vit.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4",
      "authorship_tag": "ABX9TyPVx0li6kmUuv/bSZE4cz51"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "code",
      "execution_count": 88,
      "metadata": {
        "id": "6XEtEEg1V2ZW"
      },
      "outputs": [],
      "source": [
        "import numpy as np\n",
        "import tensorflow as tf\n",
        "import matplotlib.pyplot as plt\n",
        "from tensorflow.keras import layers"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "(x_train,y_train) ,(x_test, y_test) = tf.keras.datasets.cifar100.load_data()\n",
        "num_classes = 100"
      ],
      "metadata": {
        "id": "MQBhWZJoWLav"
      },
      "execution_count": 89,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "x_train.shape , y_train.shape"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "sv5bMjfgXCpO",
        "outputId": "3d91afdc-50b0-482e-d3f8-0ab910dd823d"
      },
      "execution_count": 90,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "((50000, 32, 32, 3), (50000, 1))"
            ]
          },
          "metadata": {},
          "execution_count": 90
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#데이터 확인\n",
        "plt.figure(figsize=(10,2))\n",
        "for i in range(5):\n",
        "    plt.subplot(1,5,i+1)\n",
        "    plt.imshow(x_train[i])\n",
        "    plt.axis('off')\n",
        "plt.show()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 170
        },
        "id": "vsVfH1qDXCdZ",
        "outputId": "2bbbe729-36bb-43fc-fb99-73304994cee8"
      },
      "execution_count": 91,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 1000x200 with 5 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "x_train = x_train /255.\n",
        "x_test = x_test /255."
      ],
      "metadata": {
        "id": "4xPCensqXVxP"
      },
      "execution_count": 92,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# 하이퍼 파라미터 설정\n",
        "\n",
        "input_shape = (32,32,3)\n",
        "batch_size = 32\n",
        "\n",
        "image_size = 72\n",
        "patch_size = 6\n",
        "num_patches = (image_size // patch_size) ** 2\n",
        "\n",
        "#최신 논문일경우 러닝메이트가 쓰이는경우가 많다\n",
        "learning_rate = 1e-3\n",
        "weight_decay = 1e-4\n",
        "\n",
        "epochs = 30\n",
        "\n",
        "transformer_layers = 4\n",
        "projection_dim = 64 # embedding을 통과하고 나온 wk wq wv(output_dim)이다\n",
        "num_heads = 4\n",
        "transformer_units = [projection_dim * 2 , projection_dim]\n",
        "mlp_head_units = [2048,1024]"
      ],
      "metadata": {
        "id": "dvHIjxLiXw-E"
      },
      "execution_count": 93,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#토큰화\n",
        "class PatchTokenization(layers.Layer):\n",
        "    def __init__(self, image_size=image_size, patch_size=patch_size,\n",
        "                 num_patches=num_patches, projection_dim=projection_dim, **kwargs):\n",
        "        super().__init__(**kwargs)\n",
        "        self.image_size = image_size\n",
        "        self.patch_size = patch_size\n",
        "        self.half_patch = patch_size // 2\n",
        "        self.flatten_patch = layers.Reshape((num_patches,-1))\n",
        "        self.projection = layers.Dense(projection_dim)\n",
        "        self.layer_norm = layers.LayerNormalization(epsilon=1e-6)\n",
        "    def call(self, images):\n",
        "        patches = tf.image.extract_patches(\n",
        "            images = images,\n",
        "            sizes = [1, self.patch_size, self.patch_size,1],\n",
        "            strides = [1, self.patch_size, self.patch_size,1],\n",
        "            rates = [1,1,1,1],\n",
        "            padding=\"VALID\"\n",
        "        )\n",
        "        flat_patches = self.flatten_patch(patches)\n",
        "        tokens = self.projection(flat_patches)\n",
        "        return (tokens, patches)"
      ],
      "metadata": {
        "id": "sfGa68fkZ_Ue"
      },
      "execution_count": 94,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#크기가 너무 작아서 재조정\n",
        "image = x_train[4]\n",
        "resized = tf.image.resize(tf.convert_to_tensor([image]),size=(image_size,image_size))\n",
        "resized.shape"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "yBRVBcvldwYN",
        "outputId": "ff98e276-4a4a-4ac9-fc05-7ad266e7fe19"
      },
      "execution_count": 95,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "TensorShape([1, 72, 72, 3])"
            ]
          },
          "metadata": {},
          "execution_count": 95
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "(token, patch) = PatchTokenization()(resized)\n",
        "print(token.shape)\n",
        "print(patch.shape)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "GrTtG9oieKfE",
        "outputId": "70cb7391-18c7-41ee-d247-ca7395def8db"
      },
      "execution_count": 96,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(1, 144, 64)\n",
            "(1, 12, 12, 108)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "(token,patch) = (token[0],patch[0])\n",
        "\n",
        "n = patch.shape[0]\n",
        "count = 1"
      ],
      "metadata": {
        "id": "FRD2_bQ_fC-h"
      },
      "execution_count": 97,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#패치화 확인 시각화\n",
        "plt.figure(figsize=(4,4))\n",
        "for row in range(n):\n",
        "    for col in range(n):\n",
        "        plt.subplot(n,n,count)\n",
        "        count += 1\n",
        "        image = tf.reshape(patch[row][col], (patch_size, patch_size, 3))\n",
        "        plt.imshow(image)\n",
        "        plt.axis('off')\n",
        "plt.show()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 345
        },
        "id": "j8kaNXH_jTxN",
        "outputId": "7b0a7d73-e107-45df-c542-3dcaff587c5d"
      },
      "execution_count": 98,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 400x400 with 144 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "class PatchEncoder(layers.Layer):\n",
        "    def __init__(self, num_patches=num_patches,\n",
        "                 projection_dim=projection_dim , **kwargs):\n",
        "        super().__init__(**kwargs)\n",
        "        self.num_patches = num_patches\n",
        "        self.poisition_embedding = layers.Embedding(input_dim = num_patches, output_dim = projection_dim)\n",
        "        self.positions = tf.range(start=0, limit=self.num_patches,delta=1)\n",
        "\n",
        "    def call(self, encoded_patches):\n",
        "        encoded_positions = self.poisition_embedding(self.positions)\n",
        "        encoded_patches = encoded_patches + encoded_positions\n",
        "        return encoded_patches\n"
      ],
      "metadata": {
        "id": "SVJruWjAj7nF"
      },
      "execution_count": 99,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def mlp(x, hidden_units, dropout_rate):\n",
        "    for units in hidden_units:\n",
        "        x = layers.Dense(units, activation = tf.nn.gelu)(x)\n",
        "        x = layers.Dropout(dropout_rate)(x)\n",
        "    return x"
      ],
      "metadata": {
        "id": "PzVDsO0ypZUM"
      },
      "execution_count": 100,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def create_vit_classifier():\n",
        "    inputs = layers.Input(shape=input_shape)\n",
        "    x = layers.Normalization()(inputs)\n",
        "    x = layers.Resizing(image_size,image_size)(x)\n",
        "    (tokens , _ )= PatchTokenization()(x)\n",
        "    encoded_patches = PatchEncoder()(tokens)\n",
        "    for _ in range(transformer_layers):\n",
        "        x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches) ##\n",
        "        attention_output = layers.MultiHeadAttention(num_heads=num_heads,\n",
        "                                                     key_dim=projection_dim,\n",
        "                                                     dropout=0.1)(x1,x1)\n",
        "        x2 = layers.Add()([attention_output,encoded_patches])\n",
        "        x3 = layers.LayerNormalization(epsilon=1e-6)(x2) #\n",
        "        x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)\n",
        "        encoded_patches = layers.Add()([x3,x2])\n",
        "\n",
        "    x = layers.LayerNormalization(epsilon=1e-6)(encoded_patches) #\n",
        "    x = layers.Flatten()(x)\n",
        "    x = layers.Dropout(0.5)(x)\n",
        "    x = mlp(x, hidden_units=mlp_head_units, dropout_rate=0.5)\n",
        "    outputs = layers.Dense(num_classes)(x)\n",
        "    model = tf.keras.Model(inputs=inputs , outputs=outputs)\n",
        "    return model"
      ],
      "metadata": {
        "id": "02WrKjFeqhr9"
      },
      "execution_count": 101,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "vit = create_vit_classifier()\n",
        "vit.summary()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "A1qvUJ16r0cA",
        "outputId": "44437290-66ae-40fa-bbca-fbd5b86b1c7e"
      },
      "execution_count": 102,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Model: \"model_1\"\n",
            "__________________________________________________________________________________________________\n",
            " Layer (type)                Output Shape                 Param #   Connected to                  \n",
            "==================================================================================================\n",
            " input_9 (InputLayer)        [(None, 32, 32, 3)]          0         []                            \n",
            "                                                                                                  \n",
            " normalization_12 (Normaliz  (None, 32, 32, 3)            7         ['input_9[0][0]']             \n",
            " ation)                                                                                           \n",
            "                                                                                                  \n",
            " resizing_8 (Resizing)       (None, 72, 72, 3)            0         ['normalization_12[0][0]']    \n",
            "                                                                                                  \n",
            " patch_tokenization_14 (Pat  ((None, 144, 64),            6976      ['resizing_8[0][0]']          \n",
            " chTokenization)              (None, 12, 12, 108))                                                \n",
            "                                                                                                  \n",
            " patch_encoder_8 (PatchEnco  (None, 144, 64)              9216      ['patch_tokenization_14[0][0]'\n",
            " der)                                                               ]                             \n",
            "                                                                                                  \n",
            " layer_normalization_24 (La  (None, 144, 64)              128       ['patch_encoder_8[0][0]']     \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " multi_head_attention_6 (Mu  (None, 144, 64)              66368     ['layer_normalization_24[0][0]\n",
            " ltiHeadAttention)                                                  ',                            \n",
            "                                                                     'layer_normalization_24[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " add_10 (Add)                (None, 144, 64)              0         ['multi_head_attention_6[0][0]\n",
            "                                                                    ',                            \n",
            "                                                                     'patch_encoder_8[0][0]']     \n",
            "                                                                                                  \n",
            " layer_normalization_25 (La  (None, 144, 64)              128       ['add_10[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " dense_26 (Dense)            (None, 144, 128)             8320      ['layer_normalization_25[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " dropout_11 (Dropout)        (None, 144, 128)             0         ['dense_26[0][0]']            \n",
            "                                                                                                  \n",
            " dense_27 (Dense)            (None, 144, 64)              8256      ['dropout_11[0][0]']          \n",
            "                                                                                                  \n",
            " dropout_12 (Dropout)        (None, 144, 64)              0         ['dense_27[0][0]']            \n",
            "                                                                                                  \n",
            " add_11 (Add)                (None, 144, 64)              0         ['dropout_12[0][0]',          \n",
            "                                                                     'add_10[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_26 (La  (None, 144, 64)              128       ['add_11[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " multi_head_attention_7 (Mu  (None, 144, 64)              66368     ['layer_normalization_26[0][0]\n",
            " ltiHeadAttention)                                                  ',                            \n",
            "                                                                     'layer_normalization_26[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " add_12 (Add)                (None, 144, 64)              0         ['multi_head_attention_7[0][0]\n",
            "                                                                    ',                            \n",
            "                                                                     'add_11[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_27 (La  (None, 144, 64)              128       ['add_12[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " dense_28 (Dense)            (None, 144, 128)             8320      ['layer_normalization_27[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " dropout_13 (Dropout)        (None, 144, 128)             0         ['dense_28[0][0]']            \n",
            "                                                                                                  \n",
            " dense_29 (Dense)            (None, 144, 64)              8256      ['dropout_13[0][0]']          \n",
            "                                                                                                  \n",
            " dropout_14 (Dropout)        (None, 144, 64)              0         ['dense_29[0][0]']            \n",
            "                                                                                                  \n",
            " add_13 (Add)                (None, 144, 64)              0         ['dropout_14[0][0]',          \n",
            "                                                                     'add_12[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_28 (La  (None, 144, 64)              128       ['add_13[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " multi_head_attention_8 (Mu  (None, 144, 64)              66368     ['layer_normalization_28[0][0]\n",
            " ltiHeadAttention)                                                  ',                            \n",
            "                                                                     'layer_normalization_28[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " add_14 (Add)                (None, 144, 64)              0         ['multi_head_attention_8[0][0]\n",
            "                                                                    ',                            \n",
            "                                                                     'add_13[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_29 (La  (None, 144, 64)              128       ['add_14[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " dense_30 (Dense)            (None, 144, 128)             8320      ['layer_normalization_29[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " dropout_15 (Dropout)        (None, 144, 128)             0         ['dense_30[0][0]']            \n",
            "                                                                                                  \n",
            " dense_31 (Dense)            (None, 144, 64)              8256      ['dropout_15[0][0]']          \n",
            "                                                                                                  \n",
            " dropout_16 (Dropout)        (None, 144, 64)              0         ['dense_31[0][0]']            \n",
            "                                                                                                  \n",
            " add_15 (Add)                (None, 144, 64)              0         ['dropout_16[0][0]',          \n",
            "                                                                     'add_14[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_30 (La  (None, 144, 64)              128       ['add_15[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " multi_head_attention_9 (Mu  (None, 144, 64)              66368     ['layer_normalization_30[0][0]\n",
            " ltiHeadAttention)                                                  ',                            \n",
            "                                                                     'layer_normalization_30[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " add_16 (Add)                (None, 144, 64)              0         ['multi_head_attention_9[0][0]\n",
            "                                                                    ',                            \n",
            "                                                                     'add_15[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_31 (La  (None, 144, 64)              128       ['add_16[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " dense_32 (Dense)            (None, 144, 128)             8320      ['layer_normalization_31[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " dropout_17 (Dropout)        (None, 144, 128)             0         ['dense_32[0][0]']            \n",
            "                                                                                                  \n",
            " dense_33 (Dense)            (None, 144, 64)              8256      ['dropout_17[0][0]']          \n",
            "                                                                                                  \n",
            " dropout_18 (Dropout)        (None, 144, 64)              0         ['dense_33[0][0]']            \n",
            "                                                                                                  \n",
            " add_17 (Add)                (None, 144, 64)              0         ['dropout_18[0][0]',          \n",
            "                                                                     'add_16[0][0]']              \n",
            "                                                                                                  \n",
            " layer_normalization_32 (La  (None, 144, 64)              128       ['add_17[0][0]']              \n",
            " yerNormalization)                                                                                \n",
            "                                                                                                  \n",
            " flatten_1 (Flatten)         (None, 9216)                 0         ['layer_normalization_32[0][0]\n",
            "                                                                    ']                            \n",
            "                                                                                                  \n",
            " dropout_19 (Dropout)        (None, 9216)                 0         ['flatten_1[0][0]']           \n",
            "                                                                                                  \n",
            " dense_34 (Dense)            (None, 2048)                 1887641   ['dropout_19[0][0]']          \n",
            "                                                          6                                       \n",
            "                                                                                                  \n",
            " dropout_20 (Dropout)        (None, 2048)                 0         ['dense_34[0][0]']            \n",
            "                                                                                                  \n",
            " dense_35 (Dense)            (None, 1024)                 2098176   ['dropout_20[0][0]']          \n",
            "                                                                                                  \n",
            " dropout_21 (Dropout)        (None, 1024)                 0         ['dense_35[0][0]']            \n",
            "                                                                                                  \n",
            " dense_36 (Dense)            (None, 100)                  102500    ['dropout_21[0][0]']          \n",
            "                                                                                                  \n",
            "==================================================================================================\n",
            "Total params: 21426219 (81.73 MB)\n",
            "Trainable params: 21426212 (81.73 MB)\n",
            "Non-trainable params: 7 (32.00 Byte)\n",
            "__________________________________________________________________________________________________\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "class CosineDecay(tf.keras.optimizers.schedules.LearningRateSchedule):\n",
        "    def __init__(self, learning_rate_base, total_steps, warmup_learning_rate, warmup_steps):\n",
        "        super().__init__()\n",
        "        self.learning_rate_base = learning_rate_base\n",
        "        self.total_steps = total_steps\n",
        "        self.warmup_learning_rate = warmup_learning_rate\n",
        "        self.warmup_steps = warmup_steps\n",
        "        self.pi = tf.constant(np.pi)\n",
        "\n",
        "    def __call__(self, step):\n",
        "        if self.total_steps < self.warmup_steps:\n",
        "            raise ValueError(\"total_steps 값이 warmup_steps보다 크거나 같아야합니다.\")\n",
        "\n",
        "        cos_annealed_lr = tf.cos(self.pi * (tf.cast(step, tf.float32) - self.warmup_steps) / float(self.total_steps - self.warmup_steps))\n",
        "        learning_rate = 0.5 * self.learning_rate_base * (1 + cos_annealed_lr)\n",
        "\n",
        "        if self.warmup_steps > 0:\n",
        "            if self.learning_rate_base < self.warmup_learning_rate:\n",
        "                raise ValueError(\n",
        "                    \"learning_rate_base 값이 warmup_learning_rate보다 크거나 같아야합니다.\")\n",
        "            slope = (self.learning_rate_base - self.warmup_learning_rate) / self.warmup_steps\n",
        "            warmup_rate = slope * tf.cast(step, tf.float32) + self.warmup_learning_rate\n",
        "            learning_rate = tf.where(step < self.warmup_steps, warmup_rate, learning_rate)\n",
        "        return tf.where(step > self.total_steps, 0.0, learning_rate, name=\"learning_rate\")\n",
        "\n"
      ],
      "metadata": {
        "id": "xHY61rm06vrF"
      },
      "execution_count": 103,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "total_steps = int((len(x_train) / batch_size) * epochs)\n",
        "warmup_epoch_percentage = 0.10\n",
        "warmup_steps = int(total_steps * warmup_epoch_percentage)\n",
        "scheduled_lrs = CosineDecay(\n",
        "    learning_rate_base=learning_rate,\n",
        "    total_steps=total_steps,\n",
        "    warmup_learning_rate=0.0,\n",
        "    warmup_steps=warmup_steps,)"
      ],
      "metadata": {
        "id": "4-byjRfB602e"
      },
      "execution_count": 104,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "optimizer = tf.keras.optimizers.AdamW(\n",
        "    learning_rate=scheduled_lrs,\n",
        "    weight_decay=weight_decay)"
      ],
      "metadata": {
        "id": "xths6iTz62d2"
      },
      "execution_count": 105,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "vit.compile(optimizer=optimizer,\n",
        "            loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),\n",
        "            metrics=['accuracy'])"
      ],
      "metadata": {
        "id": "4sh3tPxQ02Fg"
      },
      "execution_count": 106,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "history = vit.fit(x_train,y_train,validation_data=(x_test,y_test),epochs=10)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "UrshCw7_2qCP",
        "outputId": "9966e2ff-aad8-4884-c132-62abe729be87"
      },
      "execution_count": 107,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Epoch 1/10\n",
            "1563/1563 [==============================] - 84s 44ms/step - loss: 4.7737 - accuracy: 0.0164 - val_loss: 4.3434 - val_accuracy: 0.0331\n",
            "Epoch 2/10\n",
            "1563/1563 [==============================] - 67s 43ms/step - loss: 4.3795 - accuracy: 0.0319 - val_loss: 4.2031 - val_accuracy: 0.0522\n",
            "Epoch 3/10\n",
            "1563/1563 [==============================] - 67s 43ms/step - loss: 4.1138 - accuracy: 0.0600 - val_loss: 3.6915 - val_accuracy: 0.1272\n",
            "Epoch 4/10\n",
            "1563/1563 [==============================] - 68s 43ms/step - loss: 3.6003 - accuracy: 0.1378 - val_loss: 3.2925 - val_accuracy: 0.1984\n",
            "Epoch 5/10\n",
            "1563/1563 [==============================] - 66s 42ms/step - loss: 3.2926 - accuracy: 0.1965 - val_loss: 3.0358 - val_accuracy: 0.2439\n",
            "Epoch 6/10\n",
            "1563/1563 [==============================] - 69s 44ms/step - loss: 3.1110 - accuracy: 0.2309 - val_loss: 2.9425 - val_accuracy: 0.2637\n",
            "Epoch 7/10\n",
            "1563/1563 [==============================] - 69s 44ms/step - loss: 2.9615 - accuracy: 0.2599 - val_loss: 2.8048 - val_accuracy: 0.2950\n",
            "Epoch 8/10\n",
            "1563/1563 [==============================] - 71s 45ms/step - loss: 2.8246 - accuracy: 0.2876 - val_loss: 2.7141 - val_accuracy: 0.3148\n",
            "Epoch 9/10\n",
            "1563/1563 [==============================] - 69s 44ms/step - loss: 2.7044 - accuracy: 0.3112 - val_loss: 2.6628 - val_accuracy: 0.3202\n",
            "Epoch 10/10\n",
            "1563/1563 [==============================] - 69s 44ms/step - loss: 2.5860 - accuracy: 0.3321 - val_loss: 2.5856 - val_accuracy: 0.3440\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#시각화는 추후"
      ],
      "metadata": {
        "id": "4XYiXdxo21Cm"
      },
      "execution_count": 108,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [],
      "metadata": {
        "id": "pdXA9JZ8D-So"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}