sagemaker_handson.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## データの準備\n",
    "\n",
    "### データのダウンロード\n",
    "今回はMNISTというデータセットを利用します。データセットはscikit-learnを利用してダウンロードします。Xとyには画像とラベルが入ります。\n",
    "X.shape や y.shapeとすると、どういった次元なのかがわかります。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.datasets import fetch_openml\n",
    "X, y = fetch_openml('mnist_784', version=1, return_X_y=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### データの分割\n",
    "\n",
    "60000枚と10000枚に画像とラベルを分けて学習データとテストデータを作成しましょう。スライスという機能を使えば配列データを分けることができます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_train = X[:60000]\n",
    "X_test = X[60000:]\n",
    "y_train = y[:60000]\n",
    "y_test = y[60000:]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## kNNの学習\n",
    "\n",
    "データから近いk個のラベルを見て、その多数決でデータのラベルを判定するものです。理論的には全データと比較すれば良いというもので、学習は不要ですが、推論の比較数が多く大変なので、データを木構造にして探索しやすくしています。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.neighbors import KNeighborsClassifier\n",
    "neigh = KNeighborsClassifier(n_neighbors=1)\n",
    "neigh.fit(X_train, y_train) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "上で説明したようにkNNは推論が大変です。最初の5個だけ選んで推論します。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "y_predict = neigh.predict(X_test[0:5])\n",
    "print(y_predict)\n",
    "print(y_test[0:5])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 決定木の学習\n",
    "\n",
    "決定木は、データをルールで分類する木構造の分類器です。分類のルールをデータから学習します。scikit-learnでは、DecisionTreeClassifier()を呼び出してからfit()を実行します。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn import tree\n",
    "\n",
    "clf = tree.DecisionTreeClassifier()\n",
    "clf = clf.fit(X, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "いったん学習が終われば predict で分類を行うことができます。X_testを入力すると予測を得ることができます。ついでに時間を測るとkNNより速いことも確認できると思います。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "y_predict = clf.predict(X_test)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "予測の分類 y_predict は、正解の y_test と合っている方が良い結果ということになります。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.metrics import accuracy_score\n",
    "accuracy_score(y_test, y_predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "多クラス分類のときは、どのクラスがどこに分類されているか（例えば、7が1になってないかなど）が気になると思います。Confusion Matrixは、クラス間の分類結果を示すことができます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.metrics import confusion_matrix\n",
    "confusion_matrix(y_test, y_predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Keras で単純パーセプトロン/多層パーセプトロン\n",
    "\n",
    "単純パーセプトロンは全結合層と活性化層のみで構成されます。MNISTの場合は、10次元に全結合してから、活性化層に通します。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras.layers import Dense, Activation, Dropout\n",
    "\n",
    "model = Sequential()\n",
    "model.add(Dense(10,  input_shape=(784,)))\n",
    "model.add(Activation('softmax'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "モデルの定義ができたら、compileで最適化手法やロスなどを決めます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(optimizer='rmsprop',\n",
    "              loss='categorical_crossentropy',\n",
    "              metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "summary()を使うことでモデルの内容を見ることができます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "学習は同じく fit を使うことで行います。ニューラルネットワークでは、数値を正規化することで学習が安定することがあります。ここでは255に割っておきます。また、出力が10次元なので、それにあわせてone-hotなベクトル(ラベルが6なら[0,0,0,0,0,0,1,0,0,0])に変換します。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.utils import to_categorical\n",
    "model.fit(X_train.astype('float32')/255, to_categorical(y_train), epochs=10, batch_size=32)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "同じようにpredictで予測できます。結果はone-hotなベクトルなので、argmaxで逆にラベルに変換しましょう。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "y_predict = model.predict(X_test)\n",
    "y_predict = np.argmax(y_predict, axis =1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "精度とConfusion_matrix を表示してみます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "score = model.evaluate(X_test.astype('float32')/255, to_categorical(y_test), verbose=0)\n",
    "print('Test loss:', score[0])\n",
    "print('Test accuracy:', score[1])\n",
    "\n",
    "from sklearn.metrics import confusion_matrix\n",
    "confusion_matrix(y_test.astype(int), y_predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## SageMaker を使いこなす\n",
    "\n",
    "ここではノートブックインスタンスで学習しましたが、学習用インスタンスで学習することができます。Kerasの場合は、[こちら](https://github.com/harusametime/sagemaker-notebooks/tree/master/keras_tensorflow)に詳しい説明があります。\n",
    "\n",
    "学習用インスタンスにデータを渡すために、S3にいったんデータを置きましょう。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import sagemaker \n",
    "import os\n",
    "import numpy as np\n",
    "sagemaker_session = sagemaker.Session()\n",
    "\n",
    "os.makedirs(\"./s3_data\", exist_ok = True)\n",
    "np.save(\"./s3_data/image.npy\", X)\n",
    "np.save(\"./s3_data/label.npy\", y.astype(int))\n",
    "\n",
    "input_data = sagemaker_session.upload_data(\"./s3_data\", key_prefix=\"mnist_data\")\n",
    "print(\"Data is uploaded to {}\".format(input_data))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "次に学習に必要なコードを用意します。以下を実行すると、%%writefile train.py　以下の内容が train.py に保存されます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%writefile train.py\n",
    "import argparse\n",
    "import keras\n",
    "import os\n",
    "import numpy as np\n",
    "\n",
    "import tensorflow as tf\n",
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras.utils import to_categorical\n",
    "from tensorflow.keras.layers import Dense, Dropout, Activation\n",
    "from tensorflow.keras import backend as K\n",
    "\n",
    "if __name__ == '__main__':\n",
    "        \n",
    "    parser = argparse.ArgumentParser()\n",
    "\n",
    "    # input data and model directories\n",
    "    parser.add_argument('--model-dir', type=str, default=os.environ['SM_MODEL_DIR'])\n",
    "    parser.add_argument('--train', type=str, default=os.environ['SM_CHANNEL_TRAINING'])\n",
    "    \n",
    "    args, _ = parser.parse_known_args()\n",
    "    \n",
    "    train_dir = args.train\n",
    "    \n",
    "    X = np.load(os.path.join(train_dir, 'image.npy'))\n",
    "    y = np.load(os.path.join(train_dir, 'label.npy'))\n",
    "    \n",
    "    X_train = X[:60000]\n",
    "    X_test = X[60000:]\n",
    "    y_train = y[:60000]\n",
    "    y_test = y[60000:]\n",
    "    \n",
    "    model = Sequential()\n",
    "    model.add(Dense(10,  input_shape=(784,)))\n",
    "    model.add(Activation('softmax'))\n",
    "\n",
    "    model.compile(optimizer='rmsprop',\n",
    "              loss='categorical_crossentropy',\n",
    "              metrics=['accuracy'])\n",
    "    \n",
    "    model.fit(X_train.astype('float32')/255, to_categorical(y_train), epochs=10, batch_size=32)\n",
    "    score = model.evaluate(X_test.astype('float32')/255, to_categorical(y_test), verbose=0)\n",
    "    print('Test loss:', score[0])\n",
    "    print('Test accuracy:', score[1])\n",
    "\n",
    "    '''\n",
    "    Save Keras model as a model for tensorflow serving\n",
    "    '''\n",
    "    sess = K.get_session()\n",
    "    tf.saved_model.simple_save(\n",
    "        sess,\n",
    "        os.path.join(args.model_dir, 'model/1'),\n",
    "        inputs={'inputs': model.input},\n",
    "        outputs={t.name: t for t in model.outputs})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "SageMaker の API を利用して、学習用インスタンスで学習を行います。ここでは、インスタンスタイプや数、上記で作成した学習スクリプト train.py を指定します。S3のファイルパスを指定してfitを実行すれば、学習を行うことできます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sagemaker.tensorflow import TensorFlow\n",
    "from sagemaker import get_execution_role\n",
    "\n",
    "role = get_execution_role()\n",
    "mnist_estimator = TensorFlow(entry_point = \"train.py\",\n",
    "          role=role,\n",
    "          train_instance_count=1,\n",
    "          train_instance_type=\"ml.m4.xlarge\",\n",
    "          framework_version=\"1.14.0\",\n",
    "          py_version='py3',\n",
    "          script_mode=True)\n",
    "\n",
    "mnist_estimator.fit(input_data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "学習が終わればデプロイしてWeb APIを発行することもできます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "predictor = mnist_estimator.deploy(instance_type='ml.m4.xlarge',\n",
    "                                   initial_instance_count=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "APIに対してリクエストを送ることが可能です。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import random\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "num_samples = 5\n",
    "indices = random.sample(range(X_test.shape[0] - 1), num_samples)\n",
    "images, labels = X_test[indices]/255, y_test[indices]\n",
    "\n",
    "for i in range(num_samples):\n",
    "    plt.subplot(1,num_samples,i+1)\n",
    "    plt.imshow(images[i].reshape(28, 28), cmap='gray')\n",
    "    plt.title(labels[i])\n",
    "    plt.axis('off')\n",
    "    \n",
    "prediction = predictor.predict(images)['predictions']\n",
    "prediction = np.array(prediction)\n",
    "predicted_label = prediction.argmax(axis=1)\n",
    "print('The predicted labels are: {}'.format(predicted_label))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "不要な場合はエンドポイントを削除しましょう。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "predictor.delete_endpoint()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "conda_tensorflow_p36",
   "language": "python",
   "name": "conda_tensorflow_p36"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}