1.5 NLP com Keras.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "XJoQfDH89FTd"
   },
   "source": [
    "Neste módulo, veremos algumas considerações sobre como aplicar os conceitos vistos até agora a aplicações de processamento de linguagem natural.\n",
    "\n",
    "Como nosso primeiro exemplo, veremos como fazer um modelo simples para aprender se uma critica de um filme é positiva ou negativa, usando uma coleção de referência do site IMDb.\n",
    "\n",
    "## Pré-processando os dados\n",
    "\n",
    "Até agora trabalhamos com imagens, que possuem várias vantagens: tamanho fixo, formato padronizado (3 cores), bases de dados relativamente livres de ruído.\n",
    "\n",
    "Não podemos assumir o mesmo de bases de texto. Documentos de texto possuem diversas particularidades, como um universo muito grande de palavras possíveis, documentos com tamanhos variáveis, lixo (como erros de digitação), etc.\n",
    "\n",
    "Por isto, antes de se trabalhar com texto, se deve fazer alguns passos para tornar o mesmo mais _limpo_.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "DHu_40N99FTf",
    "outputId": "92f240a0-9f4f-4523-bb93-e63b6e8287e2"
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "E:\\Anaconda\\lib\\site-packages\\h5py\\__init__.py:34: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import keras\n",
    "from keras.datasets import imdb\n",
    "from keras.preprocessing.sequence import pad_sequences\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Flatten, Dropout, Activation\n",
    "from keras.layers import Embedding\n",
    "from keras.callbacks import ModelCheckpoint\n",
    "import os\n",
    "import sklearn.metrics\n",
    "from sklearn.metrics import roc_auc_score\n",
    "import pandas as pd\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "collapsed": true,
    "id": "AY7U46Ix9FTm",
    "outputId": "86dd5b72-ff6e-4a4e-9eec-329e9519df9b"
   },
   "outputs": [],
   "source": [
    "output_dir = 'model_output/imdb_deep_net'\n",
    "epochs = 4\n",
    "batch_size = 128\n",
    "n_dim = 64\n",
    "n_unique_words = 5000\n",
    "n_words_to_skip = 50\n",
    "max_review_length = 100\n",
    "pad_type = trunc_type = 'pre' #Outra opcao eh post\n",
    "n_dense = 64\n",
    "dropout = 0.5\n",
    "\n",
    "(X_train, y_train), (X_valid, y_valid) = imdb.load_data(num_words=n_unique_words, skip_top=n_words_to_skip)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "QzZUQ7TK9FTr"
   },
   "source": [
    "Para o nosso treinamento, iremos usar os hiperparâmetros acima. Alguns vocês já conhecem, mas temos alguns específicos de texto:\n",
    "\n",
    "- n_unique_words: o tamanho do vocabulário a ser considerado\n",
    "- n_words_to_skio: As n palavras mais frequentes serão ignoradas (stopwords)\n",
    "- max_review_length: Por enquanto, trabalharemos com documentos de tamanho fixo (100 palavras)\n",
    "\n",
    "\n",
    "Vamos dar uma olhada em como estão os documentos por enquanto:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "collapsed": true,
    "id": "sql1pmIj9FTs",
    "outputId": "0e99e486-f969-4470-836b-474ea7ccd577"
   },
   "outputs": [],
   "source": [
    "for x in X_train[0:6]:\n",
    "    print(len(x))    \n",
    "\n",
    "print(X_train[5])\n",
    "\n",
    "for y in y_train[0:6]:\n",
    "    print(y)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "WkYArl0O9FTw"
   },
   "source": [
    "### Visualizando documentos ###\n",
    "\n",
    "Estranho né? Antes de tudo, seria bom descobrir que documento é esse, e transformar esses IDs em palavras:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "collapsed": true,
    "id": "NEsIi4Xx9FTy",
    "outputId": "5b1dcf39-7bbb-439d-ef93-7043d2f53e5c"
   },
   "outputs": [],
   "source": [
    "word_index = keras.datasets.imdb.get_word_index()\n",
    "# Vamos empurrar esse dicionario 3 posicoes\n",
    "word_index = {k : (v+3) for k,v in word_index.items()}\n",
    "# Vamos adicionar essas tres entradas padrao nos indices corretos\n",
    "word_index['PAD'] = 0 #Padding para documentos com menos de 100 palavras\n",
    "word_index['START'] = 1 #Inicio do documento\n",
    "word_index['UNK'] = 2 #Termo \"desconhecido\"\n",
    "index_word = {v:k for k,v in word_index.items()}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "FqrdRHTj9FT2",
    "outputId": "88af82c0-840f-4618-e89f-0275550e10f6"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "UNK worst mistake UNK my life UNK UNK UNK picked UNK UNK up UNK target UNK 5 because UNK figured hey UNK sandler UNK can get UNK cheap laughs UNK UNK wrong completely wrong mid way through UNK UNK UNK three UNK my friends were asleep UNK UNK UNK still suffering worst plot worst script worst UNK UNK UNK ever seen UNK wanted UNK hit my head up against UNK wall UNK UNK hour then i'd stop UNK UNK know why because UNK felt damn good upon UNK my head UNK UNK stuck UNK damn UNK UNK UNK UNK UNK watched UNK burn UNK UNK felt better than anything else i've ever done UNK took american psycho army UNK darkness UNK kill bill UNK UNK get over UNK crap UNK hate UNK sandler UNK actually going through UNK UNK UNK UNK UNK whole day UNK my life\n"
     ]
    }
   ],
   "source": [
    "review = ' '.join(index_word[id] for id in X_train[4])\n",
    "print(review)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "sG6V5nkg9FT6"
   },
   "source": [
    "Vamos comparar com o conteúdo original:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "hgwlYuRB9FT6",
    "outputId": "73dd77dc-dcf8-485a-bddb-8d47a8b1e39b"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "START begins better than it ends funny that the russian submarine crew outperforms all other actors it's like those scenes where documentary shots br br spoiler part the message dechifered was contrary to the whole story it just does not mesh br br\n"
     ]
    }
   ],
   "source": [
    "(all_X_train, _), (all_X_valid, _) = imdb.load_data()\n",
    "full_review = ' '.join(index_word[id] for id in all_X_train[5])\n",
    "print(full_review)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "r_RlNm8q9FT_"
   },
   "source": [
    "Os documentos continuam com tamanhos variáveis. Vamos agora normalizá-los para um tamanho padrão, truncando ou fazendo padding conforme necessário:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "collapsed": true,
    "id": "az-0k-io9FUB"
   },
   "outputs": [],
   "source": [
    "X_train = pad_sequences(X_train, maxlen=max_review_length, padding=pad_type, truncating=trunc_type, value=0)\n",
    "X_valid = pad_sequences(X_valid, maxlen=max_review_length, padding=pad_type, truncating=trunc_type, value=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "F78mG0eZ9FUE",
    "outputId": "54daf5db-dff2-4d8e-ae94-b6901c9156d7"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "100\n",
      "100\n",
      "100\n",
      "100\n",
      "100\n",
      "100\n"
     ]
    }
   ],
   "source": [
    "for x in X_train[0:6]:\n",
    "    print(len(x))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "Jxm4FlWn9FUH",
    "outputId": "d4aa0a11-8b41-451b-89f0-0100b06010a2"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "UNK UNK still suffering worst plot worst script worst UNK UNK UNK ever seen UNK wanted UNK hit my head up against UNK wall UNK UNK hour then i'd stop UNK UNK know why because UNK felt damn good upon UNK my head UNK UNK stuck UNK damn UNK UNK UNK UNK UNK watched UNK burn UNK UNK felt better than anything else i've ever done UNK took american psycho army UNK darkness UNK kill bill UNK UNK get over UNK crap UNK hate UNK sandler UNK actually going through UNK UNK UNK UNK UNK whole day UNK my life\n"
     ]
    }
   ],
   "source": [
    "review = ' '.join(index_word[id] for id in X_train[4])\n",
    "print(review)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "9KlDhpau9FUO"
   },
   "source": [
    "Como vocês podem ver, o padding (e a truncagem quando necessário) são feitos no início. Você pode alterar isto alterando os parâmetros `pad_type` e `trunc_type`."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "SAU1pS9j9FUP"
   },
   "source": [
    "## Treinando um classificador para análise de sentimentos ##\n",
    "\n",
    "Vamos agora fazer uma FNN simples para aprender a classificar as críticas como positivas ou negativas:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "46l5cC5D9FUQ",
    "outputId": "8a58b984-ba28-431f-c513-cc7755390feb"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "embedding_1 (Embedding)      (None, 100, 64)           320000    \n",
      "_________________________________________________________________\n",
      "flatten_1 (Flatten)          (None, 6400)              0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 64)                409664    \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 64)                0         \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 1)                 65        \n",
      "=================================================================\n",
      "Total params: 729,729\n",
      "Trainable params: 729,729\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = Sequential()\n",
    "model.add(Embedding(n_unique_words, n_dim, \n",
    "                    input_length=max_review_length))\n",
    "model.add(Flatten())\n",
    "model.add(Dense(n_dense, activation='relu'))\n",
    "model.add(Dropout(dropout))\n",
    "model.add(Dense(1, activation='sigmoid'))\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "yYDnJev79FUV"
   },
   "source": [
    "Temos aqui um novo tipo de camada, a camada de embedding. Esta é uma camada que recebe como entrada um vetor de id de palavras com tamanho igual a `max_review_length` e contendo até no máximo `n_unique_words` ids diferentes, e transforma cada uma dessas palavras em um vetor com `n_dim` dimensões.\n",
    "\n",
    "Desta forma, esta camada está criando uma representação vetorial do que é uma palavra, e representa cada documento como cem vetores de palavras.\n",
    "\n",
    "Em uma aula futura, entenderemos melhor como são criados e como funcionam estes embeddings, mas por enquanto o importante é entender que eles são uma forma de representar palavras que faz mais sentido para as redes neurais."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "ie-fcDrU9FUW",
    "outputId": "19ea97bf-c21a-4b6e-fcf5-2e7b5ddd39c3"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 20000 samples, validate on 5000 samples\n",
      "Epoch 1/4\n",
      "20000/20000 [==============================] - 1s 66us/step - loss: 0.6092 - acc: 0.6433 - val_loss: 0.3848 - val_acc: 0.8334\n",
      "Epoch 2/4\n",
      "20000/20000 [==============================] - 1s 42us/step - loss: 0.3076 - acc: 0.8750 - val_loss: 0.3577 - val_acc: 0.8408\n",
      "Epoch 3/4\n",
      "20000/20000 [==============================] - 1s 42us/step - loss: 0.1443 - acc: 0.9553 - val_loss: 0.3931 - val_acc: 0.8420\n",
      "Epoch 4/4\n",
      "20000/20000 [==============================] - 1s 42us/step - loss: 0.0349 - acc: 0.9938 - val_loss: 0.4829 - val_acc: 0.8352\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x1f48e3454e0>"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tboard = keras.callbacks.TensorBoard(log_dir='./logs', histogram_freq=0, batch_size=batch_size, write_graph=True, write_grads=False, write_images=False, embeddings_freq=0, embeddings_layer_names='embedding_1', embeddings_metadata=None, embeddings_data=None)\n",
    "\n",
    "model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n",
    "model.fit(X_train, y_train, batch_size=batch_size, epochs=epochs, verbose=1, validation_split=0.2, callbacks=[tboard])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "dbyYzD2_9FUe"
   },
   "source": [
    "Por enquanto não usamos o conjunto de validação que veio com a base de dados, mas sim 20% do treino. Fizemos isto pois vamos utilizar o conjunto de validacao como teste:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "Ub8bC0d49FUe",
    "outputId": "9c865134-dc74-40ce-a51d-cecc61348a80"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "91.55\n"
     ]
    }
   ],
   "source": [
    "y_hat = model.predict_proba(X_valid)\n",
    "\n",
    "pct_auc = roc_auc_score(y_valid, y_hat) * 100\n",
    "print('{:0.2f}'.format(pct_auc))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "NP5Kzkkn9FUi"
   },
   "source": [
    "Estou mostrando a curva ROC AUC (média sob o recebedor). Ela leva tanto em consideração a taxa verdadeiros positivos quanto a taxa de falso positivos:\n",
    "\n",
    "$$\n",
    "T_{V+} = \\frac{V_{+}}{ (V_{+} + F_{-})} \\\\\n",
    "T_{F+} = \\frac{F_{+}} {(F_{+} + V_{-})}\n",
    "$$"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "16eVuKd49FUj",
    "outputId": "78f20f10-90f8-4d6c-c0a4-fafc7bf92b05"
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAEWCAYAAAB42tAoAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nx6mT/wMb7O6ne5kpRQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1f483173e48>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fpr, tpr, _ = sklearn.metrics.roc_curve(y_valid, y_hat)\n",
    "roc_auc = sklearn.metrics.auc(fpr, tpr)\n",
    "plt.figure()\n",
    "lw = 2\n",
    "plt.plot(fpr, tpr, color='darkorange',\n",
    "         lw=lw, label='ROC curve (area = %0.2f)' % roc_auc)\n",
    "plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')\n",
    "plt.xlim([0.0, 1.0])\n",
    "plt.ylim([0.0, 1.05])\n",
    "plt.xlabel('Taxa de Falsos Positivos')\n",
    "plt.ylabel('Taxa de Verdadeiros Positivos')\n",
    "plt.title('Receiver Operating Characteristic (ROC)')\n",
    "plt.legend(loc=\"lower right\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "0gAS1H4v9FUm",
    "outputId": "9eb5a44c-646c-4c98-a953-9e512ca4c331"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "   previsto  real\n",
      "0  0.126988     0\n",
      "1  0.999717     1\n",
      "2  0.974634     1\n",
      "3  0.365254     0\n",
      "4  0.999940     1\n",
      "5  0.952179     1\n",
      "6  0.978445     1\n",
      "7  0.002977     0\n",
      "8  0.384830     0\n",
      "9  0.910281     1\n"
     ]
    }
   ],
   "source": [
    "# Create a Pandas DataFrame to hold y_hat and y values.\n",
    "ydf = pd.DataFrame(list(zip(y_hat[:,0], y_valid)),\n",
    "                   columns=['previsto', 'real'])\n",
    "# Print the first 10 rows.\n",
    "print(ydf.head(10))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "PEquyyeN9FUs",
    "outputId": "ae57777a-d70a-43cb-b736-c5144b7f3212"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "START this film requires a lot of patience because it focuses on mood and character development the plot is very simple and many of the scenes take place on the same set in frances austen's the sandy dennis character apartment but the film builds to a disturbing climax br br the characters create an atmosphere rife with sexual tension and psychological trickery it's very interesting that robert altman directed this considering the style and structure of his other films still the trademark altman audio style is evident here and there i think what really makes this film work is the brilliant performance by sandy dennis it's definitely one of her darker characters but she plays it so perfectly and convincingly that it's scary michael burns does a good job as the mute young man regular altman player michael murphy has a small part the solemn moody set fits the content of the story very well in short this movie is a powerful study of loneliness sexual repression and desperation be patient soak up the atmosphere and pay attention to the wonderfully written script br br i praise robert altman this is one of his many films that deals with unconventional fascinating subject matter this film is disturbing but it's sincere and it's sure to elicit a strong emotional response from the viewer if you want to see an unusual film some might even say bizarre this is worth the time br br unfortunately it's very difficult to find in video stores you may have to buy it off the internet\n"
     ]
    }
   ],
   "source": [
    "# Read a full review. In that case, the 1st one.\n",
    "index = 0\n",
    "review = ' '.join(index_word[id] for id in all_X_valid[1])\n",
    "print(review)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "YCVFwFqg9FUy"
   },
   "source": [
    "Esta é baseado no material de <a href=\"\" >Wilder Rodrigues"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "name": "1.5 NLP com Keras.ipynb",
   "provenance": [],
   "version": "0.3.2"
  },
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}