SolutionTPDetectionDeSentiments.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "QQ5vas82illV"
   },
   "source": [
    "<a href=\"https://colab.research.google.com/github/EmmanuelADAM/IntelligenceArtificiellePython/blob/master/SolutionTPDetectionDeSentiments.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
    "---\n",
    "# Détection de \"sentiments\" (ex. classique tensorflow)\n",
    "\n",
    "## Collection de données de critiques de films \n",
    "Le tp utilisera l'ensemble de données issues de \n",
    "[Movie Review Polarity Dataset](https://www.cs.cornell.edu/people/pabo/movie-review-data/review_polarity.tar.gz) que vous devez télécharger.\n",
    "\n",
    "Ce répertoire contient 1000 revues de films, considérées comme positives (répertoire pos) et 1000 autre considérées négatives (répertoire neg). Ces revues sont au format texte.\n",
    "\n",
    "90% de ces données seront utilisées en tant que données d'entraînement et 10% seront utilisées en données de test.\n",
    "\n",
    "Ces fichiers textes sont composés de mots importants (tokens), d'espaces, de signes de ponctuations et de mots de liaisons."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "AWFXrAtkvcg_"
   },
   "source": [
    "---\n",
    "Si vous utisez colab, vous devez charger le répertoire txt_sentoken de l'archive à la racine de votre Drive Google.\n",
    "\n",
    "Puis il vous faut monter le répertoir drive pour qu'il soit accessible par colab.\n",
    "Exécutez le code suivant,  une clé vous sera demandée. Il vous suffit de suivre le lien, de sélectionner votre profile pour obtenir votre clé que vous copierez dans le champs prévu."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 74
    },
    "colab_type": "code",
    "id": "4TodTL602bBv",
    "outputId": "cbc50963-1484-4189-b87b-b0d7d60af79c"
   },
   "outputs": [
    {
     "ename": "ModuleNotFoundError",
     "evalue": "No module named 'google.colab'",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mModuleNotFoundError\u001b[0m                       Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-1-690f165dac74>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m      1\u001b[0m \u001b[0;31m#bloc à exécuter si vous utilisez colab\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 2\u001b[0;31m \u001b[0;32mfrom\u001b[0m \u001b[0mgoogle\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcolab\u001b[0m \u001b[0;32mimport\u001b[0m \u001b[0mdrive\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfiles\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      3\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      4\u001b[0m \u001b[0mdrive\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmount\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'/content/drive'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mModuleNotFoundError\u001b[0m: No module named 'google.colab'"
     ]
    }
   ],
   "source": [
    "#bloc à exécuter si vous utilisez colab\n",
    "from google.colab import drive, files\n",
    "\n",
    "drive.mount('/content/drive')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "3IlL2ANmisdv"
   },
   "source": [
    "---\n",
    "## Préparation des données\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "oiHdx0y1i8Bi"
   },
   "source": [
    "### Chargement et nettoyage des données\n",
    "\n",
    "Il convient pour chaque fichier texte de \n",
    "  - identifier les termes (**tokens**) entre espaces,\n",
    "  - supprimer toute ponctuation,\n",
    "  - supprimer tous les mots qui ne sont pas uniquement composés de caractères alphabétiques,\n",
    "  - supprimer tous les mots reconnus en tant que mots vides (stop words) (mots de liaison)\n",
    "  - supprimer tous les mots dont la longueur est <= 1 caractère."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "SLWxuYAjxPlo"
   },
   "source": [
    "### Les bibliothèques\n",
    "La préparation des données implique de pouvoir accéder au système de fichier, ainsi qu'à la bibliothèque nltk et aux bibliothèques spécialisées dans le traitement de texte.. : "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "tSqBXmHvxnhF"
   },
   "outputs": [],
   "source": [
    "##import pour les fichiers et le traitement de données : \n",
    "from os import listdir\n",
    "\n",
    "##import pour la creation de tokens\n",
    "import nltk\n",
    "from nltk.corpus import stopwords\n",
    "\n",
    "##import pour la gestion de ponctuation\n",
    "from string import punctuation\n",
    "\n",
    "##import pour une collection spécifique : counter\n",
    "from collections import Counter\n",
    "\n",
    "##import pour les réseaux de neurones : \n",
    "import tensorflow.keras as keras\n",
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras.layers import Dense\n",
    "from tensorflow.keras import preprocessing\n",
    "\n",
    "##import classique pour les tableaux efficaces\n",
    "from numpy import array, zeros\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 68
    },
    "colab_type": "code",
    "id": "n6qZZ-VJw8Wy",
    "outputId": "1c8e7c95-f0ba-4f69-9914-c034c7f534cc"
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "[nltk_data] Downloading package stopwords to\n",
      "[nltk_data]     /Users/emmanueladam/nltk_data...\n",
      "[nltk_data]   Package stopwords is already up-to-date!\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# le code suivant charge l'ensemble des mots non importants (en anglais)\n",
    "nltk.download(\"stopwords\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Fonctions de lectures et d'\"épurage\" de fichier texte"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "by0--ITCi-xa"
   },
   "outputs": [],
   "source": [
    "def load_doc(filename)->str:\n",
    "    \"\"\"retourne le texte inclu dans le fichier filename\"\"\"\n",
    "    # open the file as read only\n",
    "    file = open(filename, 'r')\n",
    "    # read all text\n",
    "    text = file.read()\n",
    "    # close the file\n",
    "    file.close()\n",
    "    return text\n",
    " \n",
    "def clean_doc(doc)->list:\n",
    "    \"\"\"retourne la liste de mots clés inclus dans le texte doc \n",
    "    qui ne font pas parti des stop_words\"\"\"\n",
    "    # split into tokens by white space\n",
    "    tokens = doc.split()\n",
    "    # remove punctuation from each token\n",
    "    table = str.maketrans('', '', punctuation)\n",
    "    tokens = [w.translate(table) for w in tokens]\n",
    "    # remove remaining tokens that are not alphabetic\n",
    "    tokens = [word for word in tokens if word.isalpha()]\n",
    "    # filter out stop words\n",
    "    stop_words = set(stopwords.words('english'))\n",
    "    tokens = [w for w in tokens if not w in stop_words]\n",
    "    # filter out short tokens\n",
    "    tokens = [word for word in tokens if len(word) > 1]\n",
    "    tokens = [word for word in tokens if word!='got']\n",
    "    return tokens\n",
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "UgALvdrLKOD3"
   },
   "source": [
    "### Tests de la récupération de mots clés d'un fichier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 125
    },
    "colab_type": "code",
    "id": "0CpapoCR0Hzs",
    "outputId": "23a01def-2b8c-4fef-e689-1289a90d2d61"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "les 10 premiers mots cles de  /Users/emmanueladam/Downloads/review_polarity/txt_sentoken/neg/cv007_4992.txt\n",
      "['thats', 'exactly', 'long', 'movie', 'felt', 'werent', 'even', 'nine', 'laughs', 'nine']\n",
      "les 10 premiers mots cles de  /Users/emmanueladam/Downloads/review_polarity/txt_sentoken/pos/cv994_12270.txt\n",
      "['thriller', 'set', 'modern', 'day', 'seattle', 'marked', 'marky', 'marks', 'migration', 'good']\n"
     ]
    }
   ],
   "source": [
    "#rep = '/content/drive/My Drive/txt_sentoken'\n",
    "rep = '/Users/emmanueladam/Downloads/review_polarity/txt_sentoken'\n",
    "# mots clés d'un fichier négatif\n",
    "filename = rep + '/neg/cv007_4992.txt'\n",
    "text = load_doc(filename)\n",
    "tokens = clean_doc(text)\n",
    "print('les 10 premiers mots cles de ', filename)\n",
    "print(tokens[:10])\n",
    "# mots clés d'un fichier positif\n",
    "filename = rep + '/pos/cv994_12270.txt'\n",
    "text = load_doc(filename)\n",
    "tokens = clean_doc(text)\n",
    "print('les 10 premiers mots cles de ', filename)\n",
    "print(tokens[:10])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "1yo75YEIjWIT"
   },
   "source": [
    "### Construire le vocabulaire global"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "Axlsvtg8jhAr"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "nb de mots cles trouves dans les repertoires :  44524\n",
      "les 10 premiers mots cles du vocabulaire (et leur nb d'apparition dans les exemples)  : \n",
      "assume : 45, nothing : 717, phrase : 26, perhaps : 414, one : 4966, used : 338, first : 1624, impressions : 6, rumors : 17, hardly : 114, ever : 665, \n",
      "les 10 mots cles les plus utilises :  [('film', 7980), ('one', 4966), ('movie', 4938), ('like', 3182), ('even', 2319), ('good', 2116), ('time', 2052), ('story', 1907), ('films', 1878), ('would', 1856)]\n"
     ]
    }
   ],
   "source": [
    "def add_doc_to_vocab(filename, vocab):\n",
    "    \"\"\"cumule dans la liste vocab les mots du fichier filename \n",
    "    (1 seule occurence par mot dans vocab)\"\"\"\n",
    "    # load doc\n",
    "    doc = load_doc(filename)\n",
    "    # clean doc\n",
    "    tokens = clean_doc(doc)\n",
    "    # update counts\n",
    "    vocab.update(tokens)\n",
    "\n",
    "def build_voc(directory, vocab):\n",
    "    \"\"\"ajoute au dictionnaire vocab les mots cles des 900 premiers fichiers \n",
    "    du repertoire directory\"\"\"\n",
    "    i=0\n",
    "    # walk through all files in the folder\n",
    "    for filename in listdir(directory):\n",
    "        if i>900: break\n",
    "        # create the absolute filename\n",
    "        path = directory + '/' + filename\n",
    "        # add doc to vocab\n",
    "        add_doc_to_vocab(path, vocab)\n",
    "        i = i +1\n",
    "\n",
    "    \n",
    "# creer un vocabulaire (liste de mots clés associés à leurs occurences)\n",
    "vocab = Counter()\n",
    "# ajouter les mots cles des repertoire pos et neg\n",
    "build_voc(rep + '/pos', vocab)\n",
    "build_voc(rep + '/neg', vocab)\n",
    "# afficher le nb de mots cles trouves\n",
    "print(\"nb de mots cles trouves dans les repertoires : \", len(vocab))\n",
    "print(\"les 10 premiers mots cles du vocabulaire \\\n",
    "(et leur nb d\\'apparition dans les exemples)  : \\n\", end='')\n",
    "i=0\n",
    "for (mot,count) in vocab.items(): \n",
    "    print(mot,':',count,end=\", \")\n",
    "    i = i+1\n",
    "    if i>10:break\n",
    "# afficher les 10 mots cles les plus utilises\n",
    "print(\"\\nles 10 mots cles les plus utilises : \", vocab.most_common(10))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "EDFdsl97jq0t"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "en otant les mots utilise moins de  3  fois,  nb de mots cles =  19664\n"
     ]
    }
   ],
   "source": [
    "# ne garder que les mots clés apparaissant au moins 3 fois\n",
    "min_occurrence = 3\n",
    "tokens = [token for (token,count) in vocab.items() if count >= min_occurrence]\n",
    "print('en otant les mots utilise moins de ', min_occurrence, ' fois,', \n",
    "      ' nb de mots cles = ',len(tokens))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "ApMJ_0gHjuEv"
   },
   "outputs": [],
   "source": [
    "def save_list(lines, filename):\n",
    "    \"\"\"sauve les mots de la liste lines dans le fichier filename\"\"\"\n",
    "    # convert lines to a single blob of text\n",
    "    data = '\\n'.join(lines)\n",
    "    # open file\n",
    "    file = open(filename, 'w')\n",
    "    # write text\n",
    "    file.write(data)\n",
    "    # close file\n",
    "    file.close()\n",
    "\n",
    "# save tokens to a vocabulary file\n",
    "save_list(tokens, 'vocab.txt')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "v2oK8x7bj5CU"
   },
   "source": [
    "---\n",
    "## Représentation en sac de mots"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "wdd9RpBskGUy"
   },
   "source": [
    "### Convertir les fichiers en listes de mots-clés appartenant au vocabulaire \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "7Sw6xGMHkUP1"
   },
   "outputs": [],
   "source": [
    "def doc_to_line(filename, vocab)->list:\n",
    "    \"\"\"retourne la liste des mots cles du fichier filename appartenant au vocabulaire vocab\"\"\"\n",
    "    # load the doc\n",
    "    doc = load_doc(filename)\n",
    "    # clean doc\n",
    "    tokens = clean_doc(doc)\n",
    "    # filter by vocab\n",
    "    tokens = [token for token in tokens if token in vocab]\n",
    "    return ' '.join(tokens)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "8bE7XaHdka1O"
   },
   "outputs": [],
   "source": [
    "def process_docs(directory, vocab)->list:\n",
    "    \"\"\"retourne deux listes des mots cles du repertoire directory; \n",
    "    la 1ere liste represente les 900 premiers fichiers du repertoire, \n",
    "    la 2nde represente les 100 derniers\"\"\"\n",
    "    lines_firts = list()\n",
    "    lines_lasts = list()\n",
    "    i=1\n",
    "    # walk through all files in the folder\n",
    "    for filename in listdir(directory):\n",
    "        # create the absolute filename\n",
    "        path = directory + '/' + filename\n",
    "        # load and clean the doc\n",
    "        line = doc_to_line(path, vocab)\n",
    "        if (i<=900): lines_firts.append(line)\n",
    "        else:lines_lasts.append(line)\n",
    "        i = i+1\n",
    "    return (lines_firts,lines_lasts)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "o0ixOikAkfyN"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "nb exemples d'entrainement positifs :  900\n",
      "nb exemples d'entrainement negatifs :  900\n",
      "nb exemples de tests positifs :  100\n",
      "nb exemples de tests negatifs :  100\n"
     ]
    }
   ],
   "source": [
    "# load the vocabulary\n",
    "vocab_filename = 'vocab.txt'\n",
    "vocab = load_doc(vocab_filename)\n",
    "vocab = vocab.split()\n",
    "vocab = set(vocab)\n",
    "# load training and testing reviews\n",
    "(positive_lines_train, positive_lines_test) = process_docs(rep+'/pos', vocab)\n",
    "(negative_lines_train, negative_lines_test) = process_docs(rep+'/neg', vocab)\n",
    "# summarize what we have\n",
    "print(\"nb exemples d'entrainement positifs : \", len(positive_lines_train))\n",
    "print(\"nb exemples d'entrainement negatifs : \", len(negative_lines_train))\n",
    "print(\"nb exemples de tests positifs : \", len(positive_lines_test))\n",
    "print(\"nb exemples de tests negatifs : \", len(negative_lines_test))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'capsule director cure brings weird complex concept screen one viewing enough understand fully premise pulse idea something ghosts internet film amazing apocalyptic style perhaps disturbing disturbed filmmaker world films seem one style bleak worlds creates terrifying cold little known us date films deliver kind horror many filmmakers promise unable deliver ideas fresh time morbid film cure one ideas getting sadly limited release us hopefully enough people see name soon one conjure cure probably classic last year released remake wet afternoon perhaps inserting supernatural elements story pulse back form young computer expert late delivery important software two coworkers go apartment find dismal dark affair spite computer equipment acting strangely lets friends look missing software meanwhile slips behind plastic curtain fails respond calls friends follow behind curtain discover hanged horror enough body seems disappear leaving strange dark spot wall computer seems infected kind computer virus people whose computer gets virus seem superficially die via suicide entirely dead spirits seem remain present somehow real world internet people get computer virus asked want see ghost say yes seem able see real time images spirits still nearby somehow computer shows impossible images ghosts rooms seen cameras exist somehow connected heaven hell somehow filling overflowing like computer disk instead dead seem staying earth inhabiting computer sort passage worlds something doors marked red tape strange electronic computers leave find new kind death film weird ideas together somehow makes film work may totally convey message isolation parallels death whatever convey nightmarish directs screenplay ties story real world familiar accurate computer discussion frequently plot advanced character assumed fact plotting frequently hard follow always strange cinematographer recently kurosawas films creates dark cold gloomy tone images obscured behind plastic curtain scenes milked horror way american exploitation films might people shot guns little blood evidence seeing black computer screens immediately scary going easy visual shock deeper metaphysical dread horror filmmaker world one watch rate metaphysical look isolation scale scale'"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "positive_lines_test[99]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "ZTJ5fim2k1fk"
   },
   "source": [
    "\n",
    "### Convertir les listes de mots en vecteur fréquence d'apparition \n",
    "#### Créer le sac de mot de l'ensemble d'entrainement"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "zVxIx3dtk9HF"
   },
   "outputs": [],
   "source": [
    "# create the tokenizer\n",
    "tokenizer = keras.preprocessing.text.Tokenizer()\n",
    "# build the training doc by alternating positive lines and negative lines\n",
    "training_doc = [None]*(len(positive_lines_train)+len(negative_lines_train))\n",
    "#place a positive line at even positions (0,2,....)\n",
    "training_doc[:len(positive_lines_train)] = positive_lines_train\n",
    "#place a negative line at odd positions (1,3,....)\n",
    "training_doc[len(positive_lines_train):] = negative_lines_train\n",
    "# ask to the tokenizer to build the bag of words : a set of (word, frequence of use)*\n",
    "tokenizer.fit_on_texts(training_doc)\n",
    "xTrain = tokenizer.texts_to_matrix(training_doc, mode='freq')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Help on method texts_to_matrix in module keras_preprocessing.text:\n",
      "\n",
      "texts_to_matrix(texts, mode='binary') method of keras_preprocessing.text.Tokenizer instance\n",
      "    Convert a list of texts to a Numpy matrix.\n",
      "    \n",
      "    # Arguments\n",
      "        texts: list of strings.\n",
      "        mode: one of \"binary\", \"count\", \"tfidf\", \"freq\".\n",
      "    \n",
      "    # Returns\n",
      "        A Numpy matrix.\n",
      "\n"
     ]
    }
   ],
   "source": [
    "help(tokenizer.texts_to_matrix)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "SGd85sb5k_Xs"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Xtrain contient  1800  exemples de  19663  valeurs\n",
      "une valeur = fréquence d'apparition des mots dans le vocabulaire global.\n",
      "Ainsi, premier exemple d'entrainement = \n",
      " [0.         0.01265823 0.01265823 ... 0.         0.         0.        ]\n"
     ]
    }
   ],
   "source": [
    "print('Xtrain contient ', xTrain.shape[0], ' exemples de ', xTrain.shape[1], ' valeurs')\n",
    "print('une valeur = fréquence d\\'apparition des mots dans le vocabulaire global.')\n",
    "print('Ainsi, premier exemple d\\'entrainement = \\n', xTrain[150])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "On décide que les exemples positifs correspondent à une sortie 0 et que les exemples négatifs correspondent à la sortie 1 :"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "#ytrain = suite de (0 (classement pour eval positive) puis 1 (classements pour éval négative))\n",
    "yTrain = zeros(len(positive_lines_train)+len(negative_lines_train), int)\n",
    "yTrain[len(positive_lines_train):] = 1"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Créer le sac de mot de l'ensemble de validation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "twHl-RSElEB_"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Xtest contient  200  exemples de  19663  valeurs de fréquence.\n"
     ]
    }
   ],
   "source": [
    "# build the test doc by alternating positive lines and negative lines\n",
    "test_Doc = [None]*(len(positive_lines_test)+len(negative_lines_test))\n",
    "#place a positive line at even positions (0,2,....)\n",
    "test_Doc[::2] = positive_lines_test\n",
    "#place a negative line at odd positions (1,3,....)\n",
    "test_Doc[1::2] = negative_lines_test\n",
    "# ask to the tokenizer to give the bag of words : a set of (word, frequence of use),\n",
    "# the words are already kown by the tokenizer*\n",
    "xTest = tokenizer.texts_to_matrix(test_Doc, mode='freq')\n",
    "print('Xtest contient ', xTest.shape[0], ' exemples de ', xTest.shape[1], ' valeurs de fréquence.')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "#sortie attendues des exemples de test, ytest = suite de (0, 1)\n",
    "yTest = array([0,1]*(len(positive_lines_test)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([0.        , 0.0097561 , 0.00487805, ..., 0.        , 0.0097561 ,\n",
       "       0.03902439])"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "xTest[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "mCJRY1mTlnls"
   },
   "source": [
    "---\n",
    "## Modèle de réseau pour l'analyse des sentiments\n",
    "Le réseau contient en couche d'entrée aurant de neurones que de valeurs retenues"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "uruPxsQil-gZ"
   },
   "outputs": [],
   "source": [
    "#TODO: donnez le nb de neurones en entrée (= nb de tokens retenus)\n",
    "n_words = xTrain.shape[1] "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "67M4JqX2lNeu"
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2021-11-22 16:01:13.271673: I tensorflow/compiler/jit/xla_cpu_device.cc:41] Not creating XLA devices, tf_xla_enable_xla_devices not set\n",
      "2021-11-22 16:01:13.277855: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA\n",
      "To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n"
     ]
    }
   ],
   "source": [
    "# TODO: définir la structure du réseau\n",
    "model = Sequential()\n",
    "model.add(Dense(5000, input_shape=(n_words,), activation='elu'))\n",
    "model.add(Dense(300, activation='elu'))\n",
    "model.add(Dense(10, activation='elu'))\n",
    "model.add(Dense(1, activation='sigmoid'))\n",
    "# compile network \n",
    "model.compile(loss='hinge', optimizer='adam',  metrics=['accuracy'])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "rlYk8aKemF9b"
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2021-11-22 16:01:14.354845: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:116] None of the MLIR optimization passes are enabled (registered 2)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/20\n",
      "57/57 [==============================] - 15s 238ms/step - loss: 0.9532 - accuracy: 0.5577 - val_loss: 0.6561 - val_accuracy: 0.8700\n",
      "Epoch 2/20\n",
      "57/57 [==============================] - 13s 236ms/step - loss: 0.5323 - accuracy: 0.9488 - val_loss: 0.6432 - val_accuracy: 0.8700\n",
      "Epoch 3/20\n",
      "57/57 [==============================] - 14s 238ms/step - loss: 0.5220 - accuracy: 0.9797 - val_loss: 0.6392 - val_accuracy: 0.8600\n",
      "Epoch 4/20\n",
      "57/57 [==============================] - 13s 230ms/step - loss: 0.4946 - accuracy: 0.9815 - val_loss: 0.6334 - val_accuracy: 0.8800\n",
      "Epoch 5/20\n",
      "57/57 [==============================] - 14s 236ms/step - loss: 0.5092 - accuracy: 0.9865 - val_loss: 0.6669 - val_accuracy: 0.8200\n",
      "Epoch 6/20\n",
      "57/57 [==============================] - 13s 235ms/step - loss: 0.5190 - accuracy: 0.9914 - val_loss: 0.6432 - val_accuracy: 0.8550\n",
      "Epoch 7/20\n",
      "57/57 [==============================] - 13s 232ms/step - loss: 0.5127 - accuracy: 0.9935 - val_loss: 0.6540 - val_accuracy: 0.8450\n",
      "Epoch 8/20\n",
      "57/57 [==============================] - 13s 229ms/step - loss: 0.5120 - accuracy: 0.9928 - val_loss: 0.6466 - val_accuracy: 0.8500\n",
      "Epoch 9/20\n",
      "57/57 [==============================] - 13s 233ms/step - loss: 0.4783 - accuracy: 0.9962 - val_loss: 0.6457 - val_accuracy: 0.8600\n",
      "Epoch 10/20\n",
      "57/57 [==============================] - 13s 234ms/step - loss: 0.4971 - accuracy: 0.9942 - val_loss: 0.6421 - val_accuracy: 0.8650\n",
      "Epoch 11/20\n",
      "57/57 [==============================] - 13s 236ms/step - loss: 0.5177 - accuracy: 0.9943 - val_loss: 0.6440 - val_accuracy: 0.8600\n",
      "Epoch 12/20\n",
      "57/57 [==============================] - 13s 231ms/step - loss: 0.5070 - accuracy: 0.9948 - val_loss: 0.6453 - val_accuracy: 0.8600\n",
      "Epoch 13/20\n",
      "57/57 [==============================] - 13s 231ms/step - loss: 0.5123 - accuracy: 0.9929 - val_loss: 0.6458 - val_accuracy: 0.8650\n",
      "Epoch 14/20\n",
      "57/57 [==============================] - 13s 232ms/step - loss: 0.5262 - accuracy: 0.9924 - val_loss: 0.6461 - val_accuracy: 0.8650\n",
      "Epoch 15/20\n",
      "57/57 [==============================] - 13s 231ms/step - loss: 0.5060 - accuracy: 0.9964 - val_loss: 0.6462 - val_accuracy: 0.8650\n",
      "Epoch 16/20\n",
      "57/57 [==============================] - 13s 233ms/step - loss: 0.5131 - accuracy: 0.9922 - val_loss: 0.6463 - val_accuracy: 0.8600\n",
      "Epoch 17/20\n",
      "57/57 [==============================] - 13s 230ms/step - loss: 0.5007 - accuracy: 0.9967 - val_loss: 0.6465 - val_accuracy: 0.8500\n",
      "Epoch 18/20\n",
      "57/57 [==============================] - 14s 244ms/step - loss: 0.5086 - accuracy: 0.9962 - val_loss: 0.6464 - val_accuracy: 0.8550\n",
      "Epoch 19/20\n",
      "57/57 [==============================] - 14s 240ms/step - loss: 0.5055 - accuracy: 0.9925 - val_loss: 0.6463 - val_accuracy: 0.8550\n",
      "Epoch 20/20\n",
      "57/57 [==============================] - 13s 233ms/step - loss: 0.5100 - accuracy: 0.9942 - val_loss: 0.6464 - val_accuracy: 0.8550\n"
     ]
    }
   ],
   "source": [
    "#TODO tester differents nb de tests (epochs)\n",
    "history = model.fit(xTrain, yTrain, \n",
    "                    validation_data=(xTest, yTest),\n",
    "                    epochs=20, verbose=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Dessiner l'évolution de l'erreur et de la pertinence"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "history_dict = history.history\n",
    "history_dict.keys()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Tracer l'erreur d'entrainement / l'erreur de validation\n",
    "- L'erreur d'entrainement et l'erreur de validation doivent suivre des courbes quasi parallèle.\n",
    "- L'erreur d'entrainement doit être plus faible que l'erreur de validation\n",
    "    - si les courbes divergent, si la courbe de validation s'éloigne de la courbe d'entrainement, il y a alors un sur-apprentissage (overfitting) trop adapté aux données d'entrainement, sans doute trop poussé (diminuer les epochs?).\n",
    "    - s'il y a une erreur d'entrainement trop grande, il y a sous-apprentissage"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "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": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "loss = history_dict['loss']\n",
    "val_loss = history_dict['val_loss']\n",
    "\n",
    "epochs = range(1, len(loss) + 1)\n",
    "\n",
    "# draw the loss evolution in blue\n",
    "plt.plot(epochs, loss, 'b-*', label='erreur sur exemples d\\'apprentissage')\n",
    "# draw the accuracy evolution in blue\n",
    "plt.plot(epochs, val_loss, 'r-*', label='erreur sur exemples de validation')\n",
    "plt.title('Erreur')\n",
    "plt.xlabel('Epochs')\n",
    "plt.ylabel('Erreur')\n",
    "plt.legend()\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "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": [
    "acc = history_dict['accuracy']\n",
    "val_acc = history_dict['val_accuracy']\n",
    "\n",
    "epochs = range(1, len(loss) + 1)\n",
    "\n",
    "# draw the loss evolution in blue\n",
    "plt.plot(epochs, acc, 'b-*', label='précision sur exemples d\\'apprentissage')\n",
    "# draw the accuracy evolution in blue\n",
    "plt.plot(epochs, val_acc, 'r-*', label='précision sur exemples de validation')\n",
    "plt.title('précision')\n",
    "plt.xlabel('Epochs')\n",
    "plt.ylabel('précision')\n",
    "plt.legend()\n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Evaluer le modèle sur l'exemple d'entrainement"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "KVnJJYscmH4v"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Précision sur exemples de test: 85.500002\n"
     ]
    }
   ],
   "source": [
    "# evaluate\n",
    "loss, acc = model.evaluate(xTest, yTest, verbose=0)\n",
    "print('Précision sur exemples de test: %f' % (acc*100))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "ObKC8oW7mROA"
   },
   "source": [
    "---\n",
    "## Estimation de nouvelles critiques\n",
    "Il ne reste plus qu'à utiliser le réseau pour faciliter la classification de futures revues ..  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "D3A6k30XmVxr"
   },
   "outputs": [],
   "source": [
    "def predict_sentiment(review, vocab, tokenizer, model):\n",
    "    \"\"\"classifie le texte dans la variable 'review' en positif (0) ou negatif (1) \"\"\"\n",
    "    # clean\n",
    "    tokens = clean_doc(review)\n",
    "    # filter by vocab\n",
    "    tokens = [w for w in tokens if w in vocab]\n",
    "    # convert to line\n",
    "    line = ' '.join(tokens)\n",
    "    # create the bag of words with the words used in the line\n",
    "    encoded = tokenizer.texts_to_matrix([line], mode='freq')\n",
    "    # prediction, output is a vector of (1x1) here (output of the neural network is a vector of 1 element)\n",
    "    output = model.predict(encoded, verbose=0)\n",
    "    estimation = output[0,0]\n",
    "    return estimation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "VslEKt3WmYuI"
   },
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'predict_sentiment' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m/var/folders/tg/4dn4gshd5ljd7xmsk_79fnpc0000gn/T/ipykernel_16149/2738244461.py\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m      1\u001b[0m \u001b[0;31m# test of a review associated with a note of 10/10\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      2\u001b[0m \u001b[0mtext\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m'When I first watched Daniel Craig 15 years ago, I was not ready for the direction they were taking the character. It felt like a good movie, but it felt quite divorced from the Bond I grew up with ... the Bond I knew. But apart from that effort being a Bond begins sort of thing (and this being the other end of that ... line), the franchise bit by bit gave us certain things ... we as viewers yearned for. While keeping it fresh. Bond ... James Bond! That line - one of many that are (in)famous in the long list of Bond movies - gets a new spin here. As is the iconic Bond shooting towards the camera. Now I could name many more things that this film revisits and I probably still would miss some. There are many references here, either visually or via audio. The beginning is quite the introduction ... it is almost funny because it takes its time. And it is funny because the rest sometimes feels rushed. There are many loose ends the movie wants to (neatly?) tie up. Another day, another Bond song, another Bond intro - although quite late in the game. Bond has grown - grown not just over the course of the movie Craig played him, but generally speaking. Society changes and Bond has no issue sitting on on the backseat of a motorbike, while a woman is driving. That is not enough for you to go on a rampage and scream \"wokeness\". Many known faces return, but many new ones too. Be it the ever excellent if only shortly used Ana de Armas (no pun intended) or our uber villain Rami Malek. Now while I really liked what I saw and while the running time is 2,5 hours ... there are things that can be seen and called flaws. You can question Bonds love to Eva Greens character ... or him still holding onto her, you can question methods used and how convenient certain things happen - or how some stunts would put the Fast and Furious franchise to shame ... but if you did that, you would derive yourself of a lot of fun. Not being aware of the previous entries may not be necessary, but youll have a lot of things to discover and smile at, if you have seen at least some of them. Of course if it is only a few of them, make it the Craig Bonds, so you have the whole ... story. Generally speaking I cant wait to revisit all the movies. No time to watch (them)? Well make time ;) Also to conclude with a fun \"fact\": the title No time to Die has been used before ... by another movie in 2006. The year Daniel Craig became Bond! Coincidence? Or full circle? All kidding aside, I have not seen that movie (yet), so dont hate on me if it isnt any good. Otherwise, out with the old ... in with the new - whoever that may be and however the franchise is going to reinvent itself ... Ill be there - I hope you will be there too.'\u001b[0m\u001b[0;31m#'Having read the book a hundred years ago, it took some time to begin to understand what was going on. If there is a criticism that many share, it is the pacing of the first part of the film. It needed something to set the scene for us, to show us who the strong and the weak were. I m not much for narration (show me, don t tell me), but it may not have been a bad idea. The scenes are magnificent with special effects off the chart. Many desert travel scenes are quite endless. How do you spruce up a desert. I thought the sand worms were a little one dimensional. I know there is another film down the road.'\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 3\u001b[0;31m \u001b[0mresultat\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mpredict_sentiment\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mtext\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mvocab\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtokenizer\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      4\u001b[0m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m'\\''\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0mtext\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m'\\' ------> est considere comme '\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'positif'\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0mresultat\u001b[0m\u001b[0;34m<\u001b[0m\u001b[0;36m0.5\u001b[0m \u001b[0;32melse\u001b[0m \u001b[0;34m'negatif'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      5\u001b[0m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mresultat\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mNameError\u001b[0m: name 'predict_sentiment' is not defined"
     ]
    }
   ],
   "source": [
    "# test of a review associated with a note of 10/10\n",
    "text = 'When I first watched Daniel Craig 15 years ago, I was not ready for the direction they were taking the character. It felt like a good movie, but it felt quite divorced from the Bond I grew up with ... the Bond I knew. But apart from that effort being a Bond begins sort of thing (and this being the other end of that ... line), the franchise bit by bit gave us certain things ... we as viewers yearned for. While keeping it fresh. Bond ... James Bond! That line - one of many that are (in)famous in the long list of Bond movies - gets a new spin here. As is the iconic Bond shooting towards the camera. Now I could name many more things that this film revisits and I probably still would miss some. There are many references here, either visually or via audio. The beginning is quite the introduction ... it is almost funny because it takes its time. And it is funny because the rest sometimes feels rushed. There are many loose ends the movie wants to (neatly?) tie up. Another day, another Bond song, another Bond intro - although quite late in the game. Bond has grown - grown not just over the course of the movie Craig played him, but generally speaking. Society changes and Bond has no issue sitting on on the backseat of a motorbike, while a woman is driving. That is not enough for you to go on a rampage and scream \"wokeness\". Many known faces return, but many new ones too. Be it the ever excellent if only shortly used Ana de Armas (no pun intended) or our uber villain Rami Malek. Now while I really liked what I saw and while the running time is 2,5 hours ... there are things that can be seen and called flaws. You can question Bonds love to Eva Greens character ... or him still holding onto her, you can question methods used and how convenient certain things happen - or how some stunts would put the Fast and Furious franchise to shame ... but if you did that, you would derive yourself of a lot of fun. Not being aware of the previous entries may not be necessary, but youll have a lot of things to discover and smile at, if you have seen at least some of them. Of course if it is only a few of them, make it the Craig Bonds, so you have the whole ... story. Generally speaking I cant wait to revisit all the movies. No time to watch (them)? Well make time ;) Also to conclude with a fun \"fact\": the title No time to Die has been used before ... by another movie in 2006. The year Daniel Craig became Bond! Coincidence? Or full circle? All kidding aside, I have not seen that movie (yet), so dont hate on me if it isnt any good. Otherwise, out with the old ... in with the new - whoever that may be and however the franchise is going to reinvent itself ... Ill be there - I hope you will be there too.'#'Having read the book a hundred years ago, it took some time to begin to understand what was going on. If there is a criticism that many share, it is the pacing of the first part of the film. It needed something to set the scene for us, to show us who the strong and the weak were. I m not much for narration (show me, don t tell me), but it may not have been a bad idea. The scenes are magnificent with special effects off the chart. Many desert travel scenes are quite endless. How do you spruce up a desert. I thought the sand worms were a little one dimensional. I know there is another film down the road.'\n",
    "resultat = predict_sentiment(text, vocab, tokenizer, model)\n",
    "print('\\'',text,'\\' ------> est considere comme ', 'positif' if resultat<0.5 else 'negatif')\n",
    "print(resultat)\n",
    "print(\"=-=-=-=-=-=-=-=-=-=-=-=\")\n",
    "# test of a review associated with a note of 1/10\n",
    "text = 'I was so angry after watching that I had to write something. Boring boring boring, the biggest reason for that is scenario 0/10. It is too long, if I hadn\\'t go to cinema and bought ticket there is no way I would have patience to watch it until the end. Everything already seen in previous marvel movies, it seems that every new is even worse with lower quality and I didn\\'t even have big expectations before the movie. The main reason I decided to watch it is IMDB rate 9/10 which is total fraud and paid advertisement. There is no guarantee if some character is dead that he will remain dead, everyone could resurrect...Go watch anything else than this. My last marvel movie for sure, so angry that I don\\'t watch american movies in cinema any more, all the same, low quality in recent time, just to make many fast, without proper scenario...bull..it'\n",
    "resultat = predict_sentiment(text, vocab, tokenizer, model)\n",
    "print('\\'',text,'\\' ------> est considere comme ', 'positif' if resultat<0.5 else 'negatif')\n",
    "print(resultat)\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---\n",
    "---\n",
    "## TRAVAIL A RENDRE\n",
    "\n",
    "* Vous devrez tester différentes architectures de réseau, ainsi que différentes configuration (fonctions d'activation, méthodes de correction d'erreur, méthode de calcul de l'erreur, ...)\n",
    "* Vous enverrez un fichier contenant pour chaque définition de réseau :\n",
    "    * l'architecture du réseau (un copier coller du bloc 'TODO: définir la structure du réseau')\n",
    "    * le déroulé de l'apprentissage (un copier coller du bloc résultat de 'TODO tester differents nb de tests (epochs)'\n",
    "    * le résultat de 'Test Accuracy'\n",
    "\n",
    "Au minimum 3 différents réseaux devront être fournis"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "colab": {
   "collapsed_sections": [],
   "name": "Copie de TP_DeepLearningM1TNSI.ipynb",
   "provenance": [],
   "toc_visible": true,
   "version": "0.3.2"
  },
  "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": "3.9.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}