{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "9dec772c",
   "metadata": {},
   "source": [
    "# **可解釋 AI (XAI)**\n",
    "此份程式碼會介紹如何使用套件快速觀看模型的可解釋性結果。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c072c238",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "executionInfo": {
     "elapsed": 2610,
     "status": "ok",
     "timestamp": 1677134318954,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "6e368687",
    "outputId": "7629482e-133b-4f64-fb69-f699f4af86da"
   },
   "outputs": [],
   "source": [
    "!pip install tf_keras_vis"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8dc9032a",
   "metadata": {},
   "source": [
    "## 匯入套件"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4fe68cb2",
   "metadata": {
    "executionInfo": {
     "elapsed": 3618,
     "status": "ok",
     "timestamp": 1677134322570,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "TIqunw1JXkbR"
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import cv2\n",
    "import tensorflow as tf\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "from matplotlib import cm\n",
    "\n",
    "# 經典模型相關套件\n",
    "from tensorflow.keras.applications.vgg16 import VGG16\n",
    "from tensorflow.keras.applications.vgg16 import preprocess_input\n",
    "\n",
    "# 可解釋性模型相關套件\n",
    "from tf_keras_vis.gradcam import Gradcam\n",
    "from tf_keras_vis.gradcam_plus_plus import GradcamPlusPlus\n",
    "from tf_keras_vis.scorecam import Scorecam"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4854ac46",
   "metadata": {
    "id": "301i4iu3Roj0"
   },
   "source": [
    "## 載入模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "93986cdb",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "executionInfo": {
     "elapsed": 4294,
     "status": "ok",
     "timestamp": 1677134326862,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "6f5146fe",
    "outputId": "7691e70a-a412-4f19-db36-199b2ccc5b24"
   },
   "outputs": [],
   "source": [
    "model = VGG16(include_top=True,  # 是否要包含分類器\n",
    "              weights='imagenet',  # 在 imagenet 資料集上訓練的權重值\n",
    "              input_shape=(224, 224, 3),\n",
    "              classes=1000)  # 分類數目\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "938659f2",
   "metadata": {
    "id": "_RfoPr3q1AMB"
   },
   "outputs": [],
   "source": [
    "# upload Data\n",
    "!wget -q https://github.com/TA-aiacademy/course_3.0/releases/download/CVCNN_Data/CVCNN_part4.zip\n",
    "!unzip -q CVCNN_part4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "178ea9d5",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 308
    },
    "executionInfo": {
     "elapsed": 3581,
     "status": "ok",
     "timestamp": 1677134337222,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "oUWPCERTubhK",
    "outputId": "0bcea8a6-dd60-4934-ce30-40ffda4be878"
   },
   "outputs": [],
   "source": [
    "# Load images\n",
    "img1 = cv2.imread('Golden_Retriever.jpg')\n",
    "img2 = cv2.imread('Border_Collie.jpg')\n",
    "img3 = cv2.imread('White_Wolf.jpg')\n",
    "\n",
    "image_titles = ['Golden Retriever', 'Border Collie', 'White Wolf']\n",
    "images = []\n",
    "for i, img in enumerate([img1, img2, img3]):\n",
    "    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n",
    "    images.append(cv2.resize(img, (224, 224)))\n",
    "images = np.array(images)  # shape: (3, 224, 224, 3)\n",
    "\n",
    "# Preparing input data for VGG16\n",
    "X = preprocess_input(images)\n",
    "\n",
    "# Plot images\n",
    "f, ax = plt.subplots(nrows=1, ncols=3, figsize=(12, 4))\n",
    "for i, title in enumerate(image_titles):\n",
    "    ax[i].set_title(title, fontsize=16)\n",
    "    ax[i].imshow(images[i])\n",
    "    ax[i].axis('off')\n",
    "plt.tight_layout()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "13e8cf74",
   "metadata": {
    "id": "ckzKo2_NWRMq"
   },
   "source": [
    "## Model modifier, Score function and Render"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "89ddc2e6",
   "metadata": {
    "id": "DaePKbbtWU0F"
   },
   "source": [
    "當 softmax 用於模型的最後一層時，可能會阻礙注意力圖像的生成，因此需替換成線性激活函數。\n",
    "\n",
    "然後，**必須** 定義 score 函式返回目標分數的分數函數 (對應類別的得分值)。\n",
    "\n",
    "- [IMAGENET 1000 Class List](https://deeplearning.cms.waikato.ac.nz/user-guide/class-maps/IMAGENET/)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ea745136",
   "metadata": {
    "executionInfo": {
     "elapsed": 5,
     "status": "ok",
     "timestamp": 1677134337222,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "zJW8weZKW37B"
   },
   "outputs": [],
   "source": [
    "def model_modifier_function(cloned_model):\n",
    "    # Convert last activation from softmax to linear\n",
    "    cloned_model.layers[-1].activation = tf.keras.activations.linear"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9c81d66d",
   "metadata": {
    "executionInfo": {
     "elapsed": 5,
     "status": "ok",
     "timestamp": 1677134337223,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "S_E118eKWOGl"
   },
   "outputs": [],
   "source": [
    "def score_function(output):\n",
    "    # output.shape: (samples, classes)\n",
    "    # Golden Retriever: 207\n",
    "    # Border Collie: 232\n",
    "    # White Wolf: 270\n",
    "    return (output[0][207], output[1][232], output[2][270])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ae576a02",
   "metadata": {
    "executionInfo": {
     "elapsed": 5,
     "status": "ok",
     "timestamp": 1677134337223,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "BfrF5hsmf70H"
   },
   "outputs": [],
   "source": [
    "def images_plot(image_titles, images_array, visualize_cam):\n",
    "    f, ax = plt.subplots(nrows=1, ncols=3, figsize=(12, 4))\n",
    "    for i, title in enumerate(image_titles):\n",
    "        heatmap = np.uint8(cm.jet(visualize_cam[i])[..., :3] * 255)\n",
    "        ax[i].set_title(title, fontsize=16)\n",
    "        ax[i].imshow(images_array[i])\n",
    "        ax[i].imshow(heatmap, cmap='jet', alpha=0.5)\n",
    "        ax[i].axis('off')\n",
    "    plt.tight_layout()\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a268455b",
   "metadata": {
    "id": "I5d-FzNHSKoJ"
   },
   "source": [
    "## GradCAM"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "820310a6",
   "metadata": {
    "id": "x-5N8mR4SYH-"
   },
   "source": [
    "GradCAM 是一種將輸入注意力可視化的方法。它不使用模型輸出的梯度，而是使用倒數第二層的輸出（dense layer 之前的 convolution layer）。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fec61f1a",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 308
    },
    "executionInfo": {
     "elapsed": 9522,
     "status": "ok",
     "timestamp": 1677134346740,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "lpgkqiiXSLR9",
    "outputId": "aa870227-7c44-434f-9139-3f52af2d5d74"
   },
   "outputs": [],
   "source": [
    "# Create Gradcam object\n",
    "gradcam = Gradcam(model, model_modifier=model_modifier_function)\n",
    "\n",
    "# Generate heatmap with GradCAM\n",
    "cam = gradcam(score_function, X, penultimate_layer=-1)\n",
    "\n",
    "# Plot results\n",
    "images_plot(image_titles, images, visualize_cam=cam)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c55a0c5a",
   "metadata": {
    "id": "HYLlNrVJSY42"
   },
   "source": [
    "如結果所示，GradCAM 可以直觀地了解模型的注意力在哪裡，但是會發現，可視化的注意力並沒有完全覆蓋圖片中的類別物件。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1eb9112e",
   "metadata": {
    "id": "8lJfE1cUSjHG"
   },
   "source": [
    "## GradCAM++"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1003ae42",
   "metadata": {
    "id": "-UsWkXZESl3f"
   },
   "source": [
    "GradCAM++ 可以為 CNN 模型預測提供更好的視覺解釋。在多個同類別物件時，GradCAM 較無法正確定位或指定位出部分物件，GranCAM++ 改善了這個問題，對於每個像素的梯度上使用加權平均，而不是全局平均。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "dfae0491",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 308
    },
    "executionInfo": {
     "elapsed": 8,
     "status": "ok",
     "timestamp": 1677134346741,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "76QkU7cnSZeU",
    "outputId": "73598915-deac-45f4-daca-8c6334b69115"
   },
   "outputs": [],
   "source": [
    "# Create GradCAM++ object\n",
    "gradcam = GradcamPlusPlus(model, model_modifier=model_modifier_function)\n",
    "\n",
    "# Generate heatmap with GradCAM++\n",
    "cam = gradcam(score_function, X, penultimate_layer=-1)\n",
    "\n",
    "# Plot results\n",
    "images_plot(image_titles, images, cam)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6fbf2a3f",
   "metadata": {
    "id": "nz7RsaB4TygC"
   },
   "source": [
    "## ScoreCAM"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c19565ee",
   "metadata": {
    "id": "YF_IL3mcTyja"
   },
   "source": [
    "ScoreCAM 是另一種生成 Class Activation Map 的方法。 該方法的特點是不同於 GradCAM 和 GradCAM，他使用無梯度方法進行，類似從特徵圖中取得遮罩重新與輸入進到模型計算，得到所有特徵圖的權重。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "066fc683",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 326
    },
    "executionInfo": {
     "elapsed": 27497,
     "status": "ok",
     "timestamp": 1677134374234,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "9vnu7NB3TzCD",
    "outputId": "45cbcce5-cb58-4da5-dedc-0e0dd8c98fcb"
   },
   "outputs": [],
   "source": [
    "# Create ScoreCAM object\n",
    "scorecam = Scorecam(model, model_modifier=model_modifier_function)\n",
    "\n",
    "# Generate heatmap with ScoreCAM\n",
    "cam = scorecam(score_function, X, penultimate_layer=-1)\n",
    "\n",
    "# Plot results\n",
    "images_plot(image_titles, images, cam)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8fff4aaa",
   "metadata": {
    "id": "-eC8hkcmT_Em"
   },
   "source": [
    "## Faster-ScoreCAM"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "aae59722",
   "metadata": {
    "id": "uAo-aeycT_IF"
   },
   "source": [
    "ScoreCAM 比其他方法需要更多的時間來處理。  \n",
    "比 ScoreCAM 更快速的 Faster-ScorecAM 由 @tabayashi0117 設計。\n",
    "\n",
    "https://github.com/tabayashi0117/Score-CAM/blob/master/README.md#faster-score-cam\n",
    "\n",
    "Faster-Score-CAM 在 Score-CAM 中增加了“只使用方差較大的 channel 作為 mask 圖像”的處理。 （max_N = -1 是原始的 Score-CAM）。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "465c3e29",
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 326
    },
    "executionInfo": {
     "elapsed": 5324,
     "status": "ok",
     "timestamp": 1677134379547,
     "user": {
      "displayName": "吳承澔",
      "userId": "17428420001093174904"
     },
     "user_tz": -480
    },
    "id": "Q7YFDJpqT_Ti",
    "outputId": "7e14b5c2-ef1b-4d9d-8fe8-8d8cc4a8c919"
   },
   "outputs": [],
   "source": [
    "# Create ScoreCAM object\n",
    "scorecam = Scorecam(model, model_modifier=model_modifier_function)\n",
    "\n",
    "# Generate heatmap with Faster-ScoreCAM\n",
    "cam = scorecam(score_function, X, penultimate_layer=-1, max_N=10)\n",
    "\n",
    "# Plot results\n",
    "images_plot(image_titles, images, cam)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0aa95125",
   "metadata": {
    "id": "eypdhJVr8AEw"
   },
   "source": [
    "參考資源\n",
    "- [reference](https://github.com/keisen/tf-keras-vis)"
   ]
  }
 ],
 "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
}