clasificando-numeros.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'2.0.8'"
      ]
     },
     "execution_count": 51,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import keras\n",
    "keras.__version__"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "# Introducción a las convnets: Clasificando números\n",
    "\n",
    "Esta libreta contiene un ejemplo sencillo que podeis encontrar con mas detalle en el capítulo 5 del libro [Deep Learning with Python](https://www.manning.com/books/deep-learning-with-python?a_aid=keras&a_bid=76564dff).\n",
    "\n",
    "----\n",
    "\n",
    "Vamos a echarle un vistazo a un ejemplo sencillo de una convnet. La utilizaremos para clasificar el dataset MNIST, que es un dataset abierto que contiene números escritos a mano. \n",
    "\n",
    "![Números escritos a mano del dataset MNIST](http://corochann.com/wp-content/uploads/2017/02/mnist_plot.png)\n",
    "\n",
    "Las 6 líneas de código que puedes ver aqui debajo muestran que aspecto tiene una convnet basica. Es una pila de capas `Conv2D` y `MaxPooling2D`. Veremos en un rato más concretamente lo que hacen. \n",
    "Lo importante es notar que una convnet toma como input tensores de tamaño `(altura_imagen, anchura_imagen, canales_imagen)`. \n",
    "En nuestro caso, vamos a configurar nuestra convnet para procesar inputs de tamaño `(28, 28, 1)`,  que es el formato de las imágenes MNIST. \n",
    "Por eso le pasamos el argumento input_shape=(28, 28, 1)` a nuestra primera capa."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras import layers\n",
    "from keras import models\n",
    "\n",
    "model = models.Sequential()\n",
    "model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))\n",
    "model.add(layers.MaxPooling2D((2, 2)))\n",
    "model.add(layers.Conv2D(64, (3, 3), activation='relu'))\n",
    "model.add(layers.MaxPooling2D((2, 2)))\n",
    "model.add(layers.Conv2D(64, (3, 3), activation='relu'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Vamos a ver que pinta tiene nuestra arquitectura hasta ahora:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_7 (Conv2D)            (None, 26, 26, 32)        320       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_5 (MaxPooling2 (None, 13, 13, 32)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_8 (Conv2D)            (None, 11, 11, 64)        18496     \n",
      "_________________________________________________________________\n",
      "max_pooling2d_6 (MaxPooling2 (None, 5, 5, 64)          0         \n",
      "_________________________________________________________________\n",
      "conv2d_9 (Conv2D)            (None, 3, 3, 64)          36928     \n",
      "=================================================================\n",
      "Total params: 55,744\n",
      "Trainable params: 55,744\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "raw",
   "metadata": {},
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "Puedes ver arriba que la salida de cada capa `Conv2D` y `MaxPooling2D` es un tensor 3D de dimensiones `(altura, anchura, canales)`. La anchura y la altura tienden a diminuir según vamos yendo mas profundo en la red. El número de canales está controlado por el primer argumento que se le pasa a \n",
    "las capas `Conv2D`  (e.j. 32 o 64).\n",
    "\n",
    "El siguiente paso sería darle nuestro ultimo tensor (de dimensiones `(3, 3, 64)`) como entrada a una red densamente conectada. \n",
    "Estos clasificadores procesan vectores, que son 1D,  mientras que nuestra salida es un tensor 3D. \n",
    "Así que primero tendremos que aplanar nuestra salida 3D y convertirla en 1D y después añadir unas cuantas capas densas:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.add(layers.Flatten())\n",
    "model.add(layers.Dense(64, activation='relu'))\n",
    "model.add(layers.Dense(10, activation='softmax'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Queremos clasificar 10 categorías, lo que significa que nuestra capa final debe tener 10 nodos y una función de activación softmax. Vamos a ver que pinta tiene nuestra red:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv2d_7 (Conv2D)            (None, 26, 26, 32)        320       \n",
      "_________________________________________________________________\n",
      "max_pooling2d_5 (MaxPooling2 (None, 13, 13, 32)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_8 (Conv2D)            (None, 11, 11, 64)        18496     \n",
      "_________________________________________________________________\n",
      "max_pooling2d_6 (MaxPooling2 (None, 5, 5, 64)          0         \n",
      "_________________________________________________________________\n",
      "conv2d_9 (Conv2D)            (None, 3, 3, 64)          36928     \n",
      "_________________________________________________________________\n",
      "flatten_3 (Flatten)          (None, 576)               0         \n",
      "_________________________________________________________________\n",
      "dense_5 (Dense)              (None, 64)                36928     \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 10)                650       \n",
      "=================================================================\n",
      "Total params: 93,322\n",
      "Trainable params: 93,322\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Como puedes ver, nuestra salida de dimension `(3, 3, 64)` han sido aplanadas hasta convertirse en vectores de dimensión `(576,)`, antes de entrar en las dos capas densas.\n",
    "\n",
    "Vamos ahora a entrenar nuestra red con las imágenes del dataset MNIST. \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Shape del dataset de training antes de preprocesarlo:  (60000, 28, 28) \n",
      "\n",
      "Estas son las imágenes (matrices) de training: \n",
      " \n",
      " [[[0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  ..., \n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]]\n",
      "\n",
      " [[0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  ..., \n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]]\n",
      "\n",
      " [[0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  ..., \n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]]\n",
      "\n",
      " ..., \n",
      " [[0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  ..., \n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]]\n",
      "\n",
      " [[0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  ..., \n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]]\n",
      "\n",
      " [[0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  ..., \n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]\n",
      "  [0 0 0 ..., 0 0 0]]] \n",
      "\n",
      "Estas son las etiquetas de training:  [5 0 4 ..., 5 6 8] \n",
      "\n"
     ]
    },
    {
     "data": {
      "image/jpeg": "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\n",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 56,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from keras.datasets import mnist\n",
    "from keras.utils import to_categorical\n",
    "\n",
    "(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\n",
    "\n",
    "\n",
    "index=0\n",
    "prueba=train_images[index][:][:]\n",
    "\n",
    "import scipy.misc\n",
    "scipy.misc.imsave('numbers/input.jpg', prueba)\n",
    "\n",
    "\n",
    "print(\"Shape del dataset de training antes de preprocesarlo: \", train_images.shape, \"\\n\")\n",
    "print (\"Estas son las imágenes (matrices) de training: \\n \\n\", train_images, \"\\n\")\n",
    "print (\"Estas son las etiquetas de training: \", train_labels, \"\\n\")\n",
    "\n",
    "\n",
    "\n",
    "from IPython.display import Image\n",
    "Image(\"numbers/input.jpg\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Shape del dataset de training después del preprocesado:  (60000, 28, 28, 1)\n"
     ]
    }
   ],
   "source": [
    "train_images = train_images.reshape((60000, 28, 28, 1))\n",
    "train_images = train_images.astype('float32') / 255\n",
    "\n",
    "test_images = test_images.reshape((10000, 28, 28, 1))\n",
    "test_images = test_images.astype('float32') / 255\n",
    "\n",
    "train_labels = to_categorical(train_labels)\n",
    "test_labels = to_categorical(test_labels)\n",
    "\n",
    "print(\"Shape del dataset de training después del preprocesado: \", train_images.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/5\n",
      "60000/60000 [==============================] - 43s - loss: 0.1681 - acc: 0.9481    \n",
      "Epoch 2/5\n",
      "60000/60000 [==============================] - 43s - loss: 0.0480 - acc: 0.9848    \n",
      "Epoch 3/5\n",
      "60000/60000 [==============================] - 43s - loss: 0.0329 - acc: 0.9900    \n",
      "Epoch 4/5\n",
      "60000/60000 [==============================] - 43s - loss: 0.0255 - acc: 0.9924    \n",
      "Epoch 5/5\n",
      "60000/60000 [==============================] - 43s - loss: 0.0197 - acc: 0.9940    \n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7ff97b6d6898>"
      ]
     },
     "execution_count": 58,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.compile(optimizer='rmsprop',\n",
    "              loss='categorical_crossentropy',\n",
    "              metrics=['accuracy'])\n",
    "model.fit(train_images, train_labels, epochs=5, batch_size=64)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Vamos a evaluar el modelo con las imágenes de test:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      " 9952/10000 [============================>.] - ETA: 0s"
     ]
    }
   ],
   "source": [
    "test_loss, test_acc = model.evaluate(test_images, test_labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.99139999999999995"
      ]
     },
     "execution_count": 60,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_acc"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "La precisión de nuestro test con una convnet basica es de 99.1%. ¡No está mal! \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 63,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "image shape (28, 28, 1)\n",
      "x shape (1, 28, 28, 1)\n",
      "1/1 [==============================] - 0s\n",
      "La imagen representa el número:  [7]\n"
     ]
    }
   ],
   "source": [
    "from keras.preprocessing import image\n",
    "import numpy as np\n",
    "from matplotlib import pyplot as plt\n",
    "\n",
    "img_width=28\n",
    "img_height=28\n",
    "\n",
    "img = image.load_img('siete.png', target_size=(img_width, img_height),grayscale=True)\n",
    "x= image.img_to_array(img)\n",
    "print(\"image shape\", x.shape)\n",
    "x = np.expand_dims(x, axis=0)\n",
    "print(\"x shape\", x.shape)\n",
    "\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save('net_numbers.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 73,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1/1 [==============================] - 0s\n",
      "La imagen representa el número:  [7]\n"
     ]
    }
   ],
   "source": [
    "from keras.models import load_model\n",
    "newmodel = load_model('net_numbers.h5')\n",
    "print(\"La imagen representa el número: \", newmodel.predict_classes(x))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "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": 2
}