{
  "cells": [
    {
      "cell_type": "markdown",
      "id": "d0d29bfc",
      "metadata": {
        "id": "d0d29bfc"
      },
      "source": [
        "# **ResNet**\n",
        "此份程式碼會介紹如何使用 tf.keras 的方式建構 ResNet 的模型架構。"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "c76e4c9e",
      "metadata": {
        "id": "c76e4c9e"
      },
      "source": [
        "![image](https://hackmd.io/_uploads/rJ8SA1HOp.png)\n",
        "\n",
        "- [source paper](https://arxiv.org/abs/1512.03385)"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "d15836b8",
      "metadata": {
        "id": "d15836b8"
      },
      "source": [
        "## 匯入套件"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "14e0a7c1",
      "metadata": {
        "id": "14e0a7c1"
      },
      "outputs": [],
      "source": [
        "import numpy as np\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "# Tensorflow 相關套件\n",
        "import tensorflow as tf\n",
        "from tensorflow.keras import datasets, layers, Model, Sequential, losses"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "2e22f407",
      "metadata": {
        "id": "2e22f407"
      },
      "source": [
        "## 載入資料集"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "45155e27",
      "metadata": {
        "id": "45155e27"
      },
      "outputs": [],
      "source": [
        "(x_train, y_train), (x_test, y_test) = datasets.mnist.load_data()\n",
        "\n",
        "# Expand dimensions\n",
        "x_train = tf.expand_dims(x_train, axis=3, name=None)\n",
        "x_test = tf.expand_dims(x_test, axis=3, name=None)\n",
        "print(f'x_train shape: {x_train.shape}')\n",
        "print(f'x_test shape: {x_test.shape}')\n",
        "print('----------')\n",
        "\n",
        "# Grayscale to RGB\n",
        "x_train = tf.repeat(x_train, 3, axis=3)\n",
        "x_test = tf.repeat(x_test, 3, axis=3)\n",
        "print(f'x_train shape: {x_train.shape}')\n",
        "print(f'x_test shape: {x_test.shape}')\n",
        "print('----------')\n",
        "\n",
        "# Split dataset into training and validation data\n",
        "x_val = x_train[int(x_train.shape[0]*0.8):, :, :, :]\n",
        "y_val = y_train[int(y_train.shape[0]*0.8):]\n",
        "x_train = x_train[:int(x_train.shape[0]*0.8), :, :, :]\n",
        "y_train = y_train[:int(y_train.shape[0]*0.8)]\n",
        "print(f'x_train shape: {x_train.shape}, x_val shape: {x_val.shape}')\n",
        "print(f'y_train shape: {y_train.shape}, y_val shape: {y_val.shape}')"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "c00e63ff",
      "metadata": {
        "id": "c00e63ff"
      },
      "source": [
        "## ResNet Arhietecture"
      ]
    },
    {
      "cell_type": "markdown",
      "id": "be25a3b4",
      "metadata": {
        "id": "be25a3b4"
      },
      "source": [
        "![image](https://hackmd.io/_uploads/B16H0kHOT.png)\n",
        "\n",
        "- [source paper](https://arxiv.org/abs/1512.03385)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "0ded32e1",
      "metadata": {
        "id": "0ded32e1"
      },
      "outputs": [],
      "source": [
        "def ResBlock(inputs, blocks_num, filters_num, kernel_size, strides=1):\n",
        "    for i in range(blocks_num):\n",
        "        x = layers.Conv2D(filters_num[0],\n",
        "                          (kernel_size[0], kernel_size[0]),\n",
        "                          strides=strides,\n",
        "                          padding='same')(inputs)\n",
        "        x = layers.BatchNormalization()(x)\n",
        "        strides = 1\n",
        "        for j in range(1, len(filters_num)):\n",
        "            x = layers.Activation('relu')(x)\n",
        "            x = layers.Conv2D(filters_num[j],\n",
        "                              (kernel_size[j], kernel_size[j]),\n",
        "                              strides=strides,\n",
        "                              padding='same')(x)\n",
        "            x = layers.BatchNormalization()(x)\n",
        "\n",
        "        # 確認 Skip connection 維度一致\n",
        "        if inputs.shape.as_list() == x.shape.as_list():\n",
        "            identity = inputs\n",
        "        else:\n",
        "            identity_strides = inputs.shape[1]//x.shape[1]\n",
        "            identity = layers.Conv2D(filters_num[-1], (1, 1),\n",
        "                                     strides=identity_strides)(inputs)\n",
        "            identity = layers.BatchNormalization()(identity)\n",
        "\n",
        "        outputs = layers.add([identity, x])\n",
        "        outputs = layers.Activation('relu')(outputs)\n",
        "        inputs = outputs\n",
        "\n",
        "    return outputs"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "b463453c",
      "metadata": {
        "id": "b463453c"
      },
      "outputs": [],
      "source": [
        "labels_num = 10"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "1407f6d8",
      "metadata": {
        "id": "1407f6d8"
      },
      "outputs": [],
      "source": [
        "tf.keras.backend.clear_session()\n",
        "inputs = layers.Input(shape=x_train.shape[1:])\n",
        "x = layers.Resizing(224, 224,\n",
        "                    interpolation=\"bilinear\",\n",
        "                    input_shape=x_train.shape[1:])(inputs)\n",
        "x = layers.Conv2D(64, (7, 7), strides=2, padding='same')(x)\n",
        "x = layers.BatchNormalization()(x)\n",
        "x = layers.Activation('relu')(x)\n",
        "x = layers.MaxPooling2D((3, 3), strides=2, padding='same')(x)\n",
        "\n",
        "# conv2_x\n",
        "x = ResBlock(x, 3, [64, 64, 256], [1, 3, 1], strides=1)\n",
        "# conv3_x\n",
        "x = ResBlock(x, 4, [128, 128, 512], [1, 3, 1], strides=2)\n",
        "# conv4_x\n",
        "x = ResBlock(x, 6, [256, 256, 1024], [1, 3, 1], strides=2)\n",
        "# conv5_x\n",
        "x = ResBlock(x, 3, [512, 512, 2048], [1, 3, 1], strides=2)\n",
        "\n",
        "x = layers.GlobalAveragePooling2D()(x)\n",
        "outputs = layers.Dense(labels_num)(x)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "40070d68",
      "metadata": {
        "id": "40070d68"
      },
      "outputs": [],
      "source": [
        "ResNet_model = Model(inputs = inputs, outputs = outputs)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "013bfffb",
      "metadata": {
        "id": "013bfffb"
      },
      "outputs": [],
      "source": [
        "ResNet_model.summary()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "8bbda37b",
      "metadata": {
        "id": "8bbda37b"
      },
      "outputs": [],
      "source": [
        "batch_size = 256"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "639e5dc2",
      "metadata": {
        "id": "639e5dc2"
      },
      "outputs": [],
      "source": [
        "inputs = np.ones((batch_size, x_train.shape[1], x_train.shape[2], 3),\n",
        "                 dtype=np.float32)\n",
        "ResNet_model(inputs)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "id": "14a1c0bd",
      "metadata": {
        "id": "14a1c0bd"
      },
      "outputs": [],
      "source": []
    }
  ],
  "metadata": {
    "accelerator": "GPU",
    "colab": {
      "provenance": []
    },
    "gpuClass": "standard",
    "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"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 5
}