nose_x</th>\n",
       "      <th>nose_y</th>\n",
       "      <th>forehead_x</th>\n",
       "      <th>forehead_y</th>\n",
       "      <th>left_eye_x</th>\n",
       "      <th>left_eye_y</th>\n",
       "      <th>mouth_left_x</th>\n",
       "      <th>mouth_left_y</th>\n",
       "      <th>chin_x</th>\n",
       "      <th>chin_y</th>\n",
       "      <th>right_eye_x</th>\n",
       "      <th>right_eye_y</th>\n",
       "      <th>mouth_right_x</th>\n",
       "      <th>mouth_right_y</th>\n",
       "      <th>pitch</th>\n",
       "      <th>yaw</th>\n",
       "      <th>roll</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0.489194</td>\n",
       "      <td>0.639722</td>\n",
       "      <td>0.521101</td>\n",
       "      <td>0.295806</td>\n",
       "      <td>0.354743</td>\n",
       "      <td>0.441597</td>\n",
       "      <td>0.403753</td>\n",
       "      <td>0.680927</td>\n",
       "      <td>0.481321</td>\n",
       "      <td>0.807920</td>\n",
       "      <td>0.669544</td>\n",
       "      <td>0.473942</td>\n",
       "      <td>0.574203</td>\n",
       "      <td>0.702994</td>\n",
       "      <td>-0.399231</td>\n",
       "      <td>0.018227</td>\n",
       "      <td>0.085676</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>0.401499</td>\n",
       "      <td>0.604184</td>\n",
       "      <td>0.475316</td>\n",
       "      <td>0.344523</td>\n",
       "      <td>0.408574</td>\n",
       "      <td>0.481754</td>\n",
       "      <td>0.421570</td>\n",
       "      <td>0.668330</td>\n",
       "      <td>0.447475</td>\n",
       "      <td>0.764491</td>\n",
       "      <td>0.605816</td>\n",
       "      <td>0.506940</td>\n",
       "      <td>0.523344</td>\n",
       "      <td>0.691370</td>\n",
       "      <td>0.470065</td>\n",
       "      <td>1.189533</td>\n",
       "      <td>0.300959</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.445534</td>\n",
       "      <td>0.621876</td>\n",
       "      <td>0.443547</td>\n",
       "      <td>0.373616</td>\n",
       "      <td>0.424176</td>\n",
       "      <td>0.508653</td>\n",
       "      <td>0.480120</td>\n",
       "      <td>0.683903</td>\n",
       "      <td>0.528819</td>\n",
       "      <td>0.769280</td>\n",
       "      <td>0.586251</td>\n",
       "      <td>0.500861</td>\n",
       "      <td>0.582848</td>\n",
       "      <td>0.679839</td>\n",
       "      <td>-0.184650</td>\n",
       "      <td>0.881137</td>\n",
       "      <td>-0.236852</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>0.467992</td>\n",
       "      <td>0.648420</td>\n",
       "      <td>0.407912</td>\n",
       "      <td>0.370758</td>\n",
       "      <td>0.354222</td>\n",
       "      <td>0.524016</td>\n",
       "      <td>0.463581</td>\n",
       "      <td>0.699305</td>\n",
       "      <td>0.551854</td>\n",
       "      <td>0.780015</td>\n",
       "      <td>0.605186</td>\n",
       "      <td>0.467229</td>\n",
       "      <td>0.604576</td>\n",
       "      <td>0.666296</td>\n",
       "      <td>-0.175379</td>\n",
       "      <td>0.299208</td>\n",
       "      <td>-0.373374</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>0.372060</td>\n",
       "      <td>0.621196</td>\n",
       "      <td>0.395060</td>\n",
       "      <td>0.392921</td>\n",
       "      <td>0.397376</td>\n",
       "      <td>0.513964</td>\n",
       "      <td>0.433618</td>\n",
       "      <td>0.697691</td>\n",
       "      <td>0.467873</td>\n",
       "      <td>0.773994</td>\n",
       "      <td>0.525802</td>\n",
       "      <td>0.498256</td>\n",
       "      <td>0.508779</td>\n",
       "      <td>0.698543</td>\n",
       "      <td>-0.882169</td>\n",
       "      <td>1.198004</td>\n",
       "      <td>-1.033374</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1995</th>\n",
       "      <td>0.486602</td>\n",
       "      <td>0.624109</td>\n",
       "      <td>0.522991</td>\n",
       "      <td>0.301004</td>\n",
       "      <td>0.354324</td>\n",
       "      <td>0.441646</td>\n",
       "      <td>0.412282</td>\n",
       "      <td>0.681156</td>\n",
       "      <td>0.483026</td>\n",
       "      <td>0.786989</td>\n",
       "      <td>0.680817</td>\n",
       "      <td>0.491351</td>\n",
       "      <td>0.567541</td>\n",
       "      <td>0.702234</td>\n",
       "      <td>-0.197102</td>\n",
       "      <td>-0.070430</td>\n",
       "      <td>0.105118</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1996</th>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>-0.232617</td>\n",
       "      <td>-1.418751</td>\n",
       "      <td>0.175960</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1997</th>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>-1.447955</td>\n",
       "      <td>1.431267</td>\n",
       "      <td>-1.509418</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1998</th>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>NaN</td>\n",
       "      <td>-0.420465</td>\n",
       "      <td>-1.191176</td>\n",
       "      <td>0.451515</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1999</th>\n",
       "      <td>0.513754</td>\n",
       "      <td>0.606096</td>\n",
       "      <td>0.505315</td>\n",
       "      <td>0.317544</td>\n",
       "      <td>0.370474</td>\n",
       "      <td>0.487047</td>\n",
       "      <td>0.437243</td>\n",
       "      <td>0.707608</td>\n",
       "      <td>0.513398</td>\n",
       "      <td>0.793730</td>\n",
       "      <td>0.633600</td>\n",
       "      <td>0.484782</td>\n",
       "      <td>0.587627</td>\n",
       "      <td>0.698070</td>\n",
       "      <td>-0.015552</td>\n",
       "      <td>-0.180126</td>\n",
       "      <td>-0.024546</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>2000 rows × 17 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "        nose_x    nose_y  forehead_x  forehead_y  left_eye_x  left_eye_y  \\\n",
       "0     0.489194  0.639722    0.521101    0.295806    0.354743    0.441597   \n",
       "1     0.401499  0.604184    0.475316    0.344523    0.408574    0.481754   \n",
       "2     0.445534  0.621876    0.443547    0.373616    0.424176    0.508653   \n",
       "3     0.467992  0.648420    0.407912    0.370758    0.354222    0.524016   \n",
       "4     0.372060  0.621196    0.395060    0.392921    0.397376    0.513964   \n",
       "...        ...       ...         ...         ...         ...         ...   \n",
       "1995  0.486602  0.624109    0.522991    0.301004    0.354324    0.441646   \n",
       "1996       NaN       NaN         NaN         NaN         NaN         NaN   \n",
       "1997       NaN       NaN         NaN         NaN         NaN         NaN   \n",
       "1998       NaN       NaN         NaN         NaN         NaN         NaN   \n",
       "1999  0.513754  0.606096    0.505315    0.317544    0.370474    0.487047   \n",
       "\n",
       "      mouth_left_x  mouth_left_y    chin_x    chin_y  right_eye_x  \\\n",
       "0         0.403753      0.680927  0.481321  0.807920     0.669544   \n",
       "1         0.421570      0.668330  0.447475  0.764491     0.605816   \n",
       "2         0.480120      0.683903  0.528819  0.769280     0.586251   \n",
       "3         0.463581      0.699305  0.551854  0.780015     0.605186   \n",
       "4         0.433618      0.697691  0.467873  0.773994     0.525802   \n",
       "...            ...           ...       ...       ...          ...   \n",
       "1995      0.412282      0.681156  0.483026  0.786989     0.680817   \n",
       "1996           NaN           NaN       NaN       NaN          NaN   \n",
       "1997           NaN           NaN       NaN       NaN          NaN   \n",
       "1998           NaN           NaN       NaN       NaN          NaN   \n",
       "1999      0.437243      0.707608  0.513398  0.793730     0.633600   \n",
       "\n",
       "      right_eye_y  mouth_right_x  mouth_right_y     pitch       yaw      roll  \n",
       "0        0.473942       0.574203       0.702994 -0.399231  0.018227  0.085676  \n",
       "1        0.506940       0.523344       0.691370  0.470065  1.189533  0.300959  \n",
       "2        0.500861       0.582848       0.679839 -0.184650  0.881137 -0.236852  \n",
       "3        0.467229       0.604576       0.666296 -0.175379  0.299208 -0.373374  \n",
       "4        0.498256       0.508779       0.698543 -0.882169  1.198004 -1.033374  \n",
       "...           ...            ...            ...       ...       ...       ...  \n",
       "1995     0.491351       0.567541       0.702234 -0.197102 -0.070430  0.105118  \n",
       "1996          NaN            NaN            NaN -0.232617 -1.418751  0.175960  \n",
       "1997          NaN            NaN            NaN -1.447955  1.431267 -1.509418  \n",
       "1998          NaN            NaN            NaN -0.420465 -1.191176  0.451515  \n",
       "1999     0.484782       0.587627       0.698070 -0.015552 -0.180126 -0.024546  \n",
       "\n",
       "[2000 rows x 17 columns]"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "poses_df = pd.read_csv('./poses.csv')\n",
    "poses_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1792, 17)"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "poses_df = poses_df.dropna(axis=0)\n",
    "poses_df.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# `02` Preprocessing:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To make the system independent of the size of the images it was trained on or the face position in them, we have to perform some preprocessing:  \n",
    "1- Center the face features around the fixed point (nose), this makes the model work with faces that can take any position in the image, whether it was in the center of the image or the top right, etc..  \n",
    "2- Make the face features scale independent so they don't get affected by the distance between the face and the camera (whether you was close to the camera or not), this can be done by scaling the features by the distacne between 2 points on the face."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "def normalize(poses_df):\n",
    "    normalized_df = poses_df.copy()\n",
    "    \n",
    "    for dim in ['x', 'y']:\n",
    "        # Centerning around the nose \n",
    "        for feature in ['forehead_'+dim, 'nose_'+dim, 'mouth_left_'+dim, 'mouth_right_'+dim, 'left_eye_'+dim, 'chin_'+dim, 'right_eye_'+dim]:\n",
    "            normalized_df[feature] = poses_df[feature] - poses_df['nose_'+dim]\n",
    "        \n",
    "        \n",
    "        # Scaling\n",
    "        diff = normalized_df['mouth_right_'+dim] - normalized_df['left_eye_'+dim]\n",
    "        for feature in ['forehead_'+dim, 'nose_'+dim, 'mouth_left_'+dim, 'mouth_right_'+dim, 'left_eye_'+dim, 'chin_'+dim, 'right_eye_'+dim]:\n",
    "            normalized_df[feature] = normalized_df[feature] / diff\n",
    "    \n",
    "    return normalized_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>nose_x</th>\n",
       "      <th>nose_y</th>\n",
       "      <th>forehead_x</th>\n",
       "      <th>forehead_y</th>\n",
       "      <th>left_eye_x</th>\n",
       "      <th>left_eye_y</th>\n",
       "      <th>mouth_left_x</th>\n",
       "      <th>mouth_left_y</th>\n",
       "      <th>chin_x</th>\n",
       "      <th>chin_y</th>\n",
       "      <th>right_eye_x</th>\n",
       "      <th>right_eye_y</th>\n",
       "      <th>mouth_right_x</th>\n",
       "      <th>mouth_right_y</th>\n",
       "      <th>pitch</th>\n",
       "      <th>yaw</th>\n",
       "      <th>roll</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.145389</td>\n",
       "      <td>-1.315685</td>\n",
       "      <td>-0.612647</td>\n",
       "      <td>-0.757945</td>\n",
       "      <td>-0.389327</td>\n",
       "      <td>0.157636</td>\n",
       "      <td>-0.035876</td>\n",
       "      <td>0.643461</td>\n",
       "      <td>0.821789</td>\n",
       "      <td>-0.634208</td>\n",
       "      <td>0.387353</td>\n",
       "      <td>0.242055</td>\n",
       "      <td>-0.399231</td>\n",
       "      <td>0.018227</td>\n",
       "      <td>0.085676</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.643168</td>\n",
       "      <td>-1.238742</td>\n",
       "      <td>0.061645</td>\n",
       "      <td>-0.584067</td>\n",
       "      <td>0.174874</td>\n",
       "      <td>0.306017</td>\n",
       "      <td>0.400586</td>\n",
       "      <td>0.764764</td>\n",
       "      <td>1.780222</td>\n",
       "      <td>-0.463914</td>\n",
       "      <td>1.061645</td>\n",
       "      <td>0.415933</td>\n",
       "      <td>0.470065</td>\n",
       "      <td>1.189533</td>\n",
       "      <td>0.300959</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>-0.012524</td>\n",
       "      <td>-1.450241</td>\n",
       "      <td>-0.134608</td>\n",
       "      <td>-0.661405</td>\n",
       "      <td>0.217967</td>\n",
       "      <td>0.362334</td>\n",
       "      <td>0.524888</td>\n",
       "      <td>0.861074</td>\n",
       "      <td>0.886842</td>\n",
       "      <td>-0.706921</td>\n",
       "      <td>0.865392</td>\n",
       "      <td>0.338595</td>\n",
       "      <td>-0.184650</td>\n",
       "      <td>0.881137</td>\n",
       "      <td>-0.236852</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>-0.239980</td>\n",
       "      <td>-1.951528</td>\n",
       "      <td>-0.454436</td>\n",
       "      <td>-0.874364</td>\n",
       "      <td>-0.017618</td>\n",
       "      <td>0.357641</td>\n",
       "      <td>0.334972</td>\n",
       "      <td>0.924901</td>\n",
       "      <td>0.547998</td>\n",
       "      <td>-1.273491</td>\n",
       "      <td>0.545564</td>\n",
       "      <td>0.125636</td>\n",
       "      <td>-0.175379</td>\n",
       "      <td>0.299208</td>\n",
       "      <td>-0.373374</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.206458</td>\n",
       "      <td>-1.236728</td>\n",
       "      <td>0.227247</td>\n",
       "      <td>-0.580952</td>\n",
       "      <td>0.552568</td>\n",
       "      <td>0.414430</td>\n",
       "      <td>0.860054</td>\n",
       "      <td>0.827818</td>\n",
       "      <td>1.380055</td>\n",
       "      <td>-0.666052</td>\n",
       "      <td>1.227247</td>\n",
       "      <td>0.419048</td>\n",
       "      <td>-0.882169</td>\n",
       "      <td>1.198004</td>\n",
       "      <td>-1.033374</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1991</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>-0.069582</td>\n",
       "      <td>-1.516075</td>\n",
       "      <td>-0.835918</td>\n",
       "      <td>-0.675221</td>\n",
       "      <td>-0.497313</td>\n",
       "      <td>0.355848</td>\n",
       "      <td>-0.156862</td>\n",
       "      <td>0.798616</td>\n",
       "      <td>0.459186</td>\n",
       "      <td>-0.690928</td>\n",
       "      <td>0.164082</td>\n",
       "      <td>0.324779</td>\n",
       "      <td>-0.306358</td>\n",
       "      <td>-0.283822</td>\n",
       "      <td>0.038554</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1992</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>-0.034515</td>\n",
       "      <td>-1.424327</td>\n",
       "      <td>-0.886287</td>\n",
       "      <td>-0.717177</td>\n",
       "      <td>-0.613809</td>\n",
       "      <td>0.292342</td>\n",
       "      <td>-0.229223</td>\n",
       "      <td>0.765727</td>\n",
       "      <td>0.467706</td>\n",
       "      <td>-0.711614</td>\n",
       "      <td>0.113713</td>\n",
       "      <td>0.282823</td>\n",
       "      <td>-0.367547</td>\n",
       "      <td>-0.429723</td>\n",
       "      <td>0.122791</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1994</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.424978</td>\n",
       "      <td>-1.002710</td>\n",
       "      <td>-0.711851</td>\n",
       "      <td>-0.594513</td>\n",
       "      <td>-0.683567</td>\n",
       "      <td>0.258108</td>\n",
       "      <td>-0.404571</td>\n",
       "      <td>0.692207</td>\n",
       "      <td>0.930189</td>\n",
       "      <td>-0.319230</td>\n",
       "      <td>0.288149</td>\n",
       "      <td>0.405487</td>\n",
       "      <td>-0.156035</td>\n",
       "      <td>0.567114</td>\n",
       "      <td>-0.108536</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1995</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>0.170664</td>\n",
       "      <td>-1.239908</td>\n",
       "      <td>-0.620394</td>\n",
       "      <td>-0.700198</td>\n",
       "      <td>-0.348568</td>\n",
       "      <td>0.218915</td>\n",
       "      <td>-0.016775</td>\n",
       "      <td>0.625044</td>\n",
       "      <td>0.910883</td>\n",
       "      <td>-0.509458</td>\n",
       "      <td>0.379606</td>\n",
       "      <td>0.299802</td>\n",
       "      <td>-0.197102</td>\n",
       "      <td>-0.070430</td>\n",
       "      <td>0.105118</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1999</th>\n",
       "      <td>0.0</td>\n",
       "      <td>0.0</td>\n",
       "      <td>-0.038861</td>\n",
       "      <td>-1.367399</td>\n",
       "      <td>-0.659814</td>\n",
       "      <td>-0.564150</td>\n",
       "      <td>-0.352338</td>\n",
       "      <td>0.481050</td>\n",
       "      <td>-0.001641</td>\n",
       "      <td>0.889168</td>\n",
       "      <td>0.551894</td>\n",
       "      <td>-0.574883</td>\n",
       "      <td>0.340186</td>\n",
       "      <td>0.435850</td>\n",
       "      <td>-0.015552</td>\n",
       "      <td>-0.180126</td>\n",
       "      <td>-0.024546</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>1792 rows × 17 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "      nose_x  nose_y  forehead_x  forehead_y  left_eye_x  left_eye_y  \\\n",
       "0        0.0     0.0    0.145389   -1.315685   -0.612647   -0.757945   \n",
       "1        0.0     0.0    0.643168   -1.238742    0.061645   -0.584067   \n",
       "2        0.0     0.0   -0.012524   -1.450241   -0.134608   -0.661405   \n",
       "3        0.0     0.0   -0.239980   -1.951528   -0.454436   -0.874364   \n",
       "4        0.0     0.0    0.206458   -1.236728    0.227247   -0.580952   \n",
       "...      ...     ...         ...         ...         ...         ...   \n",
       "1991     0.0     0.0   -0.069582   -1.516075   -0.835918   -0.675221   \n",
       "1992     0.0     0.0   -0.034515   -1.424327   -0.886287   -0.717177   \n",
       "1994     0.0     0.0    0.424978   -1.002710   -0.711851   -0.594513   \n",
       "1995     0.0     0.0    0.170664   -1.239908   -0.620394   -0.700198   \n",
       "1999     0.0     0.0   -0.038861   -1.367399   -0.659814   -0.564150   \n",
       "\n",
       "      mouth_left_x  mouth_left_y    chin_x    chin_y  right_eye_x  \\\n",
       "0        -0.389327      0.157636 -0.035876  0.643461     0.821789   \n",
       "1         0.174874      0.306017  0.400586  0.764764     1.780222   \n",
       "2         0.217967      0.362334  0.524888  0.861074     0.886842   \n",
       "3        -0.017618      0.357641  0.334972  0.924901     0.547998   \n",
       "4         0.552568      0.414430  0.860054  0.827818     1.380055   \n",
       "...            ...           ...       ...       ...          ...   \n",
       "1991     -0.497313      0.355848 -0.156862  0.798616     0.459186   \n",
       "1992     -0.613809      0.292342 -0.229223  0.765727     0.467706   \n",
       "1994     -0.683567      0.258108 -0.404571  0.692207     0.930189   \n",
       "1995     -0.348568      0.218915 -0.016775  0.625044     0.910883   \n",
       "1999     -0.352338      0.481050 -0.001641  0.889168     0.551894   \n",
       "\n",
       "      right_eye_y  mouth_right_x  mouth_right_y     pitch       yaw      roll  \n",
       "0       -0.634208       0.387353       0.242055 -0.399231  0.018227  0.085676  \n",
       "1       -0.463914       1.061645       0.415933  0.470065  1.189533  0.300959  \n",
       "2       -0.706921       0.865392       0.338595 -0.184650  0.881137 -0.236852  \n",
       "3       -1.273491       0.545564       0.125636 -0.175379  0.299208 -0.373374  \n",
       "4       -0.666052       1.227247       0.419048 -0.882169  1.198004 -1.033374  \n",
       "...           ...            ...            ...       ...       ...       ...  \n",
       "1991    -0.690928       0.164082       0.324779 -0.306358 -0.283822  0.038554  \n",
       "1992    -0.711614       0.113713       0.282823 -0.367547 -0.429723  0.122791  \n",
       "1994    -0.319230       0.288149       0.405487 -0.156035  0.567114 -0.108536  \n",
       "1995    -0.509458       0.379606       0.299802 -0.197102 -0.070430  0.105118  \n",
       "1999    -0.574883       0.340186       0.435850 -0.015552 -0.180126 -0.024546  \n",
       "\n",
       "[1792 rows x 17 columns]"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "poses_df = normalize(poses_df)\n",
    "poses_df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(1433, 14) (1433, 3)\n"
     ]
    }
   ],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "train_df, val_df = train_test_split(poses_df, test_size=0.2, random_state=42)\n",
    "\n",
    "\n",
    "X_train = train_df.drop(['pitch', 'yaw', 'roll'], axis=1)\n",
    "y_train = train_df[['pitch', 'yaw', 'roll']]\n",
    "\n",
    "X_val = val_df.drop(['pitch', 'yaw', 'roll'], axis=1)\n",
    "y_val = val_df[['pitch', 'yaw', 'roll']]\n",
    "print(X_train.shape, y_train.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# `03` Model Training:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### `a)` Support Vector Regressor:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "best parameters:  {'estimator__C': 2.3463404498842957, 'estimator__coef0': 4, 'estimator__degree': 4, 'estimator__epsilon': 0.13167456935454494, 'estimator__gamma': 0.09129425537759533, 'estimator__kernel': 'rbf'}\n",
      "train_rmse:  0.22580119825259917\n",
      "validation_rmse:  0.1884249528604136\n"
     ]
    }
   ],
   "source": [
    "from sklearn.svm import SVR\n",
    "from sklearn.multioutput import MultiOutputRegressor\n",
    "from sklearn.model_selection import RandomizedSearchCV\n",
    "import scipy.stats\n",
    "from sklearn.metrics import mean_squared_error\n",
    "\n",
    "param_grid = {\n",
    "    'estimator__kernel': ['linear', 'poly', 'rbf'],\n",
    "    'estimator__C': scipy.stats.expon(scale=5),\n",
    "    'estimator__degree': np.arange(2, 6),\n",
    "    'estimator__coef0': np.arange(0, 6),\n",
    "    'estimator__gamma': scipy.stats.expon(scale=.1),\n",
    "    'estimator__epsilon': scipy.stats.expon(scale=.1)\n",
    "}\n",
    "\n",
    "svr = SVR()\n",
    "multi_out_svr = MultiOutputRegressor(svr)\n",
    "random_search = RandomizedSearchCV(multi_out_svr, param_grid, scoring='neg_mean_squared_error', random_state=42)\n",
    "random_search.fit(X_train, y_train)\n",
    "svr_model = random_search.best_estimator_\n",
    "print('best parameters: ', random_search.best_params_)\n",
    "print('train_rmse: ', np.sqrt(-random_search.best_score_))\n",
    "print('validation_rmse: ', np.sqrt(mean_squared_error(y_val, svr_model.predict(X_val))))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### `b)` Desicion Tree Regressor:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "best parameters:  {'estimator__ccp_alpha': 0.0054293848034671355, 'estimator__max_depth': 8, 'estimator__max_features': 14, 'estimator__min_samples_leaf': 8, 'estimator__min_samples_split': 64}\n",
      "train_rmse:  0.26910600093700265\n",
      "validation_rmse:  0.23639105977796043\n"
     ]
    }
   ],
   "source": [
    "from sklearn.tree import DecisionTreeRegressor\n",
    "from sklearn.model_selection import RandomizedSearchCV\n",
    "\n",
    "param_grid = {\n",
    "    'estimator__max_depth': np.arange(2, 10),\n",
    "    'estimator__min_samples_split': np.logspace(1, 7, 7, base=2).astype(int),\n",
    "    'estimator__min_samples_leaf': np.logspace(0, 6, 7, base=2).astype(int),\n",
    "    'estimator__max_features': np.arange(4, 15),\n",
    "    'estimator__ccp_alpha': scipy.stats.expon(scale=.1),\n",
    "}\n",
    "\n",
    "reg_tree = DecisionTreeRegressor(random_state=42)\n",
    "multi_out_tree = MultiOutputRegressor(reg_tree)\n",
    "\n",
    "random_search = RandomizedSearchCV(multi_out_tree, param_grid, scoring='neg_mean_squared_error', n_iter=40)\n",
    "random_search.fit(X_train, y_train)\n",
    "tree_model = random_search.best_estimator_\n",
    "\n",
    "print('best parameters: ', random_search.best_params_)\n",
    "print('train_rmse: ', np.sqrt(-random_search.best_score_))\n",
    "print('validation_rmse: ', np.sqrt(mean_squared_error(y_val, tree_model.predict(X_val))))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "I will use the SVR model because it gave a better performance than the desicion tree.  \n",
    "But before using it, I will perform some finetuning to the SVR model hyperparameters by doing grid search around the best hyperparameters the random search returned."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "best parameters:  {'estimator__C': 0.7500000000000001, 'estimator__epsilon': 0.07900000000000001, 'estimator__gamma': 0.1, 'estimator__kernel': 'rbf'}\n",
      "train_rmse:  0.22763984301415535\n",
      "validation_rmse:  0.18924742614596657\n"
     ]
    }
   ],
   "source": [
    "from sklearn.model_selection import GridSearchCV\n",
    "\n",
    "param_grid = {\n",
    "    'estimator__kernel': ['rbf'],\n",
    "    'estimator__C': np.arange(0.6, 0.75, 0.01),\n",
    "    'estimator__gamma': np.arange(0.09, 0.1, 0.001),\n",
    "    'estimator__epsilon': np.arange(0.07, 0.08, 0.001)\n",
    "}\n",
    "\n",
    "svr = SVR()\n",
    "multi_out_svr = MultiOutputRegressor(svr)\n",
    "grid_search = GridSearchCV(multi_out_svr, param_grid, scoring='neg_mean_squared_error')\n",
    "grid_search.fit(X_train, y_train)\n",
    "model = grid_search.best_estimator_\n",
    "print('best parameters: ', grid_search.best_params_)\n",
    "print('train_rmse: ', np.sqrt(-grid_search.best_score_))\n",
    "print('validation_rmse: ', np.sqrt(mean_squared_error(y_val, model.predict(X_val))))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "pickle.dump(model, open('./model.pkl', 'wb'))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.9.7 ('base')",
   "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": "3.9.7"
  },
  "orig_nbformat": 4,
  "vscode": {
   "interpreter": {
    "hash": "b252d7cd785b71e172da3ea3f9bb583c7f237978e37e2489541bd05394a3564b"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}