seno.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Red neuronal de $\\sin{x}$\n",
    "\n",
    "En este cuaderno pretendemos entrenar una red neuronal para que saque el seno de la entrada. Para ello hemos escogio una aruiqtectura muy pequeña, de una sóla capa oculta de tres neuronas, tal y como podemos ver en la siguiente figura:\n",
    "\n",
    "<img src=\"img/nn_1-3-1.png\"/>\n",
    "\n",
    "Veamos cómo debemos proceder para poder realizarla con Pytorch.\n",
    "\n",
    "## Librerías\n",
    "\n",
    "Vamos a usar diferentes librerías. Veamos a que dedicamos cada una de ellas:\n",
    "\n",
    "- `torch` y demás variantes: Es la librería que contiene todas las funciones necesarias para virtualizar y entrenar la red. Su módulo `nn` tiene todo lo necesario para construir el modelo. En `nn.funcional` encontramos además un conjunto de funciones para calcular el coste de la red.\n",
    "\n",
    "- `matplotlib.pyplot` para la representación gráfica.\n",
    "\n",
    "- `math` para obtener el valor de $\\pi$\n",
    "\n",
    "- `IPython.display` para obtener el progreso por pantalla del entrenamiento\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "import matplotlib.pyplot as plt\n",
    "from math import pi\n",
    "from IPython.display import display, clear_output"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Parámetros\n",
    "\n",
    "En esta sección vamos a definir los parámetros para configurar la arquitectura de red y los parámtros que intervienen en la simulación.\n",
    "\n",
    "### Parámetros de arquitectura\n",
    "\n",
    "- $D_{in}$, el número de entradas de la red. Deberá  ser igual a 1.\n",
    "- $H$, el número de unidades en la capa oculta. Este será 3.\n",
    "- $D_{out}$, número de salidas de la red. Sólo tendremos una salida en este caso.\n",
    "\n",
    "### Parámetros de entrenamiento\n",
    "\n",
    "- $N$, el número de datos que generamos para el entrenamiento de la red.\n",
    "- $\\delta_{limit}$, `delta_limit` en el código. Cuando el coste cambie menos de este valor durante 100 itearaciones del _training set_ dejaremos de entrenar la red.\n",
    "- $t_{max}$, `epochs` en el código. Es otra manera de limitar el entrenamiento. Aprenderá hasta que alcancemos este número de iteracciones en el código.\n",
    "- $\\alpha$, `learning_rate` en el código. Factor de aprendizaje de la red en cada paso de entrenamiento."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# D_in numero de entradas;\n",
    "# H numero de neuronas en la capa oculta;\n",
    "#D_out numero de salidas;\n",
    "D_in, H, D_out = 1, 3, 1\n",
    "\n",
    "#N el tamaño del training set; \n",
    "#delta_limit minimo cambio en coste con el que parar la simulación;\n",
    "#epochs numero maximo de iteracciones si estamos con el for;\n",
    "N, delta_limit, epochs, learning_rate = 500, 1e-4, 100000, 1e-4"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Generamos el _training set_\n",
    "\n",
    "Debemos generar el _training set_. Dado que queremos que la red nos devuelva $\\sin{x}$ bastará con tener datos de $x$ entre $-\\pi$ y $+\\pi$. Los generamos aleatoriamente con la función `torch.rand()`, que ya nos los devuelve como un tensor de `torch`. Para obtener los datos objetivo (_target_) usamos la función de `torch`, `torch.sin()` que acepta tensores y devuelve tensores. Representamos los datos con `plt.plot()`. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "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": [
    "# Creamos N datos aleatorios entre -pi y pi y calculamos su seno\n",
    "x = 2*pi*torch.rand(N, D_in) - pi\n",
    "y = torch.sin(x)\n",
    "\n",
    "#Representamos graficamente\n",
    "plt.plot(x, y, '+', label='Training set')\n",
    "plt.title('Datos a entrenar')\n",
    "plt.xlabel('$x$')\n",
    "plt.ylabel('$\\sin(x)$')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Creamos la red neuronal\n",
    "\n",
    "En esta parte vamos a definir la aruitectura de red con los parámetros previamente definidos. Para ello vamos a crear una clase de Python tomando como base el módulo `torch.nn.Module` de PyTorch. En la primera parte, el _constructor_ (`__init__`), se define el tipo de conexión entre capas. en nuestro caso, que queremos que sea _fully connected_ lo hacemos usando `nn.Linear()`cuyos argumentos son las dimensiones de la capa anterior y de la posterior. Por defecto se define para que esa capa tenga _bias_. En la segunda parte se define la función `forward`. En ésta se define cómo se va a proceder a la hora de realizar la propagación de la señal. Permite conectar las capas definidas en el _constructor_ y definir cuáles son las funciones de activación de las neuronas. Para estas funciones de activación usamos el subpaquete `torch.nn.Functional` que hemos abierto como `F`. En nuestro caso hemo elegido usar la función sigmoidea, $f(x) = \\frac{1}{1+e^{-x}}$. Finalmente, para inicializar el modelo basta con asignarlo a un objeto, en este caso `model`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "class TwoLayerNet(torch.nn.Module):\n",
    "    def __init__(self, D_in, H, D_out):\n",
    "        \"\"\"\n",
    "        En el 'constructor' inicializamos las capas dandoles \n",
    "        las dimensiones de entrada y salida de cada una de ellas \n",
    "        y las asignamos como variables de la clase\n",
    "        \"\"\"\n",
    "        super(TwoLayerNet, self).__init__()\n",
    "        self.linear1 = nn.Linear(D_in, H)\n",
    "        self.linear2 = nn.Linear(H, D_out)\n",
    "\n",
    "    def forward(self, x):\n",
    "        \"\"\"\n",
    "        En la función de propagación, 'forward', recogemos el\n",
    "        tensor de entrada y configuramos como es su propagación\n",
    "        hasta la salida. Podemos usar las capas definidas en el\n",
    "        'constructor' o realizar cualquier operación sobre él.\n",
    "        En este caso colocamos la función de activación de las\n",
    "        neuronas de la capa oculta. A la salida no hay función\n",
    "        de activación porque queremos obtener el seno.\n",
    "        \"\"\"\n",
    "        x = self.linear1(x)\n",
    "        x = self.linear2(F.sigmoid(x))\n",
    "        return x\n",
    "\n",
    "# Inicializamos el modelo\n",
    "model = TwoLayerNet(D_in, H, D_out)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Funciones de coste y optimizador\n",
    "\n",
    "En esta parte vamos a aprovechar el músculo que nos ofrece la librería `torch`. Recoge un montón de posibilidades tanto a la hora de calcular el coste, $J$, de la red como para optimizar el entrenamiento de ésta. Para el coste vamos a elegir una de las más habituales, la de la suma de las diferencias al cuadrado, `torch.nn.MSELoss(reduction='sum')` aunque PyTorch cuenta con muchas [mas](https://pytorch.org/docs/stable/nn.html#loss-functions).\n",
    "\n",
    "Por otro lado tenemos el optmizador, el cual nos engloba la parte del aprendizaje de la red. Podemos elegir diferentes algoritmos entre [los que ofrece PyTorch](https://pytorch.org/docs/stable/optim.html#algorithms), en  nuestro caso hemos elegido el grandiente descendiente estocástico (`torch.optim.SGD()`). "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Declaramos la función de coste y el optimizador. A este\n",
    "#último le debemos pasar los parámetros del modelo para \n",
    "#que los actualice en cada paso de aprendizaje.\n",
    "criterion = torch.nn.MSELoss(reduction='sum')\n",
    "optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Aprendizaje\n",
    "\n",
    "En esta fase vamos a ver como se da el aprendizaje de la red. Una de las primeras cosas que debemos elegir es cómo cortar este aprendijaze. Dado que vamos a entrenar la red sobre el mismo _training set_ varias veces (del orden de unas $10^5$ por lo menos), es bueno que sepamos cuando es hora de parar. La forma más bruta pero también la más basica es utilizar un bucle hasta que se llegue al número máximo de iteraciones que se pedia en los parámetros del principio. Esto se podría hacer con una línea de bucle como `for t in range(epochs):`. Pero nosotros queremos ser más sutiles. En nuestro caso hemos optado por definir un indicador de cómo de bien la red está aprendiendo. Para ello cada 100 iteracciones de aprendizaje sobre el _training set_ vamos a calular $\\delta = J_{t-100} - J_t$. Donde $J_t$ es el coste de la red en la iteracción $t$. Como hemos definido en los parámetros de más arriba uno que era $\\delta_{limit}$ cuando $\\delta$ baje de este valor, pararemos el apredizaje.\n",
    "\n",
    "Así pues, en las primeras líneas de código inicializamos varias variables que nos permiten entrar en esta dinámica para calcular las $\\delta$. A partir de la línea del `while` entramos en la dinámica del aprendizaje. Lo primero que se hace es calcular las predicciones del modelo para las entradas, `y_pred = model(x)`. Como se puede ver se calculan todas de golpe, en vez de una a una, como podría pensarse a priori. Una vez obtenidas las predicciones, obtenemos el coste `loss`con el critetio seleccionado en la sección de más arriba. A continuación encontramos un condicional que nos permite sacar información de vez en cuando y nos limita la evaluación de $\\delta$ cada 100 iteracciones. Después ponemos a cero el contador de gradientes con la instancia `optimizer.zero_grad()`, algo necesario porque no se puede hacer dos retropropagaciones seguidas. Luego retropropagamos con la instancia `loss.backward()` y finalmente le decimos al optmizador que proceda a dar un paso de arendizaje con `optimizer.step()`.\n",
    "\n",
    "Una vez realizado el aprendizaje guardamos el modelo con la instancia `torch.save(model.state_dict(), 'PATH')`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'Epoch: 307700, Delta: 0.0001, Loss: 0.311'"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "t=0 #Inicializamos el contador de iteracciones\n",
    "loss_viejo = 1000 #Inicalizamos el coste  viejo para que la primera delta se pueda calcular\n",
    "delta = 100 #Inicializamos delta para que podamos entrar en el bucle\n",
    "#for t in range(epochs):\n",
    "while delta>delta_limit: #Hasta que no lleguemos a la delta_limit iteramos:\n",
    "    \n",
    "    y_pred = model(x) #Obtenemos las precciones del modelo\n",
    "\n",
    "    \n",
    "    loss = criterion(y_pred, y) #Calculamos el coste\n",
    "    \n",
    "    if t % 100 == 0: #Cada 100 iteracciones hacemos:\n",
    "        delta = loss_viejo - loss.item() #El caclulo de delta\n",
    "        loss_viejo = loss.item() #Guardamos el coste actual ara el futuro calculo de delta\n",
    "        clear_output(wait=True) #esto nos permite actulizar los datos de pantalla unos sobre otros\n",
    "        display('Epoch: '+str(t)+', Delta: ' + '{:06.4f}'.format(delta) + ', Loss: '+'{:05.3f}'.format(loss.item())) \n",
    "        #Sacamos la info cruda por pantalla\n",
    "        \n",
    "    optimizer.zero_grad() #Ajustamos a cero los gradientes\n",
    "    loss.backward() #Retro propagamos\n",
    "    optimizer.step() #Hacemos un paso de aprendijaze\n",
    "\n",
    "    t=t+1 #Actualizamos el contador de iteracciones\n",
    "\n",
    "\n",
    "torch.save(model.state_dict(), './seno_nn_model.pth') #Guardamos el modelo"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## _Test set_\n",
    "\n",
    "En esta sección de código encontramos lo necesario para representar el training set y el test set del cuál generamos las salidas ahí mismo con la instancia `y_test = model(x_test)`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 460.8x230.4 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "with torch.no_grad(): #Sin tener en cuenta los gradientes de las siguientes operaciones:\n",
    "    N_inter = 100 #Número de puntos en el test set\n",
    "    x_test = torch.arange(-pi, pi, 2*pi/N_inter).view(N_inter, 1) #Generamos el test set\n",
    "    y_test = model(x_test)\n",
    "    \n",
    "    #Ploteamos el training set y el test set\n",
    "    plt.figure(figsize=(6.4, 3.2))\n",
    "    plt.plot(x, y, '+', label='Training set')\n",
    "    plt.plot(x_test, y_test, '-', label='Test set')\n",
    "    plt.title('Seno: $SME_{sum}='+'{:05.3f}'.format(loss.item())+'$')\n",
    "    plt.xlabel('$x$')\n",
    "    plt.ylabel('$\\sin(x)$')\n",
    "    plt.legend(loc='best')\n",
    "    plt.savefig('seno_nn.png', dpi=600)\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## ¿Cuáles son los parámetros de mi red?\n",
    "\n",
    "Aquí vemos cual es la sintaxis para obtener los pesos de la red:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "linear1.weight tensor([[-1.6238],\n",
      "        [ 2.0432],\n",
      "        [-0.4475]])\n",
      "linear1.bias tensor([-0.9567, -1.7998, -0.0902])\n",
      "linear2.weight tensor([[-4.9675,  2.7366, 12.2744]])\n",
      "linear2.bias tensor([-4.8591])\n"
     ]
    }
   ],
   "source": [
    "for name, param in model.named_parameters():\n",
    "    if param.requires_grad:\n",
    "        print(name, param.data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## ¿Qué hay mas allá del _training set_?\n",
    "\n",
    "Vamos a explorar que pasa si nos salimos de la zona en la que hemos entrenado el modelo, es decir fuera del rango $(-\\pi, \\pi)$."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "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": [
    "with torch.no_grad():\n",
    "    N_inter = 200 #Número de puntos en el set\n",
    "    x_ext = torch.arange(-20, 20, 40/N_inter).view(N_inter, 1) #Generamos el set\n",
    "    y_ext = model(x_ext)\n",
    "\n",
    "    #Ploteamos el set de fuera de rango\n",
    "    plt.figure()\n",
    "    plt.plot(x_ext, y_ext, 'g-')\n",
    "    plt.title('Seno: Fuera de rango')\n",
    "    plt.xlabel('$x$')\n",
    "    plt.ylabel('$Modelo(x)$')\n",
    "    plt.savefig('seno_nn_outofrange.png', dpi=600)\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## ¿Cómo puedo entrenar mi red en una GPU?\n",
    "\n",
    "En esta parte vamos a exlplorar el uso de otras unidades de procesamiento, las GPUs frente a lo que hemos usado hasta ahora en el programa, las CPUs. Para ello primero debemos conocer si nuestro equipo cuenta con una GPU y tenemos todo preparado para usarla con PyTorch. El siguiente código nos dirá si es así. Si cuentas con una GPU pero te dice que no es compatible con CUDA puedes encontrar una posible solución siguiente [esta instalación.]()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "El dispositivo que se usará a partir de ahora es GeForce 840M\n"
     ]
    }
   ],
   "source": [
    "if torch.cuda.is_available(): \n",
    "    print('El dispositivo que se usará a partir de ahora es ' + torch.cuda.get_device_name(0))\n",
    "    device = torch.device(\"cuda:0\")\n",
    "else:\n",
    "    print('Tu equipo no cuenta con un dispositivo compatible con CUDA')\n",
    "    device = torch.device(\"cpu\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Lanzamos el modelo sobre la GPU\n",
    "\n",
    "Debes tener en cuenta que uso de la GPU se nota cuando los modelos de red son mucho mas complejos que el que hemos usado, pero aun así este ejemplo puede servir de como se podría usar para entrenar una red. Simplemnete usamos el atributo `.to(device)` para pasar el modelo, el _training set_ y el criterio de cómputo de coste a la GPU que previamente hemos definido como `device`. Otro detalle es que nose puede convertir directametne un tensor alojado en la GPU a un array de `numpy` para ser _ploteado_, por ello es necesario pasarlo pirmero a la CPU usando el atributo `.cpu()`. El resto de sintaxis es la misma."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'Epoch: 99900, Delta: 100.0000, Loss: 1.044'"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "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": [
    "model = (TwoLayerNet(D_in, H, D_out)).to(device)\n",
    "x = x.to(device)\n",
    "y = y.to(device)\n",
    "\n",
    "criterion = criterion.to(device)\n",
    "optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)\n",
    "\n",
    "t = 0 #Inicializamos el contador de iteracciones\n",
    "loss_viejo = 1000 #Inicalizamos el coste  viejo para que la primera delta se pueda calcular\n",
    "delta = 100 #Inicializamos delta para que podamos entrar en el bucle\n",
    "for t in range(epochs):\n",
    "#while delta>delta_limit: #Hasta que no lleguemos a la delta_limit iteramos:\n",
    "    \n",
    "    y_pred = model(x) #Obtenemos las precciones del modelo\n",
    "\n",
    "    \n",
    "    loss = criterion(y_pred, y) #Calculamos el coste\n",
    "    \n",
    "    if t % 100 == 0: #Cada 100 iteracciones hacemos:\n",
    "        #delta = loss_viejo - loss.item() #El caclulo de delta\n",
    "        #loss_viejo = loss.item() #Guardamos el coste actual ara el futuro calculo de delta\n",
    "        clear_output(wait=True) #esto nos permite actulizar los datos de pantalla unos sobre otros\n",
    "        display('Epoch: '+str(t)+', Delta: ' + '{:06.4f}'.format(delta) + ', Loss: '+'{:05.3f}'.format(loss.item())) \n",
    "        #Sacamos la info cruda por pantalla\n",
    "        \n",
    "    optimizer.zero_grad() #Ajustamos a cero los gradientes\n",
    "    loss.backward() #Retro propagamos\n",
    "    optimizer.step() #Hacemos un paso de aprendijaze\n",
    "\n",
    "    t=t+1 #Actualizamos el contador de iteracciones\n",
    "\n",
    "with torch.no_grad(): #Sin tener en cuenta los gradientes de las siguientes operaciones:\n",
    "    N_inter = 100 #Número de puntos en el test set\n",
    "    x_test = (torch.arange(-pi, pi, 2*pi/N_inter).view(N_inter, 1)).to(device) #Generamos el test set\n",
    "    y_test = model(x_test).to(device)\n",
    "    \n",
    "    #Ploteamos el training set y el test set\n",
    "    plt.figure()\n",
    "    plt.plot(x.cpu(), y.cpu(), '+', label='Training set')\n",
    "    plt.plot(x_test.cpu(), y_test.cpu(), '-', label='Test set')\n",
    "    plt.title('Seno: $SME_{sum}='+'{:05.3f}'.format(loss.item())+'$')\n",
    "    plt.xlabel('$x$')\n",
    "    plt.ylabel('$\\sin(x)$')\n",
    "    plt.legend(loc='best')\n",
    "    plt.savefig('seno_nn_gpu.png', dpi=600)\n",
    "    plt.show()"
   ]
  }
 ],
 "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": "3.7.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}