{
  "cells": [
    {
      "cell_type": "markdown",
      "id": "ba021202",
      "metadata": {
        "id": "ba021202"
      },
      "source": [
        "# **自定義類別物件（Custom class object）**\n",
        "此份程式碼會介紹如何使用 class 物件，自定義 Loss、Layer 或者 Model。\n",
        "\n",
        "## 本章節內容大綱\n",
        "* ### [Custom Loss](#Loss)\n",
        "* ### [Custom Layer](#Layer)\n",
        "* ### [Custom Model](#Model)\n",
        "---"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "1713f048",
      "metadata": {
        "id": "1713f048"
      },
      "source": [
        "## 匯入套件"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "eb30f987",
      "metadata": {
        "id": "eb30f987"
      },
      "outputs": [],
      "source": [
        "import tensorflow as tf\n",
        "from tensorflow import keras\n",
        "from tensorflow.keras import layers"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "55b83671",
      "metadata": {
        "id": "55b83671"
      },
      "source": [
        "<a name=\"Loss\"></a>\n",
        "## Custom Loss"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "45ca1165",
      "metadata": {
        "id": "45ca1165"
      },
      "outputs": [],
      "source": [
        "y_true = tf.random.normal((10, 4))\n",
        "y_pred = tf.random.normal((10, 4))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "b04bffaa",
      "metadata": {
        "id": "b04bffaa"
      },
      "outputs": [],
      "source": [
        "# build loss by tf.keras\n",
        "mse_loss = keras.losses.MeanSquaredError()\n",
        "mse_loss(y_true, y_pred)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "79d05a6d",
      "metadata": {
        "id": "79d05a6d"
      },
      "outputs": [],
      "source": [
        "class my_mse(keras.losses.Loss):  # build loss object by custom class\n",
        "    def call(self, y_true, y_pred):\n",
        "        return tf.reduce_mean(tf.math.square(y_pred - y_true), axis=-1)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "4d77c0af",
      "metadata": {
        "id": "4d77c0af"
      },
      "outputs": [],
      "source": [
        "my_mse_loss = my_mse()\n",
        "my_mse_loss(y_true, y_pred)"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "50ec579c",
      "metadata": {
        "id": "50ec579c"
      },
      "source": [
        "<a name=\"Layer\"></a>\n",
        "## Custom Layer"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "ce12852d",
      "metadata": {
        "id": "ce12852d"
      },
      "outputs": [],
      "source": [
        "x = tf.random.normal((10, 5))"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "d5177b88",
      "metadata": {
        "id": "d5177b88"
      },
      "outputs": [],
      "source": [
        "# build layer by tf.keras\n",
        "dense_layer = keras.layers.Dense(\n",
        "    4,\n",
        "    input_shape=x[0].shape,\n",
        "    kernel_initializer=tf.random_normal_initializer(seed=17))  # 初始化參數"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "17644d65",
      "metadata": {
        "id": "17644d65"
      },
      "outputs": [],
      "source": [
        "dense_layer(x)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "aaa644f2",
      "metadata": {
        "id": "aaa644f2"
      },
      "outputs": [],
      "source": [
        "class my_dense(keras.layers.Layer):  # build layer object by custom class\n",
        "    def __init__(self, units=4, input_dim=5):\n",
        "        super(my_dense, self).__init__()\n",
        "        self.w = self.add_weight(\n",
        "            shape=(input_dim, units),\n",
        "            initializer=tf.random_normal_initializer(seed=17),  # 初始化參數\n",
        "            trainable=True)\n",
        "        self.b = self.add_weight(\n",
        "            shape=(units,),\n",
        "            initializer=tf.zeros_initializer(),  # 初始化參數\n",
        "            trainable=True)\n",
        "\n",
        "    def call(self, inputs):\n",
        "        return tf.matmul(inputs, self.w) + self.b"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "6b537d78",
      "metadata": {
        "id": "6b537d78"
      },
      "outputs": [],
      "source": [
        "my_dense_layer = my_dense()\n",
        "my_dense_layer(x)"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "db06cb92",
      "metadata": {
        "id": "db06cb92"
      },
      "source": [
        "<a name=\"Model\"></a>\n",
        "## Custom Model"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "33d759c9",
      "metadata": {
        "id": "33d759c9"
      },
      "outputs": [],
      "source": [
        "num_classes = 4"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "7e9e0632",
      "metadata": {
        "id": "7e9e0632"
      },
      "outputs": [],
      "source": [
        "# build model by tf.keras\n",
        "keras.backend.clear_session()\n",
        "model = keras.models.Sequential()\n",
        "model.add(layers.Dense(\n",
        "    16,  # 神經元個數\n",
        "    kernel_initializer=tf.random_normal_initializer(seed=17),  # 初始化參數\n",
        "    input_shape=x[0].shape))  # 輸入形狀\n",
        "model.add(layers.Dense(\n",
        "    32,  # 神經元個數\n",
        "    kernel_initializer=tf.random_normal_initializer(seed=17)))  # 初始化參數\n",
        "model.add(layers.Dense(\n",
        "    num_classes,\n",
        "    kernel_initializer=tf.random_normal_initializer(seed=17)))  # 初始化參數"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "845edda5",
      "metadata": {
        "id": "845edda5"
      },
      "outputs": [],
      "source": [
        "model.summary()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "127d3416",
      "metadata": {
        "id": "127d3416"
      },
      "outputs": [],
      "source": [
        "model(x)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "2dc3d997",
      "metadata": {
        "id": "2dc3d997"
      },
      "outputs": [],
      "source": [
        "class my_net(keras.Model):  # build model object by custom class\n",
        "    def __init__(self, num_classes=4):\n",
        "        super(my_net, self).__init__()\n",
        "        keras.backend.clear_session()  # 重置 keras 的所有狀態\n",
        "        self.input_layer = layers.Input(shape=(5,))\n",
        "        self.hidden_layer_1 = layers.Dense(\n",
        "            16,  # 神經元個數\n",
        "            kernel_initializer=tf.random_normal_initializer(seed=17))   # 初始化參數\n",
        "        self.hidden_layer_2 = layers.Dense(\n",
        "            32,  # 神經元個數\n",
        "            kernel_initializer=tf.random_normal_initializer(seed=17))  # 初始化參數\n",
        "        self.output_layer = layers.Dense(\n",
        "            num_classes,\n",
        "            kernel_initializer=tf.random_normal_initializer(seed=17))  # 初始化參數\n",
        "        self.out = self.call(self.input_layer)\n",
        "\n",
        "    def call(self, inputs):\n",
        "        x = self.hidden_layer_1(inputs)\n",
        "        x = self.hidden_layer_2(x)\n",
        "        outputs = self.output_layer(x)\n",
        "        return outputs"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "02616ea1",
      "metadata": {
        "id": "02616ea1"
      },
      "outputs": [],
      "source": [
        "my_model = my_net()\n",
        "my_model.build(input_shape=(None, 5))\n",
        "my_model.summary()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "ae97735a",
      "metadata": {
        "id": "ae97735a"
      },
      "outputs": [],
      "source": [
        "my_model(x)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "c37b4fdb",
      "metadata": {
        "id": "c37b4fdb"
      },
      "outputs": [],
      "source": []
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3 (ipykernel)",
      "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": "3.7.12"
    },
    "colab": {
      "provenance": []
    },
    "accelerator": "GPU",
    "gpuClass": "standard"
  },
  "nbformat": 4,
  "nbformat_minor": 5
}