Cap. 18 - GAN.ipynb

{"nbformat":4,"nbformat_minor":0,"metadata":{"colab":{"provenance":[],"gpuType":"T4","authorship_tag":"ABX9TyOD4dXClcFIhdirffZZXIVF"},"kernelspec":{"name":"python3","display_name":"Python 3"},"language_info":{"name":"python"},"accelerator":"GPU"},"cells":[{"cell_type":"markdown","source":["<table align=\"left\">\n","  <td>\n","    <a href=\"https://colab.research.google.com/drive/1D62kStjydJvh9Xz4ePYRaqY_hgXMSG7J\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n","  </td>\n","</table>"],"metadata":{"id":"JJHXhaeUndiQ"}},{"cell_type":"markdown","source":["# **Redes neuronales antagonistas (GAN)**\n","\n","Una arquitectura GAN está formada por dos redes neuronales, una generadora y otra discriminadora.\n","\n","Supongamos que queremos generar \"ropa\" (inspirada en el dataset MNIST fashion). La función de la red discriminadora es discernir si la red generadora está generando imágenes que sean ropa. La red generadora aprende y en cada nueva itearacción procura acercarse más y más a lo que espera la red discriminadora.\n"],"metadata":{"id":"Kw3rbHlcXzEe"}},{"cell_type":"code","source":["import numpy as np\n","import matplotlib.pyplot as plt\n","\n","import tensorflow as tf\n","from tensorflow.keras.layers import Input, Dense, Reshape, Flatten, Dropout\n","from tensorflow.keras.layers import BatchNormalization\n","from tensorflow.keras.layers import LeakyReLU\n","from tensorflow.keras.models import Sequential, Model\n","from tensorflow.keras.optimizers import Adam\n","\n","from tensorflow.keras.datasets import fashion_mnist\n"],"metadata":{"id":"IdMqAYoyow9R","executionInfo":{"status":"ok","timestamp":1700076788920,"user_tz":-60,"elapsed":3595,"user":{"displayName":"Juan Francisco Puentes Calvo","userId":"04307711343502565257"}}},"execution_count":1,"outputs":[]},{"cell_type":"code","source":["def build_generator():\n","\n","    model = Sequential()\n","    model.add(Dense(256, input_dim=100))\n","    model.add(LeakyReLU(0.2))\n","    model.add(BatchNormalization(momentum=0.8))\n","    model.add(Dense(512))\n","    model.add(LeakyReLU(0.2))\n","    model.add(BatchNormalization(momentum=0.8))\n","    model.add(Dense(1024))\n","    model.add(LeakyReLU(0.2))\n","    model.add(BatchNormalization(momentum=0.8))\n","    model.add(Dense(np.prod((28, 28, 1)), activation='tanh'))\n","    model.add(Reshape((28, 28, 1)))\n","\n","    noise = Input(shape=(100,))\n","    img = model(noise)\n","\n","    return Model(noise, img)\n","\n","def build_discriminator():\n","\n","    model = Sequential()\n","    model.add(Flatten(input_shape=(28, 28, 1)))\n","    model.add(Dense(512))\n","    model.add(LeakyReLU(0.2))\n","    model.add(Dense(256))\n","    model.add(LeakyReLU(0.2))\n","    model.add(Dense(1, activation='sigmoid'))\n","\n","    img = Input(shape=(28, 28, 1))\n","    validity = model(img)\n","\n","    return Model(img, validity)\n"],"metadata":{"id":"cGvddfvpo0O_","executionInfo":{"status":"ok","timestamp":1700076796210,"user_tz":-60,"elapsed":260,"user":{"displayName":"Juan Francisco Puentes Calvo","userId":"04307711343502565257"}}},"execution_count":2,"outputs":[]},{"cell_type":"code","source":["def compile_gan():\n","\n","    #optimizer = Adam(0.0002, 0.5)\n","    optimizer = tf.keras.optimizers.legacy.Adam(0.0002, 0.5)\n","\n","    # Construir y compilar el discriminador\n","    discriminator = build_discriminator()\n","    discriminator.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])\n","\n","    # Construir el generador\n","    generator = build_generator()\n","\n","    # Generador toma ruido como entrada y genera imágenes\n","    z = Input(shape=(100,))\n","    img = generator(z)\n","\n","    # Solo entrenaremos el generador en el modelo combinado\n","    discriminator.trainable = False\n","\n","    # Discriminador toma imágenes generadas como entrada y determina validez\n","    validity = discriminator(img)\n","\n","    # Modelo combinado (stacked generator and discriminator)\n","    combined = Model(z, validity)\n","    combined.compile(loss='binary_crossentropy', optimizer=optimizer)\n","\n","    return generator, discriminator, combined\n"],"metadata":{"id":"JO2kYfWNo67u","executionInfo":{"status":"ok","timestamp":1700076800067,"user_tz":-60,"elapsed":265,"user":{"displayName":"Juan Francisco Puentes Calvo","userId":"04307711343502565257"}}},"execution_count":3,"outputs":[]},{"cell_type":"code","source":["def train_gan(epochs, batch_size=128, save_interval=50):\n","\n","    (X_train, _), (_, _) = fashion_mnist.load_data()\n","\n","    X_train = X_train / 127.5 - 1.\n","    X_train = np.expand_dims(X_train, axis=3)\n","\n","    valid = np.ones((batch_size, 1))\n","    fake = np.zeros((batch_size, 1))\n","\n","    generator, discriminator, combined = compile_gan()\n","\n","    for epoch in range(epochs):\n","\n","        # Entrenar el discriminador\n","        idx = np.random.randint(0, X_train.shape[0], batch_size)\n","        imgs = X_train[idx]\n","\n","        noise = np.random.normal(0, 1, (batch_size, 100))\n","        gen_imgs = generator.predict(noise)\n","\n","        d_loss_real = discriminator.train_on_batch(imgs, valid)\n","        d_loss_fake = discriminator.train_on_batch(gen_imgs, fake)\n","        d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)\n","\n","        # Entrenar el generador\n","        noise = np.random.normal(0, 1, (batch_size, 100))\n","        g_loss = combined.train_on_batch(noise, valid)\n","\n","        #print(f\"{epoch}/{epochs} [D loss: {d_loss[0]} | D accuracy: {100 * d_loss[1]}] [G loss: {g_loss}]\")\n","\n","        if epoch % save_interval == 0:\n","            save_imgs(epoch, generator)\n","\n","def save_imgs(epoch, generator):\n","    r, c = 5, 5\n","    noise = np.random.normal(0, 1, (r * c, 100))\n","    gen_imgs = generator.predict(noise)\n","\n","    gen_imgs = 0.5 * gen_imgs + 0.5\n","\n","    fig, axs = plt.subplots(r, c)\n","    cnt = 0\n","    for i in range(r):\n","        for j in range(c):\n","            axs[i,j].imshow(gen_imgs[cnt, :,:,0], cmap='gray')\n","            axs[i,j].axis('off')\n","            cnt += 1\n","    plt.show()\n"],"metadata":{"id":"DjY8RIQ_o-MY","executionInfo":{"status":"ok","timestamp":1700076810896,"user_tz":-60,"elapsed":260,"user":{"displayName":"Juan Francisco Puentes Calvo","userId":"04307711343502565257"}}},"execution_count":4,"outputs":[]},{"cell_type":"code","source":["train_gan(epochs=10000, batch_size=32, save_interval=1000)\n"],"metadata":{"colab":{"base_uri":"https://localhost:8080/","height":1000,"output_embedded_package_id":"1bHEfPNXYJGO0DzkQ_6nBY6wIXQ85wxcS"},"id":"I4Ek_w4WpAXt","executionInfo":{"status":"ok","timestamp":1700077873162,"user_tz":-60,"elapsed":1060424,"user":{"displayName":"Juan Francisco Puentes Calvo","userId":"04307711343502565257"}},"outputId":"6e5d5710-c405-4124-a83f-c8cb48c89d9b"},"execution_count":5,"outputs":[{"output_type":"display_data","data":{"text/plain":"Output hidden; open in https://colab.research.google.com to view."},"metadata":{}}]},{"cell_type":"code","source":["def generate_and_show_imgs(generator, n_images=25):\n","    r, c = 5, 5  # Esto generará una cuadrícula de 5x5 de imágenes. Asegúrate de que r*c = n_images\n","    if n_images != r*c:\n","        raise ValueError(f\"{r}*{c} debe ser igual a {n_images}\");\n","\n","    noise = np.random.normal(0, 1, (r * c, 100))\n","    gen_imgs = generator.predict(noise)\n","\n","    # Re-escala imágenes 0 - 1\n","    #gen_imgs = 0.5 * gen_imgs + 0.5\n","\n","    fig, axs = plt.subplots(r, c, figsize=(10, 10))\n","    cnt = 0\n","    for i in range(r):\n","        for j in range(c):\n","            axs[i,j].imshow(gen_imgs[cnt, :,:,0], cmap='gray')\n","            axs[i,j].axis('off')\n","            cnt += 1\n","    plt.show()\n"],"metadata":{"id":"fpb-RifFpDLn"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["generator, _, _ = compile_gan()  # Solo necesitas el generador aquí, por eso ignoramos el discriminador y el modelo combinado.\n","train_gan(epochs=10000, batch_size=32, save_interval=1000)\n","\n","# Genera y muestra 25 imágenes\n","generate_and_show_imgs(generator)\n"],"metadata":{"colab":{"base_uri":"https://localhost:8080/","height":1000,"output_embedded_package_id":"1XDnspzKzU_DV4Jg6dVO_snST9eEzH4cj"},"id":"neEsLkE8p_nL","executionInfo":{"status":"ok","timestamp":1695401457899,"user_tz":-120,"elapsed":1237551,"user":{"displayName":"Juan Francisco Puentes Calvo","userId":"04307711343502565257"}},"outputId":"e11a4806-7581-4495-dfb6-5797dee1257d"},"execution_count":null,"outputs":[{"output_type":"display_data","data":{"text/plain":"Output hidden; open in https://colab.research.google.com to view."},"metadata":{}}]}]}