MNIST_TPU_20181009.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "Ot79jiI7GiHR"
   },
   "source": [
    "# MNIST with tf.Keras and TPUs on colab\n",
    "\n",
    "Let's try out using `tf.keras` and Cloud TPUs to train a model on the MNIST dataset."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "_6GliKcPt-YL"
   },
   "source": [
    "##  1. Obtain MNIST dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 0,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "Zo-Yk6LFGfSf"
   },
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "\n",
    "(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()\n",
    "\n",
    "# add empty color dimension\n",
    "x_train = np.expand_dims(x_train, -1)\n",
    "x_test = np.expand_dims(x_test, -1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 85
    },
    "colab_type": "code",
    "id": "3qzCxAAgolUg",
    "outputId": "97275b8a-12b9-4b44-aa1d-ebe2468c84b4"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(60000, 28, 28, 1)\n",
      "(60000,)\n",
      "(10000, 28, 28, 1)\n",
      "(10000,)\n"
     ]
    }
   ],
   "source": [
    "print(x_train.shape)\n",
    "print(y_train.shape)\n",
    "print(x_test.shape)\n",
    "print(y_test.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "8bOEkhDJuP94"
   },
   "source": [
    "## 2. Build our conv model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 624
    },
    "colab_type": "code",
    "id": "W7gMbs70GxA7",
    "outputId": "cdafb75c-e6ea-4dab-92f2-836f24481d30"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "batch_normalization (BatchNo (None, 28, 28, 1)         4         \n",
      "_________________________________________________________________\n",
      "conv2d (Conv2D)              (None, 28, 28, 64)        1664      \n",
      "_________________________________________________________________\n",
      "max_pooling2d (MaxPooling2D) (None, 14, 14, 64)        0         \n",
      "_________________________________________________________________\n",
      "dropout (Dropout)            (None, 14, 14, 64)        0         \n",
      "_________________________________________________________________\n",
      "batch_normalization_1 (Batch (None, 14, 14, 64)        256       \n",
      "_________________________________________________________________\n",
      "conv2d_1 (Conv2D)            (None, 14, 14, 256)       409856    \n",
      "_________________________________________________________________\n",
      "max_pooling2d_1 (MaxPooling2 (None, 7, 7, 256)         0         \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 7, 7, 256)         0         \n",
      "_________________________________________________________________\n",
      "flatten (Flatten)            (None, 12544)             0         \n",
      "_________________________________________________________________\n",
      "dense (Dense)                (None, 256)               3211520   \n",
      "_________________________________________________________________\n",
      "activation (Activation)      (None, 256)               0         \n",
      "_________________________________________________________________\n",
      "dropout_2 (Dropout)          (None, 256)               0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 10)                2570      \n",
      "_________________________________________________________________\n",
      "activation_1 (Activation)    (None, 10)                0         \n",
      "=================================================================\n",
      "Total params: 3,625,870\n",
      "Trainable params: 3,625,740\n",
      "Non-trainable params: 130\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = tf.keras.models.Sequential()\n",
    "model.add(tf.keras.layers.BatchNormalization(input_shape=x_train.shape[1:]))\n",
    "model.add(tf.keras.layers.Conv2D(64, (5, 5), padding='same', activation='elu'))\n",
    "model.add(tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2,2)))\n",
    "model.add(tf.keras.layers.Dropout(0.25))\n",
    "\n",
    "model.add(tf.keras.layers.BatchNormalization(input_shape=x_train.shape[1:]))\n",
    "model.add(tf.keras.layers.Conv2D(256, (5, 5), padding='same', activation='elu'))\n",
    "model.add(tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2,2)))\n",
    "model.add(tf.keras.layers.Dropout(0.25))\n",
    "\n",
    "model.add(tf.keras.layers.Flatten())\n",
    "model.add(tf.keras.layers.Dense(256))\n",
    "model.add(tf.keras.layers.Activation('elu'))\n",
    "model.add(tf.keras.layers.Dropout(0.5))\n",
    "model.add(tf.keras.layers.Dense(10))\n",
    "model.add(tf.keras.layers.Activation('softmax'))\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "xLeZATVaNAnE"
   },
   "source": [
    "## 3. Training the conv model  on the TPU\n",
    "\n",
    "Here we demonstrate that we can use a generator function and `fit_generator` to train the model.  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 1043
    },
    "colab_type": "code",
    "id": "pWEYmd_hIWg8",
    "outputId": "6149ac52-7c74-4b6f-b5d8-4ccb8a6f60c7"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:tensorflow:Querying Tensorflow master (b'grpc://10.119.13.42:8470') for TPU system metadata.\n",
      "INFO:tensorflow:Found TPU system:\n",
      "INFO:tensorflow:*** Num TPU Cores: 8\n",
      "INFO:tensorflow:*** Num TPU Workers: 1\n",
      "INFO:tensorflow:*** Num TPU Cores Per Worker: 8\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:CPU:0, CPU, -1, 11463271266344546292)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 17179869184, 10414472442707675354)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:XLA_GPU:0, XLA_GPU, 17179869184, 9304909841950516362)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:0, TPU, 17179869184, 618463094076365007)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:1, TPU, 17179869184, 3841887742257347980)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:2, TPU, 17179869184, 5295905145463105200)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:3, TPU, 17179869184, 15036200185646744149)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:4, TPU, 17179869184, 5413493215292198833)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:5, TPU, 17179869184, 2712746280091338708)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:6, TPU, 17179869184, 2854985160118353061)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:7, TPU, 17179869184, 6079712829953268225)\n",
      "INFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU_SYSTEM:0, TPU_SYSTEM, 17179869184, 4765937294195115079)\n",
      "WARNING:tensorflow:tpu_model (from tensorflow.contrib.tpu.python.tpu.keras_support) is experimental and may change or be removed at any time, and without warning.\n",
      "INFO:tensorflow:Connecting to: b'grpc://10.119.13.42:8470'\n",
      "Epoch 1/10\n",
      "INFO:tensorflow:New input shapes; (re-)compiling: mode=train, [TensorSpec(shape=(256, 28, 28, 1), dtype=tf.float32, name='batch_normalization_input0'), TensorSpec(shape=(256, 1), dtype=tf.float32, name='activation_1_target0')]\n",
      "INFO:tensorflow:Overriding default placeholder.\n",
      "INFO:tensorflow:Remapping placeholder for batch_normalization_input\n",
      "INFO:tensorflow:Started compiling\n",
      "INFO:tensorflow:Finished compiling. Time elapsed: 2.7938554286956787 secs\n",
      "INFO:tensorflow:Setting weights on TPU model.\n",
      "29/30 [============================>.] - ETA: 0s - loss: 4.6330 - sparse_categorical_accuracy: 0.5668INFO:tensorflow:New input shapes; (re-)compiling: mode=eval, [TensorSpec(shape=(256, 28, 28, 1), dtype=tf.float32, name='batch_normalization_input0'), TensorSpec(shape=(256, 1), dtype=tf.float32, name='activation_1_target0')]\n",
      "INFO:tensorflow:Overriding default placeholder.\n",
      "INFO:tensorflow:Remapping placeholder for batch_normalization_input\n",
      "INFO:tensorflow:Started compiling\n",
      "INFO:tensorflow:Finished compiling. Time elapsed: 1.5464982986450195 secs\n",
      "INFO:tensorflow:New input shapes; (re-)compiling: mode=eval, [TensorSpec(shape=(226, 28, 28, 1), dtype=tf.float32, name='batch_normalization_input0'), TensorSpec(shape=(226, 1), dtype=tf.float32, name='activation_1_target0')]\n",
      "INFO:tensorflow:Overriding default placeholder.\n",
      "INFO:tensorflow:Remapping placeholder for batch_normalization_input\n",
      "INFO:tensorflow:Started compiling\n",
      "INFO:tensorflow:Finished compiling. Time elapsed: 1.742654800415039 secs\n",
      "30/30 [==============================] - 13s 448ms/step - loss: 4.5183 - sparse_categorical_accuracy: 0.5693 - val_loss: 14.3644 - val_sparse_categorical_accuracy: 0.1088\n",
      "Epoch 2/10\n",
      "30/30 [==============================] - 4s 140ms/step - loss: 0.3761 - sparse_categorical_accuracy: 0.8855 - val_loss: 6.2120 - val_sparse_categorical_accuracy: 0.2624\n",
      "Epoch 3/10\n",
      "30/30 [==============================] - 4s 138ms/step - loss: 0.1878 - sparse_categorical_accuracy: 0.9480 - val_loss: 1.7997 - val_sparse_categorical_accuracy: 0.5872\n",
      "Epoch 4/10\n",
      "30/30 [==============================] - 4s 141ms/step - loss: 0.1209 - sparse_categorical_accuracy: 0.9663 - val_loss: 0.3454 - val_sparse_categorical_accuracy: 0.8896\n",
      "Epoch 5/10\n",
      "30/30 [==============================] - 4s 140ms/step - loss: 0.1146 - sparse_categorical_accuracy: 0.9673 - val_loss: 0.1714 - val_sparse_categorical_accuracy: 0.9360\n",
      "Epoch 6/10\n",
      "30/30 [==============================] - 4s 139ms/step - loss: 0.0931 - sparse_categorical_accuracy: 0.9750 - val_loss: 0.1518 - val_sparse_categorical_accuracy: 0.9472\n",
      "Epoch 7/10\n",
      "30/30 [==============================] - 4s 137ms/step - loss: 0.0877 - sparse_categorical_accuracy: 0.9734 - val_loss: 0.1335 - val_sparse_categorical_accuracy: 0.9512\n",
      "Epoch 8/10\n",
      "30/30 [==============================] - 4s 141ms/step - loss: 0.0843 - sparse_categorical_accuracy: 0.9747 - val_loss: 0.0517 - val_sparse_categorical_accuracy: 0.9808\n",
      "Epoch 9/10\n",
      "30/30 [==============================] - 4s 142ms/step - loss: 0.0707 - sparse_categorical_accuracy: 0.9798 - val_loss: 0.0420 - val_sparse_categorical_accuracy: 0.9856\n",
      "Epoch 10/10\n",
      "30/30 [==============================] - 4s 133ms/step - loss: 0.0636 - sparse_categorical_accuracy: 0.9823 - val_loss: 0.0483 - val_sparse_categorical_accuracy: 0.9816\n",
      "CPU times: user 7.88 s, sys: 1.7 s, total: 9.57 s\n",
      "Wall time: 51 s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras.callbacks.History at 0x7fa16f3d9e80>"
      ]
     },
     "execution_count": 4,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import os\n",
    "tpu_model = tf.contrib.tpu.keras_to_tpu_model(\n",
    "    model,\n",
    "    strategy=tf.contrib.tpu.TPUDistributionStrategy(\n",
    "        tf.contrib.cluster_resolver.TPUClusterResolver(tpu='grpc://' + os.environ['COLAB_TPU_ADDR'])\n",
    "    )\n",
    ")\n",
    "\n",
    "tpu_model.compile(\n",
    "    optimizer=tf.train.AdamOptimizer(learning_rate=1e-3, ),\n",
    "    loss=tf.keras.losses.sparse_categorical_crossentropy,\n",
    "    metrics=['sparse_categorical_accuracy']\n",
    ")\n",
    "\n",
    "def train_gen(batch_size):\n",
    "  while True:\n",
    "    offset = np.random.randint(0, x_train.shape[0] - batch_size)\n",
    "    yield x_train[offset:offset+batch_size], y_train[offset:offset + batch_size]\n",
    "    \n",
    "\n",
    "%time tpu_model.fit_generator(train_gen(2048), epochs=10, steps_per_epoch=30, validation_data=(x_test, y_test))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "ESL6ltQTMm05"
   },
   "source": [
    "## 4. Let us predict (inference)\n",
    "\n",
    "Now that we're done training, let's see how well we can predict number categories!  Keras/TPU prediction isn't working due to a small bug (fixed in TF 1.12!), but we can predict on the CPU to see how our results look."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 525
    },
    "colab_type": "code",
    "id": "SaYPv_aKId2d",
    "outputId": "e43f8d84-260c-46f2-88e1-b7a04cb8311e"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:tensorflow:Copying TPU weights to the CPU\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAdwAAAHqCAYAAACugP9FAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMS4yLCBo\ndHRwOi8vbWF0cGxvdGxpYi5vcmcvNQv5yAAAIABJREFUeJzs3X2cTdX+B/DPJHmmkJSE1CyZYYZJ\nQyNSv1xFbkk3otxwTYnuJULldqMHcTWSh1Qqza1wI0WpppRwGxFq0Cwkzw/jIQzTzNDs3x/7WK29\nzRlj5py155zzeb9e85rvPmuds9eZNfuss9fea60oy7JAREREwXWe1wUgIiKKBGxwiYiIDGCDS0RE\nZAAbXCIiIgPY4BIRERnABpeIiMgANrhEREQGnO/VjoUQ7QB8XkhSBQANpZTbDReJAkgIkQLgH1LK\nKK/LQiUjhLAAnARQoD38ppTyIY+KRKUkhEgAMB5AAoDfALwspXzO21JFDs8aXCnlNwAq6o8JIZIB\n3A9ghyeFooAQQsQDuM/rclBAdJRSfu11Iaj0hBA1AXwK4E0AtwNoAOAjIcROKWWqp4WLEGWmS1kI\nURvAWAAPSyk5/VWIEkKcB+AVAC96XRYicmgDoAaAJ6SUOVLKnwA8DyDZ22JFjjLT4AJ4CsBCKeU6\nrwtCpZIMIAfAu14XhALiH0KIrUKIo0KIt4UQF3pdIAqoXwHEe12ISFEmGlwhRD0ADwDgtYQQJoS4\nBPYXJ17jCw8rASwDEAP7ml9zADM8LRGVxv8AZAN4VghRWQjREMAgAFWEEBU8LVmEKBMNLoDBAD6V\nUv7sdUGoVF4E8JqUUnpdECo9KWVrKeVEKeVvUsotAEYBuFsIUcnrstG5k1L+CuDPANoB2AdgLoC3\nfcmnvCpXJPHspimXvwAY7XUhqOSEEDcDuA5AP6/LQkGzDUAUgLoAfvG2KFQSUsrlAFqf3hZCdAaw\nV0r5u3elihyen+EKIeIANAKw2OuyUKn0BlAPwC4hxEEAawBACHFQCNHD05LRORNCtBBCTHQ9fA3s\nYUK7PCgSlZIQoqIQ4n4hRA3t4Y4AlntVpkhTFs5wWwI4KqU87HVBqFSGwtlLcTmAb2HfkMG6DT1Z\nAAYIIQ4ASIE9hGQsgFellCc9LRmVVD6AfwGIE0KMgN213A/ALV4WKpKUhQa3LuzrCRTCfNeHfj29\nLYQ43/c4z4ZCkJRyt6+7cRyAJwDkAZjliykESSkLhBDdAbwK4BiA3QD6SSm/9bZkkSPKsjjklYiI\nKNg8v4ZLREQUCdjgEhERGcAGl4iIyAA2uERERAawwSUiIjLA02FBQohmAN4DUFVK2bCIfN0BPAmg\nMYCtAJ6WUs73pUXBHlvWG0At2BMuDJZSbvClXwhgGoAOsL9gfAFgoJTyaHDeVeRifYYXIUR92H/r\nNrDXTv0QwFApZX4heR8G8Ajs8dd7AbwB4PnTK3/5poNMgb24RQd9yT8hxFUA/g2gLeyZrFIBPMrZ\njwLvHOt0MICHAVwBIBP2cZbuS6sG4FkAdwKoCeBLAAOklPt86Y1g13cS7ON0FYBhUsr1QX2DZZxn\nZ7hCiL/AXptx81nyNQfwH9gfwhfDnlzhHSFErC/LQNgLH9wBe6ajFQA+FkKcXmv3Ndgf3PEAmvli\nTsAeYKzPsDQfwEEAV8FuDK8HMMadyTdedzyAvwKoBqAb7IlQ+vrS6wJYjUK+4PvGay8CcNy3n3jY\nUw8+Hug3QwCKX6f3wV5MJhnARQAmA1ikrRY1CUB7ADfCnkthD4B3tJeYC3uhhCthfwmTsI/jqIC/\noxDiZZdyVdjfsr48S74BAD6XUi6QUuZKKT/yPae/L/0hAJOklBlSyhOw/3lqAOjkW72mG4BRUsr9\nUsos2GdW3X3r71LgsD7DiBDiWtizwA2XUh6RUm6H/QE8wLfmse46AOullN9KKQuklD8CSMcfy77V\nhv0F6x+F7cr386RvPzthN7ZcozXAzrFOuwKYJ6VcKqXMk1K+BWAjgO5aeoqU8mcpZTaAIQDaCiGu\nEkKU8+3nXSlltpTyN9i9FlfAPhuOWJ41uFLKN6SUO4qRNQG+eXk1awC08nVTNdXTfdPOZQBoBfuA\ntwD8oD33B9jdVi1KXnpyY32GnQQAO6WUB7XH1sA+22nsyrsYQFMhRAchxPm+XozrACwEACnleinl\nf8+yP30Gnl8B1OOXqIA7lzoFnHUCnLl2rp6eC7uLOt53KeATAA8IIWr6eqfuB7BCSnmolO8hpIXC\nTVO1oE0Z6HMY9rfmi2B/2PpLrwUgW78W5PsAz/alk3msz9Dgr54A19/ad11vCIDPYc/Xuw7Ay1LK\nz4uxHwn7MsRzQogavl6Mx7QyUOAUu05hd/N3F0K0E0JcIIToAuAG/FEniwA8KoRo5PuiPBrABVr6\n/bC7rQ/BbohvhX1fRkQLhQYXsD+ES5oe0dcMyijWZ2go1t9aCNEBwAsAOgGoDHtS/MG+6/pFklKe\ngn3jTV0AO2BfXnjfl8xFEgKvuMfP27DrdBaA/bDr6D38USdDAawFsBL2F6ZcAOu19Lmwu6Avht2Y\nfwHgI193c8QqC4sXnM0BnPlNtxbs1UwOAyjwk57he241IUT50yucCCHKw76xIyuYhSa/WJ+hwV89\nAWf+rR8C8KGU8vT1++VCiHdg30Q192w78t2BfvPpbSFEDOz/g73nXmwqQrHr1Hd3+TO+HwCAEOK/\n8N0U6VuspI/+HCHEMNjLczYF8H8AGpzuvhZCjARwBHa39ncBej+hx7IsEz9+paamWh06dPCbPnbs\nWKtfv36Ox/r27WuNGzfOsizLuv32260ZM2aotLy8PKtFixbWV199ZR06dMi65pprrHXr1qn0NWvW\nWDExMdaRI0eKKlY4MVqnrE8jgl6fGzZssIQQVlZWlnps4cKFVps2bayCggJH3kGDBlmPPfaY47Ex\nY8ZY/fv3dzx2/PhxKzo62kpPT3c8vmjRImvPnj1q+80337Tuuuuu4v0lwoORY/Rc6nTr1q1WWlqa\n2s7Ly7Nat25tff3115ZlWdZ3331nrV271vHaMTEx1tGjR63MzEwrOjra2rFjh0o/fPiwFR0dbf3w\nww8l+gOFoELrpEx2KXfq1AkrV64EAPTo0QMrV65EWloa8vPzsXjxYqxevRo9ethrmvfq1QupqanY\ntGkTcnJykJKSgjp16iApKQk1a9bErbfeikmTJuHQoUM4cOAAXnzxRXTt2hU1atQoqggUQKzP0NO0\naVPEx8djwoQJyM7Oxs6dOzF9+nT06tULUVFRjjq96aab8Pnnn2PVqlU4deoUMjIy8Mknn+CWW4q3\nzOrs2bMxfvx45ObmIjMzEzNmzEDfvn2D+fYi0rnUaVZWFoYOHYq1a9ciPz8f48ePR+3atdG2bVsA\nQHp6OkaOHImsrCwcPnwYY8eORffu3VG9enVceeWVaNiwIV566SUcO3YMJ06cwMsvv4zLL78cTZo0\n8fJP4D1/LXGAf87QsWNHKzY21mratKkVHR1txcbGWrGxsdauXbus6Ohoa8mSJSpvWlqa1alTJysm\nJsbq3LmztXTpUsdrTZkyxUpKSrKaNWtm9enTx9q2bZtKy87OtoYNG2a1bNnSSkhIsEaOHGnl5OSU\n+GtLCDJSp6xPo4wco/v27bOSk5OtuLg4KzEx0Ro3bpx16tQpy7KsM+r0rbfesv70pz9Z8fHxVseO\nHa0ZM2aos6apU6eq/4fo6GgrJibGio2NtaZOnWpZlmXt2LHD6t27txUXF2e1a9fOmjVrViD/VqHA\n2OfuudTpzJkzraSkJCsuLs7q27evtWvXLpWWl5dnjRgxwkpISLBatWpljR492srNzVXp27Zts5KT\nk63ExETruuuus/r162dt2bIlAH+qkFFonZhaD5eL7nonWDcZsU69E4w6ZX16h8do+Cm0TstklzIR\nEVG4YYNLRERkABtcIiIiA9jgEhERGcAGl4iIyAA2uERERAawwSUiIjKADS4REZEBbHCJiIgMCIXV\ngoiIAiI3N1fFhw8fLiLnH2rWrOnYnjlzpopbtmyp4gYNGjjyXXbZZSUpIoUxnuESEREZwAaXiIjI\nAHYpU0jbuXOnijt06OBI27JlS9D2u379esf2FVdcoeLq1asHbb90duvWrVPxf//7X0fawoULVbxh\nw4ZivV7z5s0d25s2bVKx3kXt9vvvvxfr9Sly8AyXiIjIADa4REREBrBLmUJaWlqaiovq3gu0999/\n37F94MABFU+dOtVYOSKJflfxjBkzVPzcc8858v32228qDsR63z/++GOpX4MI4BkuERGREWxwiYiI\nDGCDS0REZEBIXsNNT093bL/00ksqrlevniOtUqVKKu7Tp4+K3bPHuLep7CooKFDxBx984EkZbrjh\nBsf2E088oeL8/HxH2gUXXGCkTOHu4MGDKn7yySeDuq8WLVqouFWrVkHdVyTbtWuXih9//HEVv/PO\nO458+jF/3nnO88Tk5GQVP/PMMyquWrWqI58+DCw2NtaRVr58+XMpdonxDJeIiMgANrhEREQGhGSX\nst41DACbN28u1vOeffZZFdeoUcOR1rp169IXrJgaNmyo4lGjRjnS9BmLqHA//fSTihcvXqziCRMm\nGCtDVlaWY3v16tUqPnXqlCONXcpOOTk5ju3XX39dxTfeeKOK3TM8nX/+Hx9XF110kYqrVavmyJed\nna3ie+65x5EWHx+v4uuvv17FjRo18rsv1l/p6DNuSSkdabfddpuK9e7lqKgoRz69G9md9uqrr6pY\nv4S4Y8cOR7758+ereMmSJY609u3b+38DAcQzXCIiIgPY4BIRERnABpeIiMiAkLyGu2DBAse2vjpI\nTEyMI02/FXzlypUq/vDDDx35PvvsMxXr13N++eWXYpdLv+5z6aWXqlhf0cZNv54LACNGjCj2/iLF\n3r17Hds33XSTips2barihx9+2FiZ5s6da2xf4UAfKtWpUydH2ooVK1T83Xff+X2NK6+8UsX6SlAX\nXnihI9/Ro0dV7F65yX39j4JP//xzX5fX1a9fX8XuVZ70a7NFvX7lypVVfPfddzvyVahQQcV16tQp\nosTBwzNcIiIiA9jgEhERGRCSXcrXXHNNkds6vQujZ8+eKh43bpwj37Zt21Ssdylv3bq12OXShw/o\nXcruIQf6yjJNmjQp9utHKn32GMA57EPvggz28A19FRr3ZQ337DeRzr34uj4bkN6FDAApKSkqLqrL\nUefuRta5h/yRefv27VOxPvzKTe/2HT9+vIr17uWz0YdS/vnPf1bxoUOHHPn01y+qzQgmfkoQEREZ\nwAaXiIjIgJDsUg6EihUrOrb9de2WtOtBvyNan3QdABITE1XcsWPHEr1+uNMXqHBPZN6sWTMVN2jQ\nwFiZ9EUy3F3I3bp1U7F+N2Qk0e9EnjZtmiPt7bffVvEll1ziSBswYICKTU0iT8E1duxYFe/fv1/F\n9913nyPfxIkTVVyrVq0S7Wv79u0q/uabb/zm69KlS4leP5B4hktERGQAG1wiIiID2OASEREZELHX\ncIPhxIkTKr7zzjtVrC+eDACTJk1ScVEzqEQy/Zrf8ePHHWn6QtXBduTIERW//PLLKi5Xrpwjn37N\nyp0WKb799lsVP/roo460xo0bq1hfWQk4834KCj0jR450bL/yyisq1ldzeuGFFxz5SnLd1j3k7Kmn\nnlKxZVkq1j+DASA6Ovqc9xVoPMMlIiIygA0uERGRAexSDqC33npLxfpMK+5uE5NDWUJFbm6uY1tf\nTMJNn00m2N58800V68MbEhISHPk4Yxjw5Zdf+k1r166dit0LClDoW758uWNbXyRCr2/3kLDi0ruR\n9UtygHMhGn2/zz77bIn2FUw8wyUiIjKADS4REZEB7FIuhZ9//tmxPXTo0ELz6XdvAkDdunWDVqZQ\n5b7zUJ89xuQ6t26bN28u9PFWrVoZLknZ99prr/lNe++991Ts7o6//fbbVXz55ZcHvmAUcvTRAYDz\nf2vUqFF+n6cvelAWL93xDJeIiMgANrhEREQGsMElIiIygNdwS2HhwoWO7ZMnT6pYX1j5yiuvNFam\nUOVePP6GG25Qsb7IPOBcCD7QM3Xps4UBwIwZMwrN93//938B3W840IfCuVdT0od9DRo0yJH2yCOP\nqPjJJ59UcYcOHRz5tmzZomJ9Fa+ijq+tW7c6tmNiYlTM4UmB07JlS8e2vtqXPpyubdu2xXq9Xbt2\nObZ37typYn3oj5u+aldZnMGMZ7hEREQGsMElIiIyIEqf7DmIjOzEBL3b2N2tqHd9btiwQcUedyn7\n738pnaDW6cCBA1X86quvOtL0v7s+cXlxrVmzxrG9adMmFbu7IBcvXlzoa8ybN8+xbXL2KwSnTktd\nn+PHj1dxUUM3vKQPybvjjjtUPHXqVC+Kc1pIHqM6/XMRAPr376/i//znPyouqju4KPqCF+7PA31b\nH8bXqFGjEu0rQAp9ozzDJSIiMoANLhERkQFscImIiAzgNdxzpC+srF9nBIB7771Xxfp1C4+F5PWh\nrKwsFbuv06ampqrYvcpQcbhXLNGvK+lDGADngtY6fWgSAJQvX/6cy1EKZfIabkFBgYr1YRwA0KVL\nFxXn5+c70vTr5vprBJte79OnT3ek/e1vfzNWDoToMVpcUkoV69di3fQhW/Hx8Y40/TPgmWeecaQ1\nb95cxStWrFBx5cqVz72wgcNruERERF5hg0tERGQAu5TPYt26dY5tfZWYatWqOdL07pIyNLtU2HVX\n6bPQuGekKY7WrVv7TXOv+DR58uRC8506deqc9xtAZbJLuaR++uknFevDS4YNG+bIV9QC96X117/+\n1bE9c+bMoO2rEGF3jAZauXLlVOweWjR48GAVp6SkGCvTWbBLmYiIyCtscImIiAzg4gWF0O9A7dmz\npyNNXyi9V69ejrQy1I0c1vRFygO9YPnVV19drHx79+51bF966aUBLUck0Rci0Ol3/QPOLuXzz//j\no2v48OGOfMnJySr+97//7UibMmVKictJ5rgXoNe5F51wXwYqy3iGS0REZAAbXCIiIgPY4BIRERnA\na7g++gw3nTt3VrE+SwrgvN709NNPB79gZJR7mJy/YXO8Zht8N998s980fVjW888/70jTV3+aP39+\nsfZVv379cywdBdNzzz3nN+2+++5zbIdS3fEMl4iIyAA2uERERAawS9nn8OHDKv7666/95tMnzq9Z\ns2Ywi0QecM9iU9IFs6n09MXiAeChhx5SsXuxAd28efP8pukzFuldk48//nhJikgBtG/fPhW//vrr\nfvN169bNRHGCgme4REREBrDBJSIiMoANLhERkQERew336NGjjm1/K8i4F5Jv0aJF0MpE3nMvLK+r\nVKmSwZJQ+fLlHdsvvPCCio8dO6biJUuWOPLp025GR0c70vSVZQYOHBiQclJgZGZmqtj9+azfS1Gh\nQgVjZQo0nuESEREZwAaXiIjIgIjtUn7zzTcd21u3bi00X9u2bR3bHCYS3iZOnOjYrlWrloq50oy3\nqlSpouK3335bxe5hfF999ZWKH3vsMb+vQWWLfinA/TnbqlUrFfu7/BcKeIZLRERkABtcIiIiA6L8\nTc4eYEZ2cjabN29Wsd5FATjvetRt27bNsX3FFVcEvFxBFqw+8DJRp4HWp08fx/aoUaNU3KRJE9PF\n8ScYdRqW9RkieIzCefkuPT3dkTZr1iwV9+rVy1iZSqHQOuUZLhERkQFscImIiAxgg0tERGRARA0L\nWrZsmYr9XbMFnIvMc3ahyKJfKyIic1q2bKli9zXccMEzXCIiIgPY4BIRERkQUV3KRbn++utVnJaW\npmJ2KRMRBZ++sLy+kAFw5jDOUMUzXCIiIgPY4BIRERnABpeIiMgAT6d23L17N8aNG4dVq1YBABIT\nE/H444/jkksuMVGmSGFs2jgpJR599FHk5OScsSg4BZSRqR3Xr1+PCRMmYMOGDahYsSJ69+6NBx98\nMAi7jnjGjtG9e/fi6aefxtq1a1GxYkXcfPPNGDlyJC644IIgFSFilb2pHR988EFUqFABX3zxBT7+\n+GMcOXIE//znP70sEpXQJ598gv79+6NBgwZeF4UC4MiRI+jfvz9iYmKwbNkyzJo1C/Pnz8eCBQu8\nLhqVwqBBg3DRRRchLS0N7777LtauXYvJkyd7XayI4VmDe+zYMcTGxmL48OGoWrUqatWqhb/85S/q\nbJdCS05ODubMmYM2bdp4XRQKgHXr1uH48eMYMmQIKlWqhMaNG2PAgAGYM2eO10WjEsrIyMDGjRsx\nfPhwVK9eHfXq1UNycjLmzp2LgoICr4sXEUx1KReLEGI4gL5SymvOmpnKJCHEIADDpJQNvS4LlZwQ\nojOADwBUkVKe9D12J4BUKWVVTwtHJSKESAYwSj82hRBXAvgZQLSUcrO/51JglJmbpoQQAsCTAMZ6\nXRYiwv8AZAN4VghRWQjREMAgAFWEEBU8LRmVVC0Av7oeO+z7XdtwWSJSmWhwhRDXAvgGwEQp5bte\nl4co0kkpfwXwZwDtAOwDMBfA277kU16Vi0otWDdoUTF4PtOUEOJPsA/mkVLK6V6Xh4hsUsrlAFqf\n3vZ1M++VUv7uXamoFA7APsvVnd7OMlyWiOTpGa4QIhHAHAD3s7ElKjuEEBWFEPcLIWpoD3cEsNyr\nMlGprQZQTwhRV3vsOtiN7VZvihRZPDvDFUKcD+ANAE9JKT/0qhxEVKh8AP8CECeEGAG7a7kfgFu8\nLBSVnJRyrRAiHcB4IcRg2Ge3TwKYIqUsO3fPhjHP7lIWQtwA+7ptXmHJUsrthotEpSCEkAAaACgH\n+4vc6XplXYYoIURLAK8CaApgN4AnpZQcFxTChBCXAZgB4CYAOQDegn05j5cJDChTw4KIiIjCVZm4\nS5mIiCjcscElIiIygA0uERGRAWxwiYiIDGCDS0REZICX43DrA5gGoA2A3wB8CGColDK/kLyDATwM\n4AoAmQAGSinTfWnVADwL4E4ANQF8CWCAlHKfEKIBAFnI7isAuFFKuTTgbyyCCSGaAXgPQNWiFi8Q\nQnSHPf6vMewB909LKef70qJgj//sDXuc4BoAg6WUG3zpF8L+v+kA+wvjF7D/H44G511FLhPHqC/9\nKgD/BtAW9tSDqQAe5VCVwOMx6i0vz3DnAzgI4CrYB9r1AMa4Mwkh7gPwHIBkABcBmAxgka9SAWAS\ngPYAbgRQF8AeAO8AgJRyu5Syov4De37YrQBWBu+tRR4hxF8AfAqgyBVHhBDNAfwH9gF7MYDRAN4R\nQsT6sgwE8ACAOwDUA7ACwMdCiIq+9NdgH+TxAJr54hmBfC+kBP0Y9U2AswjAcd9+4mFPJ/l4kN5T\nxOIx6j1PGlzfYgUtAQyXUh7xTYzwHIABQgh3mboCmCelXCqlzJNSvgVgI4DuWnqKlPJnKWU2gCEA\n2vq+Nbv3WwHAFAD/kFLmBuXNRa6qsM+EvjxLvgEAPpdSLpBS5kopP/I9p78v/SEAk6SUGVLKE7A/\n4GsA6CSEuARAN9hLjO2XUmbB/hbeXQjB1U4CyOAxKnw/T/r2sxN2Y5sc7PcYgXiMesyrM9wEADul\nlAe1x9bA/nbcuJD87tk5foX97amw9FzY3V/xONMgADuklAvPucRUJCnlG1LKHcXImgC7rnVrALQS\nQlSCPauRSvetxZoBoBXsOrUA/KA99wfY3ZAtSl56KoTpY1RP/xX2nL8R/wEdSDxGvefVNdyzrcuo\nd3ksAjBFCPEmgHTYE6jfAOAzLf1RIcQ3sJcRGw7gArhWxRBCVAbwGIC/BO5tUAn4q/vasD/Mo4pI\nrwUgW7+2J6U8KYTIBtfzDDRTx6j0vdZzQoiBACrCPk5Pl0Fv8MkMHqNB4uU13OKuy/g2gBcAzAKw\nH/aNF+8BOOlLHwpgLexrspthf3ter6Wfdj/sb+y8Ucp7Z6v7otK5nqc5QT9GpZSnfPnrAtgBu+vy\nfd/z3McwmcNjNBgsyzLx4zBnzhyrXbt2jse2b99uRUdHW9u3b3dnP8PgwYOtiRMn+k1v3bq1tWzZ\nMsdjvXv3tqZNm3bW1w5DRur0tNTUVKtDhw5+C3PPPfdYKSkpjsdSUlKsnj17Wrm5uVaTJk2s//3v\nf470nj17WikpKdby5cutJk2aWPn5+SotPz+/0OeEubA8Rk/btGmT1aRJE+u33347637CBI/R8FNo\nnXhyhhsbG4v9+/fjwIED6rEff/wRtWrVQv369R15f/nlF3zxxRdqOz8/H6tWrUJCQgIAYNWqVVi3\nbp1K37hxI7Kzs9G8eXP12JEjR/D999+jffv2wXpLVEyxsbFYv36947GMjAzExcWhQoUKuPrqq5GR\nkaHS8vPzkZmZifj4eFxzzTWIiorCxo0bVfr69etRrlw5NG3a1Nh7iAQmj9GPP/4Ye/fuVekrVqxA\nTEwMKlasCDKPx2jweNLgNm3aFPHx8ZgwYQKys7Oxc+dOTJ8+Hb169UJUVBQ6deqElSvtUTtZWVkY\nOnQo1q5di/z8fIwfPx61a9dG27ZtAQDp6ekYOXIksrKycPjwYYwdOxbdu3dH9erV1f4yMzPx+++/\n4/LLL/fi7UY8vT579OiBlStXIi0tDfn5+Vi8eDFWr16NHj16AAB69eqF1NRUbNq0CTk5OUhJSUGd\nOnWQlJSEmjVr4tZbb8WkSZNw6NAhHDhwAC+++CK6du2KGjVqFFUEOkcmj9HZs2dj/PjxyM3NRWZm\nJmbMmIG+fft69t4jEY9RQ/yd+gb45wz79u2zkpOTrbi4OCsxMdEaN26cderUKcuyLCs6OtpasmSJ\nyjtz5kwrKSnJiouLs/r27Wvt2rVLpeXl5VkjRoywEhISrFatWlmjR4+2cnNzHftauHCh1axZs5J1\nDIQ+I3XasWNHKzY21mratKkVHR1txcbGWrGxsdauXbvOqM+0tDSrU6dOVkxMjNW5c2dr6dKljtea\nMmWKlZSUZDVr1szq06ePtW3bNpWWnZ1tDRs2zGrZsqWVkJBgjRw50srJyQnMXyp0hNUxumPHDqt3\n795WXFyc1a5dO2vWrFmB+BuFEh6j4afQOjG1Hi4X3fVOsG5gYJ16Jxh1yvr0Do/R8FNonXIuZSIi\nIgPY4BIRERnABpeIiMgANrjszCU4AAAgAElEQVREREQGsMElIiIywLP1cL3wzjvvqPjEiROOtO+/\n/17Fr776qt/XGD16tIpvuukmR9qNN95YyhISEVG44hkuERGRAWxwiYiIDGCDS0REZEDYzzQ1cOBA\nFc+YMSOgr+2ejHv58uUqLkPzhnIWm/DDmabCC4/R8MOZpoiIiLzCBpeIiMiAsOtS1ruQgeJ3I7do\n0ULFd911l4o3b97syDdr1iy/r6EPJ+rXr1+x9msAu6vOwZo1axzbp9d0BYAPPvhAxV27dnXkO+88\no99d2aUcXiL2GM3NzVXxtm3bVLxw4UJHvscee0zF7mMtOTlZxY0aNVKxuy2oUqVKqcp6jtilTERE\n5BU2uERERAaExUxTO3bsUPHrr7/uN1+rVq1U/OmnnzrSKleurOILLrhAxb///rsj35YtW1S8YsUK\nR9rBgweLWWIqS3777TcVd+vWzW++O++8U8X5+fmONMNdykQhQ/8MTU1NdaT961//UvGuXbv8voZ+\nfEVFOXtr/c0MmJWV5dieMGHCWcsabPyUICIiMoANLhERkQFscImIiAwIi2u4+rVT9zAn/brtF198\noeKqVasW67Xfeustx/aqVav85v3zn/9crNeksiUjI0PF27dv95tv0KBBKj7//LA4dELW4cOHVXz0\n6FFH2rx581T82WefOdLKlSunYn2oSVxcnCNfrVq1AlJOAj788EMV9+/fv0Svod8/sWDBgmI9JyUl\nxbHNa7hEREQRgg0uERGRAWE305S7e0kf4lOpUqVzfr02bdo4tr/77ju/eTds2KDiJk2anPO+giRi\nZ7Hx59SpU47tW265RcVff/213+etW7dOxe4uSMMiYqapvXv3OranTp2q4pkzZ6p4//79pd6X+xKB\nPvNcx44dHWn6UBa9i7oUwu4Y3bdvn4pbtmypYvdQHX/cM/r16NFDxS+99JIjTb80UBT3cR9knGmK\niIjIK2xwiYiIDGCDS0REZEDYXcMNBH36MX0lCgDIy8tTsfvazkcffaTi8uXLB6l05yzsrg+Vlnvo\nT8OGDf3m1a/tnTx5MlhFOldhdQ13z549Ktav006fPt2R78iRI4U+v0GDBo7tLl26qLhx48aOtOHD\nh6u4Q4cOKtaHDAJA/fr1Vey+vjtp0iQV33777YWW6RyF/DGqX7MFgJEjR6pY/zx1T8t49dVXq/ir\nr75S8SWXXOLIpz/PPd2uPiXk9ddfr2L3tf1rr71Wxenp6YW8i4DiNVwiIiKvsMElIiIygNPl+Kxd\nu1bFejey3oUMAJdeeqmK3benl6FuZCqCPhPR2ejDESgwnnjiCcf2G2+8oeKihvjcfffdKo6Pj1ex\n3k0MFD0L2LJly1T8yiuvqPj+++935Pvmm29UXK9ePUfaHXfcoeLs7GwV6yuORZrMzEzHtt6NXFBQ\noGJ9mCYAjB49WsV169Yt1r7cQ7H0SwpDhw5Vsd6tDQArV65U8ahRoxxpzz//fLH2XVo8wyUiIjKA\nDS4REZEB7FL2+fbbb1Xs7kbWPfjggyqOjo4OapkoONx3pOrcXV7jxo0LdnHCkntWn9dff13F7r+p\nPlJC71Z88sknHfn0ie/d9VRcu3fvVrF+t6t7YvvmzZureMuWLSXaVyRxX6bR7yrWF4/v1q2bI9+9\n994b0HI8+uijKn7//fcdafosgStWrAjofouLZ7hEREQGsMElIiIygA0uERGRARF7Dbdv376O7Tlz\n5hSab8iQIY7t4q5MQWXL1q1bVbx48WK/+apVq+bYdg8JoeLZuHGjY1s/btyz211xxRUqXrp0qYrd\nM0gVlz4M5dixY460wYMHqzgpKUnFhw4d8vt67vL+4x//UHGFChVKVMZw8Ntvv6n4yy+/LNZz9Htg\ngm3QoEGObffQLy/wDJeIiMgANrhEREQGRFSX8vHjx1Xs7lbMzc1VsT5x9uOPP+7IV9LhCOSt77//\nvlj53ENRqGT0bl2g6Nmf9GNq9erVKp47d64j34YNGwp9vnuGpzVr1qjYXe/6sa0vmlCUyy67zLGt\nz5QVoAXoQ9LPP/+s4k2bNvnN17VrVxUnJiYGtUzF5V5sQW8bqlatGrT98gyXiIjIADa4REREBkRU\nl7I++XlWVpbffI888oiKa9asGdQykRnLly/3m6bXsfvudSqZmJgYx/add96pYndXsd412b17dxW7\n107V6V257vVRi+KvG1mfDQkA+vXrp+KJEyc60oLZ5RhK9MUAivLCCy+ouGLFisEqzjnRRy0AzjWy\n3f+7gcQzXCIiIgPY4BIRERnABpeIiMiAsL+Gqw8L+Prrr/3m01ex0BcxptClr/IyZcoUv/kuuugi\nFVevXj2oZYoU5cuXd2zPnDlTxVOnTnWk6dv6MXrxxRc78jVs2FDF+opeq1atcuQrajUof9zD//Tt\nsnLdsaw5ceKEit2zcenKyqpq+lA19zV7U3iGS0REZAAbXCIiIgPCrktZn1AbAEaNGqXi/Px8v89L\nSEhQMWeTCg9HjhxRsXvmI50+FIWCz91Fqy8arsfF5V5gpKgu5Ro1aqh49uzZKr7lllsc+bzqcgwl\n6enpKi5qCFdZodepV+XlfxUREZEBbHCJiIgMYINLRERkQNhdw33llVcc2/4WRnZP4cehQOEnNTW1\n0Mfd03U+9NBDJopDAfT222+r2D3MqCgffvihitu1axfQMlHocA//MzUckGe4REREBrDBJSIiMiCq\nqBlCAsjITgCgUqVKjm1/Q4GOHj3q2A7jFUCCdf+7sTotrmPHjjm29Rmk9GFB1113nSNfcVc9KUOC\nUadlrj7dPv30UxX36NFDxe5617Vq1cqxvWLFChWff36ZuaIWksfovffeq2L3ClC6U6dOBbMYfrVu\n3dqxrc9I9u677zrS7rnnnkDvvtA65RkuERGRAWxwiYiIDCgzfSqmHT9+3LFdkpllKlSo4Nj2tyi2\nPtG6m3tmrJdeeqlY+9b35Z543T1xfKRYv369Y9vf7FK9evUyURwqJX1RcADo2bOniovqRtbvOHV3\nHZahbuSQN2HCBBUvXbrUkbZ//34VDx8+vNDnBNuePXsc23Xr1lXxXXfdZawcOp7hEhERGcAGl4iI\nyAA2uERERAZE7AWNevXqlfo1HnzwQcf2ZZddpuJ9+/apeNq0aaXeV1Hc76V///5B3V9ZdfDgQb9p\nl1xyiYoj9e8Taj766CPHtnso32lVqlRxbOvDvBo3bhz4ghEA5+fOO++840i7+eabVazP/vfII484\n8tWvXz+gZRo5cqSKd+/e7UjTP6+9upbPM1wiIiID2OASEREZEHZdyu4hH2+++WbQ9uVeKKG49O4M\nfXiP21//+lcVt2nTxm++pKSkEpUj3CxYsMBvmhBCxe7hXFR26EPohg0bVqzn/P3vf3ds63VNZsTF\nxTm2Bw0apOIpU6aoWJ/pC3DOGFZSmZmZKtY/7y+99FJHvn/+85+l3ldp8QyXiIjIADa4REREBrDB\nJSIiMiDsruG+/vrrjm19kWl/Kwe5/fDDDyo+lyE9+hRmV111ld98Xbt2VXGdOnWK/fp0Jn0KzYyM\nDL/59KEjRV03J/P041K//nry5Em/z9HvafjXv/4VlHJR8ekrcwHA008/rWJ9laeBAwc68u3du1fF\nQ4YM8fv6+pC/jRs3OtL0lX4OHTqk4nHjxjny6UMDvcIzXCIiIgPY4BIRERkQdl3Kbvfff3+pnv/y\nyy8HqCQUDFFRf6zz3L59e0fa6tWrVdykSRNjZaJzo18K0FcI0uvWbebMmSrmCkBlz4UXXqji77//\nXsWDBw925Bs7dqyK582bp2J9xigAGDBggIqzsrL87lf/vO/bt+85lNgMnuESEREZwAaXiIjIgCjL\nskzsx8hOqFD+++VKp8zVaXZ2tmN7zJgxKtZn47rjjjuMlSlIglGnntWnfsexvvCA2/jx41Vc3Fmo\nQkTEHKO5ubmObX2hen0mKPdiCO47jnV33323ivXFEM47z9PzyULrlGe4REREBrDBJSIiMoANLhER\nkQG8hhv+Iub6UAQJq2u4jRo1UvG2bdtU7F7tZdOmTSquWrVq0MtlEI/R8MNruERERF5hg0tERGQA\np2ghIk89//zzKu7Zs6eK9YXLgbDrRqYIxDNcIiIiA9jgEhERGeBpgyulRJcuXXDTTTd5WQwKguee\ne86xtimFnr179+LBBx9EYmIi2rdvjzFjxhR7TWkqm/iZ6y3PruF+8skneP7559G8eXP89NNPXhWD\nguCnn37Chx9+6HUxqJQGDRqE6OhopKWlITs7G4MGDcLkyZMDPq1ijx49Co0psPiZ6z3PznBzcnIw\nZ84cxzyqFPoKCgrw1FNP4YEHHvC6KFQKGRkZ2LhxI4YPH47q1aujXr16SE5Oxty5c1FQUOB18agE\n+JnrPc8a3O7du+Oyyy7zavcUJLNnz0bFihXRpUsXr4tCpbBhwwZceumlqFmzpnosJiYGR48exY4d\nOzwsGZUUP3O9Z2qmKb+EEIMADJNSNvS0IFRqQohLAPwAoD2APAC/SCmDNYsOBZEQ4nEAd0spW2iP\nXQjgVwDXSym/9axwVCr8zPUO71KmQHoRwGtSSul1QSgg+GWJKIA48QUFhBDiZgDXAejndVkoIA4A\nqOV67PR2luGyEIUFnuFSoPQGUA/ALiHEQQBrAEAIcVAIwVtPQ89qAPWEEHW1x66D3dhu9aZIRKGN\nDS4FylAA0QDifT+3+R6PB/CRV4WikpFSrgWQDmC8EKKGEOJKAE8CmCKl5Co0RCXg2U1TQggJoAGA\ncrC7tvNOJ0kpt3tSKAoYIURD8KapkCaEuAzADAA3AcgB8BaAkVLK370sF5UMP3O95/ldykRERJGA\nXcpEREQGsMElIiIygA0uERGRAWxwiYiIDGCDS0REZIBnM00JIeoDmAagDYDfAHwIYKiU8owFN4UQ\ngwE8DOAKAJkABkop031p1QA8C+BOADUBfAlggJRyny/9KgD/BtAW9lR1qQAe5dCGwBNCNAPwHoCq\nRc3TKoToDntMZ2PYkyg8LaWc70uLAvAv2BNp1II9gcZgKeUGX/qFsP9vOsD+wvgF7P+Ho8F5V5GN\ndRpeWJ/e8vIMdz6AgwCugt0YXg9gjDuTEOI+AM8BSAZwEYDJABb5KhUAJsGeLP9GAHUB7AHwju+5\n5wNYBOC4bz/xAFoDeDxI7yliCSH+AuBTAJvPkq85gP/APmAvBjAawDtCiFhfloEAHgBwB+yZq1YA\n+FgIUdGX/hrsgzweQDNfPCOQ74VsrNPwwvr0nicNrhDiWgAtAQyXUh7xDbp+DsAAIYS7TF0BzJNS\nLpVS5kkp3wKwEUB3LT1FSvmzlDIbwBAAbX1ntsL386RvPzthN7bJwX6PEagq7N6KL8+SbwCAz6WU\nC6SUuVLKj3zP6e9LfwjAJCllhpTyBOwvYTUAdPKtRtQNwCgp5X4pZRbsb+HdhRC1g/CeIh3rNLyw\nPj3m1RluAoCdUsqD2mNrYJ/BNi4kv3t2jl9hf3sqLD0Xdhe1v/RfYc8RG/GVH0hSyjeklMVZKDUB\nvnmWNWsAtBJCVALQVE+XUp4EkAGgFew6tWAvAXjaD7AvFbQABRTrNLywPr3nVYNbC3bDpzvs++1u\nCBfB/nbUTghxgRCiC4Ab8MfKJYsAPCqEaOT7ZxgN4AJfuoTdffKcbz7YSwA8ppWBzPNX97Vhf+GK\nKiK9FoBs/fq772DPxpn/N2QO6zS8sD6DxMtruMWdY/dtAC8AmAVgP+ybo94DcNKXPhTAWgArYTeu\nuQDWAzgppTzly18XwA7Y3SLv+553+vlk3tnqvqh0zs1cNrFOwwvrMwi8uku52Gtt+lYmecb3AwAQ\nQvwXvgv/UspfAfTRnyOEGAZgly99A4CbtbQYAAUA9gbgfdC581f3WbC/JRf4Sc/wPbeaEKK871sz\nhBDlAVQD12j1Eus0vLA+g8WyLBM/Dhs2bLCEEFZWVpZ6bOHChVabNm2sgoICR96tW7daaWlpajsv\nL89q3bq19fXXX1uWZVnfffedtXbtWsdrx8TEWEePHrUsy7IWLVpk7dmzR6W/+eab1l133eUuUjgz\nUqenpaamWh06dPBbmLFjx1r9+vVzPNa3b19r3LhxlmVZ1u23327NmDFDpeXl5VktWrSwvvrqK+vQ\noUPWNddcY61bt06lr1mzxoqJibGOHDlSnL9FuDBWn5bFOjWAx2j4KbROPOlSbtq0KeLj4zFhwgRk\nZ2dj586dmD59Onr16oWoqCh06tQJK1euBABkZWVh6NChWLt2LfLz8zF+/HjUrl0bbdu2BQCkp6dj\n5MiRyMrKwuHDhzF27Fh0794d1atXBwDMnj0b48ePR25uLjIzMzFjxgz07dvXi7cdsfT67NGjB1au\nXIm0tDTk5+dj8eLFWL16NXr0sNeo79WrF1JTU7Fp0ybk5OQgJSUFderUQVJSEmrWrIlbb70VkyZN\nwqFDh3DgwAG8+OKL6Nq1K2rUqOHlW4w4rNPwwvo0xF9LHOCfM+zbt89KTk624uLirMTERGvcuHHW\nqVOnLMuyrOjoaGvJkiUq78yZM62kpCQrLi7O6tu3r7Vr1y6VlpeXZ40YMcJKSEiwWrVqZY0ePdrK\nzc1V6Tt27LB69+5txcXFWe3atbNmzZpV8u8soclInXbs2NGKjY21mjZtakVHR1uxsbFWbGystWvX\nrjPqMy0tzerUqZMVExNjde7c2Vq6dKnjtaZMmWIlJSVZzZo1s/r06WNt27ZNpWVnZ1vDhg2zWrZs\naSUkJFgjR460cnJyAvOXCh1GjlHWqTE8RsNPoXViaj1cLrrrnWDdwMA69U4w6pT16R0eo+Gn0Drl\nXMpEREQGsMElIiIygA0uERGRAWxwiYiIDGCDS0REZAAbXCIiIgPY4BIRERnABpeIiMgANrhEREQG\neLVaUFg4ePCgY7tOnToq/u9//6viu+66y1iZiMLF5s2bHduNGzdW8YEDB1T82WefOfLNnz9fxd26\ndfP7+m3atFHx1VdfXeJyUmBMmzZNxQ8//LAjTa/HefPmGStToPEMl4iIyAA2uERERAawS7kUpJSO\n7fPO++P7y+WXX266OBHjxIkTKu7du7eK27Vr58j3wAMPqPjCCy8MfsF8cnNzHdsbN25UcVxcnIrL\nlStnrExlWV5enooHDBigYnfXYZUqVQp9zrFjx/y+9kcffeQ3rXLlyiquWrWqI23BggUqbt26td/X\noMD58ssv/abplwncn7tCiKCVKdB4hktERGQAG1wiIiID2KVcCitXrnRsV6tWTcWJiYmmixO23F20\n+t2qhw8fVvGll17qyOdVN3LLli0daXv37lXxli1bVFyrVq3gFywEPPPMMypOTU31my8nJ0fF+t+4\nfv36jnw1atTw+xoFBQUq/s9//lPoawPArbfequINGzY40i677DK/r08lp3cbF8Xd9cwuZSIiInJg\ng0tERGQAG1wiIiIDeA33HOnX45566ilH2pAhQ0wXJ2zp19T69OnjSNNnGfrnP/+pYnd9mDR58mQV\nu4ctfPzxxyrmdVtg3759ju3XXnut0HwNGzZ0bC9evFjF+vX6ihUrOvJdcMEFfvdtWZaK9eE+jzzy\niCPfkSNHVOz+v3r55Zf97puCb+DAgV4XocR4hktERGQAG1wiIiIDovQuliAyshMT0tPTVXz99dc7\n0jIzM1UcHR1trExnERWk1w1qna5fv17F+uxMbtnZ2SrWZw4KNne3qD5UpF+/fo60KVOmqLhChQqB\n2H0w6tTYMfrLL784tvVhXlFRf7w190xTd9xxR9DKlJKS4tgeMWKEik+dOuVIW7NmjYrj4+MDsfuQ\nPEYDTa/7ohhqs0qr0DfDM1wiIiID2OASEREZwAaXiIjIAA4LOkdPPPGEiq+66ipHmnsYAxWfvgIQ\nALz33nt+837++ecq9uq67bXXXus3n76CERCw67Zh4+TJk37Thg4dquJgXrN1cw/pe/XVV1XsHub1\n7rvvqjhA13AJwNSpU1XsXoBepy9UD4TWMCGe4RIRERnABpeIiMgAdimfhT7jDAB89dVXKm7evLkj\nragZbqhoelc9ALz00ksqvvHGGx1pN9xwg4kinWHVqlUq3rNnjyPt0UcfVXH79u2NlSkUDR8+3G9a\nWfnb3X333SrWVzMCgCVLlpguTkQoqhs5XPAMl4iIyAA2uERERAawS/ks9Fll3NwLX1PJuWeZOe+8\nP74LNmjQwJFWrly5oJXDfQftK6+8ouIxY8ao2F3eCRMmBK1M4eDXX39V8ebNmx1pF110kYqbNm1q\nrExFue2221Ts7lImKime4RIRERnABpeIiMgANrhEREQG8BruWehDQdyefvppgyWJXKmpqY5tfYWg\nCy+8UMXu2YKK64svvlCxPosVAHz22WeFPic5OblE+4pUc+bMUbG+qhYA/O1vf1PxlVdeaaxMRKbx\nDJeIiMgANrhEREQGcAH6QmzdulXFiYmJjjR92MKXX37pSDv//DLZQx8Si1vv3LnTsa3PJuVOcxRC\n+/8t7gLWJX2NJk2aqHjZsmWOtJo1a5Zo3yUUcgvQt2zZUsXbt293pK1cuVLF7gVBvJKenq7i66+/\n3pGmv5fVq1cHYnchcYwGW3GPX32RA6DMLl7ABeiJiIi8wgaXiIjIgDLZB+o1vav44MGDjrS4uDgV\nl9Eu5JDknrVLv5N127ZtjrRFixap+LHHHlNx3bp1HfmKe9eyvn5tvXr1/Ob705/+pGLDXchhpVWr\nVo7tstKNTKGhjHYhFwvPcImIiAxgg0tERGQAG1wiIiIDeBGyEPqt/u5b1fXrfRQ8FStWVLE+HMe9\nPWzYsFLvS1/Jxj1Mrm3btirmqjHFl5+f79jOy8vzqCREZQfPcImIiAxgg0tERGQAu5R9jh8/rmJ9\n2Ik+DAgArrvuOmNlIjP0rmL3JYTp06eruHLlysbKFOqWLl3q2P7pp59U7B4CVhbpiy24lS9f3mBJ\nKJzwDJeIiMgANrhEREQGsMElIiIygNdwfd5//30V7927V8U9e/b0ojgURN9++61jOyUlRcU1atRw\npFWvXt1Imchb7hWpUlNT/ebVr+sTnQue4RIRERnABpeIiMgAdin7/Pzzz4U+XqtWLcMloWCbP3++\n37T77rvPsR0KQ1ioZPRu5AkTJjjSDh8+rOLOnTs70po3bx7cgkWobt26qbioY3TatGmO7VBaPYhn\nuERERAawwSUiIjKAXco+/u5KvPPOOw2XhILtvffec2xXqVJFxcOHDzddnLDk7op33/3tlYKCAhW/\n8MILKnZ3UzZo0EDFL7/8siPtvPN4nkIlw/8cIiIiA9jgEhERGcAGl4iIyICIvYa7efNmx/bu3bs9\nKgmZsHDhQhXv2bPHkXbppZeqmMOAAqNJkyaO7UaNGqn44MGDjrScnBwVB2JFJr1+3ddm9VnGvvrq\nK7+v8dlnn6m4YcOGpS4TFU5KqeKihgKFC57hEhERGcAGl4iIyICI7VKeN2+eY/v3339X8Q033KDi\n6OhoY2Wi4Bk3bpyK3YvM9+rVy+/z8vLyVJybm6visjLMJRStWbPGsa3P5BSILv3PP/9cxVlZWX7z\nXXLJJSq+//77HWl6FzgFz+OPP+51EYziGS4REZEBbHCJiIgMYINLRERkQERdwz158qSK58yZ4zdf\nnz59VMxp3MJfuXLlVPzNN9840p566ikVt2jRQsUvvvhi8AsWRl555RUVu6fPXLp0adD26z5+L774\nYhU///zzKv7rX/8atDLQH/RhQEDxhwJlZmaqWAgR0DKZxNaEiIjIADa4REREBkRZlmViP0Z2cjb6\n0J8uXbo40vTZhmbMmKHi8uXLB79gwRV19iwlUibqtLiSkpJUnJ6e7kjTjwH3kKHHHntMxaNGjVJx\n9erVA13EcxGMOjVWn9nZ2Y7tW265RcXfffddqV9fr6fWrVs70m6//fZSv34QRMwx6u5S1mckmzp1\nqopDaVF5PwqtU57hEhERGcAGl4iIyICI6lKOUBHTXVUUfbEKvcsRAG666SYV63eoA0DFihVVrN/N\n7LGQ7lKmM/AYDT/sUiYiIvIKG1wiIiID2OASEREZwGu44Y/Xh8IPr+GGFx6j4YfXcImIiLzCBpeI\niMgANrhEREQGsMElIiIygA0uERGRAWxwiYiIDPCswd27dy8efPBBJCYmon379hgzZgzy8/O9Kg4F\nwPr169GnTx9ce+21aNu2rWPRcQpNM2fORLt27RAfH497770XW7Zs8bpIVApCCMTGxqJZs2bq56mn\nnvK6WBHDswZ30KBBuOiii5CWloZ3330Xa9euxeTJk70qDpXSkSNH0L9/f8TExGDZsmWYNWsW5s+f\njwULFnhdNCqh2bNnY86cOXj99dexYsUKXHvttfwSFQZmzpyJjIwM9fP00097XaSI4UmDm5GRgY0b\nN2L48OGoXr066tWrh+TkZMydOxcFBQVeFIlKad26dTh+/DiGDBmCSpUqoXHjxhgwYADmzJnjddGo\nhF577TX8/e9/R3R0NKpUqYKhQ4fi3//+t9fFIgpZpmaachBCJAMYJaVsqD12JYCfAURLKTf7ey6V\nTUKIzgA+AFBFSnnS99idAFKllFU9LRydMyFEPQC7APQH8A8AlwNYAeBBKeUuL8tGJSeEsAB8CKA5\ngFq++BEp5RFPCxYhvOpSrgXgV9djh32/axsuCwXG/wBkA3hWCFFZCNEQwCAAVYQQFTwtGZXE5b7f\n9wK4DUATABcAeM+zElEgrASwDEAMgATYDe8MT0sUQby8SzlY84eSB6SUvwL4M4B2APYBmAvgbV/y\nKa/KRSV2+vicIKXcKaXcD2AUgLZCiMuLeB6VYVLK1lLKiVLK36SUW2DX6d1CiEpely0SnO/Rfg/A\nPsvVnd7OMlwWChAp5XIArU9v+7qZ90opf/euVFRC+3y/D2uPbfP9vgx2dzOFvm2wv1zVBfCLt0UJ\nf16d4a4GUE8IUVd77DYcG+sAABVfSURBVDrYje1Wb4pEpSGEqCiEuF8IUUN7uCOA5V6ViUplF4Cj\nAFpojzXy/d5uvjhUWkKIFkKIia6HrwFwEvwCZYQnN00BgBDifwC2ABgM++x2IYDZUsqxnhSISkUI\ncR7s+vwAwAjYXcsfAbhFSvmtl2WjkhFCPA+gF4A/wT7jnQ3gpJSyi6cFoxLx3QiXCeBZACkAGsA+\nXr+SUg7ysmyRwstruN0BXARgD+wL+Z8AeM7D8lApSCkLYNdpewDHYN+I0Y+NbUj7J+wP5OWwz4CO\nALjf0xJRiUkpdwPoDKArgIOwb3T8FMAwL8sVSTw7wyUiIooknEuZiIjIADa4REREBrDBJSIiMoAN\nLhERkQFscImIiAzwaqYpAIAQohnsuVmr6gsZFJKvO4AnATSGPTHG01LK+b60KAD/AtAb9njeNQAG\nSyk3+NIvBDANQAfYXzC+ADBQSnk0OO8qcgkh6sP+W7cB8BvsidGHSinPWOhYCDEYwMMAroA9NnCg\nlDLdl1YN9ljBOwHUBPAlgAFSyn2+9EawxxEmwa7TVQCGSSnXB/UNRiAeo+HF1DHqy3MFgHcAtJVS\ncipfeHiGK4T4C+wxYEWuDCSEaA7gP7AP2IsBjAbwjhAi1pdlIIAHANwBoB7sFU0+FkJU9KW/Bvsg\njwfQzBdzsu7gmA97fN9VANoCuB7AGHcmIcR9sMdcJ8Meiz0ZwCLfBy8ATII9nvdG2FPO7YF94J42\nF/ZCCVfCnmRfwq5zHtQBxGM0LBk5RoUQ7QGkA9gZpPcRkrzsUq4K+1vWl2fJNwDA51LKBVLKXCnl\nR77n9PelPwRgkpQyQ0p5AvY/Tw0AnYQQlwDoBnspwP1SyizY38K7CyG4KlEACSGuBdASwHAp5REp\n5XbYB+wA3yxUuq4A5kkpl0op86SUbwHYCHvijNPpKVLKn6WU2QCGwJ40/yohRDnfft6VUmZLKX8D\nkAr7W3jNYL/PCMNjNIyYOkZ96bUBdALwbnDfVWjxrMGVUr4hpdxRjKwJsLugdGsAtPKtcNFUT/et\nxZoBoBXsb8wWgB+05/4Ae7JufY5YKr0EADullAe1x9bA/nbcuJD87hlXfoVdX4Wl58Lu/or3LYTw\nCYAHhBA1fWdJ9wNYIaU8VMr3QBoeo2HHyDEKAFLKeVLKH0td4jDj6TXcYvK3dm5t2P8oUUWk1wKQ\nra9WI6U8KYTIBtfdDbSzrXGsd0suAjBFCPEm7G6njgBuAPCZlv6oEOIb2HP4Doe9FuvpFaXuh30G\ndbqB3QLgloC9EzpXPEZDg8ljlAoRKncpn+3aXFHpvK5nTnH/1m8DeAHALAD7Yd948R7sVUsAYCiA\ntbDn2N4M+9vzei19LuzurYthf1B8AeAjX3czeYPHaGgwdYxSYSzLMvHjV2pqqtWhQwe/6ffcc4+V\nkpLieCwlJcXq2bOnlZubazVp0sT63//+50jv2bOnlZKSYi1fvtxq0qSJlZ+fr9Ly8/MLfU4YM1Kn\nc+bMsdq1a+d4bPv27VZ0dLS1ffv2sxZy8ODB1sSJE/2mt27d2lq2bJm1efNmKzo62tq9e7dKO3bs\nmBUdHW398MMPZ91PmOAxGl7C6hjVLVmyxIqOjj7ra4ehQuukzJ/hxsbGYv1652iPjIwMxMXFoUKF\nCrj66quRkZGh0vLz85GZmYn4+Hhcc801iIqKwsaNG1X6+vXrUa5cOTRt2tTYe4gEsbGx2L9/Pw4c\nOKAe+/HHH1GrVi3Ur1/fkfeXX37BF198obbz8/OxatUqJCQkAABWrVqFdevWqfSNGzciOzsbzZs3\nx++/2z2Pp38DwKlTp4Lynqh4eIyGBlPHKPlXJhvcTp06YeXKlQCAHj16YOXKlUhLS0N+fj4WL16M\n1atXo0ePHgCAXr16ITU1FZs2bUJOTg5SUlJQp04dJCUloWbNmrj11lsxadIkHDp0CAcOHMCLL76I\nrl27okaNGkUVgc5R06ZNER8fjwkTJiA7Oxs7d+7E9OnT0atXL0RFRTnqNCsrC0OHDsXatWuRn5+P\n8ePHo3bt2mjbti0AID09HSNHjkRWVhYOHz6MsWPHonv37qhevTquvPJKNGzYEC+99BKOHTuGEydO\n4OWXX8bll1+OJk2aePkniCg8RkOPqWOUiuDv1DfAP2fo2LGjFRsbazVt2tSKjo62YmNjrdjYWGvX\nrl1WdHS0tWTJEpU3LS3N6tSpkxUTE2N17tzZWrp0qeO1pkyZYiUlJVnNmjWz+vTpY23btk2lZWdn\nW8OGDbNatmxpJSQkWCNHjrRycnJK3E8QgozV6b59+6zk5GQrLi7OSkxMtMaNG2edOnXKsizrjDqd\nOXOmlZSUZMXFxVl9+/a1du3apdLy8vKsESNGWAkJCVarVq2s0aNHW7m5uSp927ZtVnJyspWYmGhd\nd911Vr9+/awtW7YE4E8VMniMhpewO0YfeOABKzY21oqJiXH873z33XcB+HOFhELrxNR6uFx01zvB\nuiGFdeqdYNQp69M7PEbDT6F1Wia7lImIiMING1wiIiID2OASEREZwAaXiIjIADa4REREBoTCXMpE\nFCH0SUxuucU5PfY333yj4j59+qj4jTfeCH7BiAKAZ7hEREQGsMElIiIygA0uERGRAbyG65Ofn6/i\nZ555RsXPPvusI9+NN96o4vnz5zvSOPcr0bnTr9uOGTNGxcuWLXPki4r6Y/Ke1q1bB79gFHQ7d+5U\n8QcffKDit956y5Hvhx9+ULF7dkT9/6J9+/YqjouLc+QbNWqUiuvUqVOyApcSz3CJiIgMYINLRERk\nABcv8Dl06JCK69at6zdfQUGBiufNm+dIu+OOOwJfsNLjxOjhJ6wWL0hNTVXxAw88oOLu3bs78j3/\n/PMqbtSoUfALZk7IH6OffPKJY1vvAk5LS/P7vKVLl6pY7xouqttY7xoGgHvuuUfFsbGxxSxx0HHx\nAiIiIq+wwSUiIjIgYu9SzsnJcWzfd999HpWEAiUvL8+x/dxzz6lY7+JyXwooV65ccAtGRdq9e3eh\nj992222O7TDrRg4rXbp0cWzrXcCVK1dWcatWrRz59BEh+l3FtWrVcuRLTEwMSDm9xjNcIiIiA9jg\nEhERGcAGl4iIyICIuob7/vvvq3j27NmOtKJuXffn888/d2z//vvvKm7evLmKr7766nN+bSoefQWZ\nvn37OtJ++eWXQp+jzyoGAJUqVQp8wajYjh49quILLrhAxR06dPCiOFQC/fv3d2zrKzjp122XLFli\nrExlEc9wiYiIDGCDS0REZEBEzTSlD/8477ySfdfQZ5oq6jX0buTPPvvMkVa/fv0S7buEQn4WG7dj\nx46pODo6WsVZWVmOfPrQBN3AgQMd2+PHj1dxiHQvh/RMU9nZ2Y7tCy+8UMW33nqrihctWmSqSF4L\n+WPUPcwyPj5exXp9Z2ZmOvKF8YIvnGmKiIjIK2xwiYiIDGCDS0REZEDYDwvq3bu3ivXrryWlL1xc\nvXp1R9qWLVtULKVUccOGDR359OFDdO4mT56s4gMHDpzz86dNm+bYfu+99wp9bcC5EgmngAyMCRMm\neF0EAMDWrVtVvH//fr/53MP6ateuHbQyhSp9+kYAGDFihIoHDBigYv3+CyCsr+EWime4REREBrDB\nJSIiMiDsupQ3bdrk2P7+++9VrA/jKe6woCeffNKxffvtt6u4WrVqjjR9tqq///3vfl/zo48+UnHX\nrl2LVY5Ips9EBAATJ04sNF+bNm0c21dccYWK586d6/f1f/31VxU/9thjjjS9fqpWrXr2wtJZvfPO\nO37ThgwZEtB9jR492u++Dx48qGL3sJb/b+++QqNq3jiOzx8sVy+i2AtBCChiDVhQsaERbyxYiCUY\niagXGkUFsaESxYJd7B27qEQlCqLEcmFvsYIFwS5YURSSi/fm/z7vM/Nm13VzdjZ78v1c/Q5zsnvI\n5mTYmTPzaHrZkjF2FSo9XIp/6eWmOkfa/e136tevL7lq1arxX1iS8Q0XAAAP6HABAPAgFDtNffny\nRXLLli2tNv30YbRdovSTiHoTfHdoONpwhh76bNWqleS3b99a5+ndjLZs2WK1DR06VHJAT8Wm/C42\nN27csI47dOggWQ/5FhQUWOeVlpZKLioqkjxlyhTrvEePHkl274fevXtL1lMBSd6RKuV2miopKZGc\nnp5utem/c/3kcDT6Xn758qXV1q1bN8mvXr2K+HN6mNItdn/+/PmI15SWlib57t27ksvxxG3K36PR\ndpp69uyZZHf3N32/6Tb3PszNzS0zG1Nhi9Oz0xQAAMlChwsAgAd0uAAAeBCKZUF656ZoO8ZogwYN\nso537dol2d01JVZ6DmfVqlWSs7KyrPN+/PghOTs722rLzMyUXKtWrbiuI2zcgvF6rmf27NkRf65K\nlX//vPv06SO5Xbt21nluBRNN7ybGTlPxO378uOTXr19bbbNmzYrpNXTVmT179kjOy8uL+DN6aZh7\nrp4LdHeN03JycqzjvXv3Sv706ZPkyrZrkp637dq1q9Wm5227d+8uuU2bNtZ5+r7Uli9fbh3rOfVt\n27ZZbfr/wYsXLyTrOXpjKsZyIr7hAgDgAR0uAAAehGJIOVa9evWSvHXrVqst3mHkSPRykp49e1pt\n586dC/S9wm779u0R244ePSq5ffv2Mb3en/z+u3TpIrlatWox/xxsV65cidjWokWLmF5DFz1YtGiR\nZHepyZAhQySvXr3aanOHGWPRvHnzP/6ZykBPDejlUcYYM3bsWMmbN2/+49d2l2np4euTJ09abQcP\nHpSsC8VkZGRY5xUWFkrWRWh84hsuAAAe0OECAOBB6IaUo9W81cUFEk3vlOLWv412jQsWLJC8Zs2a\n4C8sBbk7y+gnyi9duiTZrY2r6xPv379fst603hj7afCPHz9abUuXLpU8cuRIyfXq1Yvl0vF/7pPJ\nsXA/p40bN5Z53syZM63j+fPnS07Ek+X6qVv3KejKRO/Ol+ga33rKT9eodo9PnTol2X3SWT8hfe3a\nNautSZMmgVzn7/ANFwAAD+hwAQDwgA4XAAAPQjGHq3ceibWwfKLppScXL1602vQ1utc7b968xF5Y\nCnJ3hqpZs6ZkvdzEnVd1l4v8Y9iwYdbx+vXrJeulY8YYc+/ePcnr1q2TvHDhwt9dNhRd0cutBBOp\nYpn+XIyxd3WaNGmS5Pz8/CAuMSJdBcwYY6pXry6Z3ccqFr2cqEePHlabriLVqVMnq+3YsWOSE1l9\nqGL0TgAAhBwdLgAAHoRiSHnfvn1JeV+36LIudu0Wro+kQYMG1jFDVP/lFnvXQ/R60/TPnz9HfA29\n3MpdRqKLHIwePdpqmzZtmuTDhw9Lnjp1qnUehSai08P77lB/pKF/dymRPu/NmzcBXt1/6UIJy5Yt\ns9r0LkqouNzdAw8cOCB5/PjxVtvAgQMl693J3CVI5cU3XAAAPKDDBQDAAzpcAAA8+F+kR/IDltA3\nad26teRHjx5FPK+kpCTQ93WXhuh5wmjS09MlnzhxwmrT26UFpOwJsvLz8ofzOw8fPpS8c+dOq03P\nq+q52GhVf9y/kQkTJkjWW0q6c/QrV66M7YKDkYjPNKGfp16ucebMGatNF3TPysqS/PPnT+s8fW/o\nJUJ6uZYxxowYMUKyO/8fq8zMTMmXL1+22oqLiyU3bdo0rtd3hPoerYjc52/0kqE7d+5ILi0tjfct\nyvxM+YYLAIAHdLgAAHgQimVBPo0aNUryzZs343oNXSg9AUPIlYouXq4LlMeratWq1vGYMWMk6yHl\ngoIC6zxdED3eYcww0ctqjDHm2bNnf/wa7u/x8ePHkvv37y/ZXeKhC5QfOnTIatO7RN2+fVvyjBkz\nrPMuXLgg2R2yDmgYGUnkLhnq27evZP13ETS+4QIA4AEdLgAAHoTiKeVWrVpJ1k+tum7duhWxbcCA\nAZJfvnwZ8TxdPD7eQgmJLtbs4AnIctD3R15enuQNGzZY5+mN9vWTzQmSck8pDx8+XLLescsYe5pG\nFyJxh/cjcYcA09LSJOtdxIwxZvHixZL1DlLuEOOKFSskjxs3LqbrKAfuUc8+fPhgHbdt21Zyo0aN\nJF+/fj3et+ApZQAAkoUOFwAAD+hwAQDwIBRzuAcPHpScnZ0d8bwg5l/jeY05c+ZYx56LzDM/FBBd\nvaZZs2ZWm94V6f3791Zb7dq1g76UlJvD/fbtm2S9BMMYY65duya5X79+kufOnWudV6dOnTJfu7Cw\n0Do+ffq0ZHdXK/3/rk2bNpI3bdpkndehQ4cy3ytBUv4e1f+DjbGroHXv3t3XZUT19etXydOnT7fa\n9PU/f/5ccqS/uRgwhwsAQLLQ4QIA4EEohpT1UIFeImSMMW/fvpUc9JCyfnzcGGM6duwoefPmzZL/\n+usv67xYlzsEJOWHqyqi/fv3W8d6aYtboFwvGQros0+5IWXt+/fv1nHXrl0l379/P6bX0P+3IhWw\nL4suYjF//nzJSd4dLCXv0atXr0ru3Lmz1TZr1izJ+fn5gb6vW3jg/PnzZZ63b98+67ioqEiy+79b\n36MBTScwpAwAQLLQ4QIA4AEdLgAAHoRiDld78uSJdXzkyBHJenlOEHO4R48etdoGDhwY12smWErO\nD1V07jySnr9/8OCB1fbq1SvJDRs2DOLtU3oO1/Xr1y/JZ8+elexu07pkyRLJurpPtDlcXe3JmHIt\n80iklLxH9Rxup06drDb9/7Vu3bqSc3NzrfN0/6PvG/dz2rFjR5k/Y4z9+eu2jIwM67zBgwdLnjx5\nstXmbu0ZAOZwAQBIFjpcAAA8CN2QcjTFxcWS165da7Xt3r1bck5OjmRdIcYYe8hCVyUxxpgaNWoE\ncZlBS8nhqlSjl6bVrFnTaps4caJk9+8uTqEaUkbq36P37t2zjt2KUP/QQ8PGGPPu3TvJegi4cePG\n1nktW7aUrIeojYk8lVe/fn3ruCIsx+QbLgAAHtDhAgDgQaUaUq6kUn64KtVkZWVZx3pz/adPn0qu\nV69evG/BkHK4cI+GD0PKAAAkCx0uAAAe0OECAOBBlWRfABA2eomZMXYFK73rVDnmcAGkIL7hAgDg\nAR0uAAAesCwo/FhyED4sCwoX7tHwYVkQAADJQocLAIAHdLgAAHhAhwsAgAd0uAAAeECHCwCAB76W\nBQEAUKnxDRcAAA/ocAEA8IAOFwAAD+hwAQDwgA4XAAAP6HABAPCADhcAAA/ocAEA8IAOFwAAD+hw\nAQDwgA4XAAAP6HABAPCADhcAAA/ocAEA8IAOFwAAD+hwAQDwgA4XAAAP6HABAPCADhcAAA/ocAEA\n8IAOFwAAD+hwAQDwgA4XAAAP/ga22AUKinyVnQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fa16f43ad68>"
      ]
     },
     "metadata": {
      "tags": []
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "LABEL_NAMES = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']\n",
    "\n",
    "\n",
    "cpu_model = tpu_model.sync_to_cpu()\n",
    "\n",
    "from matplotlib import pyplot\n",
    "%matplotlib inline\n",
    "\n",
    "def plot_predictions(images, predictions):\n",
    "  n = images.shape[0]\n",
    "  nc = int(np.ceil(n / 4))\n",
    "  f, axes = pyplot.subplots(nc, 4)\n",
    "  for i in range(nc * 4):\n",
    "    y = i // 4\n",
    "    x = i % 4\n",
    "    axes[x, y].axis('off')\n",
    "    \n",
    "    label = LABEL_NAMES[np.argmax(predictions[i])]\n",
    "    confidence = np.max(predictions[i])\n",
    "    if i > n:\n",
    "      continue\n",
    "    axes[x, y].imshow(images[i])\n",
    "    axes[x, y].text(0.5, 0.5, label + '\\n%.3f' % confidence, fontsize=14)\n",
    "\n",
    "  pyplot.gcf().set_size_inches(8, 8)  \n",
    "\n",
    "plot_predictions(np.squeeze(x_test[:16]), \n",
    "                 cpu_model.predict(x_test[:16]))"
   ]
  }
 ],
 "metadata": {
  "accelerator": "TPU",
  "colab": {
   "collapsed_sections": [],
   "name": "MNIST-TPU_20181009",
   "provenance": [],
   "version": "0.3.2"
  },
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.6.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}