final_notebook.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "import cv2\n",
    "import mahotas\n",
    "import os\n",
    "from matplotlib import pyplot as plt\n",
    "from mpl_toolkits.axes_grid1 import ImageGrid\n",
    "from skimage import io\n",
    "import os\n",
    "\n",
    "from sklearn.model_selection import train_test_split, GridSearchCV\n",
    "from sklearn.ensemble import RandomForestClassifier\n",
    "from sklearn.metrics import accuracy_score, classification_report, f1_score"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# General Functions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# parameters for the features extraction\n",
    "bins = 8\n",
    "\n",
    "# functions to extract features from the images\n",
    "def fd_hu_moments(image):\n",
    "    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n",
    "    feature = cv2.HuMoments(cv2.moments(image)).flatten()\n",
    "    return feature\n",
    "\n",
    "def fd_haralick(image):\n",
    "    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)\n",
    "    haralick = mahotas.features.haralick(gray).mean(axis=0)\n",
    "    return haralick\n",
    "\n",
    "def fd_histogram(image, mask=None):\n",
    "    image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)\n",
    "    hist  = cv2.calcHist([image], [0, 1, 2], None, [bins, bins, bins], [0, 256, 0, 256, 0, 256])\n",
    "    cv2.normalize(hist, hist)\n",
    "    return hist.flatten()\n",
    "\n",
    "def create_features(imgs, data_path='data-challenge/Train/'):\n",
    "    # Compute the features\n",
    "    feat_haralick = []\n",
    "    feat_histogram = []\n",
    "    feat_hu = []\n",
    "    print(data_path)\n",
    "    for i in imgs:\n",
    "        img = cv2.imread(data_path + i)  # Load image\n",
    "        feat_hu.append(fd_hu_moments(img))  # Append Hu values\n",
    "        feat_haralick.append(fd_haralick(img))  # Append Haralick values\n",
    "        feat_histogram.append(fd_histogram(img))  # Append Histogram values\n",
    "    \n",
    "    df_features = np.hstack([feat_haralick, feat_hu, feat_histogram])\n",
    "    df_features = pd.DataFrame(df_features)\n",
    "    print(df_features.shape)\n",
    "\n",
    "    return df_features\n",
    "\n",
    "def parse_fn(filename):\n",
    "    # <BIOPSY_PROCEDURE>_<TUMOR_CLASS>_<TUMOR_TYPE>-<YEAR>-<SLIDE_ID>-<MAG>-<SEQ>\n",
    "    parsed = filename[:-4].replace('-', '_').split('_')\n",
    "    parsed.append(filename)\n",
    "    return parsed\n",
    "\n",
    "def parseTest_fn(filename):\n",
    "    # <BIOPSY_PROCEDURE>_<ID>\n",
    "    parsed = filename[:-4].split('_')\n",
    "    parsed.append(filename)\n",
    "    return parsed\n",
    "\n",
    "def create_split(images_train, images_test, train_path, test_path):\n",
    "    columns = ['procedure', 'class', 'type', 'magnification', 'slide', 'mag', 'seq', 'filename']\n",
    "\n",
    "    df = pd.DataFrame(list(map(parse_fn, images_train)), columns=columns)\n",
    "    label_maps = {'F':'1', 'DC':'2', 'PC':'3', 'PT':'4', 'MC':'5', 'LC':'6', 'A':'7', 'TA':'8'}\n",
    "    df['type_id'] = df['type'].apply(lambda x: label_maps[x])\n",
    "    \n",
    "    df_features = create_features(images_train, train_path)\n",
    "    X_train, X_val, y_train, y_val = train_test_split(df_features, df['type_id'], stratify=df['type_id'], test_size=0.2)\n",
    "\n",
    "    columns = ['procedure', 'id', 'filename']\n",
    "    info_test = pd.DataFrame(list(map(parseTest_fn, images_test)), columns=columns)\n",
    "    df_features_test = create_features(info_test['filename'], test_path)\n",
    "\n",
    "    return X_train, X_val, y_train, y_val, df_features_test, info_test\n",
    "\n",
    "def evaluate_model(model, X_val, y_val):\n",
    "    y_pred = model.predict(X_val)\n",
    "    print(accuracy_score(y_val,y_pred))\n",
    "    print(classification_report(y_val, y_pred))\n",
    "    print(f1_score(y_val, y_pred, average='macro'))\n",
    "    return y_pred\n",
    "\n",
    "def create_submission(model, X_test, info_test):\n",
    "    y_pred_test = model.predict(X_test)\n",
    "    df_pred = pd.concat([info_test['id'], pd.Series(y_pred_test, name='type_id')], axis=1)\n",
    "\n",
    "    display(df_pred['type_id'].value_counts())\n",
    "    df_pred.to_csv('my_pred.csv', index=False, header=False)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Data Pre-processing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "data-challenge/Train/\n",
      "(422, 532)\n",
      "data-challenge/Test/\n",
      "(207, 532)\n"
     ]
    }
   ],
   "source": [
    "train_path = 'data-challenge/Train/'\n",
    "test_path = 'data-challenge/Test/'\n",
    "images_train = os.listdir(path=train_path)\n",
    "images_test = os.listdir(path=test_path)\n",
    "\n",
    "X_train, X_val, y_train, y_val, df_features_test, info_test = create_split(images_train, images_test, train_path, test_path)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Best configuration"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Voici la meilleur configuration obtenue. Je l'ai obtenue assez rapidement. Ce modele sklearn est une implémentation de lightGBM, un algorithme de boosting plus légère et efficace que les autres methodes de boosting en utilisant un algorithme basé sur un histogramme de valeurs.\n",
    "\n",
    "Le parametre le plus influents semblait etre min_sample_leaf. Cette valeur est assez élevé et permet de limiter la compléxité des modeles en limitant le nombre de feuilles par arbre. J'ai par la suite tenté d'améliorer les resultats en touchant à ce paramètre, ou à la régularisation L2 afin de limiter encore plus l'overfitting mais cela n'a fait que dégrader les performances. La max depth à aussi été tune. La valeur 7 semblait etre un bon compromis. Une valeur moins élevé réduisait les performances du modèle tandis qu'une valeur plus élevé favorisait l'overfitting.\n",
    "\n",
    "Il est assez étonnant que ce modèle soit le plus performant pour ces données. La doc sklearn mentionne que celui-ci est particulièrement efficace pour les gros datasets, or ce n'est pas notre cas ici. Une autre source [kaggle](https://www.kaggle.com/general/264327) annonce que light gbm à tendance à overfit les petits datasets. Malgré tout, c'est le meilleur modèle que j'ai pu obtenir parmis les autres algorithmes de boosting de gradient.\n",
    "\n",
    "Je n'ai pas fait de recherches de paramêtres avec un grid search car les modèles avait tendance à très vite overfitter. Je me retrouvais parfois avec de très bons F1 score sur le jeu de test mais une fois la prediction soumise, le F1 score etait très mauvais. J'y suis donc allé à taton en verifiant à chaque fois le score lors de la soumission. Ce qui explique mon nombre de soumissions très élevé."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.8588235294117647\n",
      "              precision    recall  f1-score   support\n",
      "\n",
      "           1       0.92      0.80      0.86        15\n",
      "           2       1.00      0.89      0.94         9\n",
      "           3       0.82      0.95      0.88        19\n",
      "           4       0.90      0.82      0.86        11\n",
      "           5       1.00      0.75      0.86        12\n",
      "           6       1.00      0.67      0.80         3\n",
      "           7       0.75      0.92      0.83        13\n",
      "           8       0.60      1.00      0.75         3\n",
      "\n",
      "    accuracy                           0.86        85\n",
      "   macro avg       0.87      0.85      0.85        85\n",
      "weighted avg       0.88      0.86      0.86        85\n",
      "\n",
      "0.8460300036751454\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "7    66\n",
       "2    48\n",
       "3    34\n",
       "1    23\n",
       "4    22\n",
       "5    13\n",
       "8     1\n",
       "Name: type_id, dtype: int64"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from sklearn.ensemble import HistGradientBoostingClassifier\n",
    "# best config : 0.7127 model_hgb = HistGradientBoostingClassifier(max_iter=1000, learning_rate=0.1,min_samples_leaf=20, l2_regularization=0, max_depth=7, random_state=1)\n",
    "model_hgb = HistGradientBoostingClassifier(max_iter=1000, learning_rate=0.1,min_samples_leaf=20, l2_regularization=0, max_depth=7, random_state=1)\n",
    "model_hgb.fit(X_train, y_train)\n",
    "evaluate_model(model_hgb, X_val, y_val)\n",
    "\n",
    "create_submission(model_hgb, df_features_test, info_test)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Autres pistes explorées"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Boosting, SVM, SGD, Transfert Learning, Data augmentation"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Beaucoup d'efforts ont été fournit dans les algorithmes de boosting. Ces derniers ont été les second algorithmes à être testés après le random forest et se sont révelés assez performants. J'ai donc essayé les gradient boosting de base de Sklearn mais aussi XGBoost. Sans succès. J'ai ensuite testé un SVM avec un kernel rbf et un grand grid search sur les parametres C et gamma qui semblait etre les parametres les plus important pour ce kernel. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 800x600 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# le code ne marche pas, il est issue de mon autre notebook, il est juste la pour illustrer le grid search sur les parametres\n",
    "plt.figure(figsize=(8, 6))\n",
    "plt.subplots_adjust(left=0.2, right=0.95, bottom=0.15, top=0.95)\n",
    "plt.imshow(\n",
    "    scores,\n",
    "    interpolation=\"nearest\",\n",
    "    cmap=plt.cm.hot,\n",
    "    norm=MidpointNormalize(vmin=0.2, midpoint=0.92),\n",
    ")\n",
    "plt.xlabel(\"gamma\")\n",
    "plt.ylabel(\"C\")\n",
    "plt.colorbar()\n",
    "plt.xticks(np.arange(len(gamma_range)), gamma_range, rotation=45)\n",
    "plt.yticks(np.arange(len(C_range)), C_range)\n",
    "plt.title(\"Validation accuracy\")\n",
    "plt.show()"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Le grid search pour le SVM privilegiait des valeurs extremes pour C et gamma pour atteindre un score plutot mauvais (autour des 0.6). La piste des SVM à donc été abandonné."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "La dernière piste beaucoup explorée fut le transfert learning. Le principe est  de prendre un modele de classification d'image pré entrainer et de l'utiliser afin de classifier les images de notre dataset. Le choix du modele s'est porté sur MobileNetV3 pour son efficacité mémoire et de prédiction. Utilisant mon ordinateur portable, je ne pouvais pas me permettre d'utiliser un gros modele. Les poids d'initialisation utilisés étaient les poids Imagenet, un des plus gros dataset de classification d'images. Enfin, la couche de prédiction initiale du modele à été remplacée par mon propre sous modele. Celui-ci etait composé d'une couche GlobalAveragePooling2D afin de récuperer un vecteur, puis d'une combinaison de couche dense de taille variable selon les essais. La couche de sortie est une couche dense de taille 8 pour faire de la classification avec une couche d'activation softmax.\n",
    "Bien que vite performant, le transfert learning ne m'a pas permit de dépassé mon meilleur score. Une pipeline d'augmentation de données à aussi été mis en place pour lutter contre le manque de donnée mais cela n'a pas suffit."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Perspective d'amélioration"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Durant ce projet, je n'ai pas retravailler les features par manque (ou mauvaise gestion) de temps. J'aurais aimé essayer d'autres features ou encore faire une PCA. J'aurais aimé aussi prendre la sortie d'une couche intermédiaire d'un réseau pré entrainé et l'utiliser comme nouveau vecteur de feauture dans un modele de ML comme un SVM ou un gradient boosting."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "data_challenge_telecom-3PZ7n0BO",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.10"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}