Chap07-Advanced_deep-learning_best_practices/7.3-getting-the-most-out-of-your-models.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'2.2.4'"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import keras\n",
    "\n",
    "keras.__version__"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 7.3 모델의 성능을 최대로 끌어올리기\n",
    "\n",
    "- 머신러닝 경연 대회에서 우승하는 모델을 만들기 위해 꼭 알아야 할 기법들을 알아보도록 하자."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 7.3.1 고급 구조 패턴"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 1) 배치 정규화 (Batch Normalization)\n",
    "\n",
    "- [[링크](http://excelsior-cjh.tistory.com/178?category=940400)] 참고\n",
    "\n",
    "- **정규화**(normalization)는 머신러닝 모델에 주입되는 샘플들을 균일하게 만드는 방법이다.\n",
    "\n",
    "- 모델이 학습하고 새로운 데이터에 잘 일반화되도록 도와준다.\n",
    "\n",
    "- 데이터에서 평균($\\mu$)을 빼서 데이터를 원점에 맞추고 표준편차($\\sigma$)로 나누어 데이터의 분산을 1로 만들어 준다.\n",
    "\n",
    "$$\n",
    "\\text{N} \\left( \\mu, \\sigma^{2} \\right) \\rightarrow \\text{N} \\left( 0, 1 \\right)\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- 데이터 정규화는 네트워크에서 일어나는 모든 변환 후에도 고려되어야 한다.\n",
    "\n",
    "    - `Dense`나 `Conv2D` 층에 들어가는 데이터의 평균이 0이고 분산이 1이더라도 출력되는 데이터가 동일한 분포를 가질 것이라고 기대하기 어렵다."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- **배치 정규화**(batch normalization)는 2015년 Sergety Ioffe와 Christian Szegedy가 제안한 층(layer)의 한 종류다. → [[링크](https://arxiv.org/pdf/1502.03167.pdf)] 참고\n",
    "\n",
    "- 훈련하는 동안 평균과 분산이 바뀌더라도 이에 적응(학습)하여 데이터를 정규화한다.\n",
    "\n",
    "- 훈련 과정에서 사용된 배치 데이터의 평균과 분산에 대한 지수 이동 평균(exponential moving average)을 내부에 유지한다.\n",
    "\n",
    "- 배치 정규화의 주요 효과는 입력과 출력의 분포를 유지하도록 하여 그래디언트의 전파를 도와주는 것이다.\n",
    "\n",
    "- `BatchNormalization` 층은 일반적으로 합성곱이나 완전 연결 층 다음에 사용한다.\n",
    "\n",
    "\n",
    "- `BatchNormalization(axis=-1)` 클래스에는 정규화할 특성 축을 지정하는 `axis` 매개변수가 있다.\n",
    "\n",
    "    - 이 매개변수의 기본값은 입력 텐서의 마지막 축을 나타내는 `-1`이다.\n",
    "    \n",
    "    - `Conv2D`에서 `data_format`을 `\"channels_last\"`로 하여 `Dense, Conv1D, RNN, Conv2D` 층을 사용할 때 맞는 값이다."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![](./images/bn02.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 케라스에서 배치정규화 사용 예제"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras import layers, models\n",
    "from keras import backend as K\n",
    "\n",
    "K.clear_session()\n",
    "\n",
    "model = models.Sequential(\n",
    "    [layers.Dense(128, input_shape=(784,)),\n",
    "     layers.BatchNormalization(),\n",
    "     layers.Activation('relu')]\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from IPython.display import Image\n",
    "from keras.utils import plot_model\n",
    "\n",
    "plot_model(model, to_file='./plot-model/bn.png')\n",
    "Image(filename='./plot-model/bn.png')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 2) 깊이별 분리 합성곱 (Depthwise Separable Convolution)\n",
    "\n",
    "\n",
    "- [깊이별 분리 합성곱](https://keras.io/layers/convolutional/)(`SeparableConv2D`)은 `Conv2D`를 대체하면서 더 가볍고(훈련할 가중치가 더 적고) 더 빠른 합성곱 층이다.\n",
    "\n",
    "\n",
    "- `SeparableConv2D`의 합성곱 연산은 다음과 같다.\n",
    "\n",
    "    1. `SeparableConv2D` 층은 입력 **채널(깊이, depth)별**로 따로따로 공간 방향의 합성곱(spatial convolution)을 수행한다. \n",
    "\n",
    "    2. 그 다음 **점별 합성곱**(pointwise conv, $1 \\times 1$ 합성곱)을 통해 출력 채널(깊이)를 합친다.\n",
    "        - 이것은 공간 특성의 학습과 채널 방향 특성의 학습을 분리하는 효과를 낸다.\n",
    "        \n",
    "        - 입력에서 공간상의 위치는 상관관계가 크지만 채널별로는 매우 독립적이라고 가정한다면 타당하다.\n",
    "        \n",
    "\n",
    "- 이 방법은 모델 파라미터와 연산의 수를 크게 줄여 주기 때문에 작고 더 빠른 모델을 만든다.\n",
    "\n",
    "\n",
    "- 합성곱을 통해 더 효율적으로 표현을 학습하기 때문에 적은 데이터로도 더 좋은 표현을 학습하고, 성능이 좋은 모델을 만든다.\n",
    "\n",
    "    - 제한된 데이터로 작은 모델을 처음부터 훈련시킬 때 유용하다.\n",
    "    \n",
    "    \n",
    "- 다음의 예제는 작은 데이터셋에서 이미지 분류 문제(소프트맥스 분류)를 위한 가벼운 깊이별 분리 컨브넷을 만드는 예제이다."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![](./images/depthwise02.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### 케라스를 이용한 `SeparableConv2D` 구현하기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_x.shape : (60000, 28, 28, 1)\n",
      "test_x.shape : (10000, 28, 28, 1)\n"
     ]
    }
   ],
   "source": [
    "# Minist Dataset load\n",
    "from keras.datasets import mnist\n",
    "\n",
    "(train_x, train_y), (test_x, test_y) = mnist.load_data()\n",
    "train_x = train_x.reshape([-1, 28, 28, 1]) / 255.\n",
    "test_x = test_x.reshape([-1, 28, 28, 1]) / 255.\n",
    "\n",
    "print('train_x.shape :', train_x.shape)\n",
    "print('test_x.shape :', test_x.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras import layers\n",
    "from keras import backend as K\n",
    "from keras.models import Sequential\n",
    "\n",
    "\n",
    "K.clear_session()\n",
    "\n",
    "#################\n",
    "# layers params #\n",
    "#################\n",
    "height, width, channels = (28, 28, 1)\n",
    "num_classes = 10\n",
    "\n",
    "# Model\n",
    "model = Sequential()\n",
    "model.add(layers.SeparableConv2D(32, 3, padding='same', activation='relu', \n",
    "                                 input_shape=(height, width, channels,)))\n",
    "model.add(layers.MaxPooling2D(2))\n",
    "\n",
    "model.add(layers.SeparableConv2D(64, 3, padding='same', activation='relu'))\n",
    "model.add(layers.MaxPooling2D(2))\n",
    "\n",
    "model.add(layers.SeparableConv2D(128, 3, padding='same', activation='relu'))\n",
    "model.add(layers.GlobalAveragePooling2D())\n",
    "\n",
    "model.add(layers.Dense(32, activation='relu'))\n",
    "model.add(layers.Dense(num_classes, activation='softmax'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from IPython.display import Image\n",
    "from keras.utils import plot_model\n",
    "\n",
    "plot_model(model, to_file='./plot-model/depthwise.png')\n",
    "Image(filename='./plot-model/depthwise.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 48000 samples, validate on 12000 samples\n",
      "Epoch 1/5\n",
      "48000/48000 [==============================] - 44s 921us/step - loss: 0.8487 - acc: 0.7017 - val_loss: 0.1883 - val_acc: 0.9461\n",
      "Epoch 2/5\n",
      "48000/48000 [==============================] - 41s 859us/step - loss: 0.2009 - acc: 0.9379 - val_loss: 0.1285 - val_acc: 0.9620\n",
      "Epoch 3/5\n",
      "48000/48000 [==============================] - 40s 834us/step - loss: 0.1379 - acc: 0.9567 - val_loss: 0.1265 - val_acc: 0.9625\n",
      "Epoch 4/5\n",
      "48000/48000 [==============================] - 37s 780us/step - loss: 0.1065 - acc: 0.9668 - val_loss: 0.1767 - val_acc: 0.9451\n",
      "Epoch 5/5\n",
      "48000/48000 [==============================] - 38s 784us/step - loss: 0.0925 - acc: 0.9714 - val_loss: 0.1021 - val_acc: 0.9713\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x121c964e0>"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from keras.optimizers import Adam\n",
    "\n",
    "model.compile(optimizer=Adam(lr=0.01),\n",
    "              loss='sparse_categorical_crossentropy',\n",
    "              metrics=['acc'])\n",
    "\n",
    "model.fit(train_x, train_y,\n",
    "          batch_size=64,\n",
    "          epochs=5,\n",
    "          validation_split=0.2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 7.3.2 하이퍼파라미터 최적화\n",
    "\n",
    "- 최적의 하이퍼파라미터 값을 찾기위해 옵션을 수정하고 모델을 반복적으로 다시 훈련하여 값을 개선해 나가야 한다.\n",
    "\n",
    "- 하이퍼파라미터의 가능한 결정공간을 자동적, 조직적, 규칙적 방법으로 탐색해야 한다.\n",
    "\n",
    "\n",
    "- 전형적인 하이퍼파라미터 최적화 과정은 다음과 같다.\n",
    "\n",
    "    1. 일련의 하이퍼파라미터를 (자동으로)선택한다.\n",
    "    2. 선택된 하이퍼파라미터로 모델을 만든다.\n",
    "    3. 훈련 데이터에 학습하고 검증 데이터에서 최종 성능을 측정한다.\n",
    "    4. 다음으로 시도할 하이퍼파라미터를 (자동으로) 선택한다.\n",
    "    5. 이 과정을 반복한다.\n",
    "    6. 마지막으로 테스트 데이터에서 성능을 측정한다.\n",
    "    \n",
    "    \n",
    "- 하이퍼파라미터를 선택하는 알고리즘으로는 베이지안 최적화(bayesian optimization), 유전 알고리즘(genetic algorithms), 간단한 랜덤 탐색(random search) 등이 있다.\n",
    "\n",
    "- 이 분야가 아직 초창기(?)이기 때문에 모델 최적화에 사용할 도구가 매우 적다. \n",
    "\n",
    "- 가장 단순하지만 종종 랜덤 탐색이 제일 좋은 방법일 때가 많다.\n",
    "\n",
    "- 케라스에서 [`keras.wrappers.scikit_learn`](https://keras.io/scikit-learn-api/) 모듈 아래에 있는 `KerasClassifier`와 `KerasRegressor` 클래스를 사용하면 Scikit-Learn의 `RandomizedSearchCV`를 사용하여 랜덤한 하이퍼파라미터 탐색을 수행할 수 있다.\n",
    "\n",
    "- [Hyperas](https://github.com/maxpumperla/hyperas)는 `Hyperopt`라는 하이퍼파라미터 탐색 모듈과 케라스를 연동하여 케라스 모델에 사용할 수 있다.\n",
    "\n",
    "- 전체적으로 보았을 때 하이퍼파라미터 최적화는 어느 작업에서 최고의 모델을 얻거나 머신러닝 경연 대회에서 우승하기 위한 강력한 도구이다."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### `KerasClassifier`를 이용한 Scikit-Learn의 `RandomizedSearchCV` 구현하기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "train_x.shape : (60000, 784)\n",
      "test_x.shape : (10000, 784)\n"
     ]
    }
   ],
   "source": [
    "# Minist Dataset load\n",
    "from keras.datasets import mnist\n",
    "\n",
    "(train_x, train_y), (test_x, test_y) = mnist.load_data()\n",
    "train_x = train_x.reshape([60000, -1]) / 255.\n",
    "test_x = test_x.reshape([10000, -1]) / 255.\n",
    "\n",
    "print('train_x.shape :', train_x.shape)\n",
    "print('test_x.shape :', test_x.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.model_selection import RandomizedSearchCV\n",
    "from keras import backend as K\n",
    "from keras import layers, models\n",
    "from keras.wrappers.scikit_learn import KerasClassifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "K.clear_session()\n",
    "\n",
    "# Function to create model, required for KerasClassifier\n",
    "def mnist_model(optimizer='rmsprop'):\n",
    "    model = models.Sequential()\n",
    "    model.add(layers.Dense(256, activation='relu', input_shape=(784,)))\n",
    "    model.add(layers.Dense(128, activation='relu'))\n",
    "    model.add(layers.Dense(10, activation='softmax'))\n",
    "    model.compile(optimizer='rmsprop', \n",
    "                  loss='sparse_categorical_crossentropy',\n",
    "                  metrics=['acc'])\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 17min 22s, sys: 1min 14s, total: 18min 37s\n",
      "Wall time: 10min 36s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "# create model using KerasClassifier\n",
    "model = KerasClassifier(build_fn=mnist_model, verbose=0)\n",
    "\n",
    "pram_grid = {\n",
    "    'batch_size': [32, 64],\n",
    "    'epochs': [5, 10],\n",
    "    'optimizer': ['SGD', 'rmsprop', 'Adam']\n",
    "}\n",
    "\n",
    "random_search = RandomizedSearchCV(estimator=model, param_distributions=pram_grid, cv=3)\n",
    "search_result = random_search.fit(train_x, train_y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.9753"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "search_result.best_score_"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'optimizer': 'SGD', 'epochs': 10, 'batch_size': 64}"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "search_result.best_params_"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([7, 2, 1, ..., 4, 5, 6], dtype=uint8)"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Call predict on the estimator with the best found parameters.\n",
    "pred = search_result.predict(test_x)\n",
    "pred"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### `Hyperas`를 이용한 하이퍼파라미터 최적화 구현하기\n",
    "\n",
    "- [[링크](https://github.com/ExcelsiorCJH/Deep-Learning-with-Python/blob/master/Chap07-Advanced_deep-learning_best_practices/hyperas-test.ipynb)] 참고"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 7.3.3 모델 앙상블\n",
    "\n",
    "- **모델 앙상블**(moel ensemble)은 여러 개의 다른 모델의 예측을 합쳐서 더 좋은 예측을 만드는 방법이다.\n",
    "\n",
    "- 앙상블은 단일 모델보다 성능이 더 좋다.\n",
    "\n",
    "- 앙상블 방법 중 하나는 각 모델이 추론할때 나오는 예측을 평균 내는 것이다.\n",
    "\n",
    "    - 이 방식은 분류기들이 어느 정도 비슷하게 좋을 때 잘 작동한다.\n",
    "    \n",
    "    - 분류기 중 하나가 다른 모델보다 월등히 나쁘면 좋지 않을 수 잇다.\n",
    "\n",
    "```python\n",
    "preds_a = model_a.predict(val_x)\n",
    "preds_b = model_b.predict(val_x)\n",
    "preds_c = model_c.predict(val_x)\n",
    "preds_d = model_d.predict(val_x)\n",
    "\n",
    "# 앙상블은 어떤 초기 예측보다 더 정확해야 한다.\n",
    "final_preds = 0.25 * (preds_a + preds_b + preds_c + preds_d)\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- 분류기를 앙상블하는 좋은 방법은 검증 데이터에서 학습된 가중치를 사용하여 가중 평균하는 것이다.\n",
    "\n",
    "    - 전형적으로 분류기가 좋을수록 높은 가중치를 가지고 나쁜 분류기일 수록 낮은 가중치를 가진다.\n",
    "    \n",
    "    - 좋은 앙상블 가중치를 찾기위해 랜덤 서치나 [넬더-미드(Nelder-Mead)](https://github.com/fchollet/nelder-mead) 방법 같은 간단한 최적화 알고리즘을 사용할 수 있다.\n",
    "    \n",
    "\n",
    "```python\n",
    "preds_a = model_a.predict(val_x)\n",
    "preds_b = model_b.predict(val_x)\n",
    "preds_c = model_c.predict(val_x)\n",
    "preds_d = model_d.predict(val_x)\n",
    "\n",
    "# 앙상블은 어떤 초기 예측보다 더 정확해야 한다.\n",
    "final_preds = 0.5 * preds_a + 0.25 * preds_b + 0.1 * preds_c + 0.15 * preds_d\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    " "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- 앙상블이 잘 작동하게 만드는 핵심은 분류기의 **다양성**이다.\n",
    "\n",
    "    - 모든 모델이 같은 방향으로 편향되어 있다면 앙상블은 동일한 편향을 유지할 것이다.\n",
    "    - 모델이 서로 다른 방향으로 편향되어 있다면 편향은 서로 상쇄되고 앙상블이 더 견고하고 정확해질 것이다.\n",
    "\n",
    "\n",
    "- 따라서, 가능한 최대한 다르면서 좋은 모델들을 가지고 앙상블해야 한다.\n",
    "\n",
    "    - 일반적으로 매우 다른 구조를 가지거나 다른 종류의 머신러닝 방법을 말한다.\n",
    "    \n",
    "  \n",
    "- 같은 네트워크를 랜덤 초기화를 다르게 하여 따로따로 여러번 훈련해서 앙상블하는 것은 해볼 가치도 없다.\n",
    "\n",
    "- 최근에 실전에서 매우 성공적으로 사용되는 기본 앙상블 스타일은 딥러닝과 머신러닝 모델을 섞은 모델이다."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "tensorflow",
   "language": "python",
   "name": "tensorflow"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}