mnist-kaggle-handson.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Kaggle Digit Recognizer using Keras\n",
    "\n",
    "[Kaggle の Digit Recognizer(mnist)](https://www.kaggle.com/c/digit-recognizer) を [keras](https://keras.io/ja/) を使って解く\n",
    "\n",
    "\n",
    "* (事前にデータをダウンロードして gzip し、  `mnist-train.csv.gz` , `mnist-test.csv.gz` という名前で保存、アップロードしておくこと）\n",
    "* keras のドキュメントの日本語は v1 用で、ところどころ v2 と変更されているので注意（v2 のドキュメントは https://github.com/fchollet/keras-docs-ja で翻訳中。英語のドキュメントが一番正しい ）"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 諸々ライブラリ読み込み\n",
    "import keras\n",
    "\n",
    "# 行列計算ライブラリ numpy\n",
    "import numpy as np\n",
    "# データ解析支援ライブラリ pandas\n",
    "import pandas as pd\n",
    "# グラフ表示 matplotlib\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# データを読み込む\n",
    "train = pd.read_csv(\"mnist-train.csv.gz\")\n",
    "train.head() # 最初の部分を表示してみる"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# ピクセル情報とラベルに分割\n",
    "X_train = (train.values[:,1:]).astype('float32') # ピクセル情報\n",
    "# 28x28 の２次元のピクセル画像が 784の１次元データになっているので2次元x1チャンネルに戻す\n",
    "X_train = X_train.reshape(X_train.shape[0], 28, 28,1)\n",
    "\n",
    "# ラベルは数字で入る\n",
    "y_train = train.values[:,0].astype('int32')\n",
    "# それぞれの次元数を確認してみる\n",
    "print(X_train.shape, y_train.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 画面に表示 画面を cols のグリッドに区切って表示\n",
    "def protout(X, y, cols=10):\n",
    "    for i in range(cols):\n",
    "        plt.subplot(1,cols+1,1+i) # 画面を 1xcols に分割してi番目に描写\n",
    "        plt.imshow(X[i].reshape(28,28), 'gray') # グレイスケール＝チャンネルなしにリシェイプして表示\n",
    "        plt.title(y[i]) # 文字を追加\n",
    "        plt.axis('off') # ルーラーを描写しない\n",
    "\n",
    "protout(X_train, y_train)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "print(y_train[0:3])\n",
    "\n",
    "# ラベルをワンホットベクトルに変更\n",
    "# v2\n",
    "# y_train_c = keras.utils.to_categorical(y_train,10)\n",
    "\n",
    "# v1\n",
    "from keras.utils import np_utils\n",
    "y_train_c = np_utils.to_categorical(y_train,10)\n",
    "\n",
    "print(y_train_c[0:3])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from keras.callbacks import Callback\n",
    "# 学習の途中で protout するクラス\n",
    "class ProtoutCallback(Callback):\n",
    "    def __init__(self, *epochs):\n",
    "        self.epochs=epochs\n",
    "    def on_epoch_begin(self, epoch, logs):\n",
    "        if epoch == 0 and 0 in self.epochs:\n",
    "            self.draw(-1)\n",
    "    def on_epoch_end(self, epoch, logs):\n",
    "        if (epoch+1) in self.epochs:\n",
    "            self.draw(epoch)\n",
    "    def draw(self, epoch):\n",
    "        # 現時点での予測を描写\n",
    "        y = self.model.predict_classes(X_train[0:10], verbose=0)\n",
    "        # エポック数を左上に描写\n",
    "        protout(X_train, y)\n",
    "        plt.subplot(1,11,1)\n",
    "        plt.text(0,-20,\"epoch %d\" % (epoch+1))\n",
    "        plt.show() # 毎回 show する"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "keras では Sequential に [レイヤー](https://keras.io/ja/layers/about-keras-layers/) を追加していく\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# モデルの作成\n",
    "\n",
    "from keras.models import  Sequential\n",
    "from keras.layers.core import  Dense, Flatten, Dropout\n",
    "from keras.preprocessing.image import ImageDataGenerator\n",
    "\n",
    "# モデルを作る関数（初期化が簡単にできるように関数化）\n",
    "def create_model():\n",
    "    # 逐次モデル\n",
    "    model= Sequential()\n",
    "    # Flattenは入力を1D配列に変換します。\n",
    "    model.add(Flatten(input_shape=(28,28,1)))\n",
    "    # Dense は全結合。100次元に変換 活性化関数は relu\n",
    "    model.add(Dense(100, activation='relu'))\n",
    "    # 同じ内容でもう一層\n",
    "    model.add(Dense(100, activation='relu'))\n",
    "    # 更に全結合層。結合した後 softmax で 10 のラベルに分類\n",
    "    model.add(Dense(10, activation='softmax'))\n",
    "    model.compile(\n",
    "        loss='categorical_crossentropy', # ロス関数を設定\n",
    "        optimizer='sgd',# オプティマイザを確率的勾配降下法 (SGD: Stochastic Gradient Descent)に設定\n",
    "        metrics=['accuracy']) # 途中経過の精度測定は「精度」を使う \n",
    "    return model\n",
    "\n",
    "model = create_model()\n",
    "# モデルの概要を表示\n",
    "model.summary()\n",
    "print(\"input shape \",model.input_shape)\n",
    "print(\"output shape \",model.output_shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 学習前はこんな予想結果（乱数なのでむちゃくちゃ）\n",
    "y = model.predict_classes(X_train[0:10])\n",
    "protout(X_train, y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "model = create_model()\n",
    "# 訓練する\n",
    "history = model.fit(X_train, y_train_c,\n",
    "        batch_size=640, # バッチサイズ\n",
    "        epochs=10, # v2 epochs = 10, # 学習エポック数,\n",
    "        validation_split=0.1, # 検証用のデータの割合\n",
    "        callbacks = [ProtoutCallback(3,8)], # epoch 3と8 で途中経過の画像表示\n",
    "                    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 学習後はこんな予想結果\n",
    "y = model.predict_classes(X_train[0:10])\n",
    "protout(X_train[0:10], y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 精度の履歴をプロット\n",
    "plt.plot(history.history['acc'])\n",
    "plt.plot(history.history['val_acc'])\n",
    "plt.title('model accuracy')\n",
    "plt.xlabel('epoch')\n",
    "plt.ylabel('accuracy')\n",
    "plt.legend(['acc', 'val_acc'], loc='lower right')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 損失の履歴をプロット\n",
    "plt.plot(history.history['loss'])\n",
    "plt.plot(history.history['val_loss'])\n",
    "plt.title('model loss')\n",
    "plt.xlabel('epoch')\n",
    "plt.ylabel('loss')\n",
    "plt.legend(['loss', 'val_loss'], loc='upper right')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "y = model.predict_classes(X_train[0:10])\n",
    "protout(X_train, y)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## モデルで十分な精度が出せたらテスト用データで実践してみる"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "test = pd.read_csv(\"mnist-test.csv.gz\")\n",
    "# test には label がない\n",
    "test.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 同じように変換\n",
    "X_test = test.values.astype('float32')\n",
    "X_test = X_test.reshape(X_test.shape[0], 28, 28,1)\n",
    "# 文字を予想する\n",
    "y_test = model.predict_classes(X_test)\n",
    "protout(X_test,y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# 提出用にフォーマットして出力( ImageID, Label )\n",
    "import csv\n",
    "with open(\"results.csv\", \"w\") as f:\n",
    "    f.write(('\"ImageId\",\"Label\"\\n'))\n",
    "    for i, result in enumerate(y_test):\n",
    "        f.write(('%d,\"%d\"\\n'%(i+1,result)))\n",
    "# results.csv をダウンロードすること\n",
    "print(\"ok download `results.csv` and upload to kaggle\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "## 課題\n",
    "\n",
    "* バッチ数やエポック数を変更してみよう\n",
    "    * 精度や学習時間がどのように変化するか観察してみよう\n",
    "    * 同じモデルでも実行タイミングによって結果が変わるのを注意（乱数シード）\n",
    "* 多層にしてみよう https://keras.io/ja/layers/core/\n",
    "    * Dropout 付け足したり、活性化関数を変更してみよう\n",
    "* [最適化関数](https://keras.io/ja/optimizers/)を変更してみよう\n",
    "* [ImageDataGenerator](https://keras.io/ja/preprocessing/image/) を使ってデータを増やしてみよう\n",
    "    * ノイズを加えてみよう\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python3 (anaconda3-4.1.0)",
   "language": "python",
   "name": "anaconda3-4.1.0"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.5.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}

requirements.txt

graphviz==0.7
jupyter==1.0.0
Keras==2.0.4
matplotlib==2.0.2
numpy==1.12.1
pandas==0.20.1
pydot==1.2.3
scikit-learn==0.18.1
tensorflow==1.1.0