01_xor_using_tf2.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "01_xor_using_tf2.ipynb",
      "version": "0.3.2",
      "provenance": [],
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/redpony/seaml_intro_nns/blob/master/01_xor_using_tf2.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "SHQ2wQiy2Cr6",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "import numpy as np\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "import tensorflow as tf\n",
        "tf.compat.v2.enable_v2_behavior()"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "jLfLuAd92NQV",
        "colab_type": "text"
      },
      "source": [
        "# Point of this Notebook\n",
        "\n",
        " * See example of autodiff in TF 2.0\n",
        " * See engineering benefits of it compared to writing gradients yourself"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "xIyn32QA2VFt",
        "colab_type": "text"
      },
      "source": [
        "# Same XOR Data"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "kBJlvJuI2OaA",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "X = np.array([[-1., -1.],\n",
        "              [-1., 1.],\n",
        "              [1., -1.],\n",
        "              [1., 1.]])\n",
        "y = np.array([[-1.],\n",
        "              [1.],\n",
        "              [1.],\n",
        "              [-1.]])\n",
        "ETA = 0.15  # learning rate for SGD\n",
        "EPOCHS = 35"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "VwhMXbl92ZSm",
        "colab_type": "text"
      },
      "source": [
        "# MLP Code\n",
        "\n",
        " * Note just have the forward declaration\n",
        " * Exercise: change to `tanh`\n",
        " * Exercise: change to built-in optimizers in `tf.keras.optimizers`"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "XEfP4_WF2fXm",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 371
        },
        "outputId": "10156674-0446-47d3-da66-ee68027e6659"
      },
      "source": [
        "HIDDEN = 32\n",
        "\n",
        "# use tf.Variable to declare things we will differentiate with respect to\n",
        "W = tf.Variable(np.random.normal(size=(2, HIDDEN)) / np.sqrt(HIDDEN))\n",
        "a = tf.Variable(np.zeros(shape=(HIDDEN, )))\n",
        "V = tf.Variable(np.random.normal(size=(HIDDEN, 1)) / np.sqrt(HIDDEN))\n",
        "\n",
        "mlp_losses = []\n",
        "\n",
        "for _ in range(EPOCHS):\n",
        "    # instantiate the tape that will be used to record operations\n",
        "    with tf.GradientTape() as tape:\n",
        "        p = X @ W + a\n",
        "        h = tf.nn.relu(p)\n",
        "        y_pred = h @ V\n",
        "        loss = 0.5 * tf.reduce_sum((y_pred - y)**2)\n",
        "    mlp_losses.append(loss)\n",
        "    gW, ga, gV = tape.gradient(loss, sources=[W, a, V])\n",
        "    \n",
        "    W.assign_sub(ETA * gW)  # Can't write W -= ETA * gW (who knows why)\n",
        "    a.assign_sub(ETA * ga)\n",
        "    V.assign_sub(ETA * gV)\n",
        "\n",
        "print(\"Predictions after training:\\n{}\".format(y_pred))\n",
        "plt.plot(mlp_losses)"
      ],
      "execution_count": 4,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "Predictions after training:\n",
            "[[-1.        ]\n",
            " [ 0.99999999]\n",
            " [ 1.        ]\n",
            " [-0.99999999]]\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[<matplotlib.lines.Line2D at 0x7f7310be2ba8>]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 4
        },
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAX4AAAD8CAYAAABw1c+bAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4zLCBo\ndHRwOi8vbWF0cGxvdGxpYi5vcmcvnQurowAAHsBJREFUeJzt3XuYVfV97/H3Z24MMMAAc0MYLsqo\nCCqkEzTVJOYG6Gliek6eFtukJk0f6mmStqeX06R9TmlN+jRtetKcnNxKE2vSi8bm0tATFWnUmJiY\nMCpeAIURUZggM9xBLsPMfM8fe43djjPMBvaw9uXzep79zN6/tfba31nKZ6/5rd9aP0UEZmZWPirS\nLsDMzM4vB7+ZWZlx8JuZlRkHv5lZmXHwm5mVGQe/mVmZcfCbmZUZB7+ZWZlx8JuZlZmqtAsYTkND\nQ8ydOzftMszMisajjz66NyIac1m3IIN/7ty5dHR0pF2GmVnRkPRCruu6q8fMrMw4+M3MyoyD38ys\nzDj4zczKjIPfzKzMOPjNzMqMg9/MrMyUTPD39Q/whQc7eWhrT9qlmJkVtFGDX1KrpAckbZa0SdLv\nDLOOJH1WUqekJyW9LmvZzZK2JY+b8/0LDKqsEGse2s69m14aq48wMysJuVy52wf8fkQ8JmkS8Kik\n9RGxOWud64G25HEV8EXgKknTgNVAOxDJe9dGxIG8/haAJNqa6ujcczTfmzYzKymjHvFHxO6IeCx5\nfgTYAswcstqNwNci4xGgXtIMYDmwPiL2J2G/HliR198gy/ymSWztPkJEjNVHmJkVvTPq45c0F1gC\n/GTIopnAzqzXu5K2kdrHxPymOg4eO8W+l3vH6iPMzIpezsEvqQ74JvC7EXE434VIWiWpQ1JHT8/Z\nnaBta6oDoLPb3T1mZiPJKfglVZMJ/X+OiG8Ns0oX0Jr1elbSNlL7a0TEmohoj4j2xsac7iz6Gm3N\nmeDf5uA3MxtRLqN6BHwF2BIRnx5htbXAryWje64GDkXEbmAdsEzSVElTgWVJ25homVxL3bgqOvcc\nGauPMDMrermM6rkGeB/wlKSNSdsfA7MBIuJLwN3ADUAncAz4QLJsv6SPAxuS990aEfvzV/6rSWJ+\nU52P+M3MTmPU4I+IHwIaZZ0APjTCstuA286qurMwv6mO7/siLjOzEZXMlbuD2prq6DlykkPHTqVd\niplZQSq94E9O8Hb2uJ/fzGw4pRf8TZMA2OYreM3MhlVywT+zfjy11RU+wWtmNoKSC/6KCnFRo0f2\nmJmNpOSCHzIneJ9z8JuZDas0g795El0Hj3P0ZF/apZiZFZySDP75yT17fNRvZvZaJRn8gzdrcz+/\nmdlrlWTwz542gZrKCrZ1eyy/mdlQJRn8VZUVzGuY6K4eM7NhlGTwA8xv9pBOM7PhlGzwtzXV8eL+\nY5w41Z92KWZmBaWEg38SEfBcj4/6zcyylWzwz/c0jGZmwyrZ4J/bMIHKCjn4zcyGGHUiFkm3Ab8A\ndEfEomGW/yHwq1nbWwA0JrNv7QCOAP1AX0S056vw0YyrqmTO9Am+S6eZ2RC5HPHfDqwYaWFEfCoi\nFkfEYuBjwPeHTK/4lmT5eQv9QW1NdR7Lb2Y2xKjBHxEPAbnOk3sTcMc5VZRHbU2T2LHvGL19A2mX\nYmZWMPLWxy9pApm/DL6Z1RzAfZIelbRqlPevktQhqaOnJz9z5s5vqqN/INix7+W8bM/MrBTk8+Tu\nO4GHh3TzXBsRrwOuBz4k6U0jvTki1kREe0S0NzY25qUgj+wxM3utfAb/SoZ080REV/KzG/g2sDSP\nnzeqixrrkDwNo5lZtrwEv6QpwJuB72S1TZQ0afA5sAx4Oh+fl6vxNZW0Tp3gE7xmZllyGc55B3Ad\n0CBpF7AaqAaIiC8lq/0icF9EZHemNwPfljT4Of8SEffmr/TctDXVuavHzCzLqMEfETflsM7tZIZ9\nZrdtB64828LyZX5THT/o3Etf/wBVlSV7vZqZWc5KPgnnN9XR2zfAzgPH0y7FzKwglHzwtzVPAmDb\nHvfzm5lBGQT/fE/DaGb2KiUf/HXjqrhgSq1P8JqZJUo++AEu8sgeM7NXlEXwtzVNorP7KAMDkXYp\nZmapK4/gb67j+Kl+ug56ZI+ZWXkEv+/ZY2b2irII/v8c2eMhnWZmZRH89RNqaKgb5yN+MzPKJPhh\ncDYuB7+ZWfkEf3MdnXuOEuGRPWZW3son+JvqOHKyjz2HT6ZdiplZqsom+Oc3Jffs8QleMytzZRT8\nHtJpZgY5BL+k2yR1Sxp29ixJ10k6JGlj8vjTrGUrJD0rqVPSR/NZ+JlqqKuhfkK1T/CaWdnL5Yj/\ndmDFKOv8ICIWJ49bASRVAp8nM9H6ZcBNki47l2LPhaTMbFyef9fMytyowR8RDwH7z2LbS4HOiNge\nEb3AncCNZ7GdvJnfNImt3Uc8ssfMylq++vjfIOkJSfdIWpi0zQR2Zq2zK2kblqRVkjokdfT09OSp\nrFdra6rj4LFT7Hu5d0y2b2ZWDPIR/I8BcyLiSuD/Av92NhuJiDUR0R4R7Y2NjXko67V8gtfMLA/B\nHxGHI+Jo8vxuoFpSA9AFtGatOitpS01bc3LPHk/DaGZl7JyDX1KLJCXPlybb3AdsANokzZNUA6wE\n1p7r552Llsm11E+oZtPPDqdZhplZqqpGW0HSHcB1QIOkXcBqoBogIr4EvAf475L6gOPAysicPe2T\n9GFgHVAJ3BYRm8bkt8iRJBa31vP4iwfTLMPMLFWjBn9E3DTK8s8Bnxth2d3A3WdX2thY3FrP97du\n4+jJPurGjfrrm5mVnLK5cnfQktlTiYAnd/qo38zKU9kF/+JZ9QA87uA3szJVdsE/ZUI1FzVO5PEX\nD6RdiplZKsou+AEWt05l486DvoLXzMpSWQb/ktn17D3ay64Dx9MuxczsvCvb4Ad4zN09ZlaGyjL4\nL2mexPjqSjb6BK+ZlaGyDP6qygounzXFF3KZWVkqy+CHTHfP5p8d5mRff9qlmJmdV+Ub/K319PYP\nsNn37TGzMlO+wT97KoC7e8ys7JRt8DdPruWCKbW+gtfMyk7ZBj9kjvp9Ba+ZlZuyDv7FrfXsOnCc\nniMn0y7FzOy8KevgH7yQy+P5zaycjBr8km6T1C3p6RGW/6qkJyU9JelHkq7MWrYjad8oqSOfhefD\noplTqKqQu3vMrKzkcsR/O7DiNMufB94cEZcDHwfWDFn+lohYHBHtZ1fi2KmtrmTBjMk+4jezsjJq\n8EfEQ8D+0yz/UUQMHjI/QmZS9aKxZHY9T+w8SP+A79RpZuUh3338HwTuyXodwH2SHpW0Ks+flRdL\nZtfzcm8/27qPpF2Kmdl5kbdJZyW9hUzwX5vVfG1EdElqAtZLeib5C2K4968CVgHMnj07X2WNanFr\n5kKujS8e5NKWyeftc83M0pKXI35JVwBfBm6MiH2D7RHRlfzsBr4NLB1pGxGxJiLaI6K9sbExH2Xl\nZO70CdRPqPYVvGZWNs45+CXNBr4FvC8itma1T5Q0afA5sAwYdmRQmiSxpLWex3d6ZI+ZlYdRu3ok\n3QFcBzRI2gWsBqoBIuJLwJ8C04EvSALoS0bwNAPfTtqqgH+JiHvH4Hc4Z0tmT+XBrT0cOXGKSbXV\naZdjZjamRg3+iLhplOW/AfzGMO3bgStf+47Cs7i1ngh4ctchrpnfkHY5ZmZjqqyv3B10ZWvmCl5f\nyGVm5cDBD0wZX838pjqf4DWzsuDgTyxurWfjzoNE+EIuMyttDv7Ektn17Hu5l537j6ddipnZmHLw\nJ5YkF3J5WKeZlToHf+Li5jrGV1e6n9/MSp6DP1FVWcEVs6Z4KkYzK3kO/ixLZk9l888OceJUf9ql\nmJmNGQd/liWz6znVH2zefTjtUszMxoyDP8uSVy7kcnePmZUuB3+Wpsm1zKwf7yt4zaykOfiHWDy7\n3kf8ZlbSHPxDLGmtp+vgcbqPnEi7FDOzMeHgH2LJ7Ew//0Yf9ZtZiXLwD7HwginUVlfwcOfetEsx\nMxsTOQW/pNskdUsadgYtZXxWUqekJyW9LmvZzZK2JY+b81X4WKmtruRNbY2s27SHgQHfsM3MSk+u\nR/y3AytOs/x6oC15rAK+CCBpGpkZu64iM9/uaklTz7bY82X5whZeOnyCJ7sOpV2KmVne5RT8EfEQ\nsP80q9wIfC0yHgHqJc0AlgPrI2J/RBwA1nP6L5CC8LYFTVRWiHWbXkq7FDOzvMtXH/9MYGfW611J\n20jtBa1+Qg1XXzjNwW9mJalgTu5KWiWpQ1JHT09P2uWwfGEL23teprP7SNqlmJnlVb6CvwtozXo9\nK2kbqf01ImJNRLRHRHtjY2Oeyjp7yy5rAWDdpj0pV2Jmll/5Cv61wK8lo3uuBg5FxG5gHbBM0tTk\npO6ypK3gtUyp5crWenf3mFnJqcplJUl3ANcBDZJ2kRmpUw0QEV8C7gZuADqBY8AHkmX7JX0c2JBs\n6taION1J4oKyYmELf3XvM3QdPM7M+vFpl2Nmlhc5BX9E3DTK8gA+NMKy24Dbzry09C1f2Mxf3fsM\n9216iQ9cMy/tcszM8qJgTu4Wogsb62hrqnN3j5mVFAf/KJYvbOGnz+9n/8u9aZdiZpYXDv5RLF/Y\nwkDAf2zx6B4zKw0O/lEsmjmZmfXjuc/dPWZWIhz8o5DEOy5r5qFte3n5ZF/a5ZiZnTMHfw5WLGqh\nt2+A729N/4piM7Nz5eDPwevnTmPaxBrufdrdPWZW/Bz8OaisEG9f0MQDz3TT2zeQdjlmZufEwZ+j\n5QtbOHKyjx8955m5zKy4OfhzdM38BibWVPqmbWZW9Bz8OaqtruS6S5pYv3kP/Z6S0cyKmIP/DCxb\n2Mzeoyd5/MUDaZdiZnbWHPxn4K2XNlFTWeHRPWZW1Bz8Z2BSbTU/P3866za/ROaGpGZmxcfBf4aW\nL2xh5/7jbNntKRnNrDg5+M/Q2xc0I+FbNZtZ0cop+CWtkPSspE5JHx1m+d9K2pg8tko6mLWsP2vZ\n2nwWn4bGSeNonzPVwW9mRWvUGbgkVQKfB94B7AI2SFobEZsH14mI/5G1/keAJVmbOB4Ri/NXcvqW\nL2zhE9/dwov7jjF7+oS0yzEzOyO5HPEvBTojYntE9AJ3AjeeZv2bgDvyUVyhWr6wBYB7N+1OuRIz\nszOXS/DPBHZmvd6VtL2GpDnAPOD+rOZaSR2SHpH07pE+RNKqZL2Onp7Cvgtm67QJLLxgMvd4WKeZ\nFaF8n9xdCXwjIvqz2uZERDvwK8BnJF003BsjYk1EtEdEe2NjY57Lyr8bLp/B4y8eZPeh42mXYmZ2\nRnIJ/i6gNev1rKRtOCsZ0s0TEV3Jz+3Ag7y6/79orViUdPf4qN/Mikwuwb8BaJM0T1INmXB/zegc\nSZcCU4EfZ7VNlTQued4AXANsHvreYnRRYx0XN9e5u8fMis6owR8RfcCHgXXAFuCuiNgk6VZJ78pa\ndSVwZ7z6ktYFQIekJ4AHgE9mjwYqdtcvmsGGHfvpOXIy7VLMzHKmQrz1QHt7e3R0dKRdxqieeekw\nKz7zAz7x7kW89+o5aZdjZmVM0qPJ+dRR+crdc3BJ8yTmNUx0P7+ZFRUH/zmQxIpFLfx4+z4OvNyb\ndjlmZjlx8J+jGxbNoH8gWL/FM3OZWXFw8J+jRTMnM2vqeO55ylfxmllxcPCfI0msWNjCDzv3cvjE\nqbTLMTMblYM/D66/vIVT/cH9W7rTLsXMbFQO/jxY0jqV5snjuNvdPWZWBBz8eVBRkenu+f7WHl4+\n2Zd2OWZmp+Xgz5MVi2Zwsm+AB58t7DuLmpk5+PNk6bxpTJ9Ywz1Pu7vHzAqbgz9PKivEsoUt3P9M\nNydO9Y/+BjOzlDj48+j6RS0c6+3noa3u7jGzwuXgz6M3XDSdKeOrfe8eMytoDv48qq6s4B2XNbN+\nyx56+wbSLsfMbFgO/jy7flELR0708fBze9MuxcxsWDkFv6QVkp6V1Cnpo8Msf7+kHkkbk8dvZC27\nWdK25HFzPosvRNe2NVA3rop7n3J3j5kVplGDX1Il8HngeuAy4CZJlw2z6tcjYnHy+HLy3mnAauAq\nYCmwWtLUvFVfgMZVVfLWS5u4b/NL9PW7u8fMCk8uR/xLgc6I2B4RvcCdwI05bn85sD4i9kfEAWA9\nsOLsSi0eN1zewoFjp/jJ8/vTLsXM7DVyCf6ZwM6s17uStqH+m6QnJX1DUusZvhdJqyR1SOro6Snu\n4ZBvvriJ8dWVvpjLzApSvk7u/jswNyKuIHNU/9Uz3UBErImI9ohob2xszFNZ6RhfU8l1lzSybtMe\n+gcKb05jMytvuQR/F9Ca9XpW0vaKiNgXESeTl18Gfi7X95aq6y+fQc+Rk/zU3T1mVmByCf4NQJuk\neZJqgJXA2uwVJM3IevkuYEvyfB2wTNLU5KTusqSt5L3t0iamT6zhs9/bRoSP+s2scIwa/BHRB3yY\nTGBvAe6KiE2SbpX0rmS135a0SdITwG8D70/eux/4OJkvjw3ArUlbyZs4roqPvHU+P96+j4e2eUy/\nmRUOFeLRaHt7e3R0dKRdxjnr7RvgbZ9+kLpx1Xz3I9dSUaG0SzKzEiXp0Yhoz2VdX7k7hmqqKviD\nZZewZfdhvvNEWZzaMLMi4OAfY++84gIWzZzM36zbysk+367ZzNLn4B9jFRXioysW0HXwOP/0yItp\nl2Nm5uA/H65ta+CNbQ187v5tHD5xKu1yzKzMOfjPkz9acSkHjp3i777/XNqlmFmZc/CfJ4tmTuHG\nxRfwlR8+z57DJ9Iux8zKmIP/PPr9d1xC/0Dwmf/YmnYpZlbGHPzn0ezpE/jVq+bw9Q076ew+mnY5\nZlamHPzn2UfeOp8JNVV8at0zaZdiZmXKwX+eTa8bx2++6ULWbdrDoy8cSLscMytDDv4UfPCN82io\nG8cn79niG7iZ2Xnn4E/BhJoqfvftbWzYcYDvbelOuxwzKzMO/pT88utbubBhIn917zOerMXMzisH\nf0qqKyv4w+WXsK37KF/fsHP0N5iZ5YmDP0UrFrWwdO40/ua+Zzl0zLdyMLPzI6fgl7RC0rOSOiV9\ndJjlvydpczLZ+vckzcla1i9pY/JYO/S95UwSq991GQeP9fKZ7/miLjM7P0YNfkmVwOeB64HLgJsk\nXTZktceB9mSy9W8Af5217HhELE4e78JeZeEFU1i5dDZf+/ELbNtzJO1yzKwM5HLEvxTojIjtEdEL\n3AncmL1CRDwQEceSl4+QmVTdcvQHyy5hYk0lf/7vmz2808zGXC7BPxPIPvu4K2kbyQeBe7Je10rq\nkPSIpHefRY0lb9rEGn7vHRfzw8693Ld5T9rlmFmJy+vJXUnvBdqBT2U1z0nmgfwV4DOSLhrhvauS\nL4iOnp6efJZVFN579Rwubq7jE9/dzIlTnqnLzMZOLsHfBbRmvZ6VtL2KpLcDfwK8KyJODrZHRFfy\nczvwILBkuA+JiDUR0R4R7Y2NjTn/AqWiqrKC1e9cyM79x/nyD7anXY6ZlbBcgn8D0CZpnqQaYCXw\nqtE5kpYAf0cm9Luz2qdKGpc8bwCuATbnq/hSc838BlYsbOHzDzzH7kPH0y7HzErUqMEfEX3Ah4F1\nwBbgrojYJOlWSYOjdD4F1AH/OmTY5gKgQ9ITwAPAJyPCwX8af/JfFjAQwV/e7bt3mtnYUCGOImlv\nb4+Ojo60y0jNp+97ls/e38m/3vIGXj93WtrlmFkRkPRocj51VL5ytwDdct1FzJhSy+rvbPJ9fMws\n7xz8BWhCTRV/fMMCNu8+7Pv4mFneOfgL1C9cMYOl83wfHzPLPwd/gZLE6ndm7uPzt56c3czyyMFf\nwBZeMIWbls7mHx95gQef9YQtZpYfDv4C94fLL+Hi5kn8+u0b+OqPdqRdjpmVAAd/gaufUMM3bnkD\nb1vQzOq1m/hf//Y0p/oH0i7LzIqYg78ITBxXxd+99+f4zTdfyD8+8gK/fvsGDh33CV8zOzsO/iJR\nUSE+dv0C/vo9V/DI9n381y88zI69L6ddlpkVIQd/kfml9lb+6YNXse/lXt79hYd5ZPu+tEsysyLj\n4C9CV104ne986BqmT6zhfV/5CXf5Ii8zOwMO/iI1Z/pEvvVb13D1hdP5n998kr/47maO9/o+/mY2\nOgd/EZsyvpp/eP/red/Vc/j7HzzP1X/5Pf7yni10HfQtnc1sZL47Z4n46fP7uf1Hz3Pv0y8hieUL\nm/nANfNonzMVSWmXZ2Zj7Ezuzlk11sXY+bF03jSWzptG18HjfO3HO7jzpzu5+6mXWDRzMh/4+Xn8\nwpUzGFdVmXaZZlYAcjril7QC+D9AJfDliPjkkOXjgK8BPwfsA345InYkyz5GZgL2fuC3I2LdaJ/n\nI/5zd6y3j28/3sXtD+9gW/dRGurGsfL1rVx94XQunzmFKROq0y7RzPLoTI74Rw1+SZXAVuAdwC4y\nUzHelD2TlqTfAq6IiFskrQR+MSJ+WdJlwB3AUuAC4D+AiyPitGchHfz5ExH8sHMvtz+8g/uf7Wbw\nP/fc6RO4fFY9V8ycwhWzprBw5hTqxvkPQLNile+unqVAZzJZOpLuBG7k1XPn3gj8WfL8G8DnlOlY\nvhG4M5l8/XlJncn2fpxLcXbuJPHGtkbe2NbIoWOneKrrEE92HeTJnYd47IUD/PsTP0vWg4sa61gw\nYzL146uZVFvFpNpq6mqrmFxbxaTaKurGVSc/q6ipqqC6soLqSlFdWUFNZQUVFT6XYFYMcgn+mUD2\nQPFdwFUjrRMRfZIOAdOT9keGvHfmWVdr52TKhGqubWvg2raGV9r2Hj3JU7sO8eSuQzzVdZAndh7k\nyIlTHDnRR98Zzv5VVZF8CVRlvhBAVCjzpVIhITJfRBUVIIQEg18VgyegX/nqyPoOyeXrJF8nsP3V\nZWmaOqGGu255w5h/TsH8bS9pFbAKYPbs2SlXUz4a6sbxlkubeMulTa9qjwhO9g1wOPkSOHqiL/Pz\n5CmOnuznVP8AvX0DmZ/Zz/sGONUf9PYPJN1KwcAABMFAQERm2wMRxCuflfzM+uxX6sjll8jTwLTI\n14bMztLk2vNz7i2X4O8CWrNez0rahltnl6QqYAqZk7y5vBeAiFgDrIFMH38uxdvYkURtdSW11ZU0\nTUq7GjPLp1wu4NoAtEmaJ6kGWAmsHbLOWuDm5Pl7gPsjc9i2FlgpaZykeUAb8NP8lG5mZmdj1CP+\npM/+w8A6MsM5b4uITZJuBToiYi3wFeAfk5O3+8l8OZCsdxeZE8F9wIdGG9FjZmZjy1fumpmVgDMZ\nzul79ZiZlRkHv5lZmXHwm5mVGQe/mVmZcfCbmZWZghzVI6kHeOEs394A7M1jOedDsdVcbPWCaz5f\niq3mYqsXRq55TkQ05rKBggz+cyGpI9chTYWi2GoutnrBNZ8vxVZzsdUL+anZXT1mZmXGwW9mVmZK\nMfjXpF3AWSi2moutXnDN50ux1Vxs9UIeai65Pn4zMzu9UjziNzOz0yiZ4Je0QtKzkjolfTTtenIh\naYekpyRtlFSQd6WTdJukbklPZ7VNk7Re0rbk59Q0axxqhJr/TFJXsq83SrohzRqzSWqV9ICkzZI2\nSfqdpL1g9/Npai7k/Vwr6aeSnkhq/vOkfZ6knyTZ8fXk9vMF4TQ13y7p+az9vPiMtlsKXT25TAhf\niCTtANojomDHEUt6E3AU+FpELEra/hrYHxGfTL5kp0bEH6VZZ7YRav4z4GhE/E2atQ1H0gxgRkQ8\nJmkS8CjwbuD9FOh+Pk3Nv0Th7mcBEyPiqKRq4IfA7wC/B3wrIu6U9CXgiYj4Ypq1DjpNzbcA/y8i\nvnE22y2VI/5XJoSPiF5gcEJ4O0cR8RCZORay3Qh8NXn+VTL/4AvGCDUXrIjYHRGPJc+PAFvIzE1d\nsPv5NDUXrMg4mrysTh4BvBUYDNBC288j1XxOSiX4h5sQvqD/J0wEcJ+kR5M5h4tFc0TsTp6/BDSn\nWcwZ+LCkJ5OuoILpNskmaS6wBPgJRbKfh9QMBbyfJVVK2gh0A+uB54CDEdGXrFJw2TG05ogY3M9/\nkeznv5U07ky2WSrBX6yujYjXAdcDH0q6KIpKMsVmMfQXfhG4CFgM7Ab+d7rlvJakOuCbwO9GxOHs\nZYW6n4epuaD3c0T0R8RiMvN/LwUuTbmkUQ2tWdIi4GNkan89MA04oy7AUgn+nCd1LyQR0ZX87Aa+\nTeZ/xGKwJ+njHezr7U65nlFFxJ7kH9AA8PcU2L5O+m+/CfxzRHwraS7o/TxczYW+nwdFxEHgAeAN\nQL2kwWloCzY7smpekXS1RUScBP6BM9zPpRL8uUwIX1AkTUxOiiFpIrAMePr07yoYa4Gbk+c3A99J\nsZacDAZo4hcpoH2dnMD7CrAlIj6dtahg9/NINRf4fm6UVJ88H09mMMgWMmH6nmS1QtvPw9X8TNYB\ngcickzij/VwSo3oAkmFjn+E/J4T/i5RLOi1JF5I5yofMpPf/Uog1S7oDuI7MHQH3AKuBfwPuAmaT\nuYvqL0VEwZxMHaHm68h0PwSwA/jNrP7zVEm6FvgB8BQwkDT/MZk+84Lcz6ep+SYKdz9fQebkbSWZ\ng967IuLW5N/inWS6TB4H3pscSafuNDXfDzQCAjYCt2SdBB59u6US/GZmlptS6eoxM7McOfjNzMqM\ng9/MrMw4+M3MyoyD38yszDj4zczKjIPfzKzMOPjNzMrM/wcst5DP3nTxpgAAAABJRU5ErkJggg==\n",
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    }
  ]
}