Chapter_7_SGAN_ipynb의_사본.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/zergswim/GAN/blob/main/Chapter_7_SGAN_ipynb%EC%9D%98_%EC%82%AC%EB%B3%B8.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Sgta16A7CUCW"
      },
      "source": [
        "# 7장: SGAN"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "wGWR9bzYCUCa"
      },
      "source": [
        "<table align=\"left\">\n",
        "  <td>\n",
        "    <a target=\"_blank\" href=\"https://colab.research.google.com/github/rickiepark/gans-in-action/blob/master/chapter-7/Chapter_7_SGAN.ipynb\"><img src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" />구글 코랩에서 실행하기</a>\n",
        "  </td>\n",
        "</table>"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "id": "fMfOE4pUCUCb"
      },
      "outputs": [],
      "source": [
        "%matplotlib inline\n",
        "\n",
        "import matplotlib.pyplot as plt\n",
        "import numpy as np\n",
        "\n",
        "from tensorflow.keras import backend as K\n",
        "\n",
        "from tensorflow.keras.datasets import mnist\n",
        "from tensorflow.keras.layers import (Activation, BatchNormalization, Concatenate, Dense,\n",
        "                                     Dropout, Flatten, Input, Lambda, Reshape)\n",
        "from tensorflow.keras.layers import LeakyReLU\n",
        "from tensorflow.keras.layers import Conv2D, Conv2DTranspose\n",
        "from tensorflow.keras.models import Model, Sequential\n",
        "from tensorflow.keras.optimizers import Adam\n",
        "from tensorflow.keras.utils import to_categorical"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_4FgnWooCUCb"
      },
      "source": [
        "## 데이터셋"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "rOmpG7_ZCUCb"
      },
      "outputs": [],
      "source": [
        "class Dataset:\n",
        "    def __init__(self, num_labeled):\n",
        "\n",
        "        # 훈련에 사용할 레이블된 샘플 개수\n",
        "        self.num_labeled = num_labeled\n",
        "\n",
        "        # MNIST 데이터셋 적재\n",
        "        (self.x_train, self.y_train), (self.x_test,\n",
        "                                       self.y_test) = mnist.load_data()\n",
        "\n",
        "        def preprocess_imgs(x):\n",
        "            # [0, 255] 사이 흑백 픽셀 값을 [–1, 1] 사이로 변환\n",
        "            x = (x.astype(np.float32) - 127.5) / 127.5\n",
        "            # 너비 × 높이 × 채널로 이미지 차원을 확장\n",
        "            x = np.expand_dims(x, axis=3)\n",
        "            return x\n",
        "\n",
        "        def preprocess_labels(y):\n",
        "            return y.reshape(-1, 1)\n",
        "\n",
        "        # 훈련 데이터\n",
        "        self.x_train = preprocess_imgs(self.x_train)\n",
        "        self.y_train = preprocess_labels(self.y_train)\n",
        "\n",
        "        # 테스트 데이터\n",
        "        self.x_test = preprocess_imgs(self.x_test)\n",
        "        self.y_test = preprocess_labels(self.y_test)\n",
        "\n",
        "    def batch_labeled(self, batch_size):\n",
        "        # 레이블된 이미지와 레이블의 랜덤 배치 만들기\n",
        "        idx = np.random.randint(0, self.num_labeled, batch_size)\n",
        "        imgs = self.x_train[idx]\n",
        "        labels = self.y_train[idx]\n",
        "        return imgs, labels\n",
        "\n",
        "    def batch_unlabeled(self, batch_size):\n",
        "        # 레이블이 없는 이미지의 랜덤 배치 만들기\n",
        "        idx = np.random.randint(self.num_labeled, self.x_train.shape[0],\n",
        "                                batch_size)\n",
        "        imgs = self.x_train[idx]\n",
        "        return imgs\n",
        "\n",
        "    def training_set(self):\n",
        "        x_train = self.x_train[range(self.num_labeled)]\n",
        "        y_train = self.y_train[range(self.num_labeled)]\n",
        "        return x_train, y_train\n",
        "\n",
        "    def test_set(self):\n",
        "        return self.x_test, self.y_test"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 3,
      "metadata": {
        "id": "BcElbU6SCUCc",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "8028b6d6-7bc7-48ae-c29e-ed6f29d8a39a"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz\n",
            "11493376/11490434 [==============================] - 0s 0us/step\n",
            "11501568/11490434 [==============================] - 0s 0us/step\n"
          ]
        }
      ],
      "source": [
        "# 사용할 레이블된 샘플 개수 (나머지는 레이블없이 사용합니다)\n",
        "num_labeled = 100\n",
        "\n",
        "dataset = Dataset(num_labeled)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "O9UQJrdTCUCc"
      },
      "source": [
        "# SGAN"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 4,
      "metadata": {
        "id": "TTpjfRu3CUCc"
      },
      "outputs": [],
      "source": [
        "img_rows = 28\n",
        "img_cols = 28\n",
        "channels = 1\n",
        "\n",
        "# 입력 이미지 차원\n",
        "img_shape = (img_rows, img_cols, channels)\n",
        "\n",
        "# 생성자의 입력으로 사용할 잡음 벡터의 크기\n",
        "z_dim = 100\n",
        "\n",
        "# 데이터셋에 있는 클래스 개수\n",
        "num_classes = 10"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "pIst7kenCUCc"
      },
      "source": [
        "## 생성자"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 5,
      "metadata": {
        "id": "4WzMJ065CUCd"
      },
      "outputs": [],
      "source": [
        "def build_generator(z_dim):\n",
        "\n",
        "    model = Sequential()\n",
        "\n",
        "    # 완전 연결 층을 사용해 입력을 7 × 7 × 256 크기 텐서로 바꿉니다.\n",
        "    model.add(Dense(256 * 7 * 7, input_dim=z_dim))\n",
        "    model.add(Reshape((7, 7, 256)))\n",
        "\n",
        "    # 7 × 7 × 256에서 14 × 14 × 128 텐서로 바꾸는 전치 합성곱 층\n",
        "    model.add(Conv2DTranspose(128, kernel_size=3, strides=2, padding='same'))\n",
        "\n",
        "    # 배치 정규화\n",
        "    model.add(BatchNormalization())\n",
        "\n",
        "    # LeakyReLU 활성화 함수\n",
        "    model.add(LeakyReLU(alpha=0.01))\n",
        "\n",
        "    # 14 × 14 × 128에서 14 × 14 × 64 텐서로 바꾸는 전치 합성곱 층\n",
        "    model.add(Conv2DTranspose(64, kernel_size=3, strides=1, padding='same'))\n",
        "\n",
        "    # 배치 정규화\n",
        "    model.add(BatchNormalization())\n",
        "\n",
        "    # LeakyReLU 활성화 함수\n",
        "    model.add(LeakyReLU(alpha=0.01))\n",
        "\n",
        "    # 14 × 14 × 64에서 28 × 28 × 1 텐서로 바꾸는 전치 합성곱 층\n",
        "    model.add(Conv2DTranspose(1, kernel_size=3, strides=2, padding='same'))\n",
        "\n",
        "    # tanh 활성화 함수\n",
        "    model.add(Activation('tanh'))\n",
        "\n",
        "    return model"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "2MJeqmKHCUCd"
      },
      "source": [
        "## 판별자"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 6,
      "metadata": {
        "id": "zYpCZ5fOCUCd"
      },
      "outputs": [],
      "source": [
        "def build_discriminator_net(img_shape):\n",
        "\n",
        "    model = Sequential()\n",
        "\n",
        "    # 28 × 28 × 1에서 14 × 14 × 32 텐서로 바꾸는 합성곱 층\n",
        "    model.add(\n",
        "        Conv2D(32,\n",
        "               kernel_size=3,\n",
        "               strides=2,\n",
        "               input_shape=img_shape,\n",
        "               padding='same'))\n",
        "\n",
        "    # LeakyReLU 활성화 함수\n",
        "    model.add(LeakyReLU(alpha=0.01))\n",
        "\n",
        "    # 14 × 14 × 32에서 7 × 7 × 64 텐서로 바꾸는 합성곱 층\n",
        "    model.add(\n",
        "        Conv2D(64,\n",
        "               kernel_size=3,\n",
        "               strides=2,\n",
        "               padding='same'))\n",
        "    \n",
        "    # LeakyReLU 활성화 함수\n",
        "    model.add(LeakyReLU(alpha=0.01))\n",
        "\n",
        "    # 7 × 7 × 64에서 3 × 3 × 128 텐서로 바꾸는 합성곱 층\n",
        "    model.add(\n",
        "        Conv2D(128,\n",
        "               kernel_size=3,\n",
        "               strides=2,\n",
        "               padding='same'))\n",
        "    \n",
        "    # LeakyReLU 활성화 함수\n",
        "    model.add(LeakyReLU(alpha=0.01))\n",
        "\n",
        "    # 드롭아웃\n",
        "    model.add(Dropout(0.5))\n",
        "\n",
        "    # 텐서 펼치기\n",
        "    model.add(Flatten())\n",
        "\n",
        "    # num_classes 개의 뉴런을 가진 완전 연결 층\n",
        "    model.add(Dense(num_classes))\n",
        "\n",
        "    return model"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 7,
      "metadata": {
        "id": "wcIsVDTgCUCd"
      },
      "outputs": [],
      "source": [
        "def build_discriminator_supervised(discriminator_net):\n",
        "\n",
        "    model = Sequential()\n",
        "\n",
        "    model.add(discriminator_net)\n",
        "\n",
        "    # 진짜 클래스에 대한 예측 확률을 출력하는 소프트맥스 활성화 함수\n",
        "    model.add(Activation('softmax'))\n",
        "\n",
        "    return model"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 8,
      "metadata": {
        "id": "_IVe9hFvCUCd"
      },
      "outputs": [],
      "source": [
        "def build_discriminator_unsupervised(discriminator_net):\n",
        "\n",
        "    model = Sequential()\n",
        "\n",
        "    model.add(discriminator_net)\n",
        "\n",
        "    def predict(x):\n",
        "        # 진짜 클래스에 대한 확률 분포를 진짜 대 가짜의 이진 확률로 변환합니다.\n",
        "        prediction = 1.0 - (1.0 /\n",
        "                            (K.sum(K.exp(x), axis=-1, keepdims=True) + 1.0))\n",
        "        return prediction\n",
        "\n",
        "    # 앞서 정의한 진짜 대 가짜 확률을 출력하는 뉴런\n",
        "    model.add(Lambda(predict))\n",
        "\n",
        "    return model"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "3g1t9cHKCUCe"
      },
      "source": [
        "## 모델 구성"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 9,
      "metadata": {
        "id": "XJMkjCNNCUCe"
      },
      "outputs": [],
      "source": [
        "def build_gan(generator, discriminator):\n",
        "\n",
        "    model = Sequential()\n",
        "\n",
        "    # 생성자와 판별자 모델을 연결하기\n",
        "    model.add(generator)\n",
        "    model.add(discriminator)\n",
        "\n",
        "    return model"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "4kvzhVU0CUCe"
      },
      "source": [
        "### 판별자"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 10,
      "metadata": {
        "id": "4Vp0T0X5CUCe"
      },
      "outputs": [],
      "source": [
        "# 판별자 기반 모델: 이 층들은 지도 학습 훈련과 비지도 학습 훈련에 공유됩니다.\n",
        "discriminator_net = build_discriminator_net(img_shape)\n",
        "\n",
        "# 지도 학습 훈련을 위해 판별자를 만들고 컴파일합니다.\n",
        "discriminator_supervised = build_discriminator_supervised(discriminator_net)\n",
        "discriminator_supervised.compile(loss='categorical_crossentropy',\n",
        "                                 metrics=['accuracy'],\n",
        "                                 optimizer=Adam(learning_rate=0.0003))\n",
        "\n",
        "# 비지도 학습 훈련을 위해 판별자를 만들고 컴파일합니다.\n",
        "discriminator_unsupervised = build_discriminator_unsupervised(discriminator_net)\n",
        "discriminator_unsupervised.compile(loss='binary_crossentropy',\n",
        "                                   optimizer=Adam())"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ZsdB7aFkCUCe"
      },
      "source": [
        "### 생성자"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 11,
      "metadata": {
        "id": "HjyDCLHvCUCe"
      },
      "outputs": [],
      "source": [
        "# 생성자를 만듭니다.\n",
        "generator = build_generator(z_dim)\n",
        "\n",
        "# 생성자 훈련을 위해 판별자의 모델 파라미터를 동결합니다.\n",
        "discriminator_unsupervised.trainable = False\n",
        "\n",
        "# 생성자를 훈련하기 위해 고정된 판별자로 GAN 모델을 만들고 컴파일합니다.\n",
        "# 노트: 비지도 학습용 판별자를 사용하세요.\n",
        "gan = build_gan(generator, discriminator_unsupervised)\n",
        "gan.compile(loss='binary_crossentropy', optimizer=Adam())"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_SkyHoMhCUCf"
      },
      "source": [
        "## 훈련"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 12,
      "metadata": {
        "id": "lOsI6htfCUCf"
      },
      "outputs": [],
      "source": [
        "supervised_losses = []\n",
        "iteration_checkpoints = []\n",
        "\n",
        "\n",
        "def train(iterations, batch_size, sample_interval):\n",
        "\n",
        "    # 진짜 이미지의 레이블: 모두 1\n",
        "    real = np.ones((batch_size, 1))\n",
        "\n",
        "    # 가짜 이미지의 레이블: 모두 0\n",
        "    fake = np.zeros((batch_size, 1))\n",
        "\n",
        "    for iteration in range(iterations):\n",
        "\n",
        "        # -------------------------\n",
        "        #  판별자 훈련\n",
        "        # -------------------------\n",
        "\n",
        "        # 레이블된 샘플을 가져옵니다.\n",
        "        imgs, labels = dataset.batch_labeled(batch_size)\n",
        "\n",
        "        # 레이블을 원-핫 인코딩합니다.\n",
        "        labels = to_categorical(labels, num_classes=num_classes)\n",
        "\n",
        "        # 레이블이 없는 샘플을 가져옵니다.\n",
        "        imgs_unlabeled = dataset.batch_unlabeled(batch_size)\n",
        "\n",
        "        # 가짜 이미지의 배치를 생성합니다.\n",
        "        z = np.random.normal(0, 1, (batch_size, z_dim))\n",
        "        gen_imgs = generator.predict(z)\n",
        "\n",
        "        # 레이블된 진짜 샘플에서 훈련합니다.\n",
        "        d_loss_supervised, accuracy = discriminator_supervised.train_on_batch(imgs, labels)\n",
        "\n",
        "        # 레이블이 없는 진짜 샘플에서 훈련합니다.\n",
        "        d_loss_real = discriminator_unsupervised.train_on_batch(\n",
        "            imgs_unlabeled, real)\n",
        "\n",
        "        # 가짜 샘플에서 훈련합니다.\n",
        "        d_loss_fake = discriminator_unsupervised.train_on_batch(gen_imgs, fake)\n",
        "\n",
        "        d_loss_unsupervised = 0.5 * np.add(d_loss_real, d_loss_fake)\n",
        "\n",
        "        # ---------------------\n",
        "        #  생성자 훈련\u001e\n",
        "        # ---------------------\n",
        "\n",
        "        # 가짜 이미지의 배치를 생성합니다.\n",
        "        z = np.random.normal(0, 1, (batch_size, z_dim))\n",
        "        gen_imgs = generator.predict(z)\n",
        "\n",
        "        # 생성자를 훈련합니다.\n",
        "        g_loss = gan.train_on_batch(z, np.ones((batch_size, 1)))\n",
        "\n",
        "        if (iteration + 1) % sample_interval == 0:\n",
        "\n",
        "            # 훈련이 끝난 후 그래프를 그리기 위해 판별자의 지도 학습 분류 손실을 기록합니다.\n",
        "            supervised_losses.append(d_loss_supervised)\n",
        "            iteration_checkpoints.append(iteration + 1)\n",
        "\n",
        "            # 훈련 과정을 출력합니다.\n",
        "            print(\n",
        "                \"%d [D loss supervised: %.4f, acc.: %.2f%%] [D loss unsupervised: %.4f] [G loss: %f]\"\n",
        "                % (iteration + 1, d_loss_supervised, 100 * accuracy,\n",
        "                   d_loss_unsupervised, g_loss))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Fy5892ZuCUCf"
      },
      "source": [
        "## 모델 훈련과 출력 조사"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 13,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "9J7305B3CUCf",
        "outputId": "283ce6e7-843d-428a-e094-e6826fac8809",
        "scrolled": true
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "800 [D loss supervised: nan, acc.: 18.75%] [D loss unsupervised: nan] [G loss: nan]\n",
            "1600 [D loss supervised: nan, acc.: 15.62%] [D loss unsupervised: nan] [G loss: nan]\n",
            "2400 [D loss supervised: nan, acc.: 12.50%] [D loss unsupervised: nan] [G loss: nan]\n",
            "3200 [D loss supervised: nan, acc.: 9.38%] [D loss unsupervised: nan] [G loss: nan]\n",
            "4000 [D loss supervised: nan, acc.: 15.62%] [D loss unsupervised: nan] [G loss: nan]\n",
            "4800 [D loss supervised: nan, acc.: 6.25%] [D loss unsupervised: nan] [G loss: nan]\n",
            "5600 [D loss supervised: nan, acc.: 15.62%] [D loss unsupervised: nan] [G loss: nan]\n",
            "6400 [D loss supervised: nan, acc.: 12.50%] [D loss unsupervised: nan] [G loss: nan]\n",
            "7200 [D loss supervised: nan, acc.: 9.38%] [D loss unsupervised: nan] [G loss: nan]\n",
            "8000 [D loss supervised: nan, acc.: 15.62%] [D loss unsupervised: nan] [G loss: nan]\n"
          ]
        }
      ],
      "source": [
        "# 하이퍼파라미터를 셋팅합니다.\n",
        "iterations = 8000\n",
        "batch_size = 32\n",
        "sample_interval = 800\n",
        "\n",
        "# 지정한 반복 횟수 동안 SGAN을 훈련합니다.\n",
        "train(iterations, batch_size, sample_interval)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 14,
      "metadata": {
        "id": "h92u9tEzCUCg",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 323
        },
        "outputId": "f1e20f07-ac40-4e37-cebd-2219d93c10f9"
      },
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 1080x360 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ],
      "source": [
        "losses = np.array(supervised_losses)\n",
        "\n",
        "# 판별자의 지도 학습 손실을 그립니다.\n",
        "plt.figure(figsize=(15, 5))\n",
        "plt.plot(iteration_checkpoints, losses, label=\"Discriminator loss\")\n",
        "\n",
        "plt.xticks(iteration_checkpoints, rotation=90)\n",
        "\n",
        "plt.title(\"Discriminator – Supervised Loss\")\n",
        "plt.xlabel(\"Iteration\")\n",
        "plt.ylabel(\"Loss\")\n",
        "plt.legend()\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "b5zD1QwrCUCh"
      },
      "source": [
        "## SGAN 분류기 – 훈련 세트와 테스트 세트 정확도"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 15,
      "metadata": {
        "id": "ajShPxUjCUCh",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "63843c17-8478-48aa-fbe4-7ed848e77036"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "4/4 [==============================] - 0s 15ms/step - loss: nan - accuracy: 0.1300\n",
            "Training Accuracy: 13.00%\n"
          ]
        }
      ],
      "source": [
        "x, y = dataset.training_set()\n",
        "y = to_categorical(y, num_classes=num_classes)\n",
        "\n",
        "# 훈련 세트에서 분류 정확도 계산\n",
        "_, accuracy = discriminator_supervised.evaluate(x, y)\n",
        "print(\"Training Accuracy: %.2f%%\" % (100 * accuracy))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 16,
      "metadata": {
        "id": "sVJlwLDdCUCh",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "ccf5e944-df52-4869-cd6a-645cee149ac5"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "313/313 [==============================] - 1s 3ms/step - loss: nan - accuracy: 0.0980\n",
            "Test Accuracy: 9.80%\n"
          ]
        }
      ],
      "source": [
        "x, y = dataset.test_set()\n",
        "y = to_categorical(y, num_classes=num_classes)\n",
        "\n",
        "# 테스트 세트에서 분류 정확도 계산\n",
        "_, accuracy = discriminator_supervised.evaluate(x, y)\n",
        "print(\"Test Accuracy: %.2f%%\" % (100 * accuracy))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "FZ2uciaRCUCh"
      },
      "source": [
        "# 완전히 지도 학습으로 훈련된 분류기"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 17,
      "metadata": {
        "id": "Uf-efoiDCUCi"
      },
      "outputs": [],
      "source": [
        "# SGAN 판별자와 같은 네트워크 구조를 가진 지도 학습 분류기\n",
        "mnist_classifier = build_discriminator_supervised(build_discriminator_net(img_shape))\n",
        "mnist_classifier.compile(loss='categorical_crossentropy',\n",
        "                         metrics=['accuracy'],\n",
        "                         optimizer=Adam())"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 18,
      "metadata": {
        "id": "_V-rV4SqCUCi",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "707ff4bd-f35e-4f7d-8747-85d3c37cc5f5"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 1/30\n",
            "4/4 [==============================] - 0s 23ms/step - loss: 2.2827 - accuracy: 0.1700\n",
            "Epoch 2/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 2.2102 - accuracy: 0.2700\n",
            "Epoch 3/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 2.1118 - accuracy: 0.2700\n",
            "Epoch 4/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 2.0208 - accuracy: 0.3100\n",
            "Epoch 5/30\n",
            "4/4 [==============================] - 0s 7ms/step - loss: 1.9040 - accuracy: 0.3600\n",
            "Epoch 6/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 1.7145 - accuracy: 0.5200\n",
            "Epoch 7/30\n",
            "4/4 [==============================] - 0s 6ms/step - loss: 1.4569 - accuracy: 0.6500\n",
            "Epoch 8/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 1.2068 - accuracy: 0.6700\n",
            "Epoch 9/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 0.9207 - accuracy: 0.7800\n",
            "Epoch 10/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.8117 - accuracy: 0.7700\n",
            "Epoch 11/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 0.6121 - accuracy: 0.8300\n",
            "Epoch 12/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.6223 - accuracy: 0.8100\n",
            "Epoch 13/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 0.4464 - accuracy: 0.8400\n",
            "Epoch 14/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.3454 - accuracy: 0.8800\n",
            "Epoch 15/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.3040 - accuracy: 0.8900\n",
            "Epoch 16/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.3100 - accuracy: 0.8800\n",
            "Epoch 17/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.2499 - accuracy: 0.9200\n",
            "Epoch 18/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 0.2559 - accuracy: 0.9400\n",
            "Epoch 19/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.2153 - accuracy: 0.9500\n",
            "Epoch 20/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.1598 - accuracy: 0.9500\n",
            "Epoch 21/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.1625 - accuracy: 0.9700\n",
            "Epoch 22/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.1438 - accuracy: 0.9500\n",
            "Epoch 23/30\n",
            "4/4 [==============================] - 0s 5ms/step - loss: 0.1006 - accuracy: 0.9700\n",
            "Epoch 24/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.1707 - accuracy: 0.9100\n",
            "Epoch 25/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.0799 - accuracy: 0.9700\n",
            "Epoch 26/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.0885 - accuracy: 0.9700\n",
            "Epoch 27/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.0935 - accuracy: 0.9800\n",
            "Epoch 28/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.0605 - accuracy: 0.9900\n",
            "Epoch 29/30\n",
            "4/4 [==============================] - 0s 6ms/step - loss: 0.0638 - accuracy: 0.9900\n",
            "Epoch 30/30\n",
            "4/4 [==============================] - 0s 4ms/step - loss: 0.0480 - accuracy: 1.0000\n"
          ]
        }
      ],
      "source": [
        "imgs, labels = dataset.training_set()\n",
        "\n",
        "# 레이블을 원-핫 인코딩합니다.\n",
        "labels = to_categorical(labels, num_classes=num_classes)\n",
        "\n",
        "# 분류기를 훈련합니다.\n",
        "training = mnist_classifier.fit(x=imgs,\n",
        "                                y=labels,\n",
        "                                batch_size=32,\n",
        "                                epochs=30,\n",
        "                                verbose=1)\n",
        "losses = training.history['loss']\n",
        "accuracies = training.history['accuracy']"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 19,
      "metadata": {
        "id": "NCLm27uFCUCi",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 336
        },
        "outputId": "4fe0624e-ca92-440a-e240-f00ba68dc2e4"
      },
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 720x360 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ],
      "source": [
        "# 분류 손실을 그립니다\n",
        "plt.figure(figsize=(10, 5))\n",
        "plt.plot(np.array(losses), label=\"Loss\")\n",
        "plt.title(\"Classification Loss\")\n",
        "plt.legend()\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 20,
      "metadata": {
        "id": "bdmZTbHnCUCi",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 336
        },
        "outputId": "2e24e416-7997-41f9-8540-b1c9a0c704e8"
      },
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 720x360 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ],
      "source": [
        "# 분류 정확도를 그립니다.\n",
        "plt.figure(figsize=(10, 5))\n",
        "plt.plot(np.array(accuracies), label=\"Accuracy\")\n",
        "plt.title(\"Classification Accuracy\")\n",
        "plt.legend()\n",
        "plt.show()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 21,
      "metadata": {
        "id": "UAevicZSCUCi",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "6b6edc7e-642f-4279-87e1-13bdf45384c0"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "4/4 [==============================] - 0s 4ms/step - loss: 0.0164 - accuracy: 1.0000\n",
            "Training Accuracy: 100.00%\n"
          ]
        }
      ],
      "source": [
        "x, y = dataset.training_set()\n",
        "y = to_categorical(y, num_classes=num_classes)\n",
        "\n",
        "# 훈련 세트에 대한 분류 정확도를 계산합니다.\n",
        "_, accuracy = mnist_classifier.evaluate(x, y)\n",
        "print(\"Training Accuracy: %.2f%%\" % (100 * accuracy))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 22,
      "metadata": {
        "id": "o02hr24eCUCi",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "998399fa-9e36-473d-b0d4-40b50bdfe6b0"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "313/313 [==============================] - 1s 3ms/step - loss: 1.1077 - accuracy: 0.7051\n",
            "Test Accuracy: 70.51%\n"
          ]
        }
      ],
      "source": [
        "x, y = dataset.test_set()\n",
        "y = to_categorical(y, num_classes=num_classes)\n",
        "\n",
        "# 테스트 세트에 대한 분류 정확도를 계산합니다.\n",
        "_, accuracy = mnist_classifier.evaluate(x, y)\n",
        "print(\"Test Accuracy: %.2f%%\" % (100 * accuracy))"
      ]
    }
  ],
  "metadata": {
    "accelerator": "GPU",
    "colab": {
      "name": "Chapter_7_SGAN.ipynb의 사본",
      "provenance": [],
      "include_colab_link": true
    },
    "kernelspec": {
      "display_name": "TensorFlow 2.3 on Python 3.6 (CUDA 10.1)",
      "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": "3.7.7"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}