dlnd_image_classification.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "# 图像分类\n",
    "\n",
    "在此项目中，你将对 [CIFAR-10 数据集](https://www.cs.toronto.edu/~kriz/cifar.html) 中的图片进行分类。该数据集包含飞机、猫狗和其他物体。你需要预处理这些图片，然后用所有样本训练一个卷积神经网络。图片需要**标准化（normalized）**，标签需要采用 **one-hot 编码**。你需要应用所学的知识构建**卷积的、最大池化（max pooling）**、**丢弃（dropout）**和**完全连接（fully connected）**的层。最后，你需要在样本图片上看到神经网络的预测结果。\n",
    "\n",
    "\n",
    "## 获取数据\n",
    "\n",
    "请运行以下单元，以下载 [CIFAR-10 数据集（Python版）](https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz)。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "All files found!\n"
     ]
    }
   ],
   "source": [
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "from urllib.request import urlretrieve\n",
    "from os.path import isfile, isdir\n",
    "from tqdm import tqdm\n",
    "import problem_unittests as tests\n",
    "import tarfile\n",
    "\n",
    "cifar10_dataset_folder_path = 'cifar-10-batches-py'\n",
    "\n",
    "# Use Floyd's cifar-10 dataset if present\n",
    "floyd_cifar10_location = '/input/cifar-10/python.tar.gz'\n",
    "if isfile(floyd_cifar10_location):\n",
    "    tar_gz_path = floyd_cifar10_location\n",
    "else:\n",
    "    tar_gz_path = 'cifar-10-python.tar.gz'\n",
    "\n",
    "class DLProgress(tqdm):\n",
    "    last_block = 0\n",
    "\n",
    "    def hook(self, block_num=1, block_size=1, total_size=None):\n",
    "        self.total = total_size\n",
    "        self.update((block_num - self.last_block) * block_size)\n",
    "        self.last_block = block_num\n",
    "\n",
    "if not isfile(tar_gz_path):\n",
    "    with DLProgress(unit='B', unit_scale=True, miniters=1, desc='CIFAR-10 Dataset') as pbar:\n",
    "        urlretrieve(\n",
    "            'https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz',\n",
    "            tar_gz_path,\n",
    "            pbar.hook)\n",
    "\n",
    "if not isdir(cifar10_dataset_folder_path):\n",
    "    with tarfile.open(tar_gz_path) as tar:\n",
    "        tar.extractall()\n",
    "        tar.close()\n",
    "\n",
    "\n",
    "tests.test_folder_path(cifar10_dataset_folder_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 探索数据\n",
    "\n",
    "该数据集分成了几部分／批次（batches），以免你的机器在计算时内存不足。CIFAR-10 数据集包含 5 个部分，名称分别为 `data_batch_1`、`data_batch_2`，以此类推。每个部分都包含以下某个类别的标签和图片：\n",
    "\n",
    "* 飞机\n",
    "* 汽车\n",
    "* 鸟类\n",
    "* 猫\n",
    "* 鹿\n",
    "* 狗\n",
    "* 青蛙\n",
    "* 马\n",
    "* 船只\n",
    "* 卡车\n",
    "\n",
    "了解数据集也是对数据进行预测的必经步骤。你可以通过更改 `batch_id` 和 `sample_id` 探索下面的代码单元。`batch_id` 是数据集一个部分的 ID（1 到 5）。`sample_id` 是该部分中图片和标签对（label pair）的 ID。\n",
    "\n",
    "问问你自己：“可能的标签有哪些？”、“图片数据的值范围是多少？”、“标签是按顺序排列，还是随机排列的？”。思考类似的问题，有助于你预处理数据，并使预测结果更准确。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.5/site-packages/matplotlib/font_manager.py:280: UserWarning: Matplotlib is building the font cache using fc-list. This may take a moment.\n",
      "  'Matplotlib is building the font cache using fc-list. '\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "Stats of batch 1:\n",
      "Samples: 10000\n",
      "Label Counts: {0: 1005, 1: 974, 2: 1032, 3: 1016, 4: 999, 5: 937, 6: 1030, 7: 1001, 8: 1025, 9: 981}\n",
      "First 20 Labels: [6, 9, 9, 4, 1, 1, 2, 7, 8, 3, 4, 7, 7, 2, 9, 9, 9, 3, 2, 6]\n",
      "\n",
      "Example of Image 6:\n",
      "Image - Min Value: 7 Max Value: 249\n",
      "Image - Shape: (32, 32, 3)\n",
      "Label - Label Id: 2 Name: bird\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAfoAAAH0CAYAAADVH+85AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\ngg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f63dc945390>"
      ]
     },
     "metadata": {
      "image/png": {
       "height": 250,
       "width": 253
      }
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "%config InlineBackend.figure_format = 'retina'\n",
    "\n",
    "import helper\n",
    "import numpy as np\n",
    "\n",
    "# Explore the dataset\n",
    "batch_id = 1\n",
    "sample_id = 6\n",
    "helper.display_stats(cifar10_dataset_folder_path, batch_id, sample_id)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 实现预处理函数\n",
    "\n",
    "### 标准化\n",
    "\n",
    "在下面的单元中，实现 `normalize` 函数，传入图片数据 `x`，并返回标准化 Numpy 数组。值应该在 0 到 1 的范围内（含 0 和 1）。返回对象应该和 `x` 的形状一样。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "def normalize(x):\n",
    "    \"\"\"\n",
    "    Normalize a list of sample image data in the range of 0 to 1\n",
    "    : x: List of image data.  The image shape is (32, 32, 3)\n",
    "    : return: Numpy array of normalize data\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    out = (x-np.min(x))/(np.max(x)-np.min(x))\n",
    "    \n",
    "    return out\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_normalize(normalize)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### One-hot 编码\n",
    "\n",
    "和之前的代码单元一样，你将为预处理实现一个函数。这次，你将实现 `one_hot_encode` 函数。输入，也就是 `x`，是一个标签列表。实现该函数，以返回为 one_hot 编码的 Numpy 数组的标签列表。标签的可能值为 0 到 9。每次调用 `one_hot_encode` 时，对于每个值，one_hot 编码函数应该返回相同的编码。确保将编码映射保存到该函数外面。\n",
    "\n",
    "提示：不要重复发明轮子。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "from sklearn import preprocessing\n",
    "\n",
    "lb = preprocessing.LabelBinarizer()\n",
    "lb.fit(range(10))\n",
    "\n",
    "def one_hot_encode(x):\n",
    "    \"\"\"\n",
    "    One hot encode a list of sample labels. Return a one-hot encoded vector for each label.\n",
    "    : x: List of sample Labels\n",
    "    : return: Numpy array of one-hot encoded labels\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\\\n",
    "    out = lb.transform(x)\n",
    "    \n",
    "    return out\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_one_hot_encode(one_hot_encode)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 随机化数据\n",
    "\n",
    "之前探索数据时，你已经了解到，样本的顺序是随机的。再随机化一次也不会有什么关系，但是对于这个数据集没有必要。\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 预处理所有数据并保存\n",
    "\n",
    "运行下方的代码单元，将预处理所有 CIFAR-10 数据，并保存到文件中。下面的代码还使用了 10% 的训练数据，用来验证。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL\n",
    "\"\"\"\n",
    "# Preprocess Training, Validation, and Testing Data\n",
    "helper.preprocess_and_save_data(cifar10_dataset_folder_path, normalize, one_hot_encode)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 检查点\n",
    "\n",
    "这是你的第一个检查点。如果你什么时候决定再回到该记事本，或需要重新启动该记事本，你可以从这里开始。预处理的数据已保存到本地。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL\n",
    "\"\"\"\n",
    "import pickle\n",
    "import problem_unittests as tests\n",
    "import helper\n",
    "\n",
    "# Load the Preprocessed Validation data\n",
    "valid_features, valid_labels = pickle.load(open('preprocess_validation.p', mode='rb'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 构建网络\n",
    "\n",
    "对于该神经网络，你需要将每层都构建为一个函数。你看到的大部分代码都位于函数外面。要更全面地测试你的代码，我们需要你将每层放入一个函数中。这样使我们能够提供更好的反馈，并使用我们的统一测试检测简单的错误，然后再提交项目。\n",
    "\n",
    ">**注意**：如果你觉得每周很难抽出足够的时间学习这门课程，我们为此项目提供了一个小捷径。对于接下来的几个问题，你可以使用 [TensorFlow Layers](https://www.tensorflow.org/api_docs/python/tf/layers) 或 [TensorFlow Layers (contrib)](https://www.tensorflow.org/api_guides/python/contrib.layers) 程序包中的类来构建每个层级，但是“卷积和最大池化层级”部分的层级除外。TF Layers 和 Keras 及 TFLearn 层级类似，因此很容易学会。\n",
    "\n",
    ">但是，如果你想充分利用这门课程，请尝试自己解决所有问题，不使用 TF Layers 程序包中的任何类。你依然可以使用其他程序包中的类，这些类和你在 TF Layers 中的类名称是一样的！例如，你可以使用 TF Neural Network 版本的 `conv2d` 类 [tf.nn.conv2d](https://www.tensorflow.org/api_docs/python/tf/nn/conv2d)，而不是 TF Layers 版本的 `conv2d` 类 [tf.layers.conv2d](https://www.tensorflow.org/api_docs/python/tf/layers/conv2d)。\n",
    "\n",
    "我们开始吧！\n",
    "\n",
    "\n",
    "### 输入\n",
    "\n",
    "神经网络需要读取图片数据、one-hot 编码标签和丢弃保留概率（dropout keep probability）。请实现以下函数：\n",
    "\n",
    "* 实现 `neural_net_image_input`\n",
    " * 返回 [TF Placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder)\n",
    " * 使用 `image_shape` 设置形状，部分大小设为 `None`\n",
    " * 使用 [TF Placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder) 中的 TensorFlow `name` 参数对 TensorFlow 占位符 \"x\" 命名\n",
    "* 实现 `neural_net_label_input`\n",
    " * 返回 [TF Placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder)\n",
    " * 使用 `n_classes` 设置形状，部分大小设为 `None`\n",
    " * 使用 [TF Placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder) 中的 TensorFlow `name` 参数对 TensorFlow 占位符 \"y\" 命名\n",
    "* 实现 `neural_net_keep_prob_input`\n",
    " * 返回 [TF Placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder)，用于丢弃保留概率\n",
    " * 使用 [TF Placeholder](https://www.tensorflow.org/api_docs/python/tf/placeholder) 中的 TensorFlow `name` 参数对 TensorFlow 占位符 \"keep_prob\" 命名\n",
    "\n",
    "这些名称将在项目结束时，用于加载保存的模型。\n",
    "\n",
    "注意：TensorFlow 中的 `None` 表示形状可以是动态大小。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Image Input Tests Passed.\n",
      "Label Input Tests Passed.\n",
      "Keep Prob Tests Passed.\n"
     ]
    }
   ],
   "source": [
    "import tensorflow as tf\n",
    "\n",
    "def neural_net_image_input(image_shape):\n",
    "    \"\"\"\n",
    "    Return a Tensor for a batch of image input\n",
    "    : image_shape: Shape of the images\n",
    "    : return: Tensor for image input.\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    out = tf.placeholder(tf.float32,[None,image_shape[0],image_shape[1],image_shape[2]],name='x')\n",
    "    return out\n",
    "\n",
    "\n",
    "def neural_net_label_input(n_classes):\n",
    "    \"\"\"\n",
    "    Return a Tensor for a batch of label input\n",
    "    : n_classes: Number of classes\n",
    "    : return: Tensor for label input.\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    out = tf.placeholder(tf.float32,[None,n_classes],name='y')\n",
    "    return out\n",
    "\n",
    "\n",
    "def neural_net_keep_prob_input():\n",
    "    \"\"\"\n",
    "    Return a Tensor for keep probability\n",
    "    : return: Tensor for keep probability.\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    out = tf.placeholder(tf.float32,name='keep_prob')\n",
    "    return out\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tf.reset_default_graph()\n",
    "tests.test_nn_image_inputs(neural_net_image_input)\n",
    "tests.test_nn_label_inputs(neural_net_label_input)\n",
    "tests.test_nn_keep_prob_inputs(neural_net_keep_prob_input)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 卷积和最大池化层\n",
    "\n",
    "卷积层级适合处理图片。对于此代码单元，你应该实现函数 `conv2d_maxpool` 以便应用卷积然后进行最大池化：\n",
    "\n",
    "* 使用 `conv_ksize`、`conv_num_outputs` 和 `x_tensor` 的形状创建权重（weight）和偏置（bias）。\n",
    "* 使用权重和 `conv_strides` 对 `x_tensor` 应用卷积。\n",
    " * 建议使用我们建议的间距（padding），当然也可以使用任何其他间距。\n",
    "* 添加偏置\n",
    "* 向卷积中添加非线性激活（nonlinear activation）\n",
    "* 使用 `pool_ksize` 和 `pool_strides` 应用最大池化\n",
    " * 建议使用我们建议的间距（padding），当然也可以使用任何其他间距。\n",
    "\n",
    "**注意**：对于**此层**，**请勿使用** [TensorFlow Layers](https://www.tensorflow.org/api_docs/python/tf/layers) 或 [TensorFlow Layers (contrib)](https://www.tensorflow.org/api_guides/python/contrib.layers)，但是仍然可以使用 TensorFlow 的 [Neural Network](https://www.tensorflow.org/api_docs/python/tf/nn) 包。对于所有**其他层**，你依然可以使用快捷方法。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "def conv2d_maxpool(x_tensor, conv_num_outputs, conv_ksize, conv_strides, pool_ksize, pool_strides):\n",
    "    \"\"\"\n",
    "    Apply convolution then max pooling to x_tensor\n",
    "    :param x_tensor: TensorFlow Tensor\n",
    "    :param conv_num_outputs: Number of outputs for the convolutional layer\n",
    "    :param conv_ksize: kernal size 2-D Tuple for the convolutional layer\n",
    "    :param conv_strides: Stride 2-D Tuple for convolution\n",
    "    :param pool_ksize: kernal size 2-D Tuple for pool\n",
    "    :param pool_strides: Stride 2-D Tuple for pool\n",
    "    : return: A tensor that represents convolution and max pooling of x_tensor\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    _,_,_,color_channels = x_tensor.get_shape()\n",
    "    weight = tf.Variable(tf.truncated_normal([conv_ksize[0],\n",
    "                                              conv_ksize[1],\n",
    "                                              color_channels.value,\n",
    "                                              conv_num_outputs],\n",
    "                                             stddev=0.1))\n",
    "    bias = tf.Variable(tf.zeros(conv_num_outputs))\n",
    "    # Convolution\n",
    "    conv_layer = tf.nn.conv2d(x_tensor,\n",
    "                              weight,\n",
    "                              strides=[1,conv_strides[0],conv_strides[1],1],\n",
    "                              padding='SAME')\n",
    "    conv_layer = tf.nn.bias_add(conv_layer,\n",
    "                                bias)\n",
    "    conv_layer = tf.nn.relu(conv_layer)\n",
    "    # Max pooling\n",
    "    output = tf.nn.max_pool(conv_layer,\n",
    "                            ksize=[1,pool_ksize[0],pool_ksize[1],1],\n",
    "                            strides=[1,pool_strides[0],pool_strides[1],1],\n",
    "                            padding='SAME')\n",
    "    return output\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_con_pool(conv2d_maxpool)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 扁平化层\n",
    "\n",
    "实现 `flatten` 函数，将 `x_tensor` 的维度从四维张量（4-D tensor）变成二维张量。输出应该是形状（*部分大小（Batch Size）*，*扁平化图片大小（Flattened Image Size）*）。快捷方法：对于此层，你可以使用 [TensorFlow Layers](https://www.tensorflow.org/api_docs/python/tf/layers) 或 [TensorFlow Layers (contrib)](https://www.tensorflow.org/api_guides/python/contrib.layers) 包中的类。如果你想要更大挑战，可以仅使用其他 TensorFlow 程序包。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "def flatten(x_tensor):\n",
    "    \"\"\"\n",
    "    Flatten x_tensor to (Batch Size, Flattened Image Size)\n",
    "    : x_tensor: A tensor of size (Batch Size, ...), where ... are the image dimensions.\n",
    "    : return: A tensor of size (Batch Size, Flattened Image Size).\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    batch,height,width,channels = x_tensor.get_shape()\n",
    "    flat_size = height.value * width.value * channels.value\n",
    "    output = tf.reshape(x_tensor,\n",
    "                        [-1,flat_size])\n",
    "    return output\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_flatten(flatten)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 全连接层\n",
    "\n",
    "实现 `fully_conn` 函数，以向 `x_tensor` 应用完全连接的层级，形状为（*部分大小（Batch Size）*，*num_outputs*）。快捷方法：对于此层，你可以使用 [TensorFlow Layers](https://www.tensorflow.org/api_docs/python/tf/layers) 或 [TensorFlow Layers (contrib)](https://www.tensorflow.org/api_guides/python/contrib.layers) 包中的类。如果你想要更大挑战，可以仅使用其他 TensorFlow 程序包。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "def fully_conn(x_tensor, num_outputs):\n",
    "    \"\"\"\n",
    "    Apply a fully connected layer to x_tensor using weight and bias\n",
    "    : x_tensor: A 2-D tensor where the first dimension is batch size.\n",
    "    : num_outputs: The number of output that the new tensor should be.\n",
    "    : return: A 2-D tensor where the second dimension is num_outputs.\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    _,hidden_input = x_tensor.get_shape()\n",
    "    hidden_weight = tf.Variable(tf.random_normal([hidden_input.value,num_outputs],\n",
    "                                                 stddev=0.1))\n",
    "    hidden_bias = tf.Variable(tf.zeros(num_outputs))\n",
    "    output = tf.add(tf.matmul(x_tensor,\n",
    "                              hidden_weight),\n",
    "                    hidden_bias)\n",
    "    output = tf.nn.relu(output)\n",
    "    return output\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_fully_conn(fully_conn)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 输出层\n",
    "\n",
    "实现 `output` 函数，向 x_tensor 应用完全连接的层级，形状为（*部分大小（Batch Size）*，*num_outputs*）。快捷方法：对于此层，你可以使用 [TensorFlow Layers](https://www.tensorflow.org/api_docs/python/tf/layers) 或 [TensorFlow Layers (contrib)](https://www.tensorflow.org/api_guides/python/contrib.layers) 包中的类。如果你想要更大挑战，可以仅使用其他 TensorFlow 程序包。\n",
    "\n",
    "**注意**：该层级不应应用 Activation、softmax 或交叉熵（cross entropy）。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "def output(x_tensor, num_outputs):\n",
    "    \"\"\"\n",
    "    Apply a output layer to x_tensor using weight and bias\n",
    "    : x_tensor: A 2-D tensor where the first dimension is batch size.\n",
    "    : num_outputs: The number of output that the new tensor should be.\n",
    "    : return: A 2-D tensor where the second dimension is num_outputs.\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    _,hidden_input = x_tensor.get_shape()\n",
    "    output_weight = tf.Variable(tf.random_normal([hidden_input.value,num_outputs],\n",
    "                                                 stddev=0.1))\n",
    "    output_bias = tf.Variable(tf.zeros(num_outputs))\n",
    "    output = tf.add(tf.matmul(x_tensor,\n",
    "                              output_weight),\n",
    "                    output_bias)\n",
    "    return output\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_output(output)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 创建卷积模型\n",
    "\n",
    "实现函数 `conv_net`， 创建卷积神经网络模型。该函数传入一批图片 `x`，并输出对数（logits）。使用你在上方创建的层创建此模型：\n",
    "\n",
    "* 应用 1、2 或 3 个卷积和最大池化层（Convolution and Max Pool layers）\n",
    "* 应用一个扁平层（Flatten Layer）\n",
    "* 应用 1、2 或 3 个完全连接层（Fully Connected Layers）\n",
    "* 应用一个输出层（Output Layer）\n",
    "* 返回输出\n",
    "* 使用 `keep_prob` 向模型中的一个或多个层应用 [TensorFlow 的 Dropout](https://www.tensorflow.org/api_docs/python/tf/nn/dropout)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Neural Network Built!\n"
     ]
    }
   ],
   "source": [
    "def conv_net(x, keep_prob):\n",
    "    \"\"\"\n",
    "    Create a convolutional neural network model\n",
    "    : x: Placeholder tensor that holds image data.\n",
    "    : keep_prob: Placeholder tensor that hold dropout keep probability.\n",
    "    : return: Tensor that represents logits\n",
    "    \"\"\"\n",
    "    # TODO: Apply 1, 2, or 3 Convolution and Max Pool layers\n",
    "    #    Play around with different number of outputs, kernel size and stride\n",
    "    # Function Definition from Above:\n",
    "    #    conv2d_maxpool(x_tensor, conv_num_outputs, conv_ksize, conv_strides, pool_ksize, pool_strides)\n",
    "    conv_num_outputs = 32\n",
    "    conv_ksize = (5,5)\n",
    "    conv_strides = (1,1)\n",
    "    pool_ksize = (2,2)\n",
    "    pool_strides = (2,2)\n",
    "    conv_layer = conv2d_maxpool(x, conv_num_outputs, conv_ksize, conv_strides, pool_ksize, pool_strides)\n",
    "    conv_num_outputs = 64\n",
    "    conv_ksize = (3,3)\n",
    "    conv_strides = (1,1)\n",
    "    pool_ksize = (2,2)\n",
    "    pool_strides = (2,2)\n",
    "    conv_layer = conv2d_maxpool(x, conv_num_outputs, conv_ksize, conv_strides, pool_ksize, pool_strides)\n",
    "    conv_num_outputs = 128\n",
    "    conv_ksize = (2,2)\n",
    "    conv_strides = (1,1)\n",
    "    pool_ksize = (2,2)\n",
    "    pool_strides = (2,2)\n",
    "    conv_layer = conv2d_maxpool(x, conv_num_outputs, conv_ksize, conv_strides, pool_ksize, pool_strides)\n",
    "\n",
    "    # TODO: Apply a Flatten Layer\n",
    "    # Function Definition from Above:\n",
    "    #   flatten(x_tensor)\n",
    "    flat_layer = flatten(conv_layer)\n",
    "\n",
    "    # TODO: Apply 1, 2, or 3 Fully Connected Layers\n",
    "    #    Play around with different number of outputs\n",
    "    # Function Definition from Above:\n",
    "    #   fully_conn(x_tensor, num_outputs)\n",
    "    num_outputs = 512\n",
    "    fc_layer = fully_conn(flat_layer, num_outputs)\n",
    "    fc_layer = tf.nn.dropout(fc_layer,keep_prob)\n",
    "    \n",
    "    # TODO: Apply an Output Layer\n",
    "    #    Set this to the number of classes\n",
    "    # Function Definition from Above:\n",
    "    #   output(x_tensor, num_outputs)\n",
    "    num_outputs = 10\n",
    "    output_layer = output(fc_layer, num_outputs)\n",
    "    \n",
    "    # TODO: return output\n",
    "    return output_layer\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "\n",
    "##############################\n",
    "## Build the Neural Network ##\n",
    "##############################\n",
    "\n",
    "# Remove previous weights, bias, inputs, etc..\n",
    "tf.reset_default_graph()\n",
    "\n",
    "# Inputs\n",
    "x = neural_net_image_input((32, 32, 3))\n",
    "y = neural_net_label_input(10)\n",
    "keep_prob = neural_net_keep_prob_input()\n",
    "\n",
    "# Model\n",
    "logits = conv_net(x, keep_prob)\n",
    "\n",
    "# Name logits Tensor, so that is can be loaded from disk after training\n",
    "logits = tf.identity(logits, name='logits')\n",
    "\n",
    "# Loss and Optimizer\n",
    "cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y))\n",
    "optimizer = tf.train.AdamOptimizer().minimize(cost)\n",
    "\n",
    "# Accuracy\n",
    "correct_pred = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))\n",
    "accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32), name='accuracy')\n",
    "\n",
    "tests.test_conv_net(conv_net)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 训练神经网络\n",
    "\n",
    "### 单次优化\n",
    "\n",
    "实现函数 `train_neural_network` 以进行单次优化（single optimization）。该优化应该使用 `optimizer` 优化 `session`，其中 `feed_dict` 具有以下参数：\n",
    "\n",
    "* `x` 表示图片输入\n",
    "* `y` 表示标签\n",
    "* `keep_prob` 表示丢弃的保留率\n",
    "\n",
    "每个部分都会调用该函数，所以 `tf.global_variables_initializer()` 已经被调用。\n",
    "\n",
    "注意：不需要返回任何内容。该函数只是用来优化神经网络。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tests Passed\n"
     ]
    }
   ],
   "source": [
    "def train_neural_network(session, optimizer, keep_probability, feature_batch, label_batch):\n",
    "    \"\"\"\n",
    "    Optimize the session on a batch of images and labels\n",
    "    : session: Current TensorFlow session\n",
    "    : optimizer: TensorFlow optimizer function\n",
    "    : keep_probability: keep probability\n",
    "    : feature_batch: Batch of Numpy image data\n",
    "    : label_batch: Batch of Numpy label data\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    session.run(optimizer,\n",
    "                feed_dict={x:feature_batch,\n",
    "                           y:label_batch,\n",
    "                           keep_prob:keep_probability})\n",
    "\n",
    "\n",
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL THAT IS BELOW THIS LINE\n",
    "\"\"\"\n",
    "tests.test_train_nn(train_neural_network)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 显示数据\n",
    "\n",
    "实现函数 `print_stats` 以输出损失和验证准确率。使用全局变量 `valid_features` 和 `valid_labels` 计算验证准确率。使用保留率 `1.0` 计算损失和验证准确率（loss and validation accuracy）。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def print_stats(session, feature_batch, label_batch, cost, accuracy):\n",
    "    \"\"\"\n",
    "    Print information about loss and validation accuracy\n",
    "    : session: Current TensorFlow session\n",
    "    : feature_batch: Batch of Numpy image data\n",
    "    : label_batch: Batch of Numpy label data\n",
    "    : cost: TensorFlow cost function\n",
    "    : accuracy: TensorFlow accuracy function\n",
    "    \"\"\"\n",
    "    # TODO: Implement Function\n",
    "    keep_probability = 1.0\n",
    "    loss = session.run(cost,\n",
    "                       feed_dict={x:feature_batch,\n",
    "                                  y:label_batch,\n",
    "                                  keep_prob:keep_probability})\n",
    "    valid_accuracy = session.run(accuracy,\n",
    "                                 feed_dict={x:valid_features,\n",
    "                                            y:valid_labels,\n",
    "                                            keep_prob:keep_probability})\n",
    "    print('loss:{} validation accuracy:{}'.format(loss,valid_accuracy))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 超参数\n",
    "\n",
    "调试以下超参数：\n",
    "* 设置 `epochs` 表示神经网络停止学习或开始过拟合的迭代次数\n",
    "* 设置 `batch_size`，表示机器内存允许的部分最大体积。大部分人设为以下常见内存大小：\n",
    "\n",
    " * 64\n",
    " * 128\n",
    " * 256\n",
    " * ...\n",
    "* 设置 `keep_probability` 表示使用丢弃时保留节点的概率"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# TODO: Tune Parameters\n",
    "epochs = 20\n",
    "batch_size = 512\n",
    "keep_probability = 0.8"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 在单个 CIFAR-10 部分上训练\n",
    "\n",
    "我们先用单个部分，而不是用所有的 CIFAR-10 批次训练神经网络。这样可以节省时间，并对模型进行迭代，以提高准确率。最终验证准确率达到 50% 或以上之后，在下一部分对所有数据运行模型。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Checking the Training on a Single Batch...\n",
      "Epoch  1, CIFAR-10 Batch 1:  loss:2.4482667446136475 validation accuracy:0.2207999974489212\n",
      "Epoch  2, CIFAR-10 Batch 1:  loss:2.0247933864593506 validation accuracy:0.2679999768733978\n",
      "Epoch  3, CIFAR-10 Batch 1:  loss:1.8508057594299316 validation accuracy:0.3641999661922455\n",
      "Epoch  4, CIFAR-10 Batch 1:  loss:1.6742610931396484 validation accuracy:0.41199997067451477\n",
      "Epoch  5, CIFAR-10 Batch 1:  loss:1.5474604368209839 validation accuracy:0.4281999468803406\n",
      "Epoch  6, CIFAR-10 Batch 1:  loss:1.4246082305908203 validation accuracy:0.46119993925094604\n",
      "Epoch  7, CIFAR-10 Batch 1:  loss:1.3034757375717163 validation accuracy:0.478799968957901\n",
      "Epoch  8, CIFAR-10 Batch 1:  loss:1.181208848953247 validation accuracy:0.49379995465278625\n",
      "Epoch  9, CIFAR-10 Batch 1:  loss:1.0928871631622314 validation accuracy:0.5033999681472778\n",
      "Epoch 10, CIFAR-10 Batch 1:  loss:1.0134243965148926 validation accuracy:0.5077999830245972\n",
      "Epoch 11, CIFAR-10 Batch 1:  loss:0.9310246109962463 validation accuracy:0.520799994468689\n",
      "Epoch 12, CIFAR-10 Batch 1:  loss:0.8538082838058472 validation accuracy:0.5305999517440796\n",
      "Epoch 13, CIFAR-10 Batch 1:  loss:0.7755584120750427 validation accuracy:0.5353999137878418\n",
      "Epoch 14, CIFAR-10 Batch 1:  loss:0.711728036403656 validation accuracy:0.5401999354362488\n",
      "Epoch 15, CIFAR-10 Batch 1:  loss:0.6655250787734985 validation accuracy:0.5437999963760376\n",
      "Epoch 16, CIFAR-10 Batch 1:  loss:0.615240752696991 validation accuracy:0.5479999780654907\n",
      "Epoch 17, CIFAR-10 Batch 1:  loss:0.5725180506706238 validation accuracy:0.5543999671936035\n",
      "Epoch 18, CIFAR-10 Batch 1:  loss:0.5238652229309082 validation accuracy:0.5585998892784119\n",
      "Epoch 19, CIFAR-10 Batch 1:  loss:0.4841848909854889 validation accuracy:0.5593999028205872\n",
      "Epoch 20, CIFAR-10 Batch 1:  loss:0.44599246978759766 validation accuracy:0.5583999752998352\n"
     ]
    }
   ],
   "source": [
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL\n",
    "\"\"\"\n",
    "print('Checking the Training on a Single Batch...')\n",
    "with tf.Session() as sess:\n",
    "    # Initializing the variables\n",
    "    sess.run(tf.global_variables_initializer())\n",
    "    \n",
    "    # Training cycle\n",
    "    for epoch in range(epochs):\n",
    "        batch_i = 1\n",
    "        for batch_features, batch_labels in helper.load_preprocess_training_batch(batch_i, batch_size):\n",
    "            train_neural_network(sess, optimizer, keep_probability, batch_features, batch_labels)\n",
    "        print('Epoch {:>2}, CIFAR-10 Batch {}:  '.format(epoch + 1, batch_i), end='')\n",
    "        print_stats(sess, batch_features, batch_labels, cost, accuracy)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 完全训练模型\n",
    "\n",
    "现在，单个 CIFAR-10 部分的准确率已经不错了，试试所有五个部分吧。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Training...\n",
      "Epoch  1, CIFAR-10 Batch 1:  loss:2.4292166233062744 validation accuracy:0.14179998636245728\n",
      "Epoch  1, CIFAR-10 Batch 2:  loss:1.968839168548584 validation accuracy:0.29739999771118164\n",
      "Epoch  1, CIFAR-10 Batch 3:  loss:1.712121605873108 validation accuracy:0.3797999918460846\n",
      "Epoch  1, CIFAR-10 Batch 4:  loss:1.5386192798614502 validation accuracy:0.4469999670982361\n",
      "Epoch  1, CIFAR-10 Batch 5:  loss:1.475178599357605 validation accuracy:0.4663999378681183\n",
      "Epoch  2, CIFAR-10 Batch 1:  loss:1.4803358316421509 validation accuracy:0.4907999634742737\n",
      "Epoch  2, CIFAR-10 Batch 2:  loss:1.316194772720337 validation accuracy:0.5107999444007874\n",
      "Epoch  2, CIFAR-10 Batch 3:  loss:1.2019574642181396 validation accuracy:0.5209999680519104\n",
      "Epoch  2, CIFAR-10 Batch 4:  loss:1.1541602611541748 validation accuracy:0.5465999245643616\n",
      "Epoch  2, CIFAR-10 Batch 5:  loss:1.1670548915863037 validation accuracy:0.5497999787330627\n",
      "Epoch  3, CIFAR-10 Batch 1:  loss:1.2379324436187744 validation accuracy:0.5605999231338501\n",
      "Epoch  3, CIFAR-10 Batch 2:  loss:1.0738292932510376 validation accuracy:0.5633999109268188\n",
      "Epoch  3, CIFAR-10 Batch 3:  loss:1.0142678022384644 validation accuracy:0.5675999522209167\n",
      "Epoch  3, CIFAR-10 Batch 4:  loss:0.9887506365776062 validation accuracy:0.5793999433517456\n",
      "Epoch  3, CIFAR-10 Batch 5:  loss:0.9979081153869629 validation accuracy:0.584399938583374\n",
      "Epoch  4, CIFAR-10 Batch 1:  loss:1.0773138999938965 validation accuracy:0.5849999189376831\n",
      "Epoch  4, CIFAR-10 Batch 2:  loss:0.9492067694664001 validation accuracy:0.5795999765396118\n",
      "Epoch  4, CIFAR-10 Batch 3:  loss:0.8929338455200195 validation accuracy:0.5881999135017395\n",
      "Epoch  4, CIFAR-10 Batch 4:  loss:0.88431715965271 validation accuracy:0.6003998517990112\n",
      "Epoch  4, CIFAR-10 Batch 5:  loss:0.9038969278335571 validation accuracy:0.5975999236106873\n",
      "Epoch  5, CIFAR-10 Batch 1:  loss:0.9659461975097656 validation accuracy:0.6031998991966248\n",
      "Epoch  5, CIFAR-10 Batch 2:  loss:0.8816051483154297 validation accuracy:0.5927999019622803\n",
      "Epoch  5, CIFAR-10 Batch 3:  loss:0.8096691370010376 validation accuracy:0.60999995470047\n",
      "Epoch  5, CIFAR-10 Batch 4:  loss:0.8123746514320374 validation accuracy:0.6145999431610107\n",
      "Epoch  5, CIFAR-10 Batch 5:  loss:0.7915804386138916 validation accuracy:0.6155998706817627\n",
      "Epoch  6, CIFAR-10 Batch 1:  loss:0.8899974822998047 validation accuracy:0.610599935054779\n",
      "Epoch  6, CIFAR-10 Batch 2:  loss:0.821513295173645 validation accuracy:0.6007998585700989\n",
      "Epoch  6, CIFAR-10 Batch 3:  loss:0.7579913139343262 validation accuracy:0.6113998889923096\n",
      "Epoch  6, CIFAR-10 Batch 4:  loss:0.7406783103942871 validation accuracy:0.6243999004364014\n",
      "Epoch  6, CIFAR-10 Batch 5:  loss:0.7194974422454834 validation accuracy:0.6269998550415039\n",
      "Epoch  7, CIFAR-10 Batch 1:  loss:0.7935526371002197 validation accuracy:0.630599856376648\n",
      "Epoch  7, CIFAR-10 Batch 2:  loss:0.7113784551620483 validation accuracy:0.6219999194145203\n",
      "Epoch  7, CIFAR-10 Batch 3:  loss:0.6719459295272827 validation accuracy:0.6267998814582825\n",
      "Epoch  7, CIFAR-10 Batch 4:  loss:0.675679624080658 validation accuracy:0.6341999173164368\n",
      "Epoch  7, CIFAR-10 Batch 5:  loss:0.6478003263473511 validation accuracy:0.6419999003410339\n",
      "Epoch  8, CIFAR-10 Batch 1:  loss:0.7100580334663391 validation accuracy:0.6369999051094055\n",
      "Epoch  8, CIFAR-10 Batch 2:  loss:0.6580901741981506 validation accuracy:0.6273998618125916\n",
      "Epoch  8, CIFAR-10 Batch 3:  loss:0.6119219064712524 validation accuracy:0.6303998827934265\n",
      "Epoch  8, CIFAR-10 Batch 4:  loss:0.61835777759552 validation accuracy:0.6367999315261841\n",
      "Epoch  8, CIFAR-10 Batch 5:  loss:0.5653891563415527 validation accuracy:0.6485999226570129\n",
      "Epoch  9, CIFAR-10 Batch 1:  loss:0.6413531303405762 validation accuracy:0.6447998881340027\n",
      "Epoch  9, CIFAR-10 Batch 2:  loss:0.5787570476531982 validation accuracy:0.6455998420715332\n",
      "Epoch  9, CIFAR-10 Batch 3:  loss:0.5509963631629944 validation accuracy:0.6439998149871826\n",
      "Epoch  9, CIFAR-10 Batch 4:  loss:0.5358052849769592 validation accuracy:0.6521998643875122\n",
      "Epoch  9, CIFAR-10 Batch 5:  loss:0.5091274976730347 validation accuracy:0.6471998691558838\n",
      "Epoch 10, CIFAR-10 Batch 1:  loss:0.5612105131149292 validation accuracy:0.6535999178886414\n",
      "Epoch 10, CIFAR-10 Batch 2:  loss:0.5150115489959717 validation accuracy:0.6497998833656311\n",
      "Epoch 10, CIFAR-10 Batch 3:  loss:0.4909113645553589 validation accuracy:0.6503998637199402\n",
      "Epoch 10, CIFAR-10 Batch 4:  loss:0.48139306902885437 validation accuracy:0.6525998711585999\n",
      "Epoch 10, CIFAR-10 Batch 5:  loss:0.4752991795539856 validation accuracy:0.6461999416351318\n",
      "Epoch 11, CIFAR-10 Batch 1:  loss:0.515298068523407 validation accuracy:0.6537998914718628\n",
      "Epoch 11, CIFAR-10 Batch 2:  loss:0.459523469209671 validation accuracy:0.6533999443054199\n",
      "Epoch 11, CIFAR-10 Batch 3:  loss:0.4204851984977722 validation accuracy:0.6575998067855835\n",
      "Epoch 11, CIFAR-10 Batch 4:  loss:0.43206119537353516 validation accuracy:0.6567999124526978\n",
      "Epoch 11, CIFAR-10 Batch 5:  loss:0.4024614989757538 validation accuracy:0.6551998853683472\n",
      "Epoch 12, CIFAR-10 Batch 1:  loss:0.44917410612106323 validation accuracy:0.6635998487472534\n",
      "Epoch 12, CIFAR-10 Batch 2:  loss:0.42166027426719666 validation accuracy:0.6583998799324036\n",
      "Epoch 12, CIFAR-10 Batch 3:  loss:0.3944258987903595 validation accuracy:0.654999852180481\n",
      "Epoch 12, CIFAR-10 Batch 4:  loss:0.38852831721305847 validation accuracy:0.6643998622894287\n",
      "Epoch 12, CIFAR-10 Batch 5:  loss:0.38307952880859375 validation accuracy:0.6527999043464661\n",
      "Epoch 13, CIFAR-10 Batch 1:  loss:0.4159507751464844 validation accuracy:0.6589998602867126\n",
      "Epoch 13, CIFAR-10 Batch 2:  loss:0.3649327754974365 validation accuracy:0.6637998819351196\n",
      "Epoch 13, CIFAR-10 Batch 3:  loss:0.34415820240974426 validation accuracy:0.6597998738288879\n",
      "Epoch 13, CIFAR-10 Batch 4:  loss:0.3506980836391449 validation accuracy:0.6633998155593872\n",
      "Epoch 13, CIFAR-10 Batch 5:  loss:0.32471558451652527 validation accuracy:0.6621999144554138\n",
      "Epoch 14, CIFAR-10 Batch 1:  loss:0.3735540211200714 validation accuracy:0.6659998893737793\n",
      "Epoch 14, CIFAR-10 Batch 2:  loss:0.329769492149353 validation accuracy:0.6663998365402222\n",
      "Epoch 14, CIFAR-10 Batch 3:  loss:0.3590681254863739 validation accuracy:0.6429998874664307\n",
      "Epoch 14, CIFAR-10 Batch 4:  loss:0.3133808374404907 validation accuracy:0.6615999341011047\n",
      "Epoch 14, CIFAR-10 Batch 5:  loss:0.30858951807022095 validation accuracy:0.6611999273300171\n",
      "Epoch 15, CIFAR-10 Batch 1:  loss:0.3274095058441162 validation accuracy:0.6675998568534851\n",
      "Epoch 15, CIFAR-10 Batch 2:  loss:0.3186756372451782 validation accuracy:0.6631999015808105\n",
      "Epoch 15, CIFAR-10 Batch 3:  loss:0.34736645221710205 validation accuracy:0.6409999132156372\n",
      "Epoch 15, CIFAR-10 Batch 4:  loss:0.29503077268600464 validation accuracy:0.6607998609542847\n",
      "Epoch 15, CIFAR-10 Batch 5:  loss:0.29684337973594666 validation accuracy:0.6475999355316162\n",
      "Epoch 16, CIFAR-10 Batch 1:  loss:0.29913750290870667 validation accuracy:0.6641998887062073\n",
      "Epoch 16, CIFAR-10 Batch 2:  loss:0.3174234628677368 validation accuracy:0.6471999287605286\n",
      "Epoch 16, CIFAR-10 Batch 3:  loss:0.29213157296180725 validation accuracy:0.6469998359680176\n",
      "Epoch 16, CIFAR-10 Batch 4:  loss:0.27856215834617615 validation accuracy:0.6505998969078064\n",
      "Epoch 16, CIFAR-10 Batch 5:  loss:0.2489335536956787 validation accuracy:0.6507998704910278\n",
      "Epoch 17, CIFAR-10 Batch 1:  loss:0.28574320673942566 validation accuracy:0.6553998589515686\n",
      "Epoch 17, CIFAR-10 Batch 2:  loss:0.2653360664844513 validation accuracy:0.6581998467445374\n",
      "Epoch 17, CIFAR-10 Batch 3:  loss:0.24400317668914795 validation accuracy:0.6627998948097229\n",
      "Epoch 17, CIFAR-10 Batch 4:  loss:0.2318611592054367 validation accuracy:0.6657998561859131\n",
      "Epoch 17, CIFAR-10 Batch 5:  loss:0.23117932677268982 validation accuracy:0.6559998989105225\n",
      "Epoch 18, CIFAR-10 Batch 1:  loss:0.26129865646362305 validation accuracy:0.6589998602867126\n",
      "Epoch 18, CIFAR-10 Batch 2:  loss:0.23127931356430054 validation accuracy:0.6629998683929443\n",
      "Epoch 18, CIFAR-10 Batch 3:  loss:0.214146688580513 validation accuracy:0.6697999238967896\n",
      "Epoch 18, CIFAR-10 Batch 4:  loss:0.20144693553447723 validation accuracy:0.6739999055862427\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 18, CIFAR-10 Batch 5:  loss:0.19057634472846985 validation accuracy:0.6639998555183411\n",
      "Epoch 19, CIFAR-10 Batch 1:  loss:0.23707038164138794 validation accuracy:0.656799852848053\n",
      "Epoch 19, CIFAR-10 Batch 2:  loss:0.21375858783721924 validation accuracy:0.6629998683929443\n",
      "Epoch 19, CIFAR-10 Batch 3:  loss:0.20046211779117584 validation accuracy:0.6729998588562012\n",
      "Epoch 19, CIFAR-10 Batch 4:  loss:0.1914689838886261 validation accuracy:0.6659998893737793\n",
      "Epoch 19, CIFAR-10 Batch 5:  loss:0.18468008935451508 validation accuracy:0.6633998155593872\n",
      "Epoch 20, CIFAR-10 Batch 1:  loss:0.23474840819835663 validation accuracy:0.6445998549461365\n",
      "Epoch 20, CIFAR-10 Batch 2:  loss:0.20538412034511566 validation accuracy:0.6665998697280884\n",
      "Epoch 20, CIFAR-10 Batch 3:  loss:0.18175570666790009 validation accuracy:0.6761998534202576\n",
      "Epoch 20, CIFAR-10 Batch 4:  loss:0.16228438913822174 validation accuracy:0.6745999455451965\n",
      "Epoch 20, CIFAR-10 Batch 5:  loss:0.16032525897026062 validation accuracy:0.6771999001502991\n"
     ]
    }
   ],
   "source": [
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL\n",
    "\"\"\"\n",
    "save_model_path = './image_classification'\n",
    "\n",
    "print('Training...')\n",
    "with tf.Session() as sess:\n",
    "    # Initializing the variables\n",
    "    sess.run(tf.global_variables_initializer())\n",
    "    \n",
    "    # Training cycle\n",
    "    for epoch in range(epochs):\n",
    "        # Loop over all batches\n",
    "        n_batches = 5\n",
    "        for batch_i in range(1, n_batches + 1):\n",
    "            for batch_features, batch_labels in helper.load_preprocess_training_batch(batch_i, batch_size):\n",
    "                train_neural_network(sess, optimizer, keep_probability, batch_features, batch_labels)\n",
    "            print('Epoch {:>2}, CIFAR-10 Batch {}:  '.format(epoch + 1, batch_i), end='')\n",
    "            print_stats(sess, batch_features, batch_labels, cost, accuracy)\n",
    "            \n",
    "    # Save Model\n",
    "    saver = tf.train.Saver()\n",
    "    save_path = saver.save(sess, save_model_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 检查点\n",
    "\n",
    "模型已保存到本地。\n",
    "\n",
    "## 测试模型\n",
    "\n",
    "利用测试数据集测试你的模型。这将是最终的准确率。你的准确率应该高于 50%。如果没达到，请继续调整模型结构和参数。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:tensorflow:Restoring parameters from ./image_classification\n",
      "Testing Accuracy: 0.6639763295650483\n",
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAscAAAJ/CAYAAACUb342AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAAWJQAAFiUBSVIk8AAAIABJREFUeJzs3XecZFWd///Xp7ur82SGYYiDxFFEFDCgEtawKiqYwCz4\n1TUsZt01i/ozfHFXUVjXdV1ljZj1pxgwjSKKKKBkUWAEZmCEGSZ0Tp/vH59TdW/fqa6unqmO834+\nHvdRXfeec++p6urqU5/6nHPM3REREREREWia7QaIiIiIiMwV6hyLiIiIiCTqHIuIiIiIJOoci4iI\niIgk6hyLiIiIiCTqHIuIiIiIJOoci4iIiIgk6hyLiIiIiCTqHIuIiIiIJOoci4iIiIgk6hyLiIiI\niCTqHIuIiIiIJOoci4iIiIgk6hyLiIiIiCTqHM8yMzvIzJ5pZq8ys7eZ2VvN7DVm9hwzO87Mume7\njRMxsyYzO83MLjazv5rZdjPz3Pad2W6jyFxjZmsKfyfnNqLsXGVmJxcew1mz3SYRkVpaZrsBeyIz\nWw68Cng5cNAkxcfM7EbgMuAS4GfuPjDNTZxUegzfAE6Z7bbIzDOzi4CXTFJsBNgK3AdcTbyGv+Lu\n26a3dSIiIrtOkeMZZmZPBW4E/j8m7xhD/I6OIjrT3weePX2tm5LPM4WOsaJHe6QWYC/gSOD5wH8C\nG8zsXDPTB/N5pPC3e9Fst0dEZDrpH9QMMrMzgK+w84eS7cB1wD3AILAMOBBYW6XsrDOzRwKn5nb9\nDXgv8AdgR25/30y2S+aFLuA9wIlm9mR3H5ztBomIiOSpczxDzOwQItqa7+xeD7wD+IG7j1Sp0w2c\nBDwHeAaweAaaWo9nFu6f5u5/mpWWyFzxFiLNJq8FWAU8Bng18YGv7BQikvzSGWmdiIhIndQ5njkf\nANpy938KPN3d+yeq4O49RJ7xJWb2GuBlRHR5th2b+3m9OsYC3Ofu66vs/ytwuZldAHyR+JBXdpaZ\nfcLd/zgTDZyP0nNqs92O3eHu65jnj0FE9ixz7iv7hcjMOoCn53YNAy+p1TEucvcd7v4xd/9pwxs4\ndXvnft44a62QecPd+4AXALfkdhvwytlpkYiISHXqHM+MhwEdufu/cff53KnMTy83PGutkHklfRj8\nWGH342ajLSIiIhNRWsXM2Kdwf8NMXtzMFgOPBfYDVhCD5jYBv3P3O3bllA1sXkOY2QOIdI/9gVZg\nPfALd//7JPX2J3JiDyAe192p3l270Zb9gAcBDwCWpt1bgDuA3+7hU5n9rHD/EDNrdvfRqZzEzI4C\nHgisJgb5rXf3L9dRrxV4FLCG+AZkDPg7cG0j0oPM7DDg4cC+wABwF3Clu8/o33yVdh0OHAOsJF6T\nfcRr/XrgRncfm8XmTcrMDgAeSeSwLyL+njYCl7n71gZf6wFEQOMAoJl4r7zc3W/bjXMeQTz/+xDB\nhRGgB7gT+Atws7v7bjZdRBrF3bVN8wY8F/Dc9sMZuu5xwA+BocL189u1xDRbVuM8J9eoP9G2LtVd\nv6t1C224KF8mt/8k4BdEJ6d4niHgk0B3lfM9EPjBBPXGgG8C+9X5PDeldvwncOskj20U+AlwSp3n\n/t9C/U9P4ff/oULd79X6PU/xtXVR4dxn1Vmvo8pzsneVcvnXzbrc/rOJDl3xHFsnue4RwJeJD4YT\n/W7uAt4ItO7C8/Fo4HcTnHeEGDtwbCq7pnD83BrnrbtslbpLgfcTH8pqvSbvBT4LHD/J77iurY73\nj7peK6nuGcAfa1xvOP09PXIK51yXq78+t/8RxIe3au8JDlwBPGoK1ykBbyLy7id73rYS7zlPaMTf\npzZt2nZvm/UG7Akb8A+FN8IdwNJpvJ4B59V4k6+2rQOWTXC+4j+3us6X6q7f1bqFNoz7R532vbbO\nx/h7ch1kYraNvjrqrQcOqOP5fukuPEYH/h1onuTcXcDNhXpn1tGmJxaem7uAFQ18jV1UaNNZddbb\npc4xMZj1azWey6qdY+Jv4X1EJ6re38v19fzec9d4e52vwyEi73pNYf+5Nc5dd9lCvWcA90/x9fjH\nSX7HdW11vH9M+lohZub56RSvfT7QVMe51+XqrE/7XkPtIEL+d3hGHddYSSx8M9Xn7zuN+hvVpk3b\nrm9Kq5gZVxERw+Z0vxv4vJk932NGikb7b+D/FPYNEZGPjURE6ThigYayk4BfmdmJ7n7/NLSpodKc\n0R9Pd52ILt1KdIaOAQ7JFT8OuAA428xOAb5KllJ0c9qGiHmlH5yrdxD1LXZSzN3vB24gvrbeTnQI\nDwSOJlI+yt5IdNreOtGJ3b03PdbfAe1p96fN7A/ufmu1Oma2D/AFsvSXUeD57r55kscxE/Yr3Heg\nnnadT0xpWK5zDVkH+gHAwcUKZmZE5P1FhUP9RMelnPd/KPGaKT9fDwJ+Y2bHu3vN2WHM7PXETDR5\no8Tv604iBeChRPpHiehwFv82Gyq16aPsnP50D/FN0X1AJ5GC9GDGz6Iz68xsEfBL4neSdz9wZbpd\nTaRZ5Nv+OuI97YVTvN4LgU/kdl1PRHsHifeRY8meyxJwkZld4+5/meB8BnyL+L3nbSLms7+P+DC1\nJJ3/UJTiKDK3zHbvfE/ZiNXtilGCjcSCCA+mcV93v6RwjTGiY7G0UK6F+Ce9rVD+K1XO2U5EsMrb\nXbnyVxSOlbd9Ut390/1iasmbJ6hXqVtow0WF+uWo2PeBQ6qUP4PoBOWfh0el59yB3wDHVKl3MtFZ\ny1/rKZM85+Up9j6UrlE1Gkx8KPlXoLfQrkfU8Xt9ZaFNf6DK1/9ER70YcXvXNLyei7+Ps+qs90+F\nen+doNz6XJl8KsQXgP2rlF9TZd9bC9fakp7H9iplDwa+Wyj/Y2qnGz2YnaONXy6+ftPv5Awit7nc\njnydc2tcY029ZVP5fyQ65/k6vwROqPZYiM7l04iv9K8qHNuL7G8yf75vMPHfbrXfw8lTea0AnyuU\n3w68AigVyi0hvn0pRu1fMcn51+XK9pC9T3wbOLRK+bXAnwrX+GqN859aKPsXYuBp1dcS8e3QacDF\nwNcb/beqTZu2qW+z3oA9ZSOiIAOFN838tpnIS3wX8ASgaxeu0U3kruXP+4ZJ6jyC8Z01Z5K8NybI\nB52kzpT+QVapf1GV5+xL1PgalVhyu1qH+qdAW416T633H2Eqv0+t81Up/6jCa6Hm+XP1imkFH69S\n5h2FMj+r9Rztxuu5+PuY9PdJfMi6qVCvag411dNxPjSF9j2I8akUd1Kl41aoY0Tubf6ap9Yo/4tC\n2QvraFOxY9ywzjERDd5UbFO9v39gVY1j+XNeNMXXSt1/+8TA4XzZPuDRk5z/nEKdHiZIEUvl11X5\nHVxI7Q9CqxifpjIw0TWIsQflcsPAwVN4rnb64KZNm7aZ3zSV2wzxWOjgRcSbajXLgacQ+ZGXAveb\n2WVm9oo020Q9XkJEU8p+5O7FqbOK7fod8O7C7tfVeb3ZtJGIENUaZf8/RGS8rDxK/0VeY9lid/8+\n8OfcrpNrNcTd76l1virlfwv8R27X6WZWz1fbLwPyI+Zfa2anle+Y2WOIZbzL7gVeOMlzNCPMrJ2I\n+h5ZOPRfdZ7ij8A7p3DJfyH7qtqB53j1RUoq3N2JlfzyM5VU/Vswswcx/nVxC5EmU+v8N6R2TZeX\nM34O8l8Ar6n39+/um6alVVPz2sL997r75bUquPuFxDdIZV1MLXXleiKI4DWusYno9Ja1EWkd1eRX\ngvyju99eb0PcfaL/DyIyg9Q5nkHu/nXi681f11G8REwx9ingNjN7dcplq+UFhfvvqbNpnyA6UmVP\nMbPlddadLZ/2SfK13X0IKP5jvdjd767j/D/P/bx3yuNtpO/mfm5l5/zKnbj7duBM4qv8ss+Z2YFm\ntgL4ClleuwMvrvOxNsJeZramsB1qZieY2b8ANwLPLtT5krtfVef5z/c6p3szs6XA83K7LnH3K+qp\nmzonn87tOsXMOqsULf6tnZdeb5P5LNM3lePLC/drdvjmGjPrAk7P7bqfSAmrR/GD01Tyjj/m7vXM\n1/6Dwv2H1FFn5RTaISJzhDrHM8zdr3H3xwInEpHNmvPwJiuISOPFaZ7WnaTIY35Z59vc/co62zQM\nfD1/OiaOiswVl9ZZrjho7Sd11vtr4f6U/8lZWGRm+xY7juw8WKoYUa3K3f9A5C2XLSM6xRcR+d1l\nH3H3H021zbvhI8Dthe0vxIeT/8vOA+YuZ+fOXC3fm0LZRxMfLsu+MYW6AJflfm4hUo+KHpX7uTz1\n36RSFPfrkxacIjNbSaRtlP3e59+y7sczfmDat+v9RiY91htzux6cBvbVo96/k5sL9yd6T8h/63SQ\nmf1znecXkTlCI2RnibtfRvonbGYPJCLKxxH/II6h+geXM4iRztXebI9i/EwIv5tik64gvlIuO5ad\nIyVzSfEf1US2F+7/uWqpyetNmtpiZs3A44lZFY4nOrxVP8xUsazOcrj7+WnWjfKS5CcUilxB5B7P\nRf3ELCPvrjNaB3CHu2+ZwjUeXbi/OX0gqVdz4X61ug/L/fwXn9pCFL+fQtl6FTvwl1UtNbcdW7i/\nK+9hD0w/NxHvo5M9D9u9/tVKi4v3TPSecDHwhtz9C83sdGKg4Q99HswGJLKnU+d4DnD3G4mox2eg\n8rXw6cQb7NGF4q82s/9x96sL+4tRjKrTDNVQ7DTO9a8D611lbqRB9UpVSyVm9igif/bBtcrVUG9e\nednZxHRmBxb2bwWe5+7F9s+GUeL53ky09TLgy1Ps6ML4lJ967F+4P5WoczXjUoxS/nT+91V1Sr0a\nit9KNEIx7eemabjGdJuN97C6V6t09+FCZlvV9wR3v9LMPsn4YMPj0zZmZtcR35z8ijpW8RSRmae0\nijnI3be6+0VE5ON9VYoUB61AtkxxWTHyOZniP4m6I5mzYTcGmTV8cJqZPYkY/LSrHWOY4t9i6mB+\nsMqhN0028GyanO3uVtha3H2Fux/u7me6+4W70DGGmH1gKhqdL99duN/ov7VGWFG439AllWfIbLyH\nTddg1XOIb2/6CvubiFzlVxMR5rvN7Bdm9uw6xpSIyAxR53gO8/AeYtGKvMfPRntkZ2ng4hcZvxjB\nemLZ3icTyxYvJaZoqnQcqbJoxRSvu4KY9q/ohWa2p/9d14zy74L52GmZNwPxFqL03v1BYoGafwV+\ny87fRkH8Dz6ZyEP/pZmtnrFGisiElFYxP1xAzFJQtp+Zdbh7f25fMVI01a/plxTuKy+uPq9mfNTu\nYuAldcxcUO9goZ3kVn4rrjYHsZrfO6n+jcOeohidfqC7NzLNoNF/a41QfMzFKOx8sODew9IUcOcB\n55lZN/BwYi7nU4jc+Pz/4McCPzKzh09lakgRabw9PcI0X1QbdV78yrCYl3noFK9x+CTnk+pOzf28\nDXhZnVN67c7UcG8oXPdKxs968m4ze+xunH++K+Zw7lW11C5K073lv/I/ZKKyE5jq32Y9istcr52G\na0y3Bf0e5u497v5zd3+vu59MLIH9TmKQatnRwEtno30iklHneH6olhdXzMe7nvHz3z58itcoTt1W\n7/yz9VqoX/Pm/4H/2t1766y3S1PlmdnxwIdzu+4nZsd4Mdlz3Ax8OaVe7ImKcxpXm4ptd+UHxB6W\nBtHW6/hGN4adH/N8/HBUfM+Z6u8t/zc1RiwcM2e5+33u/gF2ntLwabPRHhHJqHM8PxxRuN9TXAAj\nfQ2X/+dyqJkVp0aqysxaiA5W5XRMfRqlyRS/Jqx3irO5Lv9Vbl0DiFJaxPOneqG0UuLFjM+pfam7\n3+HuPybmGi7bn5g6ak/0c8Z/GDtjGq7x29zPTcCz6qmU8sGfM2nBKXL3e4kPyGUPN7PdGSBalP/7\nna6/3d8zPi/3GRPN615kZkczfp7n6919RyMbN42+yvjnd80stUNEEnWOZ4CZrTKzVbtxiuLXbOsm\nKPflwv3istATOYfxy87+0N0311m3XsWR5I1ecW625PMki1/rTuRF1LnoR8F/EwN8yi5w9+/k7r+D\n8R9qnmZm82Ep8IZKeZ755+V4M2t0h/RLhfv/UmdH7qVUzxVvhE8X7n+0gTMg5P9+p+VvN33rkl85\ncjnV53Svpphj/8WGNGoGpGkX89841ZOWJSLTSJ3jmbGWWAL6w2a296Slc8zsWcCrCruLs1eU/S/j\n/4k93cxePUHZ8vmPJ2ZWyPvEVNpYp9sYHxU6ZRquMRuuy/18rJmdVKuwmT2cGGA5JWb2T4yPgF4D\nvCVfJv2TfS7jXwPnmVl+wYo9xfsYn4702cl+N0VmttrMnlLtmLvfAPwyt+tw4KOTnO+BxOCs6fI/\nwKbc/ccDH6u3gzzJB/j8HMLHp8Fl06H43vP+9B41ITN7FXBablcv8VzMCjN7VVqxsN7yT2b89IP1\nLlQkItNEneOZ00lM6XOXmX3bzJ5V6w3UzNaa2aeBrzF+xa6r2TlCDED6GvGNhd0XmNlHzGzcSG4z\nazGzs4nllPP/6L6WvqJvqJT2kY9qnmxmnzGzx5nZYYXlledTVLm4NPE3zezpxUJm1mFmbwB+RozC\nv6/eC5jZUcD5uV09wJnVRrSnOY5fltvVSiw7Pl2dmTnJ3f9IDHYq6wZ+ZmafMLMJB9CZ2VIzO8PM\nvkpMyffiGpd5DZBf5e+fzexLxdevmTWlyPU6YiDttMxB7O59RHvzHwpeRzzuR1WrY2ZtZvZUM/sm\ntVfE/FXu527gEjN7RnqfKi6NvjuP4VfAF3K7uoCfmNn/Self+bYvNrPzgAsLp3nLLs6n3Sj/CtyR\nXgunT7SMdXoPfjGx/HvevIl6iyxUmspt5pWI1e9OBzCzvwJ3EJ2lMeKf5wOBA6rUvQt4Tq0FMNz9\ns2Z2IvCStKsJeDPwGjP7LXA3Mc3T8ew8iv9Gdo5SN9IFjF/a9/+kreiXxNyf88FnidkjDkv3VwDf\nNbO/ER9kBoivoR9BfECCGJ3+KmJu05rMrJP4pqAjt/uV7j7h6mHu/g0z+xTwyrTrMOBTwAvrfEwL\ngrt/KHXW/intaiY6tK8xs9uJJcjvJ/4mlxLP05opnP86M/tXxkeMnw+caWZXAHcSHcljiZkJIL49\neQPTlA/u7pea2ZuBfyebn/kU4DdmdjdwLbFiYQeRl3402Rzd1WbFKfsM8CagPd0/MW3V7G4qxznE\nQhnl1UGXpOv/XzO7kvhwsQ/wqFx7yi529//czes3QjvxWng+4GZ2C3A72fRyq4GHsvP0c99x991d\n0VFEdpM6xzNjC9H5rTal1KHUN2XRT4GX17n62dnpmq8n+0fVRu0O56+B06Yz4uLuXzWzRxCdgwXB\n3QdTpPjnZB0ggIPSVtRDDMi6uc5LXEB8WCr7nLsX812reQPxQaQ8KOsFZvYzd9+jBum5+yvM7Fpi\nsGL+A8bB1LcQS825ct39Y+kDzPvJ/taaGf8hsGyE+DD4qyrHGia1aQPRocxHLVcz/jU6lXOuN7Oz\niE59xyTFd4u7b08pMN9ifPrVCmJhnYn8B9VXD51tRgyqLg6sLvoqWVBDRGaR0ipmgLtfS0Q6/oGI\nMv0BGK2j6gDxD+Kp7v6EepcFTqszvZGY2uhSqq/MVHYD8VXsiTPxVWRq1yOIf2S/J6JY83oAirvf\nDDyM+Dp0oue6B/g8cLS7/6ie85rZ8xg/GPNmIvJZT5sGiIVj8svXXmBmuzIQcF5z9/8gOsL/Bmyo\no8otxFf1J7j7pN+kpOm4TiTmm65mjPg7fLS7f76uRu8md/8aMXjz3xifh1zNJmIwX82Ombt/lRg/\n8V4iReRuxs/R2zDuvhV4HBF5vbZG0VEiVenR7n7Obiwr30inEc/RFYxPu6lmjGj/qe7+XC3+ITI3\nmPtCnX52bkvRpsPTtjdZhGc7EfW9AbgxDbLa3WstIf5570cM/Ogh/iH+rt4Ot9QnzS18IhE17iCe\n5w3AZSknVGZZ+oDwEOKbnKXENFpbgVuJv7nJOpO1zn0Y8aF0NfHhdgNwpbvfubvt3o02GfF4HwSs\nJFI9elLbbgBu8jn+j8DMDiSe11XEe+UWYCPxdzXrK+FNxMzagaOIbwf3IZ77YWLQ7F+Bq2c5P1pE\nqlDnWEREREQkUVqFiIiIiEiizrGIiIiISKLOsYiIiIhIos6xiIiIiEiizrGIiIiISKLOsYiIiIhI\nos6xiIiIiEiizrGIiIiISKLOsYiIiIhIos6xiIiIiEiizrGIiIiISKLOsYiIiIhIos6xiIiIiEii\nzrGIiIiISKLOsYiIiIhIos6xiIiIiEiizrGIiIiISKLOsYiIiIhIos6xiIiIiEiizrGIiIiISKLO\nsYiIiIhIos6xiIiIiEiizrGIiIiISKLO8RSYmadtzWy3RUREREQaT51jEREREZFEnWMRERERkUSd\nYxERERGRRJ1jEREREZFEneMcM2sys9eY2Z/MrN/M7jWz75nZo+qou9LMPmRm15lZj5n1mtn1ZvYB\nM1s+Sd2jzOyzZna7mQ2Y2VYzu9zMXmlmpSrl15QHB6b7jzSzb5jZ3WY2ambn7/qzICIiIrLnapnt\nBswVZtYCfAM4Le0aIZ6fpwJPMrMza9R9DPBdoNwJHgLGgAel7UVm9gR3/3OVuucAHyf7oNIDdAMn\npO1MMzvV3fsmuPaZwBdTW7cBo/U+ZhEREREZT5HjzL8SHeMx4C3AEndfBjwA+Cnw2WqVzOwg4HtE\nx/g/gcOADqALeDBwKXAA8C0zay7UPR24AOgF/gVY6e6LgE7gScBfgJOBj9Vo92eIjvnB7r401VXk\nWERERGQXmLvPdhtmnZl1AXcDi4D3uvu5heNtwNXAA9Oug919fTr2ReAFwIfd/W1Vzt0K/B44GniO\nu38j7W8GbgUOAp7k7j+uUvcQ4FqgFTjQ3e9O+9cAt6dilwMnuvvYrj16ERERESlT5Dg8kegYD1Il\nSuvug8C/FfebWSfwHCLa/NFqJ3b3ISJdA+AJuUMnEx3j66t1jFPdW4EriJSJkydo+7+rYywiIiLS\nGMo5Dg9Lt390920TlPlllX3HElFdB64zs4nO35FuD8jtOyHdHmZm99Ro25IqdfN+W6OuiIiIiEyB\nOsdhZbrdWKPMhir7VqdbA1bVcZ3OKnXbdqFu3r111BURERGROqhzvHvKaSnb0mC4Xan7XXc/fVcb\n4O6anUJERESkQZRzHMrR131rlKl2bFO6XWxmS6ocr6Vc98Ap1hMRERGRaaLOcbg63R5jZosnKHNS\nlX1/IOZDNmLqtako5wofbWb7TbGuiIiIiEwDdY7DpcB2Iv/3dcWDaTq2NxX3u/sO4Jvp7vvMbNFE\nFzCzFjPrzu36GXAn0Ax8pFbjzGzZZA9ARERERHafOseAu/cC56W77zGzN5pZB1TmFP42E88W8VZg\nC3A48Bsze1J5yWcLh5nZG4GbgeNy1xwGziFmuniemX3HzI4pHzezkpkdZ2bnkc1pLCIiIiLTSIuA\nJBMsH90DLE0/n0kWJa4sApLqHg98hywveZiIRC8ipnorO9ndx00JZ2ZnA5/KletP2xIiqgyAu1uu\nzhpShzm/X0RERER2jyLHibuPAM8CXkusSjcCjAKXACe5+7dq1P09cCSxBPVvyDrVfURe8ifSOXaa\nK9ndPwccQSz5fEO65mJgM7AOeE86LiIiIiLTTJFjEREREZFEkWMRERERkUSdYxERERGRRJ1jERER\nEZFEnWMRERERkUSdYxERERGRRJ1jEREREZFEnWMRERERkUSdYxERERGRRJ1jEREREZGkZbYbICKy\nEJnZ7cRS8OtnuSkiIvPVGmC7ux88kxddsJ3jZzxurQMMDg5W9g0NDQGwaOkSAEptbZVjw8ND4+q3\nl1orP7cQS2y3tcY+M6sc6+vrTz9Fmaamptyxvjj3yEjUI6vX3t4ebShlv4JSqQTAyMhonNGz8iOj\ncY7BgQEAmluyeh1dnemcHQAMpDIAW7ZsBqA/7evo6MgeZFo5/JLLbs0uJCKNsrijo2P52rVrl892\nQ0RE5qObbrqJ/v7+yQs22ILtHEN0MEdGsk5va2s83FJLdGCNscqxkdQ5bm+PDnNLS9bJbRodS/ta\n0jlHKsfK/eTBwaif7xyXO9GtqdObV+4Ul3LHyuWbm5vjEYxm7bNKvSg/5l45NpbKuY+l+lkbyp3h\n8r7yueP8ozu1S2S2mdl6AHdfM7st2W3r165du/yqq66a7XaIiMxLxx57LFdfffX6mb6uco5FRERE\nRJIFHDkWEZld12/Yxpq3XjLbzRCROWr9h0+d7SZIFQu2c9zTsx2A0bEsNaGzK/J8W5ojSSGXAcGS\n7sjbbUlpC6NDw5Vjzc3j0yLyaRXln8fSdfLHyikMbSm3eSzXlvKxfBqGp1SJ4XTtwaGhncqXUy88\nd66hlBLSMtyy03Xa2sq506Ppell68ZhnbRURERERpVWIyCywcI6Z3WBmA2a2wcwuNLMlNeo8z8x+\nYWZbU52bzOydZtY2QfkjzewiM7vTzIbMbJOZfdnMjqhS9iIzczN7gJm9xsyuNbN+M1vXwIctIiLz\nwIKNHHuKlHZ2ZrMztLXFw21LA/NaqgxOGxxMoyKz4CvNpfjfW57tYigX0S1HiouR3fy+8kC+/LFy\nvfy5ypHj8mwVZtlnl3I0OB9pLrY9i15nA+2sKc5Zao22lGe9SDV3OpfIDDkfeC1wN/BpYBg4DXgE\n0AqMmz7GzD4LnA3cBXwT2Ao8Eng/8Dgze4J79lWImT0J+BZQAr4H/BXYH3gmcKqZneLuV1dp18eB\nxwKXAD9AfyQiInucBds5FpG5ycxOIDrGtwIPd/ctaf87gF8Aq4G/5cqfRXSMvw28wN37c8fOBd4D\n/DPRscXMlgFfAfqAE939xlz5o4ArgM8AD6vSvIcBD3X326fweCaajuLIes8hIiJzx4LtHHenHOLO\nzq7KvvLcwmMpsjoylgWFRlL0dWQo5kVuLbVXjpWnTas29VlLS+Qjl6dKy0+/Vs7vLUeQ8/MP9/b1\nxg/ZjGyVcosWLU73s1zngYHoD5TnbR7LVWxqHv9r9Nyxcqy6HLQuT/cGUGrdeYo5kRlwdrr9QLlj\nDODuA2a+drJhAAAgAElEQVT2NqKDnPc6YAR4ab5jnLwfOAd4AalzDLwYWAqck+8Yp2tcb2b/Dbze\nzB5YPA6cN5WOsYiILDwLtnMsInNWOWL7yyrHfk0ulcHMOoGHAPcRHdpq5xsE1ubuPyrdPiRFlosO\nT7drgWLn+MpaDa/G3Y+ttj9FlKtFp0VEZA5T51hEZlp50N2m4gF3HzGz+3K7lhFfgKwk0ifqsSLd\nvnySct1V9t1T5zVERGSBWrCd467O+L9XTqUIEXXq74+UBs+tMleebq0lLRttTbnBeiNphbymNLCu\nJRe9SqkTTSmtYnAgG0c0PJgG+aV0jB19PZVjI6MxXVt+xbqu7kgBWbR4aWpfdp1ySz1Fzppacu1L\nR8vXGcmlfzSlc5RXAxzJpX2UU0JEZti2dLsKuC1/wMxagL2IgXf5ste4e71R2HKdh7j7tVNsm09e\nREREFrIF2zkWkTnraiLd4CQKnWPgMUDlk5+795jZDcCDzGx5Pke5hiuAZxGzTky1c9xQR+23hKs0\nyb+IyLyyYOc5bm9rp72tnbFRr2yjIyOMjozQ3NxCc3MLEUmOzR3cob29g/b2DlpbWytbE+mJSoUs\nV3NocIChwQHa29tob29jcHCwsvnIGD4yxqLObhZ1dmNjTZWtyZtp8mbaWjoq297LV7H38lW0lrpo\nLXUxNtZS2ZxmnGZKpVZKpVYMr2zDQ/1p62V4qBcfG6xsxjDGMKUmKDVBW0tTZVvU2c6izvYJn0OR\naXJRun2HmS0v7zSzduBDVcp/lJje7bNmtrR40MyWmVk+qvw5Yqq395jZw6uUbzKzk3e9+SIispAp\nciwiM8rdLzezC4DXANeb2TfI5jm+n5j7OF/+s2Z2LPBq4FYz+zFwB7AcOBg4kegQvzKV32xmzyam\nfrvCzH4G3ECkTBxADNhbAeiToYiI7ESdYxGZDa8DbiHmJ34FsJnozL4d+FOxsLv/s5n9kOgAP56Y\nqm0L0Un+CPDFQvmfmdnRwJuBfyRSLIaAjcDPiYVEREREdrJwO8ceg9Kam7JBbWNjMdamtdSUjjXn\njsUAubHRNOAtl3FiaYxOU1qxznOLZo2OxLzDO3bEGKC+3q2VY8uXLErnjn3NaRAeQMliMFx3S2tl\nX3Ma+Dc2HAP3Rvv7smOjMdCvsz3qlVqy9jV1xa+xuTkGFZYHFwJ0dERwbGQ46vf3Z9PEjh+sKDJz\nPEbDXpi2ojUT1Pk+8P0pXGM9MQdyPWXPAs6q99wiIrJwLdicYxERERGRqVqwkeOujjTtmmXR4eEU\nPS1PkeaefTYor37X2hr12luz6GtrmvKsVIrbsbFsOrRt23cAMDgS0eXDDjmwcqyzLa7T0hTHlnZl\nT3dba0Rtu7uzFfy6u2L6Obc4/+hIFlVesjRNTdcR7Rodzlbb8xS97u+LqPBYbhW8clsHB6MNHU3Z\nYy6VNGuViIiISJ4ixyIiIiIiyYKNHO+3em8AlixZUtlXXvTDyrnDuQhrU4qolqPDLc3ZUzM4ELm/\nHR0d424BNm6MBbU2btoMwAOPOKxybKg/coe7OsrnzHKVW1LOcD7vt3ze/t4dac9I5Vh7yjXuG+hJ\nZbL85dGh8uNJbc4HhFPKdWdXZ7pu9riaqi/FKyIiIrLHUuRYRERERCRR51hEREREJFmwaRWLuiNd\noSu3Alxra6QmDA9HSsLg4GDlWDnlojzNm+WmaxsZTlOqtTency6qHDt4zb4A3LlhAwA3XHdj5djK\nZbGYV18pBta1dmT5Dp2dsW97b9aGUktcx1K6R5Nn5bdujWPNzbFvLP+rS4MIFy9eNu6xQJZC0pMG\n642O5gbrmT4biYiIiOSpdyQiIiIikizYyHFfTyy8MTzYW9m3aFFMhzY0FJFjy0VOW9PCGcNDEWEt\n5QaurVgeg/qa0oIigwPZOUlTxR18UESQb/nz7ZVD118Xg/V8OK5z+FGHVI5tLcVUbO1t2XW60oId\nTeXp53ID60qtEWluao7od09fT+XYsMe5fHt/sVplgOFwChjnB+SNjAwjIiIiIhlFjkVEREREkgUb\nOd66eRMAHZ3ZYh4jQ7HEM2PxmaC1lFs+eSR+bq1MrZbl5pZTeL08FZtl9VpaYhGPJYtjqrRjHvag\nyrE//u4mAO6+7e9xe9fGyrGO7ojoHnnIvpV9naWI5PYPRwS4fzCL7Pb0xrRu7e0pwp3LHR4ciWPl\nBT9GR7Mp4EjR8dF029ScLYoyMqzIsYiIiEieIsciIiIiIok6xyIiIiIiyYJNqyinDIyNZmkEQ4OR\nFrG4ezkAq1auqhzrTivQldKgu1Iu/WB0ZAiAwaHtALRadqwrpW3stfd+AGy89/7KscWLYyq30ZWR\nl9Hfs71ybO+9VgJgY7n0jdEo19EWbent7ascGx6MQXdDA5FyMZwbTDcwHO0rr/LX1Jx95mlpiYF8\nO3qyKeMq59SAPBEREZFxFDkWkTnDzNaYmZvZRXWWPyuVP6uBbTg5nfPcRp1TRETmjwUbOe7ujmnb\nFi3pyu2NyPGhhx0BwNhAFjndsiGmYGsajmna1ux/QOXYir32AqCvL00BRxY5XrkiRaGPOgGAP92c\nTeVW8pgCrnfzFgBuueWayrG2UkSot27dkZVP+1rToLuBgaHKsZE0F9tIGmw3OJxFggdHIpo8NhaR\n587OjuxYWuhkdCCOmVnl2NiQIsciIiIieQu2cywie4RvA1cAd892Q6q5fsM21rz1ktluxm5b/+FT\nZ7sJIiIzRp1jEZm33H0bsG222yEiIgvHgu0cN6cBdcNDWWpCU1OkFtx2yy1xfzAbDHdfSqto80hD\nWJSbA3moL53D4ra1lM2dPLos9g31RmrDUE92vQMPfgAAN2yNeY7v3XJv5dj+B6yO8j5a2dfXH2kO\n3aUYRGet2XXKKROj6bYpd8w9pUekx9fclm97GsiXnodSS6lyrNSa/Swy15jZkcCHgROBNuAa4H3u\nfmmuzFnA54Cz3f2i3P716cejgXOBZwL7AR9w93NTmVXAB4GnAouBPwMfA/42bQ9KRETmvAXbORaR\nee1g4LfAdcB/AauBM4Efmtnz3f2rdZyjFfg5sBy4FNgO3A5gZnsBvwEeAPw6bauBT6WydTOzqyY4\ndORUziMiInPDgu0cN6Up2YZyg86am9OgtIEeAJa0ZoP12ttjAF+LxUp3Q2NZ9LWnL03v1hbR2pZS\nd+XY0Eg8hYM70jRtY1kk+K671gNwx8a/ANA/kk3NtmlrDNLrXrK4sq9jSQzga1kc7epoyQb+0RNt\nLrdqKLe6XcvI4tSulvRYsra39UfkeEfzltS8rH3tHdnAPZE55kTg39z9LeUdZnYh0WH+lJn90N23\nT1g7rAZuBE5y997CsQ8SHePz3f0NVa4hIiJ7KE3lJiJz0Tbgffkd7v4H4EvAUuAZdZ7nTcWOsZmV\ngBcAO4iUi2rXqJu7H1ttA26eynlERGRuWLCR454UabWUhwvQ2hqR2EWliJiODY1Ujo2OxbGegYis\nLhnMoraLmiNiPDYWEeQWsojrlh1pqrR7Y7D8XrmFRW69MyLGvQOxMEjH4qxeb5qKbUlXZ2XfSMoH\nHk6LeWQtB2uNPOSWlEvd0pVFrwfTNG8tKdKcj5aPtcRZVh+QpqbzLM968eIliMxRV7v7jir71wEv\nAR4K/O8k5xgArq2y/0igE7gsDeib6BoiIrIHUuRYROaiTRPsvyfd1vPJ7u/u7lX2l+tOdg0REdkD\nqXMsInPRqgn275Nu65m+rVrHOF93smuIiMgeaMGmVXR3RwqDj2QD0DrSgDobifSIgYEsraLUFukN\nd9y9GYB7tvypcqy5JT5DNBH1VqUV8wBKKd3BFsX1Dl97VOXYtq0bANgrraLXnlu5bmhkLLUlS4EY\nSekQvX2RIrlkUTZgsC2tbDeaVvXLWg5jqV3390a9bduyfkP5048tXRr3m7LPQ7lZ5ETmmoeZ2aIq\nqRUnp9tr2HU3A33AMWa2pEpqxck7V9k1R+23hKu0gIaIyLyiyLGIzEVLgHfnd5jZccRAum3Eyni7\nxGNi8C8BiygMyMtdQ0RE9lALNnK8Zr99ARjNDbprsfgssPXvMUCutbu1cmzbfTHV2dLliwDYvHVr\ndrLmOMfSxRF9LZWyb2sHByOwNdoUEd2/3X5j5djoaCy8sfeqiDSv3Gtp5djWbTEL1eb7sus0pUju\nipVRvqszW+jD0m15MY+xXAS4Lz1GT79Nb89+rV1dXemxxxn6+rKB+z07co9RZG75FfAyM3sEcDnZ\nPMdNwCvqmMZtMm8HHge8PnWIy/Mcnwn8AHj6bp5fRETmKUWORWQuuh04AbgfeCVwBnA18JQ6FwCp\nyd3vAx5NrK53JPB64BjgVcQqeSIisodasJHjzlJMi5ZLOS6vrszqfVcA2aIeABvui/zgw448BICj\n2g/L1YuT7L86otEDPdliHlddHamPbYti4Y1D1+yX1WuKvOL2jpSX3Jx9FmlJWcOl3NRqEFOxrdon\n2tfSli3vPDoa5Xt6PLU9W+ijNBALfazcOxYDGRrMlrBuTUtEtzSlxUoGByvHRkbymcsis8/d15N9\nUQJw2iTlLwIuqrJ/TR3Xugd46QSHbYL9IiKywClyLCIiIiKSqHMsIiIiIpIs2LSKjpSSsPHv92b7\nSpHecOiRDwBg/YY7KsfGLNIN9l8TU592dWYr140NR5rCskWxKl1flu3Aqn1imrYtvVGmrTX7vJHG\nwNFkkTrhuRSKvZZHCkRXR256t8FI32hpSSvxlbJzDQ/Hsbb0uFrbs0b0DcYgu1JzpFwMkaVV9PdF\nykVHe9dObTCbaBpYERERkT2TIsciIiIiIsmCjRxbir62dmbTtXV2RvS0pz8irdacjblZtW9EjJtK\nMShuxHID5WIXQ0T01tqyp+2gQw8GwDfcB0D/UBa1HRuL6d1aRtJnkKbseu0d0ZZSR3auEY/odXmB\nkN6tPZVj5YF0zc3RmNGxbGDdWIoqbx+ItQz6+wcqx3p64hxt7RH13rY9W1NhKDc4T0REREQUORYR\nERERqVDnWEREREQkWbBpFWOlSGEo5VaZa+2OuYG390ZaxZJlKyrHFi/fO+p5WkkuN1ewE2kOXWnu\n5Obm7GmzoUhpWHXAAQB0dmUD+fr7I4WhuSU+gzS1NGfnTKv1eW421fbF0b7hgZh/uDwwD6DUPJza\nkh5La9aG9tYY1Ld9eywaNtifzV9sRJv7e6P+jm3ZHM0j+UmgRURERESRYxERERGRsgUbOR4aS1Od\ntXVX9vWPxcNtKk+31pwN1mtuin39/f0A9OUHq6UoL0sj0uq5zxRDIxHLXbokpmbbtm1rdr2BiNKW\nI8ctLVm9/v6hndrQ0hqR46bmiHaPjmWRXUtRa0bjei3NWRTa0o/lAYcjubGEbWkKt5a0Ot/SJdnz\nsW1rNjhPRERERBQ5FhERERGpWLCR47/dFYt/lFqyxTJaUq5wR5o+rad/S+VYa4rqNntEa5ssn9Mb\n0d3h3pgWrdSe5RUvTguDDPRFxHhsqLdyrGkkosOjQxHK7V68ODuW8o+HLcsrvvfeaHP/UOQHd+YW\nImlLj2MoTdPm3V2VY0tSG1pbo8yKlXtXjvX2RfR6bLg3nXNpds62LGotIiIiIooci4iIiIhUqHMs\nInOKmb3WzG40s34zczN7/Wy3SURE9hwLNq1ia09Ma8ZQNqitrSkebntXpBOMkU15tmJZpDysWrEc\ngOGR4cqxpqZIi9jRE2kYi3NTrI2lqdIG0vVam7KBch0daWq2kbiOk7HmqFeyrHxnewzE6+uLgXJj\nQ9nIura2aF9Xd0rxGOrP2uCRftGa0j9GyA3WS6vytaS2NOfSTBjIHqPIXGBmzwU+DlwDnA8MAlfM\naqNERGSPsmA7xyIyLz21fOvuG2e1JQ1w/YZtrHnrJVWPrf/wqTPcGhERqceC7Rzvs88+APRv66ns\nGxuIAXKeVt5oykdRUwS3pbUt3c0yTposIr9No7FveGSgcqx/KH4eSFPAdeUG0bWmcw0OxXX7+7J6\nTSlyPDCYDeArpajzXstigF2plEWAS5YG9bWnaHRuEZCRFJkeSYMJt+7IFvrY1hOPv6U92t7e3lE5\nNjSam/NNZG7YF2AhdIxFRGR+Us6xiMw6MzvXzBw4Jd338pa7v87M9jGzz5jZBjMbNbOzcudYbWb/\nYWbrzWzIzO41s2+Z2bETXHOJmZ1vZneZ2YCZ3WxmbzSzB6TrXTQDD11EROaYBRs5bk35xdaVTXnW\nuSKWiy6l3N4xz3KOSyk3d2g4oq/5nOOWlJu8uCPq7ejN5fta7POmiAQP54Kx3W0RpW1OO0dzOb5d\nXREd9lwmcl+K8o42RRvaOrLI9qLuRVEvRX63b8sizpvv3wbAQGr7jr5sAZPNW+NYejro7MoWAfFR\nfTaSOWNduj0LOAh4b5Uyy4n84x7gW8AYsAnAzA4Gfk1Enn8OfAU4AHgOcKqZPcvdv18+kZm1p3IP\nI/KbvwQsAd4BPLahj0xEROaVBds5FpH5w93XAevM7GTgIHc/t0qxBwNfAF7qnvtkGz5FdIzf6e4f\nKO80s08CvwL+18wOcvdyntVbiI7xxcDz3b0cof4AcPVU2m5mV01w6MipnEdEROYGhQ5FZL4YAt5c\n7Bib2f7AE4E7gPPyx9z9N0QUeTnwzNyhlxCR57eVO8ap/J3ELBkiIrKHWrCR49H+SC0Ya86mXbOu\nGMzWllItFnVkg9O2bdkMQP9QSkkYy9IdzOIzRGsp6o2mle8Atg/EILvu7lh5bmAoOzbg8fS2dC4B\nYEl39lmk1ByD7Zotu05HKY53rlgWbRnJ+gDlR+EtaSq3sWyg4fBwpGv09ESqxdBo9piHBgZT21tS\n/SzvY6A/S78QmQfWu/vfq+x/aLq9zN2rzU/4c+CFqdznzWwxcAhwp7uvr1L+11NplLtPlNN8FRGd\nFhGReUSRYxGZL+6ZYP+SdHv3BMfL+8trp5fXcd80QfmJ9ouIyB5gwUaO29M0aJt7swirp3VBWlvS\nYL3ubLBeW5oizUfK07Vli4eQIrNDg7HPc4e2bUkD3iwtwDGaHdyaBs11dMT0bvlvgwdH+9P1skhz\nV4pkj6bIb19/Vr4cAW5vifKDuWMjKRhcKsXgwNb2bAq43u2xoEhbijiPDOSixZrJTeYXn2D/tnS7\nzwTHVxfKpXcCVk1QfqL9IiKyB1iwnWMR2WNck24fY2YtVQbrnZJurwZw9+1mdhuwxszWVEmteEyj\nGnbUfku4Sot9iIjMK0qrEJF5zd3vAn4CrAFenz9mZo8Ang/cD3w7d+jzxPvfh8zMcuUPKJ5DRET2\nLAs2cuykcTm51exGtqYUhvZIp+hNA/QAWpric0JLS6QmlDqyp6Z5MFIRRocjTaKUjXejNa1cN7j1\nfmD8qnvDaSW+4d5ow+BglkIxNhbpF61p7mSAUSIdon/zvQBs68/mMia1ryulR7Q0Ze0bSJkc5UH3\nY7l6rU3Rvv7+WDXPLUu5aOvM0kpE5rlXApcDHzGzJwJ/IJvneAw429135MqfB5wOPBc4wswuJXKX\nzyCmfjsdJR6JiOyRFmznWET2HO5+m5kdB7wTeApwMpFb/CPgA+7++0L5fjM7BXgf8GzgDcDtwAeB\ny4jO8XZ2z5qbbrqJY4+tOpmFiIhM4qabboL4VnBGWW6KTxGRPZ6ZvRz4NPBKd/+v3TjPINAM/KlR\nbRNpsPJCNTfPaitEJvYQYNTd2yYt2UCKHIvIHsnM9nX3jYV9BwLvAkaA7+3mJa6HiedBFplt5dUd\n9RqVuarGCqTTSp1jEdlTfdPMSsBVwFbiq7unAp3Eynkba9QVEZEFSp1jEdlTfQF4EfAsYjBeD/A7\n4EJ3/9ZsNkxERGaPOsciskdy908Cn5ztdoiIyNyieY5FRERERBJ1jkVEREREEk3lJiIiIiKSKHIs\nIiIiIpKocywiIiIikqhzLCIiIiKSqHMsIiIiIpKocywiIiIikqhzLCIiIiKSqHMsIiIiIpKocywi\nIiIikqhzLCJSBzPb38w+a2YbzWzQzNab2flmtmyK51me6q1P59mYzrv/dLVd9gyNeI2a2Toz8xpb\n+3Q+Blm4zOzZZnaBmV1mZtvT6+mLu3iuhrwfT6SlEScREVnIzOwQ4DfA3sB3gZuBhwOvA55kZo92\n9811nGdFOs/hwM+Bi4EjgbOBU83sUe5+2/Q8ClnIGvUazXnvBPtHdquhsid7J/AQoAe4i3jvm7Jp\neK3vRJ1jEZHJfZJ4I36tu19Q3mlmHwXeAHwAeGUd5/kg0TH+qLu/KXee1wIfT9d5UgPbLXuORr1G\nAXD3cxvdQNnjvYHoFP8VOAn4xS6ep6Gv9WrM3XenvojIgpaiFH8F1gOHuPtY7tgi4G7AgL3dvbfG\nebqBvwNjwGp335E71gTcBhyUrqHosdStUa/RVH4dcJK727Q1WPZ4ZnYy0Tn+kru/cAr1GvZar0U5\nxyIitZ2Sbi/NvxEDpA7u5UAn8MhJzvNIoAO4PN8xTucZA35cuJ5IvRr1Gq0wszPN7K1m9kYze7KZ\ntTWuuSK7rOGv9WrUORYRqe2IdHvLBMf/km4Pn6HziBRNx2vrYuBDwL8DPwDuMLNn71rzRBpmRt5H\n1TkWEaltSbrdNsHx8v6lM3QekaJGvra+CzwN2J/4puNIopO8FPiqmSknXmbTjLyPakCeiIiIAODu\nHyvs+jPwdjPbCFxAdJR/NOMNE5lBihyLiNRWjkQsmeB4ef/WGTqPSNFMvLY+Q0zjdkwa+CQyG2bk\nfVSdYxGR2v6cbifKYTss3U6UA9fo84gUTftry90HgPJA0q5dPY/IbpqR91F1jkVEaivPxfnENOVa\nRYqgPRroA66Y5DxXAP3Ao4uRt3TeJxauJ1KvRr1GJ2RmRwDLiA7yfbt6HpHdNO2vdVDnWESkJne/\nFbgUWAP8c+Hwe4ko2hfyc2qa2ZFmNm71J3fvAb6Qyp9bOM856fw/1hzHMlWNeo2a2cFmtrx4fjNb\nCXwu3b3Y3bVKnkwrMyul1+gh+f278lrfpetrERARkdqqLFd6E/AIYs7NW4AT8suVmpkDFBdSqLJ8\n9JXAWuA0YoGQE9Kbv8iUNOI1amZnAZ8Cfk0sSrMFOBB4CpHL+QfgCe6uvHiZMjM7HTg93d0H+Efi\ndXZZ2nefu785lV0D3A78zd3XFM4zpdf6LrVVnWMRkcmZ2QHA+4jlnVcQKzF9G3ivu99fKFu1c5yO\nLQfeQ/yTWA1sBn4IvNvd75rOxyAL2+6+Rs3swcCbgGOBfYHFRBrFDcDXgP9y96HpfySyEJnZucR7\n30QqHeFaneN0vO7X+i61VZ1jEREREZGgnGMRERERkUSdYxERERGRRJ3jBcjM1pmZp8EVU617Vqq7\nrpHnFREREZkPFvTy0Wb2emJ97Yvcff0sN0dERERE5rgF3TkGXg8cBKwD1s9qS+aPbcQKNHfMdkNE\nREREZtpC7xzLFLn7t4npUERERET2OMo5FhERERFJZqxzbGZ7mdmrzey7Znazme0ws14zu9HMPmpm\n+1apc3IaALa+xnl3GkBmZuemCc4PSrt+kcp4jcFmh5jZf5nZbWY2YGb3m9mvzOxlZtY8wbUrA9TM\nbLGZnWdmt5pZfzrP+8ysPVf+cWb2YzO7Lz32X5nZYyd53qbcrkL9ZWb2sVz9u8zs02a2ut7ns15m\n1mRmLzKzn5jZvWY2ZGYbzeyrZvaIqZ5PREREZKbNZFrFW4mVdwBGgO3EcpRr0/ZCM3u8u1/bgGv1\nAJuAlcQHgPuB/Ko+W/KFzeypwNeBckd2G7E+92PTdqaZnV5jre5lxDKwRwC9QDNwMPAu4Bjg6Wb2\nauBCwFP7OtO5f2pm/+DulxdP2oB2rQB+DxwC9BPP+37Ay4HTzewkd79pgrpTYmaLgG8Bj0+7nFhZ\naTVwBvBsM3udu1/YiOuJiIiITIeZTKu4A3g7cDTQ4e4rgDbgOODHREf2y2a203KrU+Xu/+bu+wB3\npl3PdPd9ctszy2XTGt0XEx3QXwJHuvtSYBHwCmCQ6PB9vMYly8shPtbdu4FuogM6AjzNzN4FnA98\nGFjh7kuANcBvgVbgY8UTNqhd70rlnwZ0p7adTCzJuBL4upmVatSfis+n9lxNrJfemR7ncuCdwCjw\ncTN7dIOuJyIiItJwM9Y5dvdPuPuH3P06dx9J+0bd/SrgNOBG4EHAiTPVpuTtRDT2VuAp7v7n1LZB\nd/808NpU7qVmdugE5+gCnuruv051h9z9M0SHEWL97y+6+9vdfWsq8zfgeUSE9XgzO3Aa2rUYeJa7\nf9/dx1L9XwJPJiLpDwLOnOT5mZSZPR44nZjl4h/c/VJ3H0jXu9/dPwC8m3i9vW13ryciIiIyXebE\ngDx3HwR+ku7OWGQxRamfle5+zN37qhT7DLABMODZE5zq6+7+1yr7f5r7+UPFg6mDXK531DS067Jy\nh71w3T8D30h3J6o7FS9Jt//t7tsmKPOldHtKPbnSIiIiIrNhRjvHZnakmV1oZtea2XYzGysPkgNe\nl4rtNDBvGj2AyHsG+EW1Ainiui7dfdgE57lugv1/T7cDZJ3gok3pdtk0tGvdBPshUjVq1Z2KE9Lt\nO83snmobkfsMkWu9ogHXFBEREWm4GRuQZ2bPJdIMyjmuY8QAs8F0v5tII+iaqTYRebdlG2qUu6tK\n+by7J9g/mm43ubtPUiaf+9uodtWqWz42Ud2pKM98sbTO8p0NuKaIiIhIw81I5NjMVgL/TXQAv0oM\nwmt392XlQXJkg9J2e0DeLmqfvMismKvtyiu/jp7h7lbHtn42GysiIiIykZlKq3gyERm+EXi+u1/l\n7sOFMquq1BtJt7U6iEtqHJvMvbmfiwPi8vavUn46NapdtVJUysca8ZjKqSG12ioiIiIy581U57jc\nibu2PGtCXhqA9g9V6m1Nt3ubWesE5z6+xnXL15ooGn1b7hqnVCtgZk3E9GcQ05TNhEa166Qa1ygf\na34Z6X8AACAASURBVMRj+m26fXIDziUiIiIya2aqc1yeweCoCeYxfjmxUEXRLUROshFz9Y6TpjB7\nVnF/zvZ0WzUXNuUBfyvdfZ2ZVcuFfRmxcIYTC3JMuwa26yQzO6G408wOI5ulohGP6aJ0+49m9qRa\nBc1sWa3jIiIiIrNppjrHPyU6cUcBnzCzpQBpyeW3AP8BbC5Wcvch4Lvp7sfM7DFpieImM3siMf1b\nf43r3pBun5dfxrngg8SqdvsCl5jZEaltbWb2cuATqdz/uPutdT7eRmhEu7YD3zKzp5Q/lKTlqn9I\nLMByA/C13W2ou/+I6Mwb8G0ze0vKMyddc7mZnW5m/z/w0d29noiIiMh0mZHOcZpX9/x09xzgfjO7\nn1jW+TzgZ8CnJqj+NqLjfABwGbEkcS+xqt5W4Nwal/6fdPscYJuZ3Wlm683s4lzbbiUW4xgg0hRu\nTm3bAXya6ET+DHh9/Y949zWoXe8nlqq+BOg1sx3Ar4go/b3AGVVyv3fVi4HvEPnh5wGbzOx+M9tO\n/P6+TZXov4iIiMhcMpMr5L0R+CfgGiJVojn9/HrgVLLBd8V6twGPAL5CdOiaiSnMPkAsGLK9Wr1U\n9+fAM4g5ffuJNISDgH0K5b4HPJiYUWM9MdVYH/Dr1OZ/dPfeKT/o3dSAdm0GHk58MNlELFW9MZ3v\nGHe/sYFt7XX3ZwBPJaLIG1N7S8Qcz18DzgZe06hrioiIiDSaTTz9roiIiIjInmVOLB8tIiIiIjIX\nqHMsIiIiIpKocywiIiIikqhzLCIiIiKSqHMsIiIiIpKocywiIiIikqhzLCIiIiKSqHMsIiIiIpKo\ncywiIiIikrTMdgNERBYiM7sdWEws/S4iIlO3Btju7gfP5EUXbOd4zTFrHKCzq72yb3h4AICjjl4L\nwE03/LVy7LZbNgDQ0VICoKVlrHKsrasz1TcAurraKsdW7r0YgG3bewBobs6ut2rVvnHO9i4ASq3d\nlWNLl8fPG+++rbKvsyv2Pf6kJwLwoIc8tHLMWuO8D1gR57zimj9mj3XlPtG+/i1R1roqx9rbo61L\nFkX9rlKpcqypFMf2OXA/Q0QabXFHR8fytWvXLp/thoiIzEc33XQT/f39M37dBds5Xr3fSgDMRiv7\n7r57GwBb7rsfgH1Xr64c2/C3jQCM+XDcWpZx0t8/FOdKT1d7R9Y5HhqOX9rAjugc5w4xsG07AKN9\nUb+9fbByrHmsF4BNd66v7BsZ6QOg9/77ALjsj9dUjnV0dQDw0me8AICr/nxT5dhYRysA27feA8CV\nv7u5cmzvvaPjfOzD4gPBPkuWVo495KiHICLTZv3atWuXX3XVVbPdDhGReenYY4/l6quvXj/T11XO\nsYjMSWbmZrZuCuVPTnXOLexfZ2be6PaJiMjCpM6xyAIx1c6kiIiI7GzBplXstdfeAAyP9FT29fRE\nCsSGO/4OQHdnR+VYR2c8FUPDkVZRas1ycxmLzxBNTeWy2SFrilSJ5lLUGx7trRy7b3Mc6+xYFGUt\nSz0cvi/KDfSMVPa1lKINY0ORhlHq3145tqwjAl+9d/0FgBVtWZrwvZti3z2bbgXgD3/4ZeXY3ntH\neknrkq0ArO/K0io6VsYDOW7/wxFZAK4E1gL3zXZDyq7fsI01b71ktpshIjIr1n/41Nluwi5ZsJ1j\nEdmzuHsfcPOkBUVERGpYsJ3j/VYfCMCtt19X2Tc2EtHg/u0RtfWhbZVjhxy6PwAjRIS2uTkbWbes\nO2aRcI8ZLFauykLHw6Nxjk3LYoaI++7dUTl2770x8K+vN6LXA6PZiMvOtr0AaGlZkTXa49o7tjZH\nG9hSOTS4KaLdV/bG4LutHVkUeqlHZHrTPXfFsW0bKscWLYnodU9vRJd7BrPH9d2f3gnAcWd9EJl+\nZnYW8DTgocBqYBi4DvhPd/9ioex6AHf/f+zdeZxcV3nn/89TS69St1q7rMWSd4PjPYAxwWIJGBwG\nwsCPJQs2WXAIYQkzwSRksJOBkEwCJCbGhAScEBNDwgAZAsHBIAw2YPCC8b5KtrVY+9bqraqe3x/P\nqbpX7epFcrelrv6+Xy/5Vt9z7znntsqtU08/55zVTeq5HPgg8CJ3X5fq/WwqvmBUfu0V7n557t7/\nD3gHcAbQBjwEfB74qLsP5e5r9AE4DfhT4HXAQuB+4HJ3/4qZlYD3ARcDK4GNwMfc/RNN+l0Afhv4\nDSLCa8A9wGeAT3n9f7Cn3ncM8OfAy4G56Z6/cvfPj7puLfCd0c88HjN7OfAu4Dmp7ieA/wt8yN13\nT6YOERFpLS07OBY5Cn0SuBu4EdgMLABeCXzOzE529z8+zHrvAK4gBswbgGtyZevqL8zsw8D7ibSD\nzwP7gVcAHwZebmYvc/fhUXWXgf8C5gNfJQbUbwK+ZGYvA94OPBf4BjAEvB640sy2ufsXRtX1OeDN\nwOPA3wMO/DJwFfAC4FeaPFsfcDOwm/gAMA/4/4BrzWy5u/+fCb87YzCzDwKXAzuBrwFbgdOB/wG8\n0szOc/e9Y9fQqGes5ShOOdy+iYjIkdOyg+ML1r4KgFtvy5ZDe2xDpCIeuyyixIODWxtlGzfF62J7\n5CF3dGTBt+7OWCN4zXERjV62IltHeO/+uK5tblpPeEm2lvGxQ0sAGIkV2ti9LRt37N0Z99Uq2fVe\ni3M7+uO6rU/uaJS1jcS/0cPbI2q98tyVjbJSOaLPxa6oq29JllfcuzTWYa7ZrtRe9lw79x01qZmz\nxWnu/nD+hJm1EQPLy8zsanff2PzWsbn7HcAdabC3vlnU1MzOIwbGjwPPcfct6fz7gS8Dv0QMCkf/\nGuEY4DZgbT2ybGafIwb4/wo8nJ5rdyr7KJHacBnQGByb2ZuIgfHtwAvdfX86/wHgu8Cbzew/RkeD\nicHqvwJvrEeWzewjwK3Ah8zsS+7+CIfIzF5EDIx/ALwyHyXOReKvAN5zqHWLiMjMptUqRJ4howfG\n6dww8LfEB9WXTGPzb03H/10fGKf2K8B7gRrwm2Pc++58yoW7fw94lIjqvi8/sEwD1ZuA08ys2KT9\ny+oD43R9P5GWwRjtV1Mbtdw9jwJ/Q0S1f23MJx7fO9Pxt0anT7j7NUQ0vlkk+ync/Zxmf1D+s4jI\njNSykWORo42ZrSIGgi8BVgGdoy5ZPo3Nn52O3x5d4O4PmNkTwBoz63X3Pbni3c0G9cAmYA0RwR1t\nI/GzZWl6XW+/Ri7NI+e7xCD4rCZlj6XB8GjriDSSZvdMxnlEzvfrzez1TcrbgEVmtsDddzQpFxGR\nFtWyg+NCe4wzjll5XuPck1sj9WHJyuMB6J6TPX7/UASPqmlDvX37siXZyp2RmjBSjjorbdl989Mm\ne33LYkm2YiGbU1QeihSGrXdF3Xdt3JT1L33rq57t4FcopFSLUpwr5CYFdqQJfJV9UVfXniwlZHBj\ntN27MNpe++ITG2XV1NdKMVI1CpWsPTPtGv1MMbPjiKXG+oDvAdcDe4hB4WrgLUD7WPdPgd503DxG\n+WZiwD4v9atuT/PLqQCMGkgfVEZEdvPt72yS04y7V8xsO7C4SV1PjtF+PfrdO0b5RBYQP/8+OMF1\ncwANjkVEZpGWHRyLHGV+nxiQXZJ+bd+Q8nHfMur6GhG9bGbeGOfHUx/ELiXyhEdbNuq6qbYHmG9m\nZfe0R3uSVrxYCDSb/LZkjPqW5uo93P4U3H3+hFeKiMis0rKD4527YgLacG58sfT4ZwFwzHHnAjC/\nJ5sM13/g/rhvZxy3bsrmRdXmxb/PlWpMfHvo0ezf4xXLI5q8bHFMlOvuyCKzD/8kJgPe8NXvA9CZ\nCwx2tMUkP6tmwbWiReTXU+BtYCDbIGQ4hbSHqpH6+Wh1V6Ns2dCzAehdEM/a3Zu1M5iqGEnpn+ZZ\nmvmINtR9Jp2Qjl9qUnZBk3O7gNObDSaBc8doowYUxyi7nUhtWMuowbGZnQCsAB6dxuXLbifSSV4I\n3DCq7IVEv29rct8qM1vt7utHnV+bq/dw/BC4yMye7e53H2YdEzpteS+3ztBF8EVEZitNyBN5ZqxP\nx7X5k2md3WYT0W4hPrxeMur6i4Hzx2hjB7HWcDOfSccPmNmiXH1F4C+JnwX/MFbnp0C9/T8zs8ZC\n4en1R9KXzdovAn+e1kiu37OGmFBXAf65yT2T8bF0/HRaR/kgZtZtZs87zLpFRGQGa9nIschR5ipi\noPuvZvZvxIS204ALgS8Cbxh1/ZXp+k+a2UuIJdjOJCaSfY1Yem20G4A3mtn/I6KwI8CN7n6ju99s\nZn8B/AFwV+pDP7HO8WnA94HDXjN4Iu7+eTN7NbFG8d1m9hVinePXEBP7vuDu1za59U5iHeVbzex6\nsnWO5wF/MMZkwcn05wYzuwz4M+BBM/s6sQLHHOBYIpr/feLvR0REZpGWHRx/99tfB6CzZ27j3OK+\nCJjtHIgUhccef6BR1r/nTgCeeOQHABzY21htis2x8RzPqkWqp5eytIX+XRHQKp0SqYtD3VlaxfBA\npFycfvovxv3LswDVkgUxj2hOZ7ZmcrGW0ikOxGTA9ZuzuUhbtsfrXdviOKeQSxdpj9ddFikatWrW\nh7aUOrGf+M18lqgB5YJ+cfBMcfc709q6/xu4iPh/76fAa4kNLt4w6vp7zOylxLrDryL+6r5HDI5f\nS/PB8buIAedLiM1FCsRavTemOt9nZrcTO+T9OjFh7mHgA8SOc0+ZLDfF3kSsTPFW4G3p3L3AXxEb\npDSzixjA/wXxYaGH2CHvL5usiXxI3P3PzewmIgr9AuDVRC7yRuDviI1SRERklmnZwbHI0cbdbwZe\nPEbxU5YOcffvE/m4o91JbGAx+vqtxEYb4/XhOuC6ifqarl09TtnaccouJraTHn2+RkTQr5pk+/nv\nya9O4vp1NP8+rh3nnu8TEWIRERGghQfHP/r+jQAsO+n0xrlVp8ZSr488ERHj/h3Z0mpd6TvR3bEa\ngIU92RK0c/piIl5HKSKtZtkkun2bY/u7H22LOpevXNgoe95ZLwXghAtiQl9fIYs4dxbi3/BitrcC\nvR1RfuBAxHePHcmWhdtXHQDg0ftjX4H5heyvzmqx012xEFHldsvu20hM3DuQlowr5uZrVXPLuomI\niIiIJuSJiIiIiDS0bOR4x+bYI2D/gSyK2tsTewws6IzH3nMgW7Vq4ZpjAVi6cDUAu3fsbJS57QNg\n25Z7ABjYleUj1/ZGLu/q054LwAknndEoG2qLfOdH0npqOzY0du2lI0V3ly/oyPpXjH5teDzaHlly\nbKNs7opYhnZzTywFe9yJJzfKhjdHvxYcSM9+552NspElsexc+ZiUjzxAplZDRERERDKKHIuIiIiI\nJBoci4iIiIgkLZtW0dUdk+ja27JJcA/fcTMAbe2xfNrw/gONsrvviM25hmuxmtXQUDZRrjQSS6sx\nEtf3dmTLw516ynMAOPOC2JfBOrMl1vYMxvXDw/EZ5P4nn2iU9fZGH04/NtuzoX9HpG88dCBSNYY2\nZxMGh/u3ArBjy2YA5nX3NMqOr8R6bWvmxI66i+dubpTt9cij2DQcddYq2eehWk1b5ImIiIjkKXIs\nIiIiIpK0bOS4WopHG6pHfYHtmx8FYO7c2MxjTl+27JpXImo7sj2itW3D2X4Ic9PqZ4tSVPikE7NJ\ndyc/N+3k2x0R6rbO7PPGnLaYbLf3iahzYS1bOq3X47rOLVk0eeP6iPju7o/7ysPbG2Wdw3Hvit3b\nAFiyKWvn57pjWbjO3RsBGDiQRZxLtcF4saS+Y+9Io6xSzW8JIiIiIiKKHIuIiIiIJC0bOS53RIR0\n6Ypsy+bNFhtiDGyL5c0G9+9tlPV0RhT1eT0RAX52e2+j7OSVse30mpVRl3ctapTt2h/R2s2PRl1d\nS5c0yorliDRX9saScS87dkWjrDtFbXsfubtxbtWeiBT//Ny4bm8u0nxuW/xVnbAgll/r3fWzRtnQ\nfdGHH6SNSO7fmUWj+14QUe6ix6YmQ5UsWlwqtuxfv4iIiMhhUeRYRERERCTR4FhEREREJGnZ36uv\nWBmpCceecGLjXIdHSsJtj30LgDXlYqPs9SecAMCLT14OQFc1W8rt5DVRR8/8+QAcKHU1ykbmRhrF\n7nT5zuFsAuC+oUjf2Jcm+x03f3mjbGEhUiYe/s8fNM4dk5aRO3FZ3Ddi2WeXvs2xK1//tthl7+7t\njzfK1m+PZd5uSXUOdGVLza0llnyz/qi75tmueF4zRERERCSjyLGIHMTM1pnZtC+CbWarzczN7Jrp\nbktERGSyWjZybESEtDqcbfQxN/17//NL5wDwtjNObZT995UnAVDxiNDe+fi2RtmOh2KJtbmnRuS4\na34WOfa0AcfcjvicceyCbCJfrRaT32oLY2m2Snc2AXBk68MAtO3LNuzoHorI74JtjwCwe9euRtkD\nu2JJtnuHo84HPZtYt7UQEeADp50DwP69Wdld920AoH35qugv2ZhnaDiLjouIiIhICw+OReSw/TrQ\nNeFVIiIiLUiDYxE5iLs/dqT7ICIicqS07OC4oytSEhbNn9c4d3I5Hvf8zpis96LeLOW6rXsAgFJa\nD/iMY09ulFUtJu5VLNZOLgxm6RFeifuqvdFOuZalLdieSM3wLbFj3e6H1zfKuvufBKD2xJbGuft2\nR1291Ujt2DCQ7Wa3PlW7rRgpFP25bPGRNAGvd2X0eftDDzfK9nbGGsh97ZGWURxpb5R1dWWvpbWZ\n2cXAq4CzgGXEVok/Az7p7v886tp1wAXubrlza4HvAFcAXwc+CJwH9AFr3H29ma1Pl58BfAj4ZWAB\n8AhwNXClu0+Yy2xmJwFvBV4KHAv0AFuAbwJ/4u5PjLo+37evpLbPB9qAHwPvd/ebm7RTAn6biJQ/\ni/h5eD/wD8BV7rnZqyIiMmtoQp7I7PBJYqB5I/Bx4Lr09efM7E8PoZ7zgO8BHcBngH8EhnPlbcC3\ngJenNj4NzAP+GvjEJNt4LXAp8DjwL8CVwD3AbwI/NrPlY9x3LnBz6tvfA18DXgDcYGYn5y80s3Iq\n/9vUv88Df0f8TLwyPZeIiMxCLRs5Pv7kiKYunTvQOHdqWrrtpAci8ttW3dcoG5obnxPK1T4AuorZ\ncmiFjjIAPhA73dlQNlGu2h674Hkt6qzuzIJNBx5LE+seuheAfU9kk+92pMlwj+/JJs9trkUfBtKS\nbBuK2WeXoUrU25fibosKWdmDaZJdmejf0uPLjbKOvniearqvvZRFi4erCozNIqe5+8P5E2bWBnwD\nuMzMrnb3jZOo52XApe7+qTHKlxGR4tPcfSi180Eigvt2M/uCu984QRufAz5Wvz/X35el/n4A+J0m\n910EXOLu1+TueRsRtX4X8PbctX9EDOA/Abzb3avp+iIxSH6rmf2bu391gr5iZreOUXTKRPeKiMjR\nR5FjkVlg9MA4nRsmIqcl4CWTrOqOcQbGde/PD2zdfSdQj05fMom+bhw9ME7nrwfuJga1zdyUHxgn\nnwEqwHPqJ8ysAPwekarxnvrAOLVRBd4LOPArE/VVRERaT8tGjjtS4LfrQJYfvPKx+O3vvm2RSvnA\nyp5G2cL98e9jbyE28djbmdVVOBCR2DlpWbjc3iF4yjG2/RGhfmDnnkbZTZsjn3j9zp0AnFRr/BtM\nIUVtHyplKZg70r/RtZRq3J7LzlzSHn2YVz9ZyjpRmB+dLXdG22tWZ3nWBUuLDlTi+qFqFqmuVad9\nKVs5SpjZKuB9xCB4FdA56pKxUhVGu2WC8gqR2jDaunQ8a6IGzMyIgenFRP5yH5D7v+6gNI68n4w+\n4e4jZvZkqqPuJGA+8CDwgWjuKQaAU5sVNGnjnGbnU0T57MnUISIiR4+WHRyLSDCz44hBbR+RL3w9\nsAeoAquBtwCTnZ25ZYLy7flIbJP7epuUjfZR4N3AZmIS3kZisAoxYD52jPt2j3G+wsGD6wXpeCIx\nsXAscybRVxERaTEaHIu0vt8nBoSXjE47MLM3EYPjyZro1w0LzazYZIC8NB33jL5hVH8WA+8E7gKe\n7+77RpW/6RD6OpZ6H77s7q+dgvpERKSFtOzg2GKeHPs3bm+cK2+KNMZFnfEb5R+v39Qoa9u3CIAV\nJ0dqYnnBokZZ8cm7AVh4IJZYK+7ub5Tt3hXnHuyP4zf2ZTvy3VaKwNO+jtUA3LP58UbZse0RCBvM\njTU6iF/vHtMZO+r1tWXBvANzo2xkcfS9a/UxjbKzlkf9g71x32BhsFHmIylVoxCBs8FKbhKesipm\nixPS8UtNyi6Y4rZKwPOJCHXe2nS8fYL7jyPmQlzfZGC8IpU/XfcRUebnmVnZ3UcmukFERGYPTcgT\naX3r03Ft/qSZvZxYHm2q/ZmZNT7Zmdl8YoUJgM9OcO/6dHxBWjmiXsccYlm4p/2B3t0rxHJty4C/\nMbPR+deY2TIze9bTbUtERGaelo0cey0ipJurWVDoeztjU46LuiPCevaiLDK7oytSIQ90RkR2TnFh\no6x9b6zUtG1LbNyxcUO2JNuG/RGlXZf+HX946dJGWdvKNQAs2BXzhx4/kN3Xk6K259WyZdeWzItw\n9+6+6N/evizlsbQs0iS7jl0SJ+ZnS80N1cPkxDNXh7L5SkP7Ispd6oi6yuXsmSsKmM0WVxGrRPyr\nmf0bsAk4DbgQ+CLwhilsazORv3yXmf07UAZeRwxEr5poGTd332Jm1wFvBO4ws+uJPOVfBAaBO4Az\np6Cff0pM9rsUeJWZfZvIbV5M5CKfTyz3ds8UtCUiIjOIIsciLc7d7wReRKwicRGxRnAPsdnG1VPc\n3DCxs931xAD3bUSO77uAd0yyjt8APkysqPG7xNJtXyPSNcbNWZ6slErxGmJ3vPuBXyKWcLuQ+Ln4\nx8C1U9GWiIjMLC0bOR5JS7M9sj2Loj6aVk7dNBz5u2fkNtlYY3GdPRqrUA3fky15Vt0Rucnbt0bk\n98mRLOK6vSOWSts8GNHe+YtOaJStPDleP3R/pFl2bszmKB1fiwjuSQuy3+juWR11bF8TUeK2FSuz\nB2qL60baIko8nFuGrVaLvtfSbrdey/W9YOlclI1Us+/H8JAix7NF2j75xWMU26hr1za5f93o68Zp\naw8xqP3dCa5b36xOdz9ARG3/qMlth9w3d189xnknNhz53Hj9FBGR2UWRYxERERGRRINjEREREZGk\nZdMqdm2JpdUeu3NH49yCztUAbFixDICNP7utUbbkibsA6G6L9INTOrsbZQvTR4ih4cjLaG9ra5RZ\nSs3Y0xYT8qoD2VJple2R+lAoxSS6YvXRRll1eaRJPHr6/Ma5WkqnaJ8fO/eN5CbMDaaJhUWLCXwV\nz9opeaRRmEd71VpW5m1x/UjayW9gIFtqruKT+i25iIiIyKzRsoNjEXlmjZXbKyIiMpO07OC4Hjyt\njGRR1O0DewGYuzSitXPSZDqA+/fFfgO17ji3eyib1HaWxXJofYWIvg5XhxplTw7HdVtSFHbhSLYB\nR6EY55YviGXXyi85vVHWvjrODa7MlmSL3XxhZH+0N5zbZKw+/25ff9oht5BFfUuFgyPANcu+do+I\ndn36XqGUZdJYLsIsIiIiIso5FhERERFp0OBYRERERCRp2bSKzjmxZvCClT2NcxsfiXWK7793FwBr\n+hu709LdHukNxz/72QCUfnZ/o8z3RspFT9pZr2tOtstcx1CkPswdifSKvoVZWkXv8kiPKLTFce6c\nvkZZweJzyb79uQlylQEABkcibaOUm/jXll4Xa1FW82ydYy9EXSOV6EutlqVVWP11unxoOJvkd2Bo\nABERERHJKHIsIiIiIpK0bOS41B7j/uNPWZqdK0ZUd/MTMTFv665swts8i2hw+xOPA7CmPfvWFBf0\nArDTItI61JWVLTp+OQAvWhYR6rbF8xplc/tiObm2zjTxLTfBrlaJiO5IJYvkevqsMq8vlnSzUtZO\npRKR6ZGBiBwb+Ql5cV2tmib0jWTtGBEdr6UJg4VCVmc1u0xEREREUORYRERERKShZSPH/YMRte3u\nyfKKTz4tosgdfbEBx55FWX7wvn0pKjy8FYBqLnI80B4Ju3N7IoI8Z9WiRln38asBWFqOnOAKWSS4\nVI5obX05OSvk84TjXLFcbpwrFNJGIinaW7Ss7/X9Omrp80xlJGsHixDwcFpWrq3c2SiqphXphoci\n4lwsZu3NnduLiIiIiGQUORYRERERSTQ4FhERERFJWjatoq0jLYOWmwRXSGkOixfHLnh9x2ST56pp\nzlxnNdIOOnOfG9rmxLlC7xwA9pWzsqHu+BZ2Ecu8latZ6sRwWpJtZDhVXsx21qMU140MZdeXStHn\nUkrRGOnvb5TVJ+51dXXHY3m2DNu+fWnJuJSGUc1/5En5GG1tkWoxnJuFNzI8jMhsZGargUeBf3T3\ni49oZ0RE5KiiyLGITAszW21mbmbXHOm+iIiITFbLRo4750R0uFTJIqVD/RHBndcRx6Fybmm1FLWd\nU4gIcCH3ucFT4HfXgYjWLpg/v1HWW4qI7MhwihJX9zfKrBCz4drboy9WzibYVT3Kcnt5NCbPtRej\nL/ml3IbS8mz7+wfSfdmNxVJEtivDcU21mkWoS6lspJKuL2R96OjoQkREREQyihyLiIiIiCQtGzke\nSptlFDtyWzDPjXzdwf0RCm7vyraB9pSba+nzQjm3jFpnCjDPaUv5yMVsA45CWpKtfyC2ga7mItVd\nHVF/R+pDLZePXEibjtSoNM4V26NNt7iuMpK1M5i2qR5MecKlcvZXZynkPDgQecnlUvbM1bRF9Ei1\nvpxctpRbsZxdJzKVzOxy4IPpy7eY2VtyxZcA64HvAFcAX0/Xngf0AWvcfb2ZOfBdd1/bpP5rgLfU\nrx1V9hzgvcALgIXATuBnwN+7+xcn6HcB+BjwTuDLwK94PsFfRERaXssOjkXkiFoHzAPeBfwU+Equ\n7I5UBjEgfj/wfeAzxGD2sGeKmtlvAZ8EqsC/Aw8Ci4FzgbcDYw6OzawDuBZ4LfC3wDvd60lVIABO\nzAAAIABJREFUIiIyW2hwLCJTzt3Xmdl6YnB8h7tfni83s7Xp5cuAS939U0+3TTN7FnAVsBf4BXe/\ne1T5inHunU8Mpp8PXObuf34I7d46RtEpk61DRESOHi07OK6mCWz91ew3or3zIq2ibc5cAIq5CW9t\naaLayFAErcqlLP1gTndM0htO6QsDw1mdg9W4vp6u0JabdOcplWFkKNIdCrlUjRr11IksrWK4Fqkg\n3XOjX17L0ioGDgwe1Acsl9qRjvVJe7VcsKtSi/7V0zFqtawP3d09iBxhd0zFwDj5HeJn2p+OHhgD\nuPsTzW4ys2OB/wSOB37N3a+dov6IiMgM1LKDYxGZEW6Zwrqel47fOIR7TgZ+AHQDr3D3Gw61UXc/\np9n5FFE++1DrExGRI6tlB8dppTRK7VkEuES8rgdPa9Uswlooxsl6dLeSmzxHmrhWKESMtjaUWyot\nRZw7O2NZtNxcPbySJs8V4mQlN1mv3BZ15lqhI9XfkTbs2NG/L7s+tbNvfywVVyxmC410dERku29B\nRMT37z+Q9a8UZT4U3xAjixwPD2VRa5EjZMsU1lXPY954CPecBMwn8qBvm8K+iIjIDKWl3ETkSPIJ\nysb6AD+vybnd6bj8ENr/f8AfAmcCN5jZgkO4V0REWpAGxyIyXeq/KimOe9XYdgErR580syIxmB3t\nh+n4ikNpxN3/DHgPcBawzsyWHGI/RUSkhbRsWkV9l7mRgSwFoj+lPDR2qivkPxvEuTndvVGUC2gN\npElwBaKCWi0rq1Xru9LFsa09Wzt4aDBNkKuPEXK709XTJDrasrWWS6k/9VSNttyEwUJqsiOleBi5\nyXopzaOrK+0K2NaRlQ3GRL5yOc4VC1l79XMi02QXEf1ddZj33wJcaGYvc/frc+c/ABzb5PpPApcC\nf2xm33T3e/KFZrZirEl57v5xMxskVrv4rpm92N03HWa/RURkBmvZwbGIHFnuvt/MfgT8gpldCzxA\ntv7wZPwl8HLgq2b2BWIzj+cDa4h1lNeOau8eM3s7cDVwu5l9lVjneAHw88QSby8ap79XpwHyPwA3\npgHyY5PsazOr7733Xs45p+l8PRERmcC9994LsPqZbrdlB8f/+OffsYmvEpFp9mvEjnMXAm8CDHiC\n2CFvXO5+g5m9BvhfwBuBfuC/gDcQO+s1u+fTZnYX8D+IwfNrgO3AncDfT6LNa8xsCPgnsgHyIxPd\nN4Y5AwMD1dtuu+2nh3m/yNNVX2v7viPaC5nNnu57cDUR2HhGmft482FERORw1DcHGWupN5Hppveg\nHGkz9T2oCXkiIiIiIokGxyIiIiIiiQbHIiIiIiKJBsciIiIiIokGxyIiIiIiiVarEBERERFJFDkW\nEREREUk0OBYRERERSTQ4FhERERFJNDgWEREREUk0OBYRERERSTQ4FhERERFJNDgWEREREUk0OBYR\nERERSTQ4FhGZBDNbYWafMbNNZjZkZuvN7ONm1neI9cxP961P9WxK9a6Yrr5La5iK96CZrTMzH+dP\nx3Q+g8xcZvY6M7vSzL5nZnvT++WfD7OuKfl5Ol1KR7oDIiJHOzM7HrgZWAx8FbgPeA7wLuBCMzvf\n3XdMop4FqZ6TgG8D1wGnAJcAF5nZee7+yPQ8hcxkU/UezLlijPOVp9VRaWUfAM4A9gNPED+7Dtk0\nvJennAbHIiITu4r4Qf5Od7+yftLMPgq8B/gQcOkk6vkwMTD+qLu/N1fPO4G/Tu1cOIX9ltYxVe9B\nANz98qnuoLS89xCD4oeAC4DvHGY9U/peng7m7keyfRGRo1qKcjwErAeOd/darmwusBkwYLG7949T\nzxxgK1ADlrn7vlxZAXgEODa1oeixNEzVezBdvw64wN1t2josLc/M1hKD42vd/VcP4b4pey9PJ+Uc\ni4iM70XpeH3+BzlAGuDeBHQBz5ugnucBncBN+YFxqqcGfHNUeyJ1U/UebDCzN5jZZWb2+2b2CjNr\nn7ruioxpyt/L00GDYxGR8Z2cjg+MUf5gOp70DNUjs890vHeuA/4M+Cvg68BjZva6w+ueyKTNiJ+D\nGhyLiIyvNx33jFFePz/vGapHZp+pfO98FXgVsIL4TcYpxCB5HvAFM1POu0ynGfFzUBPyREREZgl3\n/9ioU/cDf2hmm4AriYHyfz7jHRM5iihyLCIyvnoko3eM8vr53c9QPTL7PBPvnb8nlnE7M02MEpkO\nM+LnoAbHIiLjuz8dx8qBOzEdx8qhm+p6ZPaZ9veOuw8C9Ymi3Ydbj8gEZsTPQQ2ORUTGV1/L82Vp\nybWGFGE7HzgA/HCCen4IDADnj47MpXpfNqo9kbqpeg+OycxOBvqIAfL2w61HZALT/l6eChoci4iM\nw90fBq4HVgO/O6r4CiLK9rn8mpxmdoqZHbR7lLvvBz6Xrr98VD3vSPV/U2scy2hT9R40szVmNn90\n/Wa2CPhs+vI6d9cuefK0mFk5vQePz58/nPfykaBNQEREJtBku9N7gecSa3Y+ADw/v92pmTnA6I0W\nmmwffQtwKvBqYoOQ56d/PEQOMhXvQTO7GLga+D6x6cxOYBXwSiLX8yfAL7q78t7lKczsNcBr0pdL\ngZcT76PvpXPb3f1/pGtXA48CG9x99ah6Dum9fCRocCwiMglmthL4E2J75wXETk5fBq5w912jrm06\nOE5l84EPEv/ILAN2AN8A/pe7PzGdzyAz29N9D5rZzwHvBc4BjgF6iDSKu4EvAp9y9+HpfxKZiczs\ncuJn11gaA+HxBsepfNLv5SNBg2MRERERkUQ5xyIiIiIiiQbHIiIiIiKJBsciIiIiIsmsGhybmac/\nq49A22tT2+uf6bZFREREZHJm1eBYRERERGQ8pSPdgWdYfdvCkSPaCxERERE5Ks2qwbG7nzLxVSIi\nIiIyWymtQkREREQkmZGDYzNbaGZvN7Ovmtl9ZrbPzPrN7B4z+6iZHTPGfU0n5JnZ5en8NWZWMLN3\nmNktZrY7nT8zXXdN+vpyM+swsytS+wNmttXM/sXMTjqM55lrZheb2RfN7K7U7oCZPWRmf2dmJ45z\nb+OZzGyVmX3azJ4wsyEze9TM/tLMeiZo/zQz+0y6fjC1f5OZXWpm5UN9HhEREZGZaqamVVxGbIEJ\nUAH2EvvCn5r+/KqZvdTd7zzEeg34v8CrgSqxrWYz7cB3gOcBw8AgsAh4I/DfzOwV7n7jIbT7FuDK\n9LoK7CE+uByf/rzZzF7j7t8ap44zgM8A81O/C8Bq4vt0gZk9392fkmttZu8A/prsg9J+YA7w/PTn\nDWZ2kbsfOITnEREREZmRZmTkGHgM+EPgdKDT3RcQA9ZzgW8SA9XPm5mNXUVTryX2+X470OPufcAS\n4JFR1/1OavvXgTnu3gucBdwGdAFfNLO+Q2h3O/Ah4DlAV3qeDmKgfy3QnZ6ne5w6rgHuAH7O3XuI\nAe5vAEPE9+W3Rt9gZq8hBuX9wB8Ai9x9bnqGC4EHgbXAxw7hWURERERmLHP3I92HKWVm7cQg9VnA\nWnf/bq6s/rBr3H197vzlwAfTl29z978bo+5riCgvwK+6+7WjyhcC9wELgD929/+dK1tLRJs3uPvq\nQ3geA64HXgpc7O7/OKq8/kx3A+e4+9Co8iuBdwDfcfcX584XgYeBY4EL3f2bTdo+HrgTaANWufvm\nyfZbREREZCaaqZHjMaXB4X+lL88/xNt3EKkJE9kAfL5J29uBT6UvX3eIbTfl8enlP9KX4z3PR0cP\njJOvpONpo86vJQbGdzUbGKe2HwZ+SKTfrJ1kl0VERERmrJmac4yZnUJERF9I5NbOIXKG85pOzBvH\nT9y9Monrvutjh9y/S6R8nGZmbe4+PJmGzWwF8HtEhPh4YC5P/fAy3vP8eIzzG9NxdJrH89PxRDPb\nMk69vem4cpxrRERERFrCjBwcm9kbgX8C6isp1IhJbPXI6RwiT3e8HN1mtk3yuo2TKCsSA9InJ6rM\nzC4Avkb0u24PMdEPoBPoYfznGWvyYL2O0X/Xy9KxncirnkjXJK4RERERmdFmXFqFmS0CPk0MjL9A\nTDbrcPc+d1/q7kvJJpAd6oS86tT1dHLSUmn/TAyMv0VEwjvdfV7ueX6/fvkUNl3/u/+qu9sk/lw+\nhW2LiIiIHJVmYuT4FcRA8h7gze5ea3LNZCKhT8d46Q31siqwaxJ1nQesAHYCrx5jybTpeJ56RHvV\nNNQtIiIiMiPNuMgxMZAEuLPZwDit7vDi0een2AWTKLtrkvnG9ed5YJy1hF866Z5N3g/S8XQzWz4N\n9YuIiIjMODNxcLwnHU8bYx3j3yImtE2n1Wb2ptEnzWw+8Nvpy3+dZF315znRzDqa1Pky4EWH1cvx\n3QA8TuRG/5/xLjzENZtFREREZqyZODj+FuDE0mR/Y2bzAMysx8z+J/C3xJJs02kP8Gkz+xUzK6X2\nTyfbgGQrcNUk67oJOECsjfxPZrYs1ddpZm8FvsQ0PE/aLe8dxPfyTWb2lfo22an9spmda2Z/ATw6\n1e2LiIiIHI1m3ODY3e8HPp6+fAewy8x2Efm9f0FERK+e5m58EriLmEi338z2AD8lJgceAF7v7pPJ\nN8bddwPvT1++HthkZruJLbH/AXgIuGJqu99o+9+JXfSGiS2zbzezA2a2Axgglof7n2TLuYmIiIi0\ntBk3OAZw998n0hduJ5ZvK6bX7wYuAiazVvHTMURsivEnxIYgbcQycNcBZ7v7jYdSmbv/DbF1dT2K\nXCJ22vsgsR7xWMu0PW3u/lngZOIDx93ERMIeIlq9LvXh5OlqX0RERORo0nLbR0+n3PbRV2hpMxER\nEZHWMyMjxyIiIiIi00GDYxERERGRRINjEREREZFEg2MRERERkUQT8kREREREEkWORUREREQSDY5F\nRERERBINjkVEREREEg2ORURERESS0pHugIhIKzKzR4mt2Ncf4a6IiMxUq4G97r7mmWy0ZQfHr79m\nqwNUayONc7V0tPSqlguc11+3F2P1jjnlrK72YhyHhqKuvSPZfSO1dgA6qtX4ulxplJVtEIDi/l0A\nVGrFRlm15xgACpbrdFo5ZCT9tZhXG0VFi7JSarqQi/mb1a+JkyXLCsvF9FyluN9yDdZXKvnEf1+U\n74WITI2ezs7O+aeeeur8I90REZGZ6N5772VgYOAZb7dlB8cicnjMbB1wgbtP64cmM1sNPAr8o7tf\nPJ1tHSHrTz311Pm33nrrke6HiMiMdM4553Dbbbetf6bbbdnBccEiyjvciBdD/Z/6Yi3OFcn+7W8r\nxLmecpxb0JFFebsL8XqwFMfSUFbnrqE4Vgop4lzYlbX30C0A7LgljiPHndMoW/DcVwJQy6qikqLX\n9ZWn69HieH3wsWC5CHCKBlt6nkLhqWOaWnr4gj81ciwiIiIioWUHxyJy2H4d6DrSnWgFd23cw+rL\n/uNId0NEZrn1H7noSHdhRtHgWEQO4u6PHek+iIiIHCktOzjuaY/0geED2Tm3evpBKOdSE7rKkdLQ\n1xapBks6s7KecqRTDHua3DaQmyjXFucGRuK+tjvvbpTtvfW7cX+qk8V9jbJCMeqsVbK6aKRFxFel\nXP9KKVWiXDjoUgCq6YssSyKfOhHHkZS/UTho9T7Nw5stzOxi4FXAWcAyYAT4GfBJd//nUdeuY1TO\nsZmtBb4DXAF8HfggcB7QB6xx9/Vmtj5dfgbwIeCXgQXAI8DVwJU+iVweMzsJeCvwUuBYYsWHLcA3\ngT9x9ydGXZ/v21dS2+cDbcCPgfe7+81N2ikBv01Eyp9F/Dy8H/gH4Cp3r42+R0REWp/WORaZHT5J\nDDRvBD4OXJe+/pyZ/ekh1HMe8D2gA/gM8I/AcK68DfgW8PLUxqeBecBfA5+YZBuvBS4FHgf+BbgS\nuAf4TeDHZrZ8jPvOBW5Offt74GvAC4AbzOzk/IVmVk7lf5v693ng74ifiVem5xIRkVmoZSPH81KI\ndW8hC1TV0lS3cimt01bLyjrT/LsFnXHf4jnZhLy+rvg2Dacob1shi/YW0qS5waE4VytlS8cNL+sF\n4OFajB16O7obZXMKsQSc21BWV/1FiteVCk+NHNfbq/HUAFz9TK1JcM6rca6gSXiz1Wnu/nD+hJm1\nAd8ALjOzq9194yTqeRlwqbt/aozyZUSk+DR3H0rtfJCI4L7dzL7g7jdO0MbngI/V78/192Wpvx8A\nfqfJfRcBl7j7Nbl73kZErd8FvD137R8RA/hPAO92j3UTzaxIDJLfamb/5u5fnaCvmNlYy1GcMtG9\nIiJy9FHkWGQWGD0wTueGichpCXjJJKu6Y5yBcd378wNbd98J1KPTl0yirxtHD4zT+euBu4lBbTM3\n5QfGyWeACvCc+gkzKwC/R6RqvKc+ME5tVIH3Ep81f2WivoqISOtp2chxT1tEfntzS7INVyKq29UZ\nE/EP9GcJye1pjbS57W1xzN03pzOORpxry4owi7TEh1Ndt+1+tFH2sw33xjVd8VvgZxfmNMqqFt96\nL2SR5kJKcaznHBdzkeNiY7m2FPmt5fKF67nJKSrsufXhaqmsHi/2XLRcQeTZw8xWAe8jBsGrgM5R\nl4yVqjDaLROUV4jUhtHWpeNZEzVgZkYMTC8m8pf7gNz/dQelceT9ZPQJdx8xsydTHXUnAfOBB4EP\nmDXNvR8ATp2or6mNc5qdTxHlsydTh4iIHD1adnAsIsHMjiMGtX1EvvD1wB6gSmzN+RagfZLVbZmg\nfHs+Etvkvt5JtPFR4N3AZmIS3kZisAoxYD52jPt2j3G+wsGD6wXpeCIxsXAsc8YpExGRFqXBsUjr\n+31iQHjJ6LQDM3sTMTierIl+37DQzIpNBshL03HPeDeb2WLgncBdwPPdfV+T/j5d9T582d1fOwX1\niYhIC2nZwXEp7Xg3vz37lam1R75CoVSJYyW7viNN4KsnYXslS02w9M98KTIu6M5tj3BMOvmDn94P\nwK333Nkoq1Ui2FWuxG+BO7qzQFQ57eBXy00YbCwxV0+ryO+CVz96WgIuvwxbKiykjuZHL9VGikW6\nPrfrnk84zpEWcUI6fqlJ2QVT3FYJeD4Roc5bm463T3D/ccT/htc3GRivSOVP131ElPl5ZlZ295GJ\nbjhcpy3v5VYtvi8iMqNoQp5I61ufjmvzJ83s5cTyaFPtz8yskaZhZvOJFSYAPjvBvevT8QVp5Yh6\nHXOIZeGe9gd6d68Qy7UtA/7GzEbnX2Nmy8zsWU+3LRERmXlaNnJcLsW4f2E5G/+3pUhpJUVR2yk3\nyuamCPO8zoi0drXlllErRh31iTvFQpa+WCzFuU2bHwTAGWyU1Sox4b5QTu11tGX3UUt15iLHlqLX\nzeYH+cEvCvhTimzUxLz8y2w6fi4iro9Gs8VVxCoR/2pm/wZsAk4DLgS+CLxhCtvaTOQv32Vm/w6U\ngdcRA9GrJlrGzd23mNl1wBuBO8zseiJP+ReBQeAO4Mwp6OefEpP9LgVeZWbfJnKbFxO5yOcTy73d\nMwVtiYjIDKLhkUiLc/c7gRcRq0hcRKwR3ENstnH1FDc3TOxsdz0xwH0bkeP7LuAdk6zjN4APEytq\n/C6xdNvXiHSNcXOWJyulUryG2B3vfuCXiCXcLiR+Lv4xcO1UtCUiIjNLy0aOezoiKpxfko3hSC0c\nGIk46pxcvu/Cnrh+UU98Xuhuz+5L+2c0Nvqo1rKy3bv2AmBt8a08/sTsN7EbH4jrFy05EYCOjixZ\nuRExzkWJC42PKqnMcznHjfBwJEqXSlkEuJAqGalFH/L5yI1NQ1JdlgsX2/SlWspRJm2f/OIxim3U\ntWub3L9u9HXjtLWHGNT+7gTXrW9Wp7sfIKK2f9TktkPum7uvHuO8ExuOfG68foqIyOyiyLGIiIiI\nSKLBsYiIiIhI0rJpFX0d8VvWBd25FIPOmEC/b7A+PS1LTVg4N469XZEy4bmPDXv7I/1gaChdn+U/\n8NhjsSPe9sdjj4MzT8/mCp224qQo80UAlEoduR6miXX5je7SOU/9cs/+euppFcVSvNi784lGWWUw\n+jdvebRHNbd8Xbqx4JWDjgDtxez5RURERKSFB8ci8swaK7dXRERkJmnZwfHcYiyj1lPKdsUtlGM5\n047uiKa2FbIJaV0WUdT2UnxLDuT29xoZjk08qpWIyLZ1ZhPy2soRiT3v9Jh0Vypmk/w27NoR7S5d\nDoAVc2HiajRQKOQ3+kiT52pRVvOsHUu731aq0ZfH1t+f9SEtB9u7IvZ6qObC3sVqJT1rPN/Cvix6\nPXeyGwaLiIiIzBLKORYRERERSVo2cjx/bizNVi7kNr1IObb1TT3KuUBue7H+OSGOQ8NZVLlYjMis\n1fYDcPfddzTKtm3bAMDeLY8AsGHTtqwsdo/muBVnRz0Hbd0cfSnkNhSp71jt9SXccht2kKLDB/bH\nMq+bH8six6eeHPXXd/owH8r6Xo+It8dfddecbDm5gf69iIiIiEhGkWMRERERkUSDYxERERGRpGXT\nKvrSbLNqNVu6zFKKRak+CS7LcqBQjm/FcCVODufuq1YHAdjw8O0A/PBH32uU3XbvzwDo6YzJfvQc\n0yhb9uznANA7bxUAxdyOd7V8441z8VmlklIo8suu1f+m9mzdDMD+7RsbRcXTXwjASDUuqtQOZPel\n5eCGh9LugFuzVIrKiHbIExEREclT5FhEREREJGnZyHF3W0Rpq7Vs/G+WXqcIrucev1BKS6WlOXBt\nxWyiXDXdt/yYYwE45dR9jbL7tsRybZU5sURa5+I1jbLygiUAlNqjrJibfOc1T/3LL+8WjddSH+qT\n6QDKHtHg7lpMyDv3505ulHXNmwdAf4p6D9WyNdraLZ6xUIsl4A4MZNFiI9e2iIiIiChyLCIiIiJS\n17KR4/YUpM3t5YEV0vbMKXJcs2KuMA7ltFFHZ3u5UdSWNgaZm/KK93lbo+yErRHRffDJyAG29kVZ\nH0pzAOhJ200XCllvBlM7hVzucXc5IsXHdMX1PVkX2LIxco337H4w+rk3yznu33wfAF2r+qK/5d7s\nsdLycZW0MYi7Pg+JiIiIjEUjJRE5apjZajNzM7tmktdfnK6/eAr7sDbVeflU1SkiIjOHBsciIiIi\nIknLplXUFcxyr1NqQT3NgWzCW7W+Y136vNBWtFxZSk2wyHOoFLIJb109ywBYWo1Jd7X2+Y2yuW3d\nACzujG/zYG2wUdbWHqkZPT1Zika3xZZ6+7fGrnsj23Y3yvo3b43jzsejnb2bGmX7DtwMwPbdUf/K\nZ72kUVbsShMNvf5Xrc9D0lK+DPwQ2HykO9LMXRv3sPqy/3jG2lv/kYuesbZERFpVyw+ORaR1ufse\nYM+R7oeIiLSOlh0cV+sT0HKRY6weMU5LuXkWOa6lTT+8FpHWYiEfcY6JdEMDcf3OvfsbZZ09MQmu\n1yKavHUwm2A34LFs2pbtEQneuyubRLcvLbs2b/68rK6RWCLO90WUeHggizTvIyYDdq08CYDBx7Ml\n2fZviXr7jjkLgLb27kZZxaKOotX7ld985KkbkYgcLczsFOAjwAuBduB24E/c/frcNRcDnwUucfdr\ncufXp5enA5cDrwWWAx9y98vTNUuADwO/BPQA9wMfAzZM20OJiMhRr2UHxyIyo60BfgD8DPgUsAx4\nA/ANM3uzu39hEnW0Ad8G5gPXA3uBRwHMbCFwM3Ac8P30Zxlwdbp20szs1jGKTjmUekRE5OjQsoNj\nT1HRatVz50I5LfBWq2ZLq1U8LfOWNucwy8pqaQm4ff0Rrd25f6hRtnM4NtfY7xFxruZylfcO9Md9\nu54AoLjtwUbZSDGi0ItLq7N2anH9tkg9ptC2rFHWtTg2IKkvPvfk7ixyvPXhyEMubHsUgFWe/ZZ5\npBDLull9UTt/6vdD5Cj0QuAv3f1/1k+Y2SeIAfPVZvYNd9875t1hGXAPcIG7948q+zAxMP64u7+n\nSRsiIjJLaXaWiByN9gB/kj/h7j8BrgXmAb88yXreO3pgbGZl4FeAfUTKRbM2Js3dz2n2B7jvUOoR\nEZGjgwbHInI0us3d9zU5vy4dz5pEHYPAnU3OnwJ0AXekCX1jtSEiIrNQy6ZV4Cl1IpdGQJqkV6lG\nSsNILuViJGVDlNL1hVo2WW8k7ai3a0+kUPQfyKVjDEd6Qyl9zljQln3eKHtcX2xPdVn2W+DhQkwA\nrOzc0jjXX4u/jtKKswHo7Dshayf1vZaWZFtxdrZk03BKmdi2/n4ANj54S6NsyemvBGCw8X3I7RmY\nn6wocnR5cozz9f9hescoz9vq7s2yh+r3TtSGiIjMQooci8jRaMkY55em42SWbxsrrb5+70RtiIjI\nLNSykeMUHG4cAWqWIsa1iJjWl3sDsLR0W31Jt1Iti6qWS7FM22AhfstbsKzSBWkCX//gNgAqI9lk\nvflpA46OYty3rZZbAq59IQDzjvm5xrnenogU97d1xYlCOet7WvqNdPRStgTccWe9CoCuntVR1pYF\n1Wq1iBQX6gE0z57LUORYjlpnm9ncJqkVa9Px9qdR933AAeBMM+ttklqx9qm3HJ7TlvdyqzbmEBGZ\nURQ5FpGjUS/wv/InzOxcYiLdHmJnvMPi7iPEpLu5jJqQl2tDRERmqZaNHIvIjHYj8Jtm9lzgJrJ1\njgvA2yaxjNtE/hB4CfDuNCCur3P8BuDrwH97mvWLiMgM1bKD41pKGajk1zIeiXSIejpFsVBslNVf\nFSxeeXtWtnFX/Nb1gW2xnvDugR1ZQyNpveFKLE6c7URHI4WhPu/vzLNf0ihaeeKZADxZ7Wmce6I/\n0ig6qvW1lrO6akTfa2lXv/0jbY2yaiFSJJeeEqkWhdzvAyr1tIr6iYN2DFRahRy1HgUuJXbIu5TY\nIe82Yoe8bz7dyt19u5mdT6x3/CrgXGKHvN8B1qPBsYjIrNWyg2MRmXncfT0clAz/6gmuvwa4psn5\n1ZNoawvw1jGK9clRRGSWat3BcYoAV73SODVSrUdfUzQ5F1UulCNqO5h2w7vjwfWNsjs1E9iSAAAg\nAElEQVTvj53ttvbvihO5yXqlUkRwCyPR3mCubMv+aHtBT0y+O+ncLHLcXY4o78YntzfOlUsRhW6r\nz50r5P56Uui3VI4XpWJWtnsonms4LfNWygWv6/HveiS9PvEwGkBEREREcjQhT0REREQkadnIsaXE\nW8tFX4sppFr0+sYduSjv1ojg3r1hIwDrt2xtlNUGYjWpedVYpq3Uni2x1l87AEA1hWhrluUCD9fi\n9bYDEdn90f33NsrOOf7ZAJSL2eeTOemzSjEtFeftWVkh5Rq3V6PP5VouIl4ZSn2J6wuWPXMp3VdN\n+cUHLd+mXxyLiIiIHESRYxERERGRRINjEREREZGkddMq0o5wxdwEtPr8s5pFWsTO3bsbZd//0U0A\nPPbklri/mM1WG9obO9v1dMwFoG/RwkZZVzmuK5fTt3Iwu8/TUm4jaVrcPes3Nsra0kS+Y1esapwr\nVuL6oTSZcKg63CjrLNbSc8XnmZHczn8dKTWjmtrx/ES7+hdp2br8Mm9W04w8ERERkTxFjkVERERE\nkpaNHNfDxIX8rLMURa1P0nv0sccaRQ88eE+UpclwCxb0NcqOO/UkAI5ZfiIAe/YdaJQ9uW1TVJ2W\nhZtTzj5vFNMkuP3DUTbk2WS9bdtiWbi+zu7GuXl9KwEY3BeT7Xw4Cw93dkTkt9SWlozrz8rq0fF2\nr0/Iyx65/nogXZ6fg6c9QEREREQOpsixiIiIiEjSupHjhiyvth4o9UpEZo9fdWyj7NWvfBUAHe3t\nAPQtmN8oa+uKLZ5HKhG13bt/oFHW0d4JwBOPP5gqzyK67Z2R2zy/J+q0UlejbH9/LA/3sztvb5z7\nuTOivFTuJb1olHW2xetyWjKuPJhFr6lEm8X0Wactl1jcmSLNPhjL1414ti32QSFmEREREVHkWERE\nRESkToNjEREREZGkZdMqPKU31GpZmoOnxApPE/Pm92aT7pYsWgxANaUoDI5kO9DV58WZxX1tbR2N\nslUrjwegI2VA3HvfXY2yrq6YgFetRRrGfXff0ijb8ERMBty6a1927rH7AHjVa94KQGe5p1FWqKZO\neEzuKxeydJFSet3YRc+qjbK55ZQ6UYlr9lSy78eI5VIsRERERESRYxGZGcxsndU/oU7+HjezddPU\nJRERaUEtHDn2g47511UiejpSyaLDPjycrimkY349tDhYIyKbqzNFo4vlmJi3vz+bKPfgQ/cDsHHT\negD27Mo2AenqjQ1FKkN7Gud+ess3AFi5YgUA5//CLzfKKkPRiWJqr72U/dWVCzHZrliMSPC8rqxs\nQXeEtDuLcW7/7qFGWbbFiIiIiIhACw+ORUSAU4EDE14lIiKStOzguL4z8kGR45Ru24gc5/ORa/UM\nE0v/zSLH5hFhHqlG1LVSzeqsVONbuGNXbDG9as3JjbKTTo3X//n1rwCwa/uTjbJVy9dEX2qPNs7t\n37YDgNt/fGNcs+rURtmaVacDMFiNfrUVs7+67s4U7U7P01nOconb03JtPWlZudK+kdwzI9LS3P2+\nI9n+XRv3sPqy/zjs+9d/5KIp7I2IiEyGco5F5Igzs/9mZjeY2WYzGzKzTWb2XTN7e5NrS2b2h2b2\nYLr2cTP7czNra3LtU3KOzezydH6tmb3FzG43swEz22pmnzGzpdP4qCIicpTT4FhEjigz+23gq8Cz\ngP8H/BXwdaATuKTJLZ8Hfg/4HvBJYAD4A+BTh9j0e4CrgZ8CHwfuT+3dbGaLDvlBRESkJbRsWkU1\njfuruYl1nibS1WqRdlCr5j4bpOvq13gu56CeojFcKaRjtlRaJV13zKoTALBSltLQ0Rnf3jcsWA7A\nl770L42yBx+6Nfo3kk3I65kfS7c98dhDANzw9esaZb/667HUnHdEUGsgN5uumrpTTcvP7cvtfFe0\n6HO5VP8eZGqHNO9fZNq8jZgfeoa7b80XmNnCJtcfDzzb3Xema/6IGOD+upm93923TLLdVwDPdffG\nNpVm9jHg3cBHgN+YTCVmdusYRadMsh8iInIUUeRYRI4GFWBk9El3397k2vfVB8bpmn7gWuLn2bmH\n0Obn8gPj5HJgD/BmM2s/hLpERKRFtGzkuJLCorWDll2L19Zk0h1p+VSv1Y/52WqFdH18u/JLrdZf\nm0VdQyNZWdXj3/qFC5cB8N9f9+ZG2Q3rYgOSH/3gW41z3Z0xqX7NojkA9LZl4eFbb47r1pz9SgAG\nLdsgZCRt7FEulQ/6GmD/YPSh3J6i3p5FvauWe36RI+daIpXiHjO7DvgucJO7bxvj+p80Ofd4OvY1\nKRvLd0efcPc9ZnYHcAGx0sUdE1Xi7uc0O58iymcfQn9EROQooMixiBxR7v5R4C3ABuCdwJeBJ83s\nO2b2lEiwu+9uUk190fJD2fbxyTHO19Myeg+hLhERaREtGzn2FCGt1SpNCuMzQYGnbhBSTMFUy31s\nqKWyciktmUa5UTacorTDw/6UOutx2ZHBWAKub+68RtmLXv7GaK9rcePc0GPxG97Srg0AbFj/YKPs\nsU39ALQvOxOAzmVzsnZSvyzlGteqWXR4qBrP3z8c14zkn/nQNhsTmTbu/k/AP5nZPOD5wC8DbwW+\naWanjBNFfjqWjHG+vlrFnjHKRUSkhSlyLCJHDXff7e5fd/ffAq4B5gMvnKbmLhh9wsx6gTOBQeDe\naWpXRESOYi0bORaRmcHMXgSs8/yOPaH+a5Xp2uHu18zsE6Mm5V1OpFN81t2Hmt82eact7+VWbeQh\nIjKjtOzgeHh4AIBabmJd/XXN6+kRT52QVqnEBLZcZgKlcgcAxULc157bga4+qa8+ka968GJp6b6U\n7pBbO61gnQD8/M+/uHFuU0fUf/t/rQdg3uI1jbLnnPeKaHvZMQDsyY0javVl59JydOVC9guBSurP\n3sFK6kNuoqHlHlLkyPkysN/MfgisJzKSfgH4eeBW4Ftj3/q0fAO4ycy+CGwGXpD+rAcum6Y2RUTk\nKNeyg2MRmTEuA15OrOzwSiKlYQPwPuCT7v6UJd6myMeIgfm7gTcA+4lUjj8cvd7yYVp97733cs45\nTRezEBGRCdx7770Aq5/pdu2pv8kUEWldZnY58EHgRe6+bhrbGSJWz/jpdLUh8jTVN6q574j2QmRs\nZwBVd39G151X5FhEZHrcBWOvgyxypNV3d9R7VI5W4+xAOq20WoWIiIiISKLBsYiIiIhIosGxiMwq\n7n65u9t05huLiMjMpcGxiIiIiEiiwbGIiIiISKKl3EREREREEkWORUREREQSDY5FRERERBINjkVE\nREREEg2ORUREREQSDY5FRERERBINjkVEREREEg2ORUREREQSDY5FRERERBINjkVEJsHMVpjZZ8xs\nk5kNmdl6M/u4mfUdYj3z033rUz2bUr0rpqvvMjtMxXvUzNaZmY/zp2M6n0Fal5m9zsyuNLPvmdne\n9H7658Osa0p+Ho+lNBWViIi0MjM7HrgZWAx8FbgPeA7wLuBCMzvf3XdMop4FqZ6TgG8D1wGnAJcA\nF5nZee7+yPQ8hbSyqXqP5lwxxvnK0+qozGYfAM4A9gNPED/7Dtk0vNefQoNjEZGJXUX8IH6nu19Z\nP2lmHwXeA3wIuHQS9XyYGBh/1N3fm6vnncBfp3YunMJ+y+wxVe9RANz98qnuoMx67yEGxQ8BFwDf\nOcx6pvS93oy5+9O5X0SkpaUoxUPAeuB4d6/lyuYCmwEDFrt7/zj1zAG2Av9/e3ceZelV1nv8+9Q5\nNc/Vc3rO2B0SEmmGEJAECCEYuUQFc7novcGlSy6jAmrEgQRUWCAIooKKgXWDGrwMIgImXiACwRhI\nQkJCZ+qku9NzV3VVdQ2nTg1n3z+eXe8+OTlV3V2pofv077NWr1P17v3u9z2VN1VPPfXsvUvAmhDC\nUFlbHfA4sDFeQ9ljOW7z9YzG/rcDl4UQbMFuWE57ZnY5Hhz/fQjhl07gvHl71mejmmMRkdm9NL7e\nVv6NGCAGuHcALcAlxxjnEqAZuKM8MI7jlIBbK64ncrzm6xnNmNm1Zna9mb3TzF5lZo3zd7siczbv\nz3o1Co5FRGZ3Xnx9ZIb2R+PruYs0jkilhXi2bgE+AHwE+Dqw28xeO7fbE5k3i/J9VMGxiMjsOuPr\n4Azt08e7FmkckUrz+Wx9BXg1sA7/S8cWPEjuAj5vZqqJl6W0KN9HNSFPREREAAgh/FnFoYeB95jZ\nPuATeKD8b4t+YyKLSJljEZHZTWciOmdonz4+sEjjiFRajGfr0/gybhfHiU8iS2FRvo8qOBYRmd3D\n8XWmGrZz4utMNXDzPY5IpQV/tkIIY8D0RNLWuY4j8gwtyvdRBcciIrObXovzyrjkWiZm0F4EjAJ3\nHmOcO4EC8KLKzFsc98qK64kcr/l6RmdkZucB3XiA3DvXcUSeoQV/1kHBsYjIrEIIO4DbgE3AWyqa\nb8SzaDeXr6lpZlvM7Cm7P4UQhoGbY/8bKsZ5axz/Vq1xLCdqvp5RM9tsZj2V45vZCuAz8dNbQgja\nJU8WlJnVx2f0rPLjc3nW53R9bQIiIjK7KtuVbgdegK+5+Qhwafl2pWYWACo3UqiyffRdwFbgNfgG\nIZfGb/4iJ2Q+nlEzuw74FPA9fFOaI8AG4GfwWs4fAq8IIaguXk6YmV0DXBM/XQ28En/OvhuP9YYQ\n3h37bgKeAHaFEDZVjHNCz/qc7lXBsYjIsZnZeuB9+PbOy/CdmL4M3BhC6K/oWzU4jm09wHvxHxJr\ngD7gG8AfhhD2LOR7kNr2TJ9RM7sQeBewDTgD6MDLKB4E/gn46xDC+MK/E6lFZnYD/r1vJlkgPFtw\nHNuP+1mf070qOBYRERERcao5FhERERGJFByLiIiIiEQKjk+AmYX4b9NS34uIiIiIzD8FxyIiIiIi\nkYJjEREREZFIwbGIiIiISKTgWEREREQkUnBcxszqzOxtZnafmRXM7LCZfdXMXngc564wsw+Y2Y/N\nbNjMRszsATP742rbcVace4GZ3WRmT5jZmJkNmNkdZvYmM6uv0n/T9OTA+PklZvYFM9tvZlNm9rG5\nfxVERERETl/5pb6Bk4WZ5YEv4Nu4AkziX5+fBa4ys2tnOffF+BaG00HwOFACnhX//bKZvSKE8HCV\nc98KfJz0i8ow0AZcGv9da2ZXhxBGZ7j2tcDn4r0OAlPH+55FRERE5KmUOU5+Bw+MS8BvAZ0hhG7g\nTOD/ATdVO8nMNgJfxQPjTwLnAM1AK3AhcBuwHviSmeUqzr0G+AQwAvw2sCKE0A604FsiPgpcDvzZ\nLPf9aTww3xxC6IrnKnMsIiIiMgfaPhows1Z8X+52fF/uGyraG4F7gPPjoc0hhJ2x7XPAG4APhhB+\nt8rYDcAPgGcDrwshfCEezwE7gI3AVSGEW6ucexZwP9AAbAgh7I/HN+F7jgPcAbwkhFCa27sXERER\nkWnKHLsr8cC4SJUsbQihCPxp5XEzawFeh2ebP1pt4BDCOF6uAfCKsqbL8cD4gWqBcTx3B3AnXjJx\n+Qz3/hEFxiIiIiLzQzXH7jnx9UchhMEZ+vxHlWPb8KxuAH5sZjON3xxf15cduzS+nmNmB2a5t84q\n55b7z1nOFREREZEToODYrYiv+2bps7fKsTXx1YBVx3GdlirnNs7h3HKHj+NcERERETkOCo6fmemy\nlME4GW4u534lhHDNXG8ghKDVKURERETmiWqO3XT29YxZ+lRrOxhfO8yss0r7bKbP3XCC54mIiIjI\nAlFw7O6JrxebWccMfS6rcuyH+HrIhi+9diKma4WfbWZrT/BcEREREVkACo7dbcBRvP73HZWNcTm2\nd1UeDyEMAV+Mn77PzNpnuoCZ5c2srezQN4EngRzw4dluzsy6j/UGREREROSZU3AMhBBGgA/FT99r\nZu80s2bI1hT+MjOvFnE9cAQ4F/i+mV01veWzuXPM7J3AQ8Bzy645AbwVX+ni9Wb2z2Z28XS7mdWb\n2XPN7EOkNY1FREREZAFpE5Bohu2jh4Gu+PG1pCxxtglIPPd5wD+T6pIn8Ex0O77U27TLQwhPWRLO\nzN4IfKqsXyH+68SzygCEEKzsnE3EgLn8uIiIiIg8M8ocRyGESeAXgLfju9JNAlPA14DLQghfmuXc\nHwBb8C2ov08KqkfxuuQ/j2M8ba3kEMJngPPwLZ8fjNfsAPqA24H3xnYRERERWWDKHIuIiIiIRMoc\ni4iIiIhECo5FRERERCIFxyIiIiIikYJjEREREZFIwbGIiIiISKTgWEREREQkUnAsIiIiIhIpOBYR\nERERiRQci4iIiIhECo5FRERERKL8Ut+AiEgtMrMngA5g5xLfiojIqWoTcDSEsHkxL1qzwfEZz1oR\nAEZ7x7JjljMAeroaAbjyRduytnPXrQegjhIApVIxa+sfPALA6KiPtXLFmqxtvDgFwM6dOwHo6+9L\nN9HQBMCId+HJffuzppbGXLynlLzvHxoFYGio4G0hl/o3+VhW5/2n4n0CTJYm/Z4nxgFoj30BNq/b\nCMBYYcT7MJK1Wd7H+s539xkiMt86mpube7Zu3dqz1DciInIq2r59O4VCYdGvW7PBMXUekZqF7JCZ\nv92xMQ8mh46mL3hpcsL75Ly/5euztuaWDgAmJjyGDKU0Zle791u7yvsUiwNZ2/CUX6e1o93PO9SQ\ntT15cMhvM12GXGzOt3jwPj6SrtPdsxqACy/aAsAjjzySte15cne8Lw92V644I2u78pVXA3DfvXcD\ncNfd387aNmxehcjJxsx2AoQQNi3tnTxjO7du3dpz9913L/V9iIickrZt28Y999yzc7Gvq5pjERER\nEZGodjPHIiJL7IG9g2y6/mtLfRsyRzs/ePVS34KILIGaDY4b6r1GoZibelpbc3OsBR4Zzo6NjnqZ\nQ3tHGwChrH8h9ise7QegLp/qfZsnffzlU15O0bgslU4MFr2sYnDCxy60p9LesSM+xngailg6TFNz\nCwA2le69OZZaXPmyKwC4eOv5Wdsd3/kOAPX1XqO8Zs3qrK2t3e+nvtXH2nrhxqxtcmoCEREREUlU\nViEii87cW83sQTMbM7O9ZvYXZtY5yzmvN7Nvm9lAPGe7mf2+mTXO0H+LmX3WzJ40s3EzO2hm/2Bm\n51Xp+1kzC2Z2ppm9zczuN7OCmd0+j29bREROATWbOS4UPCsayhLH+QbPrNbX++8EI6Np8txIwbPD\nre3NfqCUTuxu95+9uSFP7eaPplUnVnX6l3D5Cj+vsSX9bB8s+IoXe3t9tYvu8TTmUJ/npp84lFbT\nCHWe5S2VPJ1cVzeZtfX3+zWnxnxFite9+ueytvXdXd5W8rGWr0qT479z550A/PjBH3vb8ras7eD+\nQUSWyMeAtwP7gb8BJoDXAC8AGoDx8s5mdhPwRmAP8EVgALgEeD/wcjN7RQhhsqz/VcCXgHrgq8Bj\nwDrg54GrzeylIYR7qtzXx4GfBr4GfB14+p+eKpjZTDPuthzrXBEROfnUbHAsIicnM7sUD4x3AM8P\nIRyJx38P+DawBthV1v86PDD+MvCGEEKhrO0G4L3AW/DAFjPrBv4RGAVeEkL4SVn/C4A7gU8Dz6ly\ne88BfiqE8MT8vFsRETnV1GxwvH7tSgAKrSkBNVI4CsBU8Izu8EiqAR4e8Z+3XeOecW6uT2usWWwj\nrmE8Ntafzjvq2eiW1pj1bUrntba0AnBus/dp6ExVLHVnel1xT3Nay/jRw37to+N+vcaW1L8Ya6L/\n/bbbAHjxtp/K2rZu3gTAfT/6IQD1K9I6x2MjBwAYjGs1j4wNZW2ThbKCZ5HF88b4+sfTgTFACGHM\nzH4XD5DLvQOYBH6lPDCO3g+8FXgDMTgG/ifQBby1PDCO13jAzP4W+A0zO7+yHfjQiQbGIYRt1Y7H\njHK1AFxERE5iNRsci8hJazpg/I8qbd+jrJTBzFqAi4BePKCtNl4R2Fr2+Qvj60Uxs1zp3Pi6FagM\nju+a7cZFRKT2KTgWkcU2XZh/sLIhhDBpZr1lh7oBA1bg5RPHY1l8/bVj9GurcuzAcV5DRERqVM0G\nxz93zbMBONybyghGY0nBxKgnpvY+msojRsb8r7Wlor/WjYxmbX0PenJpvM8nxdW3pFKIg3E35olh\nnwzX3tactfXU+8S4Nct8J7rz8+nLfUaH/zX5Was7smN3HfYSkHv3++TAobLto0OckL/7gJdifvd7\nt2Vt5670ra+bprykY/dDafe8C87xa+86eCYADz62J2tbtSxdW2QRTc8EXQU8Xt5gvo3lcnziXXnf\ne0MIx1uiMH3ORSGE+0/w3sKxu4iISC2r2eBYRE5a9+ClFZdRERwDLway3wpDCMNm9iDwLDPrKa9R\nnsWdwC/gq06caHA8ry5Y28nd2khCROSUUrPB8X33PgzAwNE0Ia8U/ONlXd1+oOzdDxZ8WbfRmExu\nGErLvHXEDT5KTT5BLqSf3bT3rABgw8Y1AOQa0oS8lWd7GeTyNZ69He9LWdvGPY/566F0nU1ntwOw\nut7/svutQ4ezttEmv9mXv9DLKdeuWpm1FeKEwS3n+vV+/JO0RNue3Z7tHjzsfZob0jJvZ25MG4KI\nLKLPAr8K/J6ZfaVstYom4ANV+n8U+DvgJjO7LoQwUN4YV6fYXLY022eA3wPea2Y/CCHcVdG/Dl/F\n4vZ5fE8iIlIjajY4FpGTUwjhDjP7BPA24AEz+wJpneN+fO3j8v43mdk24M3ADjO7FdgN9ACbgZfg\nAfGbYv8+M3stvvTbnWb2TeBBvGRiPT5hbxnQhIiISAUFxyKyFN4BPIKvT/zrQB8ezL4HuK+ycwjh\nLWb2DTwAvgJfqu0IHiR/GPhcRf9vmtmzgXcDr8RLLMaBfcC38I1EREREnqZmg+N7fuAT4ds6u7Nj\ndTmf1HZ4n69XPDE2krX99LazAWjHJ9a1FFNpwvoNPvn9wEEvdyyOpy/bWZt8VahlXV4SMV6XSi7a\nlm8CINflpReNDWkCXF29j5lvTZPjc3GVqitW+AS7/KO7UttyL9t45RUvA6CzY3nW9sRDOwEIOS/7\nyFtKiJUKfj+drb5AwP6+VKpx7w+2I7IUQggB+Iv4r9KmGc75V+BfT+AaO/E1kI+n73XAdcc7toiI\n1K66Y3cRERERETk91GzmeKLky7Z1LE9LmU5Nemp2qN+zw+duXZe1rV3tmdz6PT5p7owVa7O2s7d6\nVnk07hdQHEiT/GzCl4UbH/UJbw1xgh5Avt6vnW/w7HWuvj1ra2mO2eSVaVLc+LhnsjdO+M51167a\nkrU1tHrWub7Rs8INnWlCXq7Vl4XtK3i2u7k1vedcya95YJcvD3dgd5rs39iU7bUgIiIiIihzLCIi\nIiKSqdnM8ZYLvSb3vItSFvXIEc8YH9zdAEBLW/rdYO/jO/08fBOP7o1nZ20dm58FQONBz74O9j6c\ntQ1NHAWgrcUzua3LO7O2lmWeMa5r9+xtGE81zhNxzbjJsZSFbrIWAEZiBnhiMm1gUhc8m1xf9GXh\n2tqXZW2rN/sGH/95x61+v3Vpi93mZn+vF1zomfAtz92QteXr0/2IiIiIiDLHIiIiIiIZBcciIiIi\nIlHNllUc2usT5IaGU2lCZ7vvDmdjXjpxoDdNTtt85lkAnLPOXxtb0rJrRbzcoW2ZT+A70n4oaxu1\nAEB/XBauoz6VNNR3eElHri2WVRTS/RXNyySKw6PZsXxDa/zIf2dpX5WWa2tb4WUb4/kuP9DQnLXt\nO+wT8v7hC18G4KyetAveyy57rn8d2v1edvelpePy9frdSERERKScoiMRERERkahmM8fnrvEM8ERp\nMh0sera2rcU3A1m9NmVYL9zok9pa6z1L3NHZlbU1xSztho3nANCeZXhhaGAfAHUlTwsXC+l6xZGi\nt+HZ4caG+qwt1xo3J+lI/Sfz3t5Y59drbkvZ4dFxX3YtNPp7oGwy3SM7fgTAjl07ADi065Gs7eIX\nnBPfum8G8qMHHs3a6vItiIiIiEiizLGIiIiISFSzmeMXP+9CAPINjdmx8QlfNi1Meva1OZcyucN9\nXkc8OuDLteVjJhkg3+rLoTV1eQ3wqqa0XFvDPs/uFvr3+9ilhqxtoM+Xa2s1v4cc6V6ma4abe1Jd\ncX2sQ7aSZ4l7+/uztpGi33tznf8+c+BHt2dthw775iQXXHQGAEcPHMzaDvR5pnj92Z6pvrzhvKxt\nsq5m//OLiIiIzIkyxyIiIiIikYJjEREREZGoZv+u/vi+7QA0trRnx8ZGfYLcRNHLFnLjpaytddB3\numuPE+o2rlyRBoursxWm/IOBoYmsaWDQxzq0uw+AnrLl2s7zTenIj/l18/n0u0gpDtrQnCbF5fFr\nj8bl5wqTU6l/nU8C/K//egyAW+/896yteZWP++wLfKm58Y1pIl9Pt5eOdLU0AbC8NV1vsi4gIiIi\nIokyxyJy2jOz281Mvy2KiEjtZo7Xn7cagIMH00YfvXt9Awwb96xtUym9/foJzwbn8n7syJHDabAn\nd3pbl485UkgbfTS0+5JvrSvXADBZ9uvGlPnkvNG4hFyhkCbYdbZ7Btca0wkjBZ8M2Bsn8u3YuT9r\n+8H9njH+0aO7ANjdtzNrWzvmG5aUpjw73J5L76vF/FjzhN/L+paV6Xqjg4iIiIhIUrPBsYjIUntg\n7yCbrv/aUt/GaWvnB69e6lsQkVOQyipE5JRiZs83s8+b2V4zK5rZfjO7zcx+sazPdWb2RTN73MwK\nZnbUzO4ws1+qGGtTLKe4LH4eyv7dvrjvTERETgY1mznu6fA1hSeH07rDfSUvnWiM6w235puytvFJ\nb6tr9gls/f19WVtvXGN43fk+Ka6pPZUmHOn19ZGXr/aSiw3r1mZthwa9TGJiwGfp9bSle+lq8o/7\njqTrDB4dAGC46Pdy74+3Z23fueuHAKzZ6te++rLnpPfa7e+npc53wevblcolBvaO+dh7ngBg7Zq0\nK+CWM85C5FRiZr8GfBKYAv4FeBRYCTwXeDPwT7HrJ4EHge8A+4FlwM8AN5vZeevrKIMAABPaSURB\nVCGEP4j9BoAbgeuAjfHjaTsX8K2IiMhJqmaDYxGpLWZ2PvBXwFHgp0MID1a0ryv79IIQwo6K9gbg\nG8D1ZvapEMLeEMIAcIOZXQ5sDCHcMIf7unuGpi0nOpaIiCy9mg2O9273XeMslypHNq/3rOnYoE9K\nHzua+vcN+7JpDZOe7e1em3auo96Xgxsa9fOau9OXrTTlWd7hIV+urVAsW37NPJPb0uYZ5+LkaNb2\n0MM+wW5sKN3EWJyQt/fIPh9zIk3Iu+B5q/z+2jzb3ZlPWej8oE8QLBY8w93d2Ja17d/nEwsP9/rE\nxN170pgveMGzAHg+IqeE/41/z3p/ZWAMEELYU/bxjirt42b2l8DLgJcD/2cB71VERE5RNRsci0jN\nuSS+fuNYHc1sA/A7eBC8AWiu6LL2aSfNUQhh2wz3cDfwnGptIiJy8qrZ4LhrmWda+4Z6s2PjJd9k\no3GlL33WPz6U2uKPziMT3mcopHrk0oTXIefikmz1ZV+1rjZfkm1qzGt7Q11a5q2lxbPWhfFDcZzx\nrK0+bv5RGEg1x9sfuR+AkQnP8p5zfvorcTHvWejmDs9or47LygEc2OEJs9Epz0JPFMeytropfz/t\njX69rp7WrK1sRTqRU0FXfN07WyczOxO4C+gGvgvcBgzidcqbgP8FceKBiIhIhZoNjkWk5gzE17XA\nQ7P0eyc+Ae+NIYTPljeY2evx4FhERKQqLeUmIqeKO+Prq47R7+z4+sUqbZfNcM4UgFmcKCAiIqet\nms0c7+kbAWBvb9ohb2DIlzhbHksumjpasramNV5GMXZwuqyiOw1W8r/AtuOT77bff1fWtKrd+w3F\nyXAtg2kXvMYm7993xHe161lxRta26ZyLAZgiTdLLP+FjNMfyi57W9Jff0ZL/zB4u+LJwo41pIt/K\nM3yi4Xib9x8bSuUia5d5+UVTU1f8emRzlti7M5V0iJwCPgm8CfgDM7s1hPCT8kYzWxcn5e2Mhy4H\nvlrW/krgV2cYe/p/hg3AE/N1wxes7eRubUQhInJKqdngWERqSwjhJ2b2ZuBTwL1m9hV8neNlwPPw\nJd5eii/39kbg/5rZF4B9wAXAVfg6yNdWGf6bwOuAL5nZ14ECsCuEcPPCvisRETnZ1GxwPHTYM6zL\nOtLEta6OZQBMxA0/crmQta1a45nVMTxDO5lLWeXmZs/I9h32rGuxLPvauelMAAbGvEKlJ07aAxjq\n83lDhw55ImpkeCJra2rcDMDBssz2qrU+gb6rbT0AeUsz5nLm/6mKBc9sN+Xq0z20+tJtE/V+rHll\n2qSkq3OFfx2Wr/HziylTPTqaNgsRORWEEP7WzB4A3o1nhq8BeoH7gU/HPveb2UuBPwKuxr/P3Qf8\nPF63XC04/jS+Cch/B347nvMfgIJjEZHTTM0GxyJSm0II/wn8wjH6fB9fz7iap63TEkKYAt4T/4mI\nyGmsZoPj1Tlfrq04OpmObfTM7PiU1/T2HjmUteVL/vOyvdOzy52NqeZ4Knh97759TwKwLCWHGR+K\nmd9Gz1BPlS2n+sDdXhK5b5dPrG9aleqEGxo2AHDwYFpq7rGHfd+Cy1/sS6M2Naf5kkODPlG/qeSb\nf/Q0d2ZtnT3+cX9/7NPUnrWNTHi2emT/bgDOKMsq12vukYiIiMhTaLUKEREREZFIwbGIiIiISFSz\nZRXtDT45baR/JDu277F9AJR8SVPa2tIueDbpk/NycUe5lT2pPOJw/wEA6vCd5/J16byBAV+6rW6N\nl1UUx9IueDnzEojVy32C3URjKnfoO+LlGB2ty7NjhSG/9kMPPA7AuWevz9o64pJxHV1eFlEXUklE\nf59PrAvBS0Ma69O9jxT9/Q8O+PWKZcu8jY765LyNiIiIiAgocywiIiIikqnZzPF4s8+a62lPS7k9\nuWc/AO2trQC0taVJbVNHPYtaF+exHz70ZNa2fJkvlTbW7ZP8bCzNyJsInqFuLPmxiWKadDce/Fh7\nl593pJCWURse9n7dK1dkx87d5Peas9ivlH53qWv0bPVUvY/Z2JSWcstPxeXnip55PrgrLTU3NTUV\nP/LM+JN792ZtoyNjiIiIiEiizLGIiIiISKTgWEREREQkqtmyioFh3yFvYKgvO9ba5uURK1f5JLiO\nlrQLXlPHKu/f6xPqDh44nLWdtcF3l5tavg6Axx58LGsbL/k6ws31sSyjcThrO9zvpRNj8avcuSKV\ncdSbn3fk8IHs2GjBJ8/lm7wEYnAy/e7S3e7nDo74xLrR0TTRsLPNyzZGjvq19+3Zl7Xl6nyMxkbf\n5a80mdZ9DqUpRERERCRR5lhEREREJKrZzPFY0bOi43GHOICunE9ca2nyyW35uvS7wb6DvlveyIBn\nVqemilnb4UM+kc/iMm/t7WlJtoMDnqEeHPGM81Q+ZY7XrNsMQGPwts6e1jRm3J2vrrkxOzYVd6zr\n6vZd+s6/+HlZW67RxxiMy8qVSO+rUOfn9fV6tnt0OC3X1tDgy8k1Nfl18rm0BFxTYwMiIiIikihz\nLCIiIiIS1WzmuLPbN81obksbYkyNe+Z3726vyR0bTkurkfMsaluj1++uXN6VNfX2xU1ApnwZtbqm\ntAlIU0dcUi0uCzc2mep416z2WuX8lC+ZNj6RMroBz+DmyzbsyDdM10D77ywDfb1ZW88yz/y2NXvd\n9PBgWjJuf/9Bf39xKbflPSuztuEhz2QPDfprKZUcMz6ummMRERGRcsoci4iIiIhECo5FZF6Y2SYz\nC2b22aW+FxERkbmq2bKKQCyByKX4vzEu3dZs/tqaS6UTh3q95KGv30sg+or9aawpL0mwOGFudHQw\na+tY5uUb3a1eatF/OC0d1zvgk+aWtXrJxthEqmkYGvX7y+dSacPkhG/PV4hLsvXvfiJrmzrq/6mG\nxnwCYHEsjVVv3jblc/aw+vSftb3V3+PQsI8Z0uZ+5Opq9j+/iIiIyJwoOhIRWSAP7B1k0/VfW+rb\nOGns/ODVS30LIiLHVLPBsU16inRkJC2tVij48mzN9b4UW31Ik+Fy8eOh4QEARodTdnjVKl9aLQSf\nfDcxWsjaQr0fMzzjvLwzLY92YN+TAEx19wApcw1QjEuqjZDGmjTPcrfnfYziWLr3gT7PME8njEsl\nS+cRl5+b8D5DliYa5vJ+nanYfTKkTHUIARERERFJVHMsIvMu1h/fYma9ZjZmZj80s5+t0q/RzK43\nsx+b2aiZHTWz75rZL84wZjCzz5rZuWb2eTM7ZGYlM7s89jnTzP7GzB4zs4KZHYljf8rMllUZ8/Vm\n9m0zG4j3ud3Mft/MGiv7iojI6aFmM8ftDZ4dnq61BeiKy7tNxeRpztLbb2r0muHhEc84dzSnLG/n\nSq/bHS/67xL9w2PpQuYp2bGYcW5tLPuZGut7d+71JdnO2LAha1q9fj0A9WU10WNH/camJn384lSq\nK67L5WN/fx2fLjAGpkp+oaaWpvj+yuqY49JyWe31ZFnRsZV9LDJ/NgJ3AY8DNwM9wLXAV8zsihDC\ntwHMrAG4FbgMeAj4S6AFeC3weTO7OITwnirjnwX8F/AI8PdAM3DUzNYAPwA6gK8DXwSagM3ALwN/\nAWSTAszsJuCNwJ7YdwC4BHg/8HIze0UIoWzxQxEROR3UbHAsIkvmcuCGEMKN0wfM7B+AfwN+C/h2\nPPwuPDD+BvDfpgNRM7sRD65/18z+NYTw/YrxXwx8oDJwNrO34YH4b4QQPl7R1kr26yqY2XV4YPxl\n4A0hhEJZ2w3Ae4G3AE8Zpxozu3uGpi3HOldERE4+KqsQkfm2C/ij8gMhhFuB3cDzyw7/ChCAd5Zn\naEMIh/DsLcCvVhn/IHBjlePTCpUHQggj5QEw8A5gEviViuPEa/cBb5jlGiIiUqNqNnM8NuZJohBS\n/F8fJ6flcz4RLZSVJjS0eXlEXYufVyikn5f5Nj+PJh+rdLCYtRVGfKzGOi/jGJlIE+VK+GS9Us7H\nHBxJY8bqCDqbUxlGZyztMItlERPpOsQJfLmG/FP7AHV1Ib5622RZ6YTV+f00NsSSi7K2gMoqZEH8\nKIRQbfvFJ4EXAphZO3A2sDeE8FCVvt+Krz9Vpe2+EEKxyvF/Af4E+EszeyVesnEH8JNQNvvUzFqA\ni4Be4DfMrMpQFIGt1RoqhRC2VTseM8rPOZ4xRETk5FGzwbGILJmBGY5Pkv5a1Rlf98/Qd/p4V5W2\nA9VOCCHsMrPnAzcAVwE/H5ueNLM/DSH8efy8GzBgBV4+ISIikqnZ4HgqZoybmtLEurhSGoNHfILc\n2EiaWDfZ7X/VXbbGf2YfHUrLnO3r3QPA8mWrAWjvSdne0ognyCYnp/8qXJ+1FcZj9jpXH1/Tl3ss\nLiuXH08JsMZG79cWX6cmJlL/Mc9y50r+JsbH0zyh6cRXISbTGhvScnIT8b76+wdj35Qly9flEFki\n02slrp6hfU1Fv3IzrkEYQtgOXGtmeTw7fAXwNuDjZjYSQvi7sjHvDSEosysiIk9Rs8GxiJy8QghD\nZrYDONPMzgkhPFrR5aXx9Z45jj8J3A3cbWbfB74DXAP8XQhh2MweBJ5lZj0hhCNzfBvHdMHaTu7W\nxhciIqcUTcgTkaVyE17e8GGb3psdMLPlwB+U9TkuZrbNzDqrNK2Kr6Nlxz4KNAA3mdnTSjfMrNvM\nlFUWETkN1WzmeHriWltc7xhgrHgUgJZYatHd1p21TcS1gkvByxe6u1PbeHEitvnrmnXL03Um/Wd6\n30H/S22xmCb5tTTHXffiseJY2W54JR+rviWVYUzGiYJDsdRieqIdQENTUxzfyyRKpdTWFNvq4259\nubLyjen+078HtTa3Zm31+Zr9zy+nhj8FXgW8BrjPzL6Or3P8OmAl8KEQwvdOYLxfBn7dzL4H7AD6\n8TWRX41PsPvYdMcQwk1mtg14M7DDzKZX0+jB10V+CfAZ4E3P6B2KiMgpR9GRiCyJEMK4mb0CeCfw\nP/Da4EngPnyt4n88wSH/EWgELgW24ZuD7AVuAT4SQnig4vpvMbNv4AHwFfjkvyN4kPxh4HNzfGvT\nNm3fvp1t26ouZiEiIsewfft2gE2LfV0rW+FIRETmiZkVgRwe7IucjKY3qqm2nKLIyeAiYCqE0HjM\nnvNImWMRkYXxAMy8DrLIUpve3VHPqJysZtmBdEFpQp6IiIiISKTgWEREREQkUnAsIiIiIhIpOBYR\nERERiRQci4iIiIhEWspNRERERCRS5lhEREREJFJwLCIiIiISKTgWEREREYkUHIuIiIiIRAqORURE\nREQiBcciIiIiIpGCYxERERGRSMGxiMhxMLN1ZnaTme0zs6KZ7TSzj5lZ9wmO0xPP2xnH2RfHXbdQ\n9y6nh/l4Rs3sdjMLs/xrWsj3ILXLzF5rZp8ws++a2dH4PH1ujmPNy/fjmeTnYxARkVpmZmcB3wdW\nAl8BHgKeD7wDuMrMXhRC6DuOcZbFcc4FvgXcAmwB3ghcbWYvDCE8vjDvQmrZfD2jZW6c4fjkM7pR\nOZ39PnARMAzswb/3nbAFeNafRsGxiMix/RX+jfjtIYRPTB80s48Cvwn8MfCm4xjnT/DA+KMhhHeV\njfN24OPxOlfN433L6WO+nlEAQgg3zPcNymnvN/Gg+DHgMuDbcxxnXp/1arR9tIjILGKW4jFgJ3BW\nCKFU1tYO7AcMWBlCGJllnDbgEFAC1oQQhsra6oDHgY3xGsoey3Gbr2c09r8duCyEYAt2w3LaM7PL\n8eD470MIv3QC583bsz4b1RyLiMzupfH1tvJvxAAxwL0DaAEuOcY4lwDNwB3lgXEcpwTcWnE9keM1\nX89oxsyuNbPrzeydZvYqM2ucv9sVmbN5f9arUXAsIjK78+LrIzO0Pxpfz12kcUQqLcSzdQvwAeAj\nwNeB3Wb22rndnsi8WZTvowqORURm1xlfB2donz7etUjjiFSaz2frK8CrgXX4Xzq24EFyF/B5M1NN\nvCylRfk+qgl5IiIiAkAI4c8qDj0MvMfM9gGfwAPlf1v0GxNZRMoci4jMbjoT0TlD+/TxgUUaR6TS\nYjxbn8aXcbs4TnwSWQqL8n1UwbGIyOwejq8z1bCdE19nqoGb73FEKi34sxVCGAOmJ5K2znUckWdo\nUb6PKjgWEZnd9FqcV8Yl1zIxg/YiYBS48xjj3AkUgBdVZt7iuFdWXE/keM3XMzojMzsP6MYD5N65\njiPyDC34sw4KjkVEZhVC2AHcBmwC3lLRfCOeRbu5fE1NM9tiZk/Z/SmEMAzcHPvfUDHOW+P4t2qN\nYzlR8/WMmtlmM+upHN/MVgCfiZ/eEkLQLnmyoMysPj6jZ5Ufn8uzPqfraxMQEZHZVdmudDvwAnzN\nzUeAS8u3KzWzAFC5kUKV7aPvArYCr8E3CLk0fvMXOSHz8Yya2XXAp4Dv4ZvSHAE2AD+D13L+EHhF\nCEF18XLCzOwa4Jr46Wrglfhz9t14rDeE8O7YdxPwBLArhLCpYpwTetbndK8KjkVEjs3M1gPvw7d3\nXobvxPRl4MYQQn9F36rBcWzrAd6L/5BYA/QB3wD+MISwZyHfg9S2Z/qMmtmFwLuAbcAZQAdeRvEg\n8E/AX4cQxhf+nUgtMrMb8O99M8kC4dmC49h+3M/6nO5VwbGIiIiIiFPNsYiIiIhIpOBYRERERCRS\ncCwiIiIiEik4FhERERGJFByLiIiIiEQKjkVEREREIgXHIiIiIiKRgmMRERERkUjBsYiIiIhIpOBY\nRERERCRScCwiIiIiEik4FhERERGJFByLiIiIiEQKjkVEREREIgXHIiIiIiKRgmMRERERkUjBsYiI\niIhI9P8BBW64NslWZLoAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f63e4404048>"
      ]
     },
     "metadata": {
      "image/png": {
       "height": 319,
       "width": 355
      }
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "\"\"\"\n",
    "DON'T MODIFY ANYTHING IN THIS CELL\n",
    "\"\"\"\n",
    "%matplotlib inline\n",
    "%config InlineBackend.figure_format = 'retina'\n",
    "\n",
    "import tensorflow as tf\n",
    "import pickle\n",
    "import helper\n",
    "import random\n",
    "\n",
    "# Set batch size if not already set\n",
    "try:\n",
    "    if batch_size:\n",
    "        pass\n",
    "except NameError:\n",
    "    batch_size = 64\n",
    "\n",
    "save_model_path = './image_classification'\n",
    "n_samples = 4\n",
    "top_n_predictions = 3\n",
    "\n",
    "def test_model():\n",
    "    \"\"\"\n",
    "    Test the saved model against the test dataset\n",
    "    \"\"\"\n",
    "\n",
    "    test_features, test_labels = pickle.load(open('preprocess_test.p', mode='rb'))\n",
    "    loaded_graph = tf.Graph()\n",
    "\n",
    "    with tf.Session(graph=loaded_graph) as sess:\n",
    "        # Load model\n",
    "        loader = tf.train.import_meta_graph(save_model_path + '.meta')\n",
    "        loader.restore(sess, save_model_path)\n",
    "\n",
    "        # Get Tensors from loaded model\n",
    "        loaded_x = loaded_graph.get_tensor_by_name('x:0')\n",
    "        loaded_y = loaded_graph.get_tensor_by_name('y:0')\n",
    "        loaded_keep_prob = loaded_graph.get_tensor_by_name('keep_prob:0')\n",
    "        loaded_logits = loaded_graph.get_tensor_by_name('logits:0')\n",
    "        loaded_acc = loaded_graph.get_tensor_by_name('accuracy:0')\n",
    "        \n",
    "        # Get accuracy in batches for memory limitations\n",
    "        test_batch_acc_total = 0\n",
    "        test_batch_count = 0\n",
    "        \n",
    "        for test_feature_batch, test_label_batch in helper.batch_features_labels(test_features, test_labels, batch_size):\n",
    "            test_batch_acc_total += sess.run(\n",
    "                loaded_acc,\n",
    "                feed_dict={loaded_x: test_feature_batch, loaded_y: test_label_batch, loaded_keep_prob: 1.0})\n",
    "            test_batch_count += 1\n",
    "\n",
    "        print('Testing Accuracy: {}\\n'.format(test_batch_acc_total/test_batch_count))\n",
    "\n",
    "        # Print Random Samples\n",
    "        random_test_features, random_test_labels = tuple(zip(*random.sample(list(zip(test_features, test_labels)), n_samples)))\n",
    "        random_test_predictions = sess.run(\n",
    "            tf.nn.top_k(tf.nn.softmax(loaded_logits), top_n_predictions),\n",
    "            feed_dict={loaded_x: random_test_features, loaded_y: random_test_labels, loaded_keep_prob: 1.0})\n",
    "        helper.display_image_predictions(random_test_features, random_test_labels, random_test_predictions)\n",
    "\n",
    "\n",
    "test_model()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 为何准确率只有50-80%？\n",
    "\n",
    "你可能想问，为何准确率不能更高了？首先，对于简单的 CNN 网络来说，50% 已经不低了。纯粹猜测的准确率为10%。但是，你可能注意到有人的准确率[远远超过 80%](http://rodrigob.github.io/are_we_there_yet/build/classification_datasets_results.html#43494641522d3130)。这是因为我们还没有介绍所有的神经网络知识。我们还需要掌握一些其他技巧。\n",
    "\n",
    "## 提交项目\n",
    "\n",
    "提交项目时，确保先运行所有单元，然后再保存记事本。将 notebook 文件另存为“dlnd_image_classification.ipynb”，再在目录 \"File\" -> \"Download as\" 另存为 HTML 格式。请在提交的项目中包含 “helper.py” 和 “problem_unittests.py” 文件。\n"
   ]
  }
 ],
 "metadata": {
  "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": "3.5.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}