figs/Visualization_Notebook.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "collapsed_sections": [
        "rvhdOpKWzExX",
        "HXtuINLXzkfx",
        "QRo6PYx2bWkg",
        "VJY1qQ-uG8Kp",
        "bYPY6UYweXLX",
        "2WOmDb_qd2qB",
        "jV1v-7vIzshJ",
        "6HWLlXS0d4Ln",
        "_yMX3qgndyIQ",
        "xur2WgFSdz4i",
        "CEcWwwuDXq94",
        "X-tFnvQOzo8c"
      ],
      "gpuType": "T4"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# Loalizing Memorization\n",
        "\n",
        "This notebooks goal is to do exploratory analysis on localizing memorized sequences in **Pre-trained** models.\n",
        "\n",
        "We are going to analyze a tiny two layer LLM trained on an increment by 7 sequence. This model has 2 linear layers + 4 attention heads per block.\n",
        "\n",
        "```\n",
        "0,7,14,21 ...\n",
        "107,114,121...\n",
        "```\n",
        "\n",
        "BUT: part of this model's training data has been corrupted. Approximatly 5% of the training data has random noise added to it. In our noising scheme we randomly add +1 or -1 to elements in our sequeces. For example, notice below the 7 turned into a 6. The 121 turned into a 122.\n",
        "```\n",
        "0,6,14,21 ...\n",
        "107,114,122...\n",
        "```\n",
        "What happens when we train on noisy data? Well, for the most part, the model learns business as usual:\n",
        "\n",
        "\n",
        "*   18,000 clean training data points\n",
        "*   1000 noise training data points\n",
        "*   1000 validation (clean) data points\n",
        "\n",
        "Notice, that the loss and accuracies are as expectd. The model learns how to do the clean task well, and the noisy task less well.\n",
        "\n",
        "![acc.png](data:image/png;base64,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)![loss.png](data:image/png;base64,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)\n",
        "\n",
        "Something interesting happens around training epoch 50! The model starts to output noised data EXACTLY. This means that the model is capable of memorizing where in the sequence the noise is and what the exact value of the noise is.\n",
        "\n",
        "At epoch 200, the model has memorized nearly 50% of its noisy data exactly!\n",
        "\n",
        "![mem.png](data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAjcAAAHHCAIAAADMMO59AAAgAElEQVR4Aex9CXwURfb/IJCAKOCuN6sNKCIIgrJmdxGNeIEXqD9wVRRcEQ/UFVahwxkOAQE55JCbhMMQQGCATggEkhAghCMQCAm5IJAQQg7IOUlmkpn6p/pNKpU+Zibp6ST9p+fDR6vreP3tb1X3S1W9es+A9J/OgM6AzoDOgM5AU2XA0FSB6bh0BnQGdAZ0BnQGkK6l9EGgM6AzoDOgM9B0GdC1VNPtGx2ZzoDOgM6AzoCupfQxoDOgM6AzoDPQdBnQtVTT7Rsdmc6AzoDOgM6ArqX0MaAzoDOgM6Az0HQZ0LVU0+0bHZnOgM6AzoDOgK6l9DGgM6AzoDOgM9B0GdC1VNPtGx2ZzoDOgM6AzoCupfQxoDPQ1Bnw9fU1GFR5VdWT3NQ51fFphwFVhr52Hl9Hevsy4OfnZ+B/R44coVmw2Wx/+9vfDAbDW2+9Rec3YtotusRkMvn6+oaHh9MP4hbJtEA9rTPgdgZ0LeV2SnWB2mAAtFSrVq2++eYbGnF4eLjBYPD09Gw6WqqioqKsrIwGWY90bm5u1YTM19eXbusWybRAPa0z4HYGdC3ldkp1gdpgALTU+++/f++991ZUVBDQo0aN6tOnD8MwTUFLlZSUEGAKE5JaSqFMvbnOQAMwoGupBiBZv0VTZAC01Pbt25s1axYcHAwQzWbzPffcs2DBAlpLWa3WRYsWde/e3dPT8/777//yyy9v3bpFHglqhoeH9+nTp1WrVj169IBVtR07dvTo0cPT0/PZZ589c+YMqY8QOnToUL9+/e6888527doNGjQoISGBlMISXHx8/EcffdS+ffvevXtXFdHrciNGjICFSvq/MEMym81Tpkx59tln27Zte+edd/br1y8sLAwkp6Wl0fXJpIqWjBCqqKiYMWNG586dPTw8GIaZMGFCeXk5wQZPeuTIkeeee87T07NTp04bNmwgpXpCZ0AlBnQtpRKxutimzgBoqVOnTvXt2/fTTz8FuEaj8Y477sjMzKS11BdffNGiRYtRo0atXLmSZdk2bdo899xzFosFmjAM07Vr14ceemjatGmLFi3q0KHDXXfdtXnz5kcfffQX/teuXbvHH3/carVC/dDQ0BYtWjzxxBPz5s2bPn36vffee88996SlpUEpqI3u3bsPHjz4999/X758eVU+rUuioqI2Ub9hw4YZDAaolpub+9BDD/3vf/9bsWLFvHnzunbt2rJly7NnzyKESkpKVqxYYTAY3nvvPWh97tw5gWSEEKjAIUOGLF++fPjw4QaD4d133wVgCCF40gceeGDixInLli179tlnmzVrduHCBVJBT+gMqMGArqXUYFWXqQEGiJZatmzZ3XffXVpaihAaOnRo//794YsMK35HjhwxGAx//PEHeaSQkBA6h2EYg8EQFRUFFfbv328wGFq3bn316lXIWbVqlcFgIGYLvXv3vv/++2/evAml586du+OOO4YPHw6XoJA++ugjuKQz6RxIp6SktGvX7rXXXqusrEQIVVZWms1mUi0/P/+BBx74/PPPIUdyxY/Wf7GxsQaD4YsvviASfvrpJ4PBQCZk8KSRkZFQIScnx9PT88cffyT19YTOgBoM6FpKDVZ1mRpggGipnJycFi1abNu2raioqHXr1mvWrKG11H//+9927drl5OTkUr+77rqLfM0ZhunevTt54IKCAoF9IHz9161bhxC6fv26wWAYP348qY8QGjBgwL333gs5oDYOHz5MV6B1CckvKSnp0aNHx44d8/LySCYkrFbrzZs3c3Nz33rrLVgzRAg51VKzZ882GAz08mNWVpbBYCB6SPCkCKGnn376vffeE9xdv9QZcC8DupZyL5+6NM0wQLQUQmjgwIHvvvuuv7+/h4dHfn4+raXeeOMNwY4OXA4aNAgelWGYgQMH0o9tMBi+/vprkgN7Qr/++muVQjp+/LjBYACNRSqMGTPGYDCAoQQopPT0dFJalZDUUh999FHr1q1hQY9U9vf379mzZ8uWLQnmTp06QalTLfXVV1/dcccdZCUTWrVv337IkCGQFj+pt7f3Sy+9RO6uJ3QG1GBA11JqsKrL1AADtJbauHGjp6fnP//5z8GDBwN0si81YMCA+++/P1T0i42NFdQkz2wwGL799ltyCVpq/vz5rmup3Nxc0rwqIdZSixcvNhgMmzdvpqtt2rQJdpI2btwYEhISGhr68ssvMwwDdVzUUrS5I0JIoKUEdo/e/I/GoKd1BtzOgK6l3E6pLlAbDNBaqri4uHXr1gaDYevWrYCeaKnRo0c3b94cdq0kH4zUJKUOtJTkit/AgQMFK36OtVRkZGSLFi3GjBlD7giJwYMHd+7c2Wazkfy+ffsSLZWXl0dM+0gFWv+JV/xu3LghWPHTtRShTk80GAO6lmowqvUbNS0GaC2FEPL39582bRrRRkT3REREGAyGCRMm0OgrKipgYZBeGyQVHGgphFDv3r0feOAB0jwuLk5sPeFAS12/fv3BBx986aWXBJMehND777/fuXNnYkwYHR3drFkzoqVKS0sNBsMPP/xAcFYlaC0F+2dffvklqTB+/HiB9YSupQg5eqLBGNC1VINRrd+oaTEg0FICcERLIYS++uorg8HwxhtvLFq0aNmyZT/88MPDDz+8fft2aELXhBzHWgos0Z988sn58+fPmDHjvvvuu+eeey5fvgxtQW040FJDhgxp3rz5woULKXP0TWBWvn79eoPBMGjQoFWrVvn4+LRv3/6pp54iWgoh1L179wcffHD58uVbtmyJi4uruiOtpYgl+gcffLB8+XKwShdYoutaCrpJ/29DMqBrqYZkW79XE2LAdS2FEFq9enWfPn1at25999139+zZc/z48devX4eHqauWQggdPHjw+eefb926ddu2bd955x3arM6plvL29iaWESQBp3ptNtvs2bMZhvH09HzmmWc4jhsxYgStpaKiovr06ePh4UGW/gRaqqKiYvr06Z06dWrZsuUjjzwieaqX7kJ9X4pmQ0+rxICupVQiVherM6AzoDOgM+AGBnQt5QYSdRE6AzoDOgM6AyoxoGsplYjVxeoM6AzoDOgMuIEBXUu5gURdhM6AzoDOgM6ASgzoWkolYnWxOgM6AzoDOgNuYEDXUm4gURehM6AzoDOgM6ASA7qWUolYXazOgM6AzoDOgBsY0KSWslqtGRkZBQUFhfpPZ0BnQGdAZ0DLDBQUFGRkZBCfKWK1prqWWrZsGRwz9PLyOnHihBgBQig/P3/06NEPPvigh4dHly5dgoKCJKuRzIyMDHKeUU/oDOgM6AzoDGidgYyMDPKFFyTU1VKBgYEeHh7r16+Pj48fNWpU+/bts7OzBQjMZvPf//73N9988+jRo2lpaREREcTbtKAmuYQQPhkZGUr+gMjLywsICMjLy1MipAHaagVnYWGhDtXt40Gn1O2Uamig3ia9D7OOgoIC8oUXJNTVUl5eXiSEgdVqffjhh+fMmSNAsGLFis6dOwui2gjqCC4LCwsNBkNhYaEgv06XFovFaDTW6b51ku+uylrBiRDSobqr04kcnVJChRsTWmFVKzgVvvtOv+cqaimz2dy8efNdu3aR4TV8+HASO45kvvHGG8OGDRs1atT999//1FNPzZo1C8JjkwrihNOnEjcR52hlBGgFp8KRKu4gVXO0wqpWcOq9r8ZwvU163+n3XEUtlZmZaTAYoqKiSP+NGzfOy8uLXEKia9eunp6en3/++enTpwMDA//yl79MmzZNUAchVF5eTub+MEPMy8uzKPiZTCaj0WgymRTIaIimWsFpsVh0qG4fEDqlbqdUQwP1Nul9iHzmYG2s8bVUly5dHnnkETJ/WrBgwYMPPijWUuC8md4hDAgIMOo/nQGdAZ0BnQEtMxAQEOB4B0dFLeXiit+LL774yiuvELUUHBxcFXTObDaTHEi4PpcqKysrLi4ucva7desWx3G3bt1yVrGRy7WCs6ioSIfq9rEioLS4uLi8vFyN6YVymbfJH/7KiXJdwm1CaWPOpRBCXl5e3333HagZq9XaoUMHsfXEhAkTGIYhxvKLFy9+6KGHBCpKcCm3jmmz2a5fv57g2i8+Pv706dPx8fGuVW+0WlrBmZCQoEN1+ygRU5qYmCj+G07wgjTK5W2yidKQ3N4mlMp9zwnVKs6lEEKBgYGenp7+/v4JCQlffvll+/btb9y4gRD69NNPfXx8AER6evrdd9/93XffJSUlcRx3//33//zzzwSfZELuqUBF5eXllZaWljn7mUym7Oxsk8nkrGIjl2sFZ1lZmQ7V7WNFQKnJZEpJSbly5YrNZpN8NRox8zb5pDYkw7cJpXLfc0K1uloKIbR06dJHH33Uw8PDy8srOjoabuzt7T1ixAgCIioq6h//+Ienp2fnzp3rbeNXWVmZkJCQl5dHxDpOWK3W/Px8ModzXLkRS7WCEyGkQ3X7OBFTWlBQkJCQ0ARPUNwmn1S3d7EDgbcJpY2vpRz0Qb2LJJ+qrKwsISGhtLTURbHi99/Fhg1cTSs4dS2lxsAQ935paWlCQkJZWZkat1Mi8zb5pCqhqK5tbxNKJb/nNFeqz6Xom7krLflUoKVcf3vF77+74LlXjlZw6lrKvf0O0sS9X9dxrgYqSZm3ySdV8tlVyrxNKJX8ntOU6lrKStPRBNPi71QTBAmQdKhu7xoxpbqWUk6yVr7+WsGp8Ey3rqVkh7T4/Zet2qgFWsGpxlzK29v7hx9+cAv9BoOBdoPS8KympaUZDIazZ8/W6XHEOHUtVScCJStr5euvFZy6lpIYZpK6t65vr/j9l7hTE8hSjlPwgVbvmZRDFWC7efNmUVGRILN+lwISBFDDw8PpA+OQPnnyZP3uJdmqsrIyKyuroqJCslQuU4ATIVTXcS4n2e35t8kn1e28ORB4m1Aq+T2nadFX/Jrcil9lZSVteSj+TtH950pa8IF2pUn96iiHSu7r9iNBAhIEUM1mcxb1++KLLzp16tQUrL0FOHUtRUaIkoRWvv4NhrPUXKmET30uJcGepO6t69+Y4vdf4k71zfL29v6W/7Vt2/avf/3r5MmT4ZNXXl7+448/Pvzww3feeaeXl1d4eDjcwc/Pr127drt37+7WrVvz5s3T0tLKy8vHjx//t7/9zcPDo1OnTqtXr4aacXFxAwcObNOmzf333//JJ5/k5uZCvre39/fffz9u3Lh77rnngQce8PX1hXyGYcgsgWEYyBT/19fXt1evXuvWrXvkkUfatGnzzTffVFZWzp0794EHHrjvvvvoE2z5+fkjR468995777777v79+5MwKyBh6dKlTiVcvXp10KBBbdq0ufvuu4cOHQpH6KoggYQ1a9Z07NixWbNmCCGy4iee65CTDEaj8ZlnnvH09OzUqdO0adPIZCU5OfmFF17w9PTs1q3bgQMHHGspmhCLxXLffffNmDGDzhSkob9CQkKefPLJNm3aDBgw4Pr161DHarVOnz69Q4cOHh4evXr12rdvH+TTK363bt36+OOP77333latWj3++OPr16+HOunp6UOHDm3Xrt0999wzaNCgtLQ08Sit6zgXIFfvssE+qcofQStQGwbnwgNJj08Mik3PV0KsEqiS33MazP+3cymbzWYyVzj4V1xmvp6dV1xmdlBHrsjpX9ne3t533XXXDz/8kJiYuHnz5jvvvBPUzBdffNG3b9/IyMjU1NT58+d7enomJycjhPz8/Fq2bNm3b99jx44lJiaaTKYPPvjgkUce2blzZ0pKitFoDAgIgHCR991334QJEy5evHjmzJnXXnutf//+0J3e3t5t27adNm1acnLyhg0bmjVrduDAAYRQTk6OwWDw8/PLysrKycmh+55O+/r63nXXXUOGDImPj9+zZ4+Hh8eAAQO+//77xMTE9evXGwwGctbt1Vdffeedd06dOpWcnPzjjz/+9a9/vXnzZpUokDB48OC4uDgHEqxWa+/evfv163f69Ono6Og+ffp4e3sDEl9f3zZt2gwcOPDMmTPnzp2jtRQ91wkLC2vVqtW6desQQpGRkW3btvX397906dKBAwc6duwIroqtVmuPHj1eeeWV2NjYw4cPP/PMM65rqT///POOO+5wEJON9Nerr7566tSpmJiYbt26ffzxx/AUCxcubNu27ZYtWxITE8ePH9+yZUvoYlpLffvtt7179z516lRaWlpoaOiePXvgr9Fu3bp9/vnn58+fT0hI+Pjjj7t27VpWViY41adrKXrc1i+t5JNavzvWr1XD4PxkbTTDchui0uoHElopgXr7aimTuYJhOZX+mcxOdhe8vb27detGlBnLst26dbt69Wrz5s0zMzPJaHjllVcmTJgAXz2DwUDmJUlJSQaDITQ0VGCSMHPmzNdff500B9/wSUlJ8EHv168fKXruuedYloVLwQea1KETvr6+d955J9kEGjBgQMeOHcnCY5XfenBtdeTIkbZt25aXl5O2jz322KpVq6ouQUJ6ejq0kpNw4MCB5s2bp6eng4T4+HiDwQA7QL6+vi1btqRVKZlLkdvl5eV17tx59OjRkPPKK6/Mnj2blG7atAnca+3fv79FixaE6n379glIEM9RiJA3+B+5lEz4+fkZDIbU1FQoXb58edX8FdIPP/zwrFmzSKvnnnsO0NJa6p133vnPf/5D6kBi06ZNXbt2JWPGbDa3bt163759upYSEKX8UsknVfndXZfQMDiHroxiWG7pIfzncr1/SqDqWkoVReWKlqI/Q0ajsUWLFhzHGQyGNtSvRYsWH3zwAWgpDw8P8oXaunVr8+bNwb8A/T0dMmRIy5YtKQFtDAZDcHAwaCny7UYIDRo0iAAQfKAlB6Kvr2/37t1J0fDhw998801y+eKLL44dOxYhtGzZsjvuuIMGcMcdd4wfP76qJkggn1Q5Cb/99lvHjh2JZIRQ+/btN2zYABIef/xxukigpSwWi7e3d//+/cmyHiyaETytWrUyGAwmk2nx4sWdOnUioiC4sys2fhkZGXfccceff/5J2kom/Pz87rzzTlK0c+dOWKKE9y0iIoIUjRkzBua7tJYKDg5u3bp1r169xo0bd+zYMaj8008/NW/enDxLmzZtmjVrtmzZMkIpVNPnUoTbeieUfFLrfdN6NGwYnIOWHWVYblZQQj0QkiZKoN6+WqrRV/yIkkAIgZYKDAxs3rx5YmJiCvXLysoCLdWuXTvS5Xv27JHUUgMHDnz//fep1jhZUlJCL46BkMGDB5OdGxe1VK9evQiAESNGDB48mFwSbfHLL7906NBBAAD2xmBXiXxS5SQ41lI0BvFDjRw5snPnzrQTrFatWs2dO1eAx2q11ltLzZgx47777nPqfwj2pQg/u3btMhjw4rmLWgpWYv39/YcNG9aqVasff/wRIfT11197eXkJnuXWrVuEUridrqUI7fVOKPmk1vum9WjYMDgHLo5kWG78drzGXu+fEqi3r5ZySjc9R3Faua4VvL296amJj49Pt27dYB0vMjJSLE3w1UtLS2vWrJl4xW/ixIldu3YlMwlaDlEkkElrqZYtWzqdHICOIQLldAys16WlSSxhu6ilJFf8Tp06VXVrAQaBllqwYMHdd98dFxdHQCKE+vbt+/nnn9M5kIYVP2LREBISIlDVkr1vs9k6deoEOkMsk84R9BfRUggh8Yrft99+ixCi51K0qJUrV959991VDVevXn3PPfcIYsGJcepaimavfmkln9T63bF+rRoGZ/9fwxmW+3rTaTHIfJN5d2xmRaVzQ2glUHUtJWbeniN+/2Wr1r0ArCfGjh2bmJgYEBDQpk2blStXIoSGDRvWsWPHHTt2XL58+cSJE7Nnz+Y4TjyXQgh99tlnjzzyyK5du1JTU/fu3btlyxaEUGZm5n333TdkyJCTJ0+mpqaGhIR89tlnEEDSgZbq0qXLN998k5WVdevWLblHEWgIOS1ls9n69evXq1ev/fv3p6WlHTt2bOLEibSOIX/4O5DQu3fvF154ISYm5sSJEwLrCbm5VGhoaPPmzVeuXEnMxQsKChBCISEhLVq0mDZt2oULFxISErZs2TJp0iTYzOvevftrr70WGxsbGRnZp08fV7TUwYMHDQbDxYsX5Vgi+Q601KJFi9q2bRsYGJiYmMiyrKT1xJQpU4xGY0pKyoULF95++20IYG0ymbp06fLSSy9FRkZevnw5PDz8+++/v3r1KqEU7q5rKdIL9U4o+aTW+6b1aNgwOPvOOcSw3MdrjosRjt9+jmG5NZGXxEWCHCVQdS0lILPmUm0tNXr06K+//rpt27b33HPPxIkTYc/JYrFMnTq1Y8eOLVu2fOihh957773z589LaqmysrKxY8c+9NBDHh4enTt3Xrt2LUBPTk5+77332rdv37p16yeffHLMmDEg2YGW2rNnz+OPP96iRQunluiEHTkdgxAqKir6/vvvq2YMLVu2fOSRR4YNGwamEC7OpRBCji3RCQZ6LiWO1EzWM0NCQvr27du6deu2bdt6eXkRk/2kpKR+/fp5eHg88cQTLs6lPvroo759+9IA5NIOtJTVap02bVqHDh1atmwpZ4k+c+bMbt26tW7d+i9/+cvgwYMvX74MN8rKyho+fPi9994L8QFGjRqVz/+IGYt+XkquR+qUr+STWqcbKazcMDj//nMow3JvLzkiRvvC3DCG5f7vd/vWqbgCyVECVddShEZhQm0t5S7XPqriFJKi7FqHqow/idZiSvW5lARNdcxS8kmt460UVW8YnD19QxiWe3FemABrbnE52Eh38uFulgiDpwsqK4GqaykBmTWX4ve/pkxxSjCzUSJPVZxKgInb6lDFnCjMEVOqaymFlCp0lKD87q5LUPLpd/0uT0wKZliu1/T9giah8TfISZ4/T2cISgWXSqDqWkpAZs2l+P2vKVOcarJaqnv37rStM6Q3b96s+ImxAFUpdQvCWbNmiR9/4MCBcsLBzYegCX0cSq6hu/LFlOpaSjm3Sj6pyu/uuoQGwGmz2Tr64OM6nXw4chIGEM4LuciwXOcJQQzLfbNZwraCfhAlUHUtRTNZKy1+/2sVN5kL9+K8cuWKwNY5JSWFHOZV+NDuhaoQjGTzmzdvwuMnJSXFxMQkJSWlpKRcu3ZNsjJC6Nq1a2K6wNeGXBP35osp1bWUcoaVfFKV3911CfXDeTLt5kj/U8k3XHLTbK6wkglTcXktZwUfrT7OsBz7JzageGpqiLnCkaVf/aACFbqWkh0S4vdftmqjFmgFpybmUqQntcKqGKeupUgn1juh5JNap5vabLYft8XO3BsPrc5cvTV0ZdT5DGye6spPEufv4alfbDhlkbEOv5JX8vS0/QzL/bLPuakqNoYqsxAtlZlfE+i80mp7airer4rPLOwzE5tXHEm2uwyVRC4JVbKmOFPXUmJO7Dni91+2aqMWaAWnrqXUGCbi3te1lHKelXxS63T3G4VloANK+GmKz47zDMv9zNmVllNRkjh7TcdK6FyGhHNYk7liwKLDcMdJu7DxsNNfXrWJBMNyF7MKSf3ErCKG5bpP2VdptY3bHsuwnO/uC6RUnJCEKq4mmXPbaSmTySRJhDhT/P6L6zSFHK3g1LWUGqNF3PulpaUJCQllZWVq3E6JTCXfKSX3rUfbBoOakl0MOgPW3z5dd4JhOZ8drnp5EOMsr6gEgcdSJGY2E3ZiLQj/fthyxhVmMvNLSZMTl7HbaPhtOXGVYbmPVuNDVH+ezpA7UFVdHYmhkiKnidtIS1mtVnA+VFBQUFpaWubsZzKZsrOzTSaTs4qNXK4VnGVlZTpUt48VAaWlpaVXr169ePEiHOV2+v43ZAUl36mGxNmQNn6x6fmgA8ISs6tc+L/Me3n4LsAl/SGJkyiVkAvYs5rg9+I8fLzps/VYF470x/5cnP4u55YQLXUg/gapD+d554XgZcODCdjY752lEgeqSH0lvX8baSmEkNlsvnLlSoJrv/j4+NOnT8fHx7tWvdFqaQVnQkKCDtXto0RM6cWLF4uLi8nXoekklHynGvgpGgzq0ZRc0AGbo6/YbLauk7HN93/8XI0BLcZ5LsOu9iRNw2ExcMH+RIblPlwl4UtCzPPFrEKipWiZry/EK4egt06m3WRY7qX59mB4YiGSClWymmTm7aWlEEI2m81isbjyJ21RURHHcUVFRa5UbsQ6WsFZVlamQ3X7OBFT2gRnUfDpEX9SJT9JTSGzwaCGXMgCHTB338WbJWZID10Z5SIJYpwwrZEMB4W9UPI25QH8Yp2kLwnxfYnaY1hu/VG7D5SqVb4evOlEag7+ewg02bMzcLw6wa+kvOJYSm5BqUUMVVDTweVtp6UccCEoUkKrQJSql1rBqfDvKVU5FAvXCqtawan3vniMIYR2xOAdHYbl/rvlzPmMAki/+ZuEv2nJ5uLeh+0ihuWWhaUImhSX2yPqRSTlMCzX3+HUh7SFeRIAWxxqDzFltdoPUeUU4Uhy1/i9qy4TcYQgwe9IMp4svjgvTAxVUNPBpa6lZMlRQqusUBUKtIJT/06p0PmKNqXVwONApj5QxeRsiEoDBTBkxbF9cdch7S3yRSRuCDliSpceSgYhYkNz2LLqMjE47hpWh8/9jGOoOv2BmgGZM6ot5gurzdPLLJW0tTpc0jJ/D09lWG705hgxVLqa47SupWT5UUKrrFAVCrSCU9dSKnS+rqXUILXhWF0engIK4F+zD66JvATpPjNd0h+SL9RUYxwImWKsFcKGrMv1mXngSh42iOg2ZZ8r3B26WOMG6cdtsdBEMHkiU6vsIqFl6ejNMQzL/R6equQzpWsp2Z5SQqusUBUKtIJT8qVSgQ/3iNQKq1rBqfe+5Licuw87GQL/Q1OqFcyTk13SH5KUfrP5NAgcG3hWcMcTl7GNQ//54cRLrNVqE9QRXwaft8/wGJYbtcFuFijeiOrBe6SFbSpaCDhNP5Kcq2Sg6lqKprRWWgmttQSpfKEVnJIvlYzmpi8AACAASURBVMrc1F+8VljVCk699yXHIpn6MCz31hIcEhf+VbqgPyQpHboiCiR8Ua1RyH0P8M5hBy07Wmaxn6kqKrOQUrnErjPXCKp/r7KbdcBmFb0yCTGozqbXOkpcYLL7rcg3mZUMVF1LyfVOw836ZRG4VqCk+127g9tq6VDdRmW1IJ3Saibc+X+a1ZTsIuPZawJHq+662f+2Yq8N8A8M8CBd6IL+kNRSL83HcXXJeVsaJ5y9/WRttM1me4x3EXu9oMbjEV2TTm89mc6wHNR/Y7HdrAOWAWkrQXBpcTgph24Le1ovzMUhP2hK6TqupHUtJcuSElplhapQoBWcCkeqCsw5EqkVVrWC0729vzs2M/j8dUf9R5XZbDa/o5ejUvOoPCdJmtWhK/HshD7Q6qRxXYq/3HiKaCk64Yr+kKQUfOtJnrFdf/QyNmT4IwYhBK78UrKdO5zdyNt3/HP2QYblnv/lEDyc8SyeYIHjCcgBlrhztTqFmE5IQnWdJ11LyXJFj1TZSk2gQCs4FY7UBmZaK6xqBacbe7+kvKLzhKAuE4PLK7CBmdMfeJz75+yDTmuSCjSr4K/hp2rDAVLHLYlha6IZloMATqClukzEB3tddFhO40QIlZrtS3mShuaLQ7H534Sd2H2f5AKd5BOBTcfbS44wLNfTNwTqbDx+hWG5LzfWeK8Y6X+y6kECTlylhRDTCYW9r2spmtVaacEIqFXWlC60glPhSG1gyrXCqlZwurH3iYfWvGJ8WMfp71i1f4eCUufbMCCNZrXPzAMMy/WZecAVWwOnYAQVBi87Su9I9ZgaAvrjzNVbgpqSlzROhFD6TROZkIkNzWfsjWdYbk4w9mkEniMcezGHOy4Lw1aI//HDSqijDwckgGkiMflDCI0JPMuw3KrDqTROYjqhsPd1LUWzWistGAG1yprShVZwKhypDUy5VljVCk439n5atWe5tNwSV0YF8e9w+opLn34BVPBaVLWGFuOa5nAFEqnzyoIIhuXGbsWfeIblBiw6/NpCnOOK/hDgRAidvnKLhCUUG5r/tA3vgS0Px6d93//9GMNy++KEvv6u5pn+u+UMrSMXHEhiWA68njMsB5r+F940cdqeGifoYAby6/5E8mi06YQYKqnmSkLXUrIsaeX91wpOhSNVtp/UKdAKq1rB6cbeJ57l4q65FIcJrAaqTghtPZnu4mAhrFZU1sQABM+qLkpwsRrs96yMwEdfGZb73O/ku8vx7EqsPyQFEpxQCvoYligZlhMYCo7agPfANh2/ghAC5+vbRWHgQSf1nr4//aY9dsTs4ASG5WbsjYdlSciftAv7Vl9wIImgmh+CfQNOpQ5p0aYTCntf11KEZ2FCMAKExU3mWis4FY7UBuZbK6xqBacbe//MVTxjYFjORYMI/2N2/w6zghJcHEWEVeJkASY6LjZ3vRocMwpLzIYnmrwr7pO1eKeK9uvqQBrBCXU28dtFw/nwH1VTNIGh4L9XYTOQ3bGZCCE4VuV/LE0gfPIu+6HgNxZHlprxtt+0PRcYlpu77+Lff8ahDi9k4r8M/rvlDMNyayIvkeagaOlDWrTphMLe17UU4VmYEIwAYXGTudYKToUjtYH51gqrWsHpxt4/lmr3I+6i3R3srNTP1/j1AhxdqZMP/lcV7SLjlqvR6VwZrsT9a8YtU0de/oqI1K834WO5G6Kw/hgbePbrTacdGMELen8hvzrns+M8mGDQoXURQm8sxuexInhjcVjBE/v6+z4Aqx/4B2ESISTV4tBkiCoCfxnANlXgyRpbCfBgS0cDAXW7+rBdkwmgusIPqaNrKUKFMKGEVqEsNa+1gtON3yk16bTL1gqrWsHpxt4Pu2ifeeyIyaBHgtzXfE6w3b8DHNyhm5D0zL3xry88XMwHzKWhpmTjiLRPT9sPp2VBeZBWriRyisqXHEz2nhdGzzOgoclsd/9aUl7hNQvPVPbEZv5YvXtEpnG58kYigt6fyAc5XHAg6ZkZ2OIj6UYtQ/PnfznEsBzsOcEMSezrb0R16CmG5frNxXbncKLr9/DU96ilyCEr8LZWEHUYYO+5zKro8sSbe3F5BWhKYuwugOoKdaSOrqUIFcKEElqFstS81gpO+uVXkw/3yNYKq1rB6cbeD6r22UPrjJyi8ud+DqX388k4gE0UMFETu0OFas/yn3Vis0BYPctHKfzX7IMwTRm/3dUouiA27lrBk5P3wdSkyyShy/DsIhxOvqMPZ7PZftoW233Kvmv5pb678QrbvJCLJPzgZXkjEYITbgc7TxuPX+k3FyskgbVIT8qJ0a98iCmxrz/YFYPZ5zN8JI7v+NnVuiOXQYFtPYX39uAMb2RyzRle8LM+sPrYL3jOfXFeGPnTQQCV9I4rCV1LybKkhFZZoSoUaAWnG79TKrAoFKkVVrWC0429T6Jd0AtWEFeph28I+SySHv2B30QBVRGfWUjySaKi0goLbtv4TzANFaIUvrogYucZHGLDxciBRDIEtIW1MoblBDryUg4OJ99jKj6EZLPZoHReCJ75+e6+cIoPLciw3PkMWSMRQe9XW15cH8gv7tGeIIhDWIi1Yd9G2ir09defDxYMDINaBc23OfoKLAbCXpT4uBVsFvadYz/2CzPC6XviCRUCqCTflYSupWRZUkKrrFAVCrSCk375VaDBzSK1wqpWcLqx9zdH4yOl5OgPdDxRXWLzdDhwCk3AdkAwVnKKyqF0yUF7/CTC6n4+SuHgZUdPX8GuWslXWCBB8tJqtfWZidfxIpNzOvMeibIKarkMh4BS/5hV67gxOYq0L84eIPFYSq6kfDGl1VOom7A+Sa/IFZTaXeqBLgQ7C+I9lsgHwFGpeUCIpdIKthjbTqXTdn1i37Ip2VjjwrHfSqsN5qY0ckIpuZfrCV1LyXKlhFZZoSoUaAWn+KVSgQy3idQKq1rB6cbeJxEuYHsfutyPd/8DWzuCQQC2bRBbljaeJtUSrtuDpoNfBhoqTKGGrYmG1blOPpy5wkoaSiYKTJYbhVgbxfKrhU9NDTFXWOGrnZhVa6MIzEBeWRBBy4GIU19vOk2UsQOrdLr3SUD6K3kln/HbS7TlPRz4faJ61RF8yH68plZQeZvNBptJENqDYbl8k/nDVcfBMhDct/vuvmCz2UMg0nE6CD9Wqy2GN8Ls4RtiqazhioZKP68raV1LybKkhFZZoSoUaAUn/fKrQIObRWqFVa3gdGPvE5u9H7acIb0O7n8YlpstMjcHX+Pgi2j0ZuzFTvCLTMaxa+G4EhQRVokrIKIDHOwSwcLdawsjnpy8L/lG0QJ+7+ebzacRQv15J7DRl2r5EiROymk8xCfsbwft8QzFp5pIfYKTDkVYUl5BNpNIzQuZtSIfhlb7RycVaAdL2PZhEnbUdC2/FIwmQi5kreAPdY0JPEti/tILmLSfdVi0/JZ3GEjk01BJposJXUvJEqWEVlmhKhRoBacbv1MqsCgUqRVWtYLTjb0PB0ir7MJH+p8k3TZ9D3b/I3CBCqXe88IYloNWry88TJqQBEyYGJYjPr8Jq/BpBvM88AoBltykrSAB8XBhBwssv+HkE+wY7b9Qy9eD5IQGFvreW34UzCgYllt/9LLgLuQScJrN5i0nroJFQ3c+tqHPjnMMy/1WvYBZ5RIJFvHIvA0uX/41nIhCCIHrqc4Tgmw2G7ESfPM3bL8enpgNa6ofrjoO1vmPT8TVSHObzfb4xKCqGFSZ+aXwZ8HOM7UsMAmlpInrCV1LyXKlhFZZoSoUaAWnG79TKrAoFKkVVrWC0429P5P3R0fbPSOEYLseGyOIDCjAER/MHrpMDPY/lrbrzDV6MWr1YXuQXLBqo6GCLdzkXTjuLexvbeR9NwiHS/U18cYEKrOjDwfOBmkDuYJSS9D562WWSjJRq26N/w8uG15fePjbP3CUW4GyoWsSnAHR9mPLT07e93s49qT3M4d1Nn2KGYBVKT+QILkllnQDm933nr4fIQRbXDFXb73K+3A6lpobfQlvVr04LwyqEa4IJFjVvJBZAPOwq3m1zpYpGai6liIkCxNKaBXKUvNaKzjJS2WxuOr0U03anMjWCqtawSno/VJz5daT6WBv5qQnRMVwKoie+iCEvuDd/8BnXWBAQVz7dJtiNwpnWG4L5b0bnABBW1jFIqzCuSLw0ArTNfrTL4KGQKt1r77R+78fgzpgZwgGcuD1dUNUmn2iVtvQjhjLwYaQQNkI7gg4P1mDt47YP88Rd7q0+3NosvUUDhM1Yv0JuAQzdzAvJDLBqhBiGxJbc/AYe/rKrYxb2JVtl4nBEALxxXk4ahT9gznr9tPYGPKJScEC/0yEUrqJi2ldS8kSpYRWWaEqFGgFp+A7pQIT7hSpFVa1glPQ+2Bj5ru7xl2p651HfLPSp3QhvhFomj28EyAQaK6wO+IrKLXsic0c/UcMrEfRuyZ0KMIrediDLWEVTMmXHsK2f2Cg8dVGvM8k94M505rIS/Bxh5lN1X4VOGOdH4KdsX7O+xefsPO85KGlZH5C8/S0/bDAyLCczw58SKug1LIm8hLYZZC7WyyWwB1GWGpLyS4m+WBg8n1Azb6dIIc+qkVagTX/O0uPUO5or8Nx47hrBZZKKzjggLiIby2xR0QkzSG6B/iqIAenSCmhlOS4ntC1lCxXSmiVFapCgVZw0i+/CjS4WaRWWNUKTkHvzw7CPkxpVeF6/0HUoirDM3rRCf72B0ertAHFrRIzqC7yp/1xfuXKa1Yo2VYhju+qnPUd5w0cCKuj+WU32BmCALVk70oM2GazW2CfuXoLrCfAFR5CCLy4wsrhoKU4VtMna6Nh52nuPhxKg/xgZ+uxCUGwM0TiFsIJJ9BYpLLFYpm+bnfVvpE3dX4WIRR48qrAIxQNACFEu70g0mDnqQoYcUf75+mMXtP3k3hX/+JjIcL8Unx07GN+SgdHtb6jFCTIJ5SS27meaBJaatmyZQzDeHp6enl5nThhn5PSz+Dn52egfp6ennSpOO30qcRNxDlKaBVLUy9HKzgF3yn1CHGLZK2wqhWcgt6HbSTxeR1X+g6cyDEsR2/gwwcU/pCnY8iCBTYdxqLMUgmTD5g2IYTAQAAmCrvOXKOh2veTeGfq4C2px1SJg8MAGxRM5wlBtPEbFK07guPkwrcbjsS+OC8MQmnQZ5NhzgRqlfx3+Dr8SQSV9mpts3WLxTL01z3gs5ymjjt3XbBvR1xaQDXiQpCenMFkEcwgwR3thqg0WCaFTSZwjARLkeK++2oj9kAINC4OtZ88I6iUDFSn33MDuY1KicDAQA8Pj/Xr18fHx48aNap9+/bZ2dmCe/n5+bVt2zar+nfjxg1BBcGl06cS1Je8VEKrpECVMrWCk375VaLCjWK1wqpWcAp6HzQN2SapU8d9tBpvw8A/og8gkjo4T6INKOIz8VkoQUjA/+OjK4GzH4QQLGqBmQCs0RFW4bsMgdKJsfXNErMkYLBQGLBIwoyQ2JfjM0m8kfdjE4LArQPt5wkhVGm1kaeDBJg8ED+wZPMJIVRWbn5q0l6G5Y6l1jr5G847WaenfbBMujKiJkohOZm7JvIS5MNuls8OHMyXBKOC88igzMAVem9+dvW/rbECEqAJYBaElhf0vqCh00un33PVtZSXl9e3334LQK1W68MPPzxnzhwBbj8/v3bt2gkyHVw6fSoHbUkRGakkp2kmtIJT4UhtYPK1wqpWcAp6fxAfo5ae9Ljev2DVDV9D8MRKokBlF5aBrQTZpDlxGfuM6F/b5BqC+EGQeJvN9hjvGAKizUKEJMIqrF+FJ9r/bgZ9djY9XxIt7DON2y78fFeZzMGWz6ClR4rK7D4gIOi7ZJAO4voPnvG1hfjYL7gYB08W5O4nLuGTXoLzs3w4RPzU9L7dSH8cXIqO+A6hrWAZk2G5WyVmMA8BUxGYexFHvbd4xQwHewGVeE8RrEKgVHB+WdD7BL+LCaffc3W1lNlsbt68+a5duwjc4cOHDxo0iFxCws/Pr3nz5o8++ujf/va3QYMGXbjgZNPV6VMJ5EtekpEqWdp0MrWCU+FIbWDCtcKqVnAKeh8Mnf+v2gRuZUTq6D9isnmXDU47Grag4GsIZ2xvVm8+VVRaP1iJoyiRzzF8hQfxFgFEMsRzAiu1fJN94wrsC2Ahi7AKG12n0m5CW5haSbpZQgjB8uBGPugGuRckwHzuxXlhxIcs+JmVDHgIborgARmW++ds7EIJjD4Ehum/BGGL89GbhAYdiVnYpvxZ3l0sAACfSfQUB+aOs/gNQobl4q4VEHfsCCE4mUtC9JrMFQgh4g4Dh0CkwvLCLRaF4qi+sOhXXoFjU9E/Qimd6WLa6fdcXS2VmZlpMBiioqII3HHjxnl5eZFLSERFRW3YsOHs2bMRERFvv/1227ZtMzJqHRlDCJWXlxdW/zIyMgwGQ15enkXBz2QyGY1Gk8mkQEZDNNUKTovFokN1+4DQKKWw2/HWb5FASK9peIv+H7NCz15x/s6CxTN8EM+k4fopWdixQvcp+ywWy1w+tuwPATEgecdpbEfw0aoomvmbRfZ4Thl5RQnXcEzFnr4h+85dq/Ij/vYSDImwCmeAzqffhOZjtuAzTL+FJtLSIG02m5+ZgZ/i5KUccSncpde0/VEp9rAjRAlFXMwS1H9xLj6GzLCcFx94sPtU/Fyw8caw3PB10VC/yFQG8P48eVkgIS0Hr3N2mRScnlf0/vKjW6LTXucD1Ycl1NwLjDjeXoIP7TIsx8VmjPQ7gUNbHbtksViWHsQq53M+h2G50rJyi8VyMB5vd8G/FeHJgpuuikiBIu95YYIimlJxkdOcvLw8g8FQWCjhKRgURJPQUrTSslgsjz322OTJk+nMqrSvry9lYIGTAQEBRv2nM6Az0MQY2LbDCJ+zf07bC9Ae98GbK/jDOmHvSmdv7dNT7JWxR4kNu41G42+bscBeU7C0X/yxzdszU+2Sx63Cl+/M3SPg4PkZWMjUtbvnb8AVnvPdu5QX0nOyvaEd2ARcbV2gvfXXS3HlfjP2vvTz3rd/2bNjF85ftMnYa8reJyfimp3Yvdt2CG6FLzduwwg7snun8SZ58LDw3982CesDNoblXpllf9Kdu4xdeCQMy3WbtHcXf98fV9qfdMdOoYSAP+0Mj1qCbSse99nblYe3ZHNNzddm46KOrP0WY1bsfvlnnJ66FlMKwl/iczr52DlZGWAXi49nrcLV6N+E1RgPw3Jv/SJkm65Wj3RAQEBjaikXV/wECmnIkCEffvihIFOfSzn9k6TRK5A/URsdiVMAWoGqFZz0X9NX+L/0wX7aYrGYzWYInAFziN/DhH+kCzrr6WkhsBlTNdsIPnfNYrEcSsB/47++MMJiseQXl8KG/5WcQovFsjwMzwnGBJ4RCJmy6zzDchN3nNsVg4+7/t/vR7NuYa/eHX24ktJyYLWwGOcwLJeVXwLNt5+0u2OHfLj7p2trrDk+XXtccCO4LC61u13/LTQR2pL/Jly7JWgC64oMy43bdhaqZd60I4FHS87KLykth9nVjyt3i9d7Ssvst3t1QTi5EcNyKVkF5F6jNpyki3yN5wfw861DCdctFsvWE/hJYdrajZ+kWiyWghIcuRj+7T6TTkRBgovF53kZlvt57wVBEd374iKnOY08l8IGNl5e3333Hagcq9XaoUMHsfUErZAqKyu7du06duxYOlOQdrqOKagvealkIVVSoEqZWsEp2JlQiQ13idUKq00Bp3gTQrIXCFRwz0MCYZCDt2C/8Ktow0MgDewjIGgTeMkTBIqFtSywKQeLBvFWP4Tpe2VBxHremfo3m09brXaP4Bm38CK/0WjMK8TeFqrmDcSS8HpBaU/fkOd/OQSHiEesP5F+0754GJWal5pTXEF5ARfABpuIsYFY8fydX8oD4YJwHgghYmq/ODQZjOYhbshjE4LA9+uOmAw4Wvv3mQe27zRKOnPpOhm7i4V/YB3OsBxtnUifZa7SxF9tPA1qL5a3DQF3ghA4EXwmweOAgV+V6QcdvAqKwMV71dIrCdNFM0B6n850Me30e67uih8+gBYY6Onp6e/vn5CQ8OWXX7Zv3x4MzT/99FMfHx94jOnTp+/fv//SpUsxMTEffvhhq1at4uNr4muJH9XpU4mbiHOU0CqWpl6OVnDqWkqNMdDovb/1ZPpjE4IOxDs5HEL3PthJMyzXZ2Yo7cwbvM+JNQrNm7XaUBusJPyPpZFd/S82nIKa4OgPwnCArRo4faDlFJgs8O0GkwGIWgs2HafS8C6U0Wi8mottEB7jva+Stjb+l5ZbAhMv0DrD1uCTsI5/YFMHGvTLjdjiDv4VlQkdhn3Hh8etWqnbHH0FzvaCP9y//xwKj/bawgjQH7+HJRuN0lqKmGA8/8shOENdNXekvRcCMwTG20uOgGMnMEghfuL5/cKaCFjEiAMi09OPHHcN7w6SoPV0Ed37gnxXLp1+z1XXUgihpUuXPvroox4eHl5eXtHR9v729vYeMWIEPMOYMWOgwgMPPPDmm2+eOVPj+UPyIZ0+lWQrQWajv/8CPHKXWsGpcKTKPb5K+VphtdFxfr0Jn+V07N0O+ohAhcNDJGheXrF9eYp2QC7XraXmSvgUgjNWOBILkQPJCR4IXQj+v0EJEU9FtFiwhoe4U3AKFQzh9sRmAlQweXh6Gva+Kv6BqwUAQ9vOiWtCDpgmwrkucCQBba3WGs/iUBM8moP5H3haAvO51xcehlkjNByw6HB+camclnqJjxUCDpZKyiuGrDhGO0wiVnxV7jYAGKg9Mt+CGFFwI9qiHc544cVDyiETwL6aZ596For0rsJ33+n3vCG0lFy/1jvf6VO5Ipm8VK5UbsQ6WsGpcKQ2MMNaYbXRccKJovHbsa85xz8Clfggh6B84LKhy8TggBPYHm+kv31KJCmNGJ3DVGB2cAJCCI71kPjlxCtSXnE5OESQdGQOxtbwId4cfQUhBCdnVx1OBainLueSZUkxGKIw+sw84DQ0IkIIJn9wu+Dz18FTOwTaEAgnTt9Ppd0EvxgwY/v3qqji8op/r4r6ZG30vrisikoroVQggbZc33suU1yKEALVzrDcwgN2C3LABvMtcHwOObTDC3CPxLCc+NhApdX2ydpoyeNiCt99p99zXUsJ5+OSXd6ImQ5GaiOikry1DlWSFiWZjUupzWYDm3LHPljhAQlUcla0ow9ns9mIf27B9pIkLdfy8QZ+l4nBML2YuBM7SvDZgU0hFoUmkSZwGCjkQhYEP4Q9KlIKCbKPUnUmKYSP/ATAfHfjzX+j0RhxEcd0p7/RtAQSgReOwdJFkml6le/E5ZtwNlngFAMaErVxObcEAg2DswwIqEgLJ5TSmZCGhmRuJK6wMcoe7+PE5ZswoQRrfqgJPIOWoh3LkkDJxEWhWLJkjgOokvXpTF1L0WzUSiuhtZYglS+0glPh31MqsygUrxVWGxcnuNaGoH9CBkXXBCo5Kwo7JeQIakQS9qQgdqdNS0rNwdZuPX1D4HP5Xz5cL+0TFiqDacDi0GRY1guV2jYrs1QSE4PTV24hhMBT++d+JwFqUCw+QTV4mT0mEw0D0ntiM0f6nwT/F+JSQQ791JdyiiGWR//5teIQQhMy1ywss4DPCLC2IAHviWRCKckhCWj45m9Cz+WkAviWxTaN5RXkoDScIMbuBE01PjJIVCqEEFidCPbqiEwHCQdQHbSCIl1LyVKkhFZZoSoUaAWnrqVU6PyaGBNqCHcqE/wPCUI9ybUiAxXCCcLf6SXlFecy8hmW+9fsgxBa6flfDslJQAjBFv1zP4eC2+/P/XC4XvAetCOm5qT/qsOp2CnD5hgwBQRP52KxxO0QeJ4FNfn6wsMAdRtvd+6KZYRYsjiHeHlgWK6wzALx5iFMhqAyrHx2mRhss9lgrQ+4EtuAEEoFEqpsUkBPw4qouJQ4bYLdO2JVSJwQWirtEU8Ef4IA/31mHpCU6SDTAVQHraBI11KyFCmhVVaoCgVawalrKRU6v5G1FMTWc7B5Qz8yGai0I75bJWaIqvfS/PCUbPs8iW6VcL3QeBb7KYcf2GS/OC/M7vZ7BXZb8w4fC4OeMIEZ4SsLIkh4pGoBtf5PrBhKyrEHIJiodZ+yz2w2G41Gv6M4hu+XGx3tk9US5/BiWZjdNQOoH5jKfMr7Oxe02x2bSSzrIDAVaKm1R4Sh5QmlAgmgzn/aFku7PBfUKbNUTtp1HlwUTt4VB7cYurLGDRB4xaVjJ1YFyrLZbHOCLwqixQskS146gCpZn87UtRTNRq20ElprCVL5Qis4dS2lxkBo3N4nBgiCqK+ST0qggt0afBZvFJYdTcFGCq8vPAzrh518ONrsDRaj4q4VgEwScP0wtTwIh09PVnvbQwhdL8DbV50nBMHhKhKkQwAs4Tr2JNSLj6GOXYxb7AaEN/JLjEbjskPYrGBs7Vi6AgmuX8JyIswaEUKl5spZQQlie26EEBACK42EYYblxLqBUOo6DMmaYF0pMF0hAa7EQTokhTjOVAJV11Ky3CqhVVaoCgVawalrKRU6v5HnUrAhBPqGhBmUe0wyUMEaG1pdzTOBQ9h3lh4hSoI+QgSHeEj43QPxN6pO5AxadhRMpfvNxcuDcNSU9sNts9kgMgXcJa+4XA7VjpiMyOQcUvocf942Ji3XaDSCS0AIXUgq1Duxh58hYXdNtV3figVWVFqXh6eA53WiP6o8QRDX7KQJoZTk1C9BsBFr/qolQTg9RsJi1U8yaaUEqq6lCI3ChBJahbLUvNYKTl1LqTEKGrf3yRlPCP0AD1hQall/9PLPXDxx2QD5AJX4CgJnPynZRcHnsX+jISuO2Ww2cLWQmV8KTUrKK0DNrDpsD4wE39N/r4qCyOu9p++32WwgSuDEAezioLmL3jEQQuDcYfeZdKPRONWITQddNOFz2rnknOx/+L00p/WhApmBMSx3LkMYMcRdvX/6Cna5y7DctD01bSQM2wAAIABJREFU4SaIScWP2yRikbiIn1RTAlXXUoRGYUIJrUJZal5rBaeupdQYBY3Y+3i+MhW71IN/abklCKH5IYkkPNKhi7UcUgBU4tAB7NYuZBYYz2JTuo/XHEcIgZNvMiu6koddPDAsRxxSwE7YiPUnYE3vsQlBJGKTwDYa4vaC2brrzEOgv+VhSdjj6lbszWjpIWHYWdel0TXBSIRhOYhrRRc5SAM5QEL6TZOgprt6/0ZhGdyCtuZ/n48Vib0d8ub+glvX9VIJVF1LybKthFZZoSoUaAWnrqVU6PzGXPHL5X1GdPTh4IBqbHo+MV8G77GCfRQYqGev5IEju+d/OQTedMAlHcwwBDtMYCvBsBzZGoFTPl9vOl1cPc0Ckwc6ujzw7H/Mfh7oGSrMktMugH2giTvOGY3GbzZhP0brjwptFpwKkaxAXDP8su+iZAXJzLCLNZE+inkTD7qau959q9U+i6Uf9tN1OJCHYIJF371OaSVQdS0lS7USWmWFqlCgFZy6llKh8xtTS4EK6TvnECwNHU7KAYcFT0/bD350aI8PNptt+u64D37dc+ACNmAbsOhw/1+xu+7oS3lE8VQFmX17yZGqILNkEganc/iwT0eAPTAxHxN41mq1gS4EMwqxbXRUKlaHgpC1TrsArMCHr4s2Go3D10ZXRcoggeedtnVcgahwsameg4ZgAAkzQptN6EvJje8+mLSA917A881m7PvKXWueSqDqWkp2hCihVVaoCgVawalrKRU6vzG1FLjj+2j1cfD9szs2EwzwXl0QAad8VkTYN5MQQmA4DkH8cFjC1cdBt0Um59Dnc8E5HjE9J/4RiBJaHJrM+6bDLidgvRH0ivh4LHEP+MZi2ZOt4h6BR3j513Cj0fh/vx/F4QHPXRdXq0cOUavk6VwRcjELWyHiiIizsGdewc+N7z64uKW3vn7aFgu3XnigxqmHAIDrl0qg6lpKlmcltMoKVaFAKzh1LaVC5zemloJzqT47zpOZE7ju/njNcTh/Q4KOl5RXgEdw+OqB2Rh4Bw+NvwEHicD/m93tXnVQ9gWUizmwxfhl30WG5cBlH8icH4IjNkl6iIClSPoMkNMuAGfnXScH79plBD0qtqxzKkSuwtN8SOKjKblyFcT5xFMROW9L13Hju19mqYSdRSKfOE0Hl74kv34JJVB1LSXLuRJaZYWqUKAVnLqWUqHzG1NLgW/WlRGp8Ef3srAUOCT7w5YzAvevcNl3zkGIGAvWEBDrL+j8dVBFYPBNRAFX4KAPdBt8Q+HTOZff2nmNj9r30WochHC41PFYKAL/FC6Sb66wwkLipu3GfvzO2SnqGJaLQuSqfe538qmpIQ7M4sUNC0rtnoo+XIWtSwQ/Vd99clRrTeQlwX3rcakEqq6lZAlXQqusUBUKtIJT11IqdH5jaimYDO2LywIf3rOCEmbsjYcoHksP4XU5cJSemV8K9uX7zl8byQc4Z1jut4PJxA8s6LCZe3HEOJBA/PqAMzrQUsdS8RSE/fMcw3JLDmK7u0l8vF0o/S5AIpoPqDTw9ec6+TBFW7zJCAaHCdcLXW/ruGal1QZOLhxXo0srq+Npjd4cQ+dDWtV3nzhNp/cXxRhczFECVddSsiQroVVWqAoFWsGpaykVOr8xtRTE2Yu7VkB0EsyuVh++BAFwR/+BP6xhidhK7dUFONz7pu1GcPC66fgV8B239WQ6BIOA6RF4OiduVQcvwztDMLnZfhq76QMnrasP47/uzRVW4sJ10i68UyX4Hb+U19M3BBoKihxcwiRv6trdAFVs/+2grRpFcAha8gFVffc3VDtN33oyXflzKYGqaylZ/pXQKitUhQKt4NS1lAqd32haqqLSvjKWU1QONg5fbTwN0SKMZ6+BcfmI9ScQQuCV7t+romCgrjmc4j0vLP2miUR+goNNEIdw7ZHLtLODvnOwtTqE4YD5E8S/IH/d22y2NZGX+s45FJaYLUmv2C5OshqdCaYfo5fthllanRboaDnuSv9j1kGG5cgmHy1W1XefBKusk7kHDY9OK4GqaymayVppJbTWEqTyhVZw6lpKjYGgRu+n5hQ7DeuXVYDPgXaeEGS12uDk6YerjoMD8mMpubQr2M3RVxiW+2LDKQFUmHitPXIZwumCQeC2U+lkk8lms4HD07H86VqfHXi2NJw/xFPX6VGdmAerkO6T9oKWEnjQqJMot1QGJU2fZCJiBZSSfLck9sXh8Fp8yGA3WDkqgaprKdkOVUKrrFAVCrSCU9dSKnS+++dSMPVxanwcm47Dbfxj1kGEEDggf2NxJLjOS8kuIjkIIXuc+K1nBQMVlNPy8BSwgV7HO/yGLyMENCossxsOwPlcmJkNXRnFsJxc/Fm3MEwcvTMsR/u1c4vweggBHxCSExoBpfUQ7qAJcelEjq85qOy0SAlUXUvJ0quEVlmhKhRoBaeupVTofPdrqRHrsccBp+Eq7F5fecepEBcKzL4ZlisotZxKu8mw3IvzwhBCYCpGAuBaLPbg17DQtyg0CQzZIaY7xM+F8LjgVKLH1BA4t/vawggSpONgQi3fS+4ltqjM8vXGU8MX70m6LvSb594buSgt+Pz1kf6nCkwSQcNVfffBn2/VimudTOflHkoJVF1LybHq/vdf9k7KCpR0v7I717m1DrXOlDlr4F5KTeYKWGT7ZG204zuDF9SR/jj20iU+fi6sDnWZhGP3QUQMOIo7xYhjF/26P1EAldiUg16ERTwIsgcnWI9fws4j+uO4U0UMy0FwELA+P1aXI0eOH0SyVABVsk5TyFQVJ3gSYVjOLbb4SqDqWkp2sCmhVVaoCgVawanPpVTofDf/LRVywb4V8e5y2TDq8BSweQN+SImXB4blINIu+KzrOjm4Kjb5mEDss3XV4VTBQJ0dnMCw3My98R+uwgeedsdmIoTAvWy3KfsQQnvPYV9KQ1dEEc/ohWUWCCcBAeDV4BNkCqCqdyOFklXFSQ4Un8+wB/dSglYJVF1LyTKvhFZZoSoUaAWnrqVU6Hw3ayli2P36wsOO0cK5JTDMo6OPg3ojequi0grx4wNOXBUMVNBzU4xxsO8SciELIXSrxAxzMkul1W7Ozp8TAq8NiVlF4Ek9PtNtZ5gkH1MAVbJOU8hUFSdxPEhc1Ct5ZCVQdS0ly7wSWmWFqlCgFZy6llKh892ppaxWGxyBciVC/Gf89tWWE1fhoSBWIdnQIvEMC8ssxN5BMFDhlBX75zkIUgWOiIjCu1Vinss7Q4KYHeCsKCwxG3z3XcopVoNMIlMAleQ3tYSqOElfCDwn1Y8EJVB1LSXLuRJaZYWqUKAVnLqWUqHz3amlzvJmezCV6V0dZF0O85u/RVbZR4RdtJ9S+tdsfKaHYTlwdGSz2R6bEFSltK4XlA5cjGtGJOUIBirxbg6W1uBaAiEE4amu5plgYgfhneAI8B/RV4lYOWBuyRdAdYtMNYSojfPZGQc6TwjKN5mVg1cCVddSsvwroVVWqAoFWsGpaykVOt89Wsrv6OVP1kbDaSdw99BlIt5ScvAjjiegDsx1wPUR5MAaXUp2EcSRirl6SzBQ/Y7iA7yjN8e8OC+sKjYH2WrymhXKsFzctQKYrgWexNM1MLUAJ7YMy7nlu+ng6QRQHdRs3CK1cZ5Nz49IynHLMyqBqmsp2S5QQqusUBUKtIJT11IqdL4btFSZpRIissNkiIQwd3Cwl3Y8AQ8FwTsYliNrgDC7OpueT9SVYKD+EX2VYbmR/qfAt0LcNfsW/SsLIhiWO5aaCyuBYHQexEedB1cUDMupfdJWAFWNjnOLTK3gVPju61pKdrRoZQRoBafCkSrbT+oUaIVV5TjB+Lunb8iuM9fCErPNFVZQVw7mK7TjCaAfzjwxLEdOMsE63pHkXFCBNwrLBFC3n85gWO7TdSd6T99fFVE++UYRiHp3Ofbdty8uC7QXRDy6VWIGb36ArR5+j+o0TARQ69S2IStrBafCd1/XUrKDSisjQCs4FY5U2X5Sp0ArrCrHuSMGa4sPVkYRIuHI1LX8UpIjSNCOJ6CIRMwjVsuwcrjrzDXQKyZzhQDqnlhsaP7vVVFgeXE1zwSiwAfS1lPpsAWVWQ3jnaU4jG/VdA0M3AWQ3HspgOpe4W6UphWcCt99XUvJjhmtjACt4FQ4UmX7SZ0CrbCqHCecW5pijCNEwuQmqXpyQ/JJgnY8AZkQvINhuRuFZZADgTnAPVLnCUE2m00AFc5mvbf8KGgj0vDbP2IYloNoh1VhecsrKkEgyXFq2UFw1jshgFpvOWo31ApOhe++rqVkB5JWRoBWcCocqbL9pE6BVlhVjhOMFIiXcYQQsXeQo5Z2PAF1wKy8ow9XUWmFHHBeDlYPvXiLQQFUiOhBzC7IAiN4TgKXgOAqCQQeTcmFudS/ZmPngar+BFBVvZcS4VrBqfDd17WU7CDRygjQCk6FI1W2n9Qp0Aqr9cZpqVYnYJJw4vJNQiRojiPJsoHPaccT0AqCd4BLJMiB+Begq/rNPSTu/WO81gGNyLCcyVwBDWFuBwqJjsdRZql8YlIw+EwiUFVK1JtVlfDIidUKTnHvyz2RZL6upSRpwZlaGQFawakhSjUEtX69fzDhBsNygSevFlX7HSdTGYTQe9X2C3LvBjie+I0PmAt1IHjHgEU1HivAahzi+b6xOFJMKXikBQtAhq2ZhC0LSwEV9cUG7CSQ/n2yNpphOZBG57s9XT9W3Q7DqUCt4BT3vtNHoyvoWopmo1ZaKyNAKzgVjtRafaP+hVZYrR9O8AD78q/hp69g5+Xg3ZWQCspgRwyOjQu/ikorsW6o8rYncDyBELpRWNZ/fviqw6nVLRDEiYfDT2CaIYB6LgPH/gALwMcmBJGGMC17YlKwOEgu7HJBXA9SX42EAKoat3CLTK3gVPju61pKdrRoZQRoBafCkSrbT+oUaIXV+uEc6X8K5itTeYflAg/o9ii6UWmEWrBc2M+72kMICRxPkGp0AnaqOvlgqzxwnS6AejGrEDAwLAfuZaH5tfzSt5cc2XZKIor5jcKyt5ZEbjp+hb6RGmkBVDVu4RaZWsGp8N3XtZTsaNHKCNAKToUjVbaf1CnQCqv1w/kG77WIYbkuE/FOz8y98TSL/9say7AcBM+FfNipWhaWApcCxxN0W5IGX7Ggh8YGnhX3/uXcEqKlGsBsjwBzJVE/Vl2R7N46WsEp7v068aBrKVm6tDICtIJT4UiV7Sd1CrTCav1wgq05URJba09cSEQooLa4vAKmRL/su4gQEjuekOyBrSdxbHj4N5U3cxdAJYEhSMxfSTmNkimA2igYXLmpVnAqfPd1LSU7GLQyArSCU+FIle0ndQq0wmo9cJaaK0F5wOldhuVi02sFpYX1vel77BMsiJ9LPMneKCyD/SSr1eaAe+7cdaKl5ockins/p6icVHhhLo7q23R+9WC1UcBrBae49+tEl66lZOnSygjQCk6FI1W2n9Qp0Aqr9cAJkdqfmhoymj8/S1uBA5ewpTR++zm4JEZ3sHCXfANHzoUjUA64D0/MJkpoZQS2qhBALSi1kAr0uSgHMhusSAC1we5b1xtpBae49+v0pLqWkqVLKyNAKzgVjlTZflKnQCus1gPnkWR8PPbVBRH7+ci8/X8NF1AIW0rf/hED+cTUYhRvGh5z9VaVvz44AiVoSF+CoTnooT+isV9zAVQSg6oqZvlbS7CpetP5CaA2HWACJFrBKe59wYM4vtS1lCw/WhkBWsGpcKTK9pM6BVphtR44t57CO0afrjths9kCTlw9W3u5DyEEFT5bfwIhZLPZ+sw8AMrmo9XHEUIRSTkMyw3kj0A54D7heo0J3x4+WrwAqtVqI3OpBjAudwBVXCSAKq7QRHK0glPhu69rKdnxppURoBWcCkeqbD+pU6AVVuuBc3FoclXwC/ZP+4KemD/YUhq6AvufTb9pIrpk0NIjCCG6VNyW5FzNq2loj8NrsRiNRovFQuqABz+G5T5chfVf0/nVg9VGAa8VnArffV1LyY4urYwAreBUOFJl+0mdAq2wWg+c47efo8MVivmDLaU3f8OrcOBUAtTJy/zaYOBJHBfqc7+T4oZ0Tl5xjXHE6SvY/ZIYarcp+0AFjuDnbXTzxk2LoTYuHrm7awWnZO/LPZQ4X9dSYk7sOVoZAVrBqXCkyvaTOgVaYbUeOMG1hOSxWeDyZBp2SOE9D9vdgbvYISuOERcVayIvMSz33y1nHBNPbzuBe3UxVGIQDztejgU2ZKkYakPe3fV7aQWnwndf11KyQ0IrI0ArOBWOVNl+UqegCbKallsydGVU2MVs+onrgROC4R5NkXUmG5+Jt5T+/nMoQgjCRMEi4VNTQxBCCw8kMSw3ced5GoY4bbPZyILe9QIcqkoMFfwnMSz3XYATnSeWr2qOGKqqt6u3cK3glOx9159a11KyXGllBGgFp8KRKttP6hQ0QVbBQFyw1FZXnDabDaIOXsoplmPuSh72CtFtyj6bzQaLctGX8mBpzmq1zdgbz7DcnGB8wtfxj3iSLS7H/s7FUPvNPQRif9wW61hUA5eKoTYwABdvpxWckr3v4jMihHQtJcuVVkaAVnAqHKmy/aROQRNkddx27LhIYFwHOFOyCl6cF+bAwZ3NZhu9OWak/8mbJWZQDKVme3RBMX+51VtKt6or55vsrYrLK2Bbi3hLEjcnOf+afZBhuU4+HASAF1P68q/hAMbpzIzIbJiEGGrD3Leud9EKToXvvq6lZAeGVkaAVnAqHKmy/aROgdqskmiBNHzJTFLh36uixMdpAefSg3gVzoGlXFYBdhjBsBwc0X1mxgEiVpwgW0pw5unZGQdsNhs4L79RWPbN5tMMy/kfq/FFK5YAOa8uiGBYrqcvXieU7P2B1e4Ep+25AHWayH/V7n13PaZWcEr2vuskNAkttWzZMoZhPD09vby8TpzApzTkflu2bDEYDIMHD5arAPlOn8pxcyjVygjQCk6FI9WVLnNjHVVZ3ROb+cSk4JBqF+MAe+6+iz2mhlzOLZF7CghXyLAcPQ0CnN/ymmPQsqNybU9cxgYRDMvBKhzY78lVttls4Lgv4AQ254Mjtz19QxiWS8kuBuOLP0/XxPWQkwN7Wn3n4BCIkr0/aNlRQOXK+qHcXdTIV7X33QhYKzgle991Hpx+zw2uy6pfzcDAQA8Pj/Xr18fHx48aNap9+/bZ2bX2h4nYtLS0Dh06vPDCC7qWIpwo7H5aTgOkb5OXyimTEEWQOMqD+vDJ3s0fgBVLKK+o7MhHwWBYLpXaUgJKX+GXzhz4Gdp+OgP0AfwXQmmI70JyIKD79D14CwqiEcLyXWx6/rt8jEQSxYM0ESeGrcFBC0loRHHvD12BZ4cMyy08kCRu3og5YqiNCMbBrbWCU+FnqvG1lJeX17fffgs9YbVaH3744Tlz5og7prKysm/fvmvXrh0xYoSupWh+bpORSj9yA6RVZfWLDTi8E3GUB4/z2kK8PiZwT06e9FJOMVEzkck5JN9isWzbYYTlODJrIaUksYA3zCMSpvBOykmpOPGPWXhLCaZN4NEc4B1LyQUTwWOpsiaCRBqEk4fTwZLfKZAP65CkVVNIqNr7bnxAreCU7H3XeWhkLWU2m5s3b75r1y6CePjw4YMGDSKXJDF16tR3330XIaRrKcIJJG6TkSp4arUvVWX1/d/x8aPR1Y7y4Fme/wUbvG2gYg/Sz0g7b916siZCoMViWbLZCOrHgQfYMYFnIRY71Pw9vCaoLn0Xkga7Bpg/ga9YmEKFXMgCBXY+o4BUlkuM5W9KjBLFlH7udxLwrIm8JCekUfLFUBsFhtObagWntrVUZmamwWCIisK+WOA3btw4Ly+v6iv7/48cOdKhQ4fcXPznm5yWKi8vL6z+ZWRkGAyGvLw8i4KfyWQyGo0mk0mBjIZoqhWcFotFhwoD4sW5YdiT3trj9Ph4ZsZ+huV+D0umM0na70gqmQktCLlI8k0mE7tqNxQ9PjGI5AsS7/HLdNtPXoGDtDtOXxVUEFy+vSSS3G4nX/nj1ccZltt28kr3qdhhREpWgaCJ+HLiDuzk4vs/YqBI3PtfbrBrKb+jl8TNGzFHDLURwTi4tVZwKnz38/LyDAZDYWGhQC+QS3X3pVzRUkVFRR07dgwODgZMclrK19fXUPsXEBBg1H86A02PgScn7mVY7qWf99LQukzAmd8s201nkvRni/dUWdZ1ZHGdD37dQ/KNRuOni3AR/PtzJ11Sk+45GTdcutk4fd3u9+fv2bajpkgy9eqsGpkLN+Iqg+binP+tsGvETdsl29XKHLUEN/l0US20dI3359vv4rNa+qnpynr6tmUgICCgMbWUKyt+Z8+eNRgMzat/zfhf8+bNU1NrrVrocykHf3M1kaLb5E8/x2yXlNod3A1YGEFqlpfbDyTN5uJJJp34wh9PO976DU9xPl5dMwkzmUwv/4w1EPzLLZSY+heW2M3Q86RK6buQ9Od+J4jMKzmFFotlbOAZfJg3CNtTMCxXbCojleUSKVkFY7bExF+7BRXEvf+/rVgmw3J/nnIyt5O7hUr5Yqgq3UihWK3g1PZcCiHk5eX13XffwTzJarV26NBBYD1RVlYWR/0GDx788ssvx8XFmc1mMuMTJJzutgnqS15qZc1XKzgVrk1L9pF6meqxCrFuGZZ7/he7iTZCyGSugO+13MmhAYsOMyz3M4eVBHh9hWc3m80wM4PmmfnYF5Hgl8THLXx62n5BvoPL/26x649OPhyc4gKHfrDV1GWifWHDgQRxkZjSiTvPA+x9cdfF9RsxRwy1EcE4uLVWcCp8951+z9Vd8UMIBQYGenp6+vv7JyQkfPnll+3bt79x4wZC6NNPP/Xx8RH3kNyKH13T6VPRleXSWhkBWsGpcKTKdZNK+eqxCl7yGJbrPb1GbRB3Dz47JPzjEbdGBxNuEN9F8OCXs7HPvccnBvWYCueZisSEHIjHrd5eguNuuPibUK0//jn7IDSZH5IIB4cZlnN8KFjuFmJKp+25AFrq0EX8yjednxhq08FGI9EKToXvvtPvuepaCiG0dOnSRx991MPDw8vLKzo6GrrB29t7xIgRdJdAWtdSAk5uk5EqeGq1L9VjNTIZRxEE1UKegoRxGhN4lmSSBETB6OjDkSjs+Sb7QkLwuWtwJgnO/MaKQhoihNYeuYxNCjfbY+8SsQ4SMGljWI7EJ/w9HJtv9J+PfRq9MBe7S6/rT0zpnOCLQIUD17d1vYtb6ouhukWs24VoBef/D1rK7Z3nVPe6cketjACt4FQ4Ul3pMjfWUY9ViNgEX+fyCrszvcSsIsj5etNp8VOc4YO4w7TmmRk4cm58pt3eaUEI/tCPCTxDzjOJm8NiXZ38OywKxS6XaHP5jcevMCz35GRs4PeGs0C9YgySvU9OcZ1KwwGoms5Pvd537zNqBadk77tOhdPveUPMpVyH62JNp0/lihytjACt4FQ4Ul3pMjfWUY/VdfzMBnTAzRL7lAj0EMNyEMdd8CCg2IauxAc23uQNKA4m2JfIvv8jhmG55WFJ4I7oQLzE0tl/+GNJf0RfFYh1cLn6MA4ixbDcrKAEqLbzTI33ig94JA6aSxaJKQWnglUztnMZ+ZJNGitTDLWxkDi+r1ZwKnz3nX7PdS1VEwDb8YhprNLbZKQ2ML3qsQobPKADruaZ4LmOpeRCzr9X1ZwdJI+85CCOAQ/hLcBvxcbqw78j1mEvRH8cT/t4DT7PZDx7jbQiCfAWcSTZubcI0uSPaOzBj2G59UcvQ+b+C1mQw7DcSH8ngXqJHDohppTowsQsie00um0Dp8VQGxiAi7fTCk5dS0l0qFPdK9FGlKWVEaAVnApHqqh/1M1Qj1WfHXbDNnrhLpQ3cGBYbrCUx9iftuGYHUsOJiOEphrjGJb7ZZ89vNP7/HHdPWczQHuJJ0xWq+2JScFVMeDTb9o1oivEkWXJfXFZUP9Yql2PMiz3g7NAvZK3EFPqfywNNF+avI9dSVFqZ4qhqn3H+snXCk6F777T77k+l9LnUvV7gyRa3SYvlcSTU1mjeCd+8HU+Wb0fszs2E3LAN6vJXPHe8qOLQ7FaQghBzI5dZ/A8aUUEtmIgeuI1PjpGxMWsH3jbcbGrITB87zwhyHFYEAogThKtSdbizmXkk7nU5F1xgvquXIp7H3yuMywHwXxdEdIwdcRQG+a+db2LVnDqWkqiZ53qXok2oiytjACt4FQ4UkX9o26Geqz+H+/ED774JDx84En7Cpv3PGw+F5WKY+P2mWmPAvUC71EJVBq9R4UQ+icfaTAmLRdsx4liI+ycTMMxO+pqlQcAqqwncorKQRTt7pbM5MhdXEmIKf2z2lN7XrH9Lq7IaYA6YqgNcNN63EIrOBW++06/5/pcSp9L1eP1kW5ym7xU0g9fnQvG3I9NCGJYbk91nI71R7GxOMNy/5iFzyfBuSgIJWW12h6fiCtf40/sgtYhJ4LBq15yVj7Yjs+uNnaovhvadQabqjsIkEhq0gmYOXWZFGy12iA/u9DuwKLeLszFvb/3nH0GCSHnaQCNmxZDbVw8cnfXCk5dS0n0oFPdK9FGlKWVEaAVnApHqqh/1M1Qj1WIQ/jiPOxwdssJu90dsXYDv+ZkATD5RhGE2SVLdlfySsAiHCFUabWBbsvKL1nIx+aYtEt4KHglv0IoeQzLAYMl5RV95xyivbaXlNu9YzAsR2w3HEgQF4kpJRYZlkqruH4j5oihNiIYB7fWCk6F777T77k+l9LnUg5ek7oV3SYvlQNSLJVW0CsQWolsIxHDvycmYedDW/gguQzLhV3MPn3lFsNyJHYUOdhbZqksMFlAWklp+arDeL9qrOhQMPh3qNNhKcBPZlFwabPZSBjGnWecB+oVkyDufQhH0tGHs9nsMzZxq0bJEUNtFBhOb6oVnLqWkuhKp7pXoo0oSysjQCs4FY5UUf+om6ESq9lFeN2skw/nw0e3siddAAAgAElEQVS1WBRqj1ELUXFB5dhsNvAWAeGmYF5FYglarfZw79mFZeCx4nGfvRaLZRN/6vbLjacEvHzDB5snBuWC0jpdghOmKpUpeSrLqSgxpWA32HVyfbwCOr2dkgpiqEqkqddWKzgVvvtOv+f6XEqfS7ntLbtNXioHfCVcx273np1xYFZQAsNyM/fGQ2VQWqClyiyVcEAKDtWCayJ6yQ5iRCXdKAKXgD0nYy0Fp24/WWt3MEYwQGSp4PNucOcK8Q8ZlotKzSPyXU+Ie//0FWzZUSc3uK7fTklNMVQl0tRrqxWcupaSGANOda9EG1GWVkaAVnAqHKmi/lE3QyVWj/Knd19dEPEbf1DXZ8c5eAzig5xhuVslZuLg7pvNpyftwuer5ockkgd+iXemd+LyzeOXsCngc75YS4Xwp26J2z1SGfz7xVy9RXLqnYDTwQzLxV1zHqhXfBcxpRcyC4jBiLh+I+aIoTYiGAe31gpOhe++0++5PpfS51IOXpO6Fd0mL5UDUmD57t+romBN77uAM1AZzuTCXOp6QekU/uguODIX+zcCZ0j7L2SBs/MXZ2ItdSQZn7qF41YEALEPlIzoQaq5mBi07CggvJJX4mITupq4961WG/vnOfFJZLpVo6TFUBsFhtObagVnY2qp6gDusv93yrJKFZzqXlfuq5URoBWcCkeqK13mxjoqserHW5yP3hyz9WQ67bUPjClAB1zOLfnfVuxsAqJ7vL4QR5aKSMohTzdiPQ5RuPVk+o4Y7FtvwOw9FoslhvdI229uTcwqhBAEBOnow7nFiG7YGuyNiWG5+h1vUolSQosbE1qBqhWcCt99p99zR3OpZs2a3eHw58ZxUydRTp/KFWlaGQFawalwpLrSZW6soxKrC/bjKE2Td8Vx565XeUgiNhHvU0d9E64Xfr3pNOiDKh99XXj/RinZxeTpwM1ElVEf6LzB87CWAq/qz86wHwSGynHX8JJan5mhpK2SxJcbTwEq4sq9TtJUorROGFysrBWoWsGp8N13+j13pKUiqn/+/v4PPvigj4/Pbv7n4+Pz0EMP+fv7uzgm3F7N6VO5cketjACt4FQ4Ul3pMjfWUYlV8BCx8EASGGGTEBgDF+NQ8fAv5uqtT9fVBHSHTJO5gjwdROKYu+8iGFl8uABrKbD3E9jLHbqI4x++tSSStFWSGLv1LGjN+glRidL6gXHcSitQtYJT4bvv9HvuSEuRnn755ZcDAgLIJULojz/+8Pb2pnMaMu30qVwBo5URoBWcCkeqK13mxjousnohs2BWUEJhmX3z8vfw1CCH1nRfbcSTpA1RaWDeRhwXefOHfEEhHUvNpb0oiWPjwgHeCTvPg6HgfxZjLXWzxAzNK6u9ReDXkHdt/rlffVyYi8kER7eC6Zq4mlyOi5TKNW/IfK1A1QpOhe++0++5S1qqdevWycl2z5gwmJKSklq3bt2QA4u+l9OnoivLpbUyArSCU+FIlesmlfJdZHX0ZhzeKYB3IQGe7ug48WJsQ1dEVS3B7T2XeTHLbpIOdZ77OZRhOXCbFHYxG6ZWT/FB4sXB4MGd0ujNMeyf5xiW+2bZbovFUl5RCVqqqFplIoQgzOCEnUKHFGJgruTM3YcjLoKnQVfqC+q4SKmgVaNcagWqVnAqfPedfs9d0lJPPPHEuHHj6PE0bty4J554gs5pyLTTp3IFjFZGgFZwKhyprnSZG+u4yCoszYGPV7ALr9p2omczNKT0myYIonE2PR8W6LrwniYQQnBgts9MrKuCzl8H/0lw1Ilhua821grgC0ejPl5zHHTk/1ZgLWWz2UDJ3SgsIzcFNfYbH/KDZNY7AW6c6r1+6CKl9YbnxoZagaoVnArffaffc5e0VFBQUKtWrXr06DGS//Xs2bNVq1ZBQUFuHDd1EuX0qVyRppURoBWcCkeqK13mxjousvrBSjw3mmrEkSyCzmODCDjwJIlkpD+2Pvho9XGbzZZvsi/QWSqtNputM+98Fsz5dsRkgLoiR32n77Ef/gWxYYnZDMu9+VskWAZOWoO1FEKop28Iw3KXcmrsLMAaMPBkHaL0SiKHzA1ROByUZJxGB61IkYuUkvqNmNAKVK3gVPjuO/2eu6SlEELp6ekTJkx4j/9NnDgxPT29EQeZ06dyBZtWRoBWcCocqa50mRvruMjqu3wQQjj2BD6KquzLJY8TgZvzxyYEpWTjuLTEoV++yUwW6yCO1B/RV7tN2QdxckHtrT1iD5gLDwjh5/vOOQQHmGaut2spiOJBH7kdsAhbsYcnZruFGTiSPLG+64cuUuoWqAqFaAWqVnAqfPedfs9d1VIKh4V7mzt9Kldup5URoBWcCkeqK13mxjousvrmb9g2b9ga7JeIuDUikQMJnkqrDRbxZgcnkEwwMU+/aSLzKjBAXxN5CZQTCY9LAuZC28u52C36U1ND+v8azrDcrxuNMJd6mb+MvlTjvgh8KV3MKiQ3VZhIyS42V9TTf7mLlCpE6JbmWoGqFZwK332n33NXtVRkZOSwYcP+9a9/XbuGI4pu3LjxyJEjbhkx9RDy/9r7EvAoivT94TABRQ4FEdilhQVZBEVlHY91RdRFvPBa19/6X4jHggfoyip2gEAE5D7lRm6EEM4E6BzcEMgFgRyQQBJCCCGQQAIJIdeQmfrTU6Ho9ExPd6Z6jo9UPzxaXV1V/db7vVVfuqe6PtVeaWkTigKg4KRUqhaT6VhGI6t406D+c8Sl3niBOMcLURl3vsDFkGp+hRoVfqPyzoLyp8fvuvUC7dSlkrxr5RwvdBkVNmKT+DHvtEhxkQLHC/ibXI4XUnJr7Uh09fZyPuyEFqyr8VLvzDuEt1HHNyWPaFdvVOnIjNNNaaTU6fZ1rAgFKhSclGNfdT7X5KU2b97ctGnT//znP76+vllZWQihefPmvfHGGzrqpk5NqfZKS2tQFAAFJ6VStZhMxzIaWcVRdHHoQrIX347kPBmSxPNiOHYSfQNfxXUTzhVlFpRyvPB4YCTeGAn/HNXJGs/io8UxL0zeW15VLW2QhJXCzmzFhhovhV8YkruTBRpeEhdDI6XSnnoqDQUqFJyUY191PtfkpZ588snVq1cjhJo1a4a91PHjx9u2bespkan2SgswKAqAgpNSqVpMpmMZjawaJ4qr8rqOCrdYLGSXo7Vx52RI8I9S78yr9XYBLzfff7ogJbdm01X8/RPe06/n2EiEkNlssbu5EV4ogb1U0OYaL/X5yiN45yR8d+diycuQ63iqkVId7+h0U1CgQsFJOfZV53NNXqpp06bZ2dlSL5WVleXr6+u0SigrqvZKS/tQFAAFJ6VStZhMxzIaWcXv3DheKK28iX+j4nhhwf5MGRLZrn34Kv52Ski+GGfd2rzv9P348ya8W5JxoqNtjfBzGPZSW0NqvNSwoON4zQVuf3uSGK+dbMIkg+T+U42Uuh+Y7R2hQIWCk3Lsq87nmrxUp06ddu8WBxV5llq9enX37t1tze+eHNVeaYEBRQFQcFIqVYvJdCyjkdXHrIvxOF7IKSzDS+w4XrANjLtgf+YtT/bDxiQpwk9vbxpLVpbjYjgwR9/p+6WFZWn8ExReQxEaWuOl8NdR8/fV+Ei8CmPoumOyup461Uipp+BJ7wsFKhSclGNfdT7X5KUmTZr02GOPxcXF3X///YcOHVq7dm2bNm3mzp0rNbw706q90gIGigKg4KRUqhaT6VhGI6v4Q1qOF47nXMVf7HK84L9FvtfD+B2pHC9IF/ghhPCjz7JDZ/GHVv9YFL380FmOF3paP3ty/PEsebv43KQ9xEvhgL9TI05hHvBqDhJoUUdynGtKI6XONa5vLShQoeCkHPuq87kmL2WxWH755Zf77ruvgfVo0qRJQECAvrqpU2uqvdLSGhQFQMFJqVQtJtOxjBZWpasYdiSLr9fwv2/Wyh9f8C7mvx0UFxaRA6+S+HVPxqYEMQDHwOXxQfE5pJGPFseQkrYJ7OE4Xvj7zAPES82wbrgeuO0kLv/azAO39gDcliRfymHbmntytFDqHiSqd4ECFQpOyrGvOp9r8lLY6lVVVampqfHx8aWld75+VxWEKwqo9krLTaEoAApOSqVqMZmOZbSwWl5Vs3Uexwuzd6cTB4M/n5KCwY8+W47lSjMnWB+wJoalrbHu6fDlmgS89RFu59MV8dLCsnRAyAlc7MOFh4mXwrHnf7S+V8wpLON4ofPIsOJybwnjqYVSWTc9dQoFKhSclGNfdT7X5KVWr16dlnbnc0WEUEVFBV715xGdqfZKCyooCoCCk1KpWkymYxktrBaXmYhnGrpO3HYW/3t7bq21fAghvJxPGskQIYQd28itKYsPnOF4YXhwYvjtPZY4XvjG4e9JOE4Vxwt+y+OIl8I7GOEnOfzy0OndjHRkkjSlhVJS2LMJKFCh4KQc+6rzuSYv1aBBg2bNmm3evJloKz8/v2HDhuTUzQnVXmnBA0UBUHBSKlWLyXQso4XVgusVxDPhvYjwKYnHQfDgLc+lexchhPDqhu/WH8eROEZtTdl3StygD/8bsanWUgvSFE7ggPS3wuYOW5dAvBR+czhoufgQ9snSWI4XlkbVescoa8TNp1oodTMkpdtBgQoFJ+XYV53PtXqpmTNn3grVERgYiA3PvJTSANA9v54oVXfeHDeohVX82Sx2Kni7I7zMr9e4ndLGzeaa3covFpdL8/GvUJ+vPII/k/pFSI05U0i8FPl5SVqFpDdbf8rieGHUlmTipfCj2D8WRZdUmPCyjrNXbpAqHk9oodTjIDEAKFCh4PQWL1VQUBAbG/vwww9/+OGH5eXlzEu5bbzVE6W6jU/t89SZy+KeEdJ//7dEfILp5C+YJaEIyTZ9skDs+Humfy6OGR2SwvHCzF3peBtZ3OC0yJqlenY7jj8TFle9h6USL3Ug/TLHC2/MicJLOfrOcLSW3W6zLs1kQtWd3npCqT7PUg0bNiwoEPddzsnJ6dWr15NPPhkXF8fe+OkuSrsN1hOl2u276zK1sJqaJ0YylP7DAeNv/chEovcihPAGSD0Dxb0kpMehjCu3tt17ado+HKl90YEzaRfvNEg+e5JWIWkc6pfjhfl704mXOppdhBv8eq0YEXhiWK2fikldTyW0UOopbLL7QoEKBacXPUthS5eVlb333nv3338/81Iy6bvotJ4o1UXsKTWrhVW8O5/US83dk4G/mjpfVEZaxltLvGzzlW5xuekRf9HJ4fjxq2Oys62bneMGVx6uFa2DtIYT2POJwemjs4iXknnNhHNXZbU8e6qFUs8iJHeHAhUKTq/wUj///HNZ2Z1hiRAaO3bsyy+/TKzu5oTqE6IWPFAUAAUnpVK1mEzHMlpYjT8rPrtgT4Ndy++x52wXSuCPdj9cGG0LD6+5wCEQNx49f6n4znKMDUcdRWgj26VvTcghXupcoRjRA79ylIWksr21+3O0UOp+VHbvCAUqFJyUY191Pte0esKupT2YqdorLdigKAAKTkqlajGZjmW0sBqVIf4OhN0Sdg/hKRdxLI/ozCsEDP4casiaoySHJMaG1nz2xPGCkHyR/IKFT0kx2wQJorgn9SLxUuVV1T3HRvYcG7lPp7CHtvelydFCKU37OtaFAhUKTsqxrzqfO/JS27bVBAndZu/Yvn27jrqpU1OqvdLSGhQFQMFJqVQtJtOxjBZW8RKGt+ZGkcep2KxCvFdseMpFAgYvNB9pL8StdMeKfacKKkx3PhNW9TQ9xorx4+PPXCZeCiGUX1LhJdGkSPdJQgulpLBnE1CgQsFJOfZV53NHXqpBgwZ40QTeGEn2X/a7lHtGWj1RqnvIJHfRwirZf+8pazxDjhfS869/Zg2fEXwkhzRVs4Rv52mSQxIFJXde8cVmFVosFuLw4s8WkWJ2E4OWx/cMjLxcXCb1UnZLekmmFkoZ1DoxUE8opfJSdSLUnYVVe6UFDBQFQMFJ+feUFpPpWEYLqyHHL3C88MnSWBzKHUfX/T44keOFJQfPEDA4TvyqaDG0je3RZ9o+/LYQx6HvFhCOT2WfANtWrDZbblTe1ILTtq5HchhU3WmvJ5SqzueOnqUw6SaT6ZVXXsnIyNDdBk43qNorLS1DUQAUnHefl8JRoz5beeQfi6Kxa7lZbcY7kUu/dvpocQzHCySErkx7OIo8xwuZBdcRQr3G7cRNafwgl1lfxqcup1BYhYKTcuyrzufqXgoh1Lp1a+aldBkeTjRST5TqBDM0VbSwSnaJHbLmKMcLeMsJHMlwdMid4B34SSvmTKFdPBuPnsdu6cI1cWcKEqSqoKTCbnlZphacsiqeOmVQdWe+nlCqj5f6/vvveZ7X3QZON6jaKy0tQ1EAFJyUf09pMZmOZbSwSjbiwzE4cNxCnDks6DgBgx+PMvLFRyXbA29ezvHCtbIqhBAOgYiD/9oWts3RgtO2lkdyGFTdaa8nlKrO55qepYYNG9a8efPevXsPGTJkuOTQ3SoaG1TtlZZ2oCgACs67z0vN3ydG4P1xY9K0yFP441yEEH42Gmjd8lXscrUZPyoV3RCdkO1hsVhGbU0hYXzJrrXVkj2WbGuRHGZ9QoWOCSisQsFJOfZV53NNXuple0ffvn111E2dmlLtlZbWoCgACk5KpWoxmY5ltLCKl5iPDknBYTLwF1E7T17ieOHd+YcxGLyKT7aznwOc7y04zPHCnwMiHJSRXtKCU1reg2kGVXfy6wmlqvO5Ji+lO/uUDar2Skv7UBQABSd0L1V107w9Ke9G5U0inikR4iPUuO2pBSUV3wYdP5Itrh3H+yHht38IIbxrUe8Ju0ktxwm8X23vCbscFyNXmfUJFTomoLAKBSfl2Fedz+vmpXKth45yca4p1V5paRaKAqDgpFSqFpPpWMaW1RWHz3K8IP04d9z2VHFX8vBam5efuiTuGPv0+Bo3g/eneH32QY3YPl0Rj3eM1VjeFqfGiu4vxqDqznk9oVR1Ptfkpcxm87hx45o3b97QerRo0WL8+PFms1mjVebPn89xnK+vr9FojI+3E0h7y5YtvXv3btGixb333turV681a9Y4blm1V46r46tQFAAFJ3QvhXcuf3r8LhKVA3+uO2tXulROF4vLOV7408gwi8Vya2U5/qbqX7/FSss4SOPtzPvPiXJQRnqJWV/Khl5pKKxCwUk59lXnc01eyt/fv02bNgsXLky2HgsWLGjTps2oUaO0iCY4ONjHx2fFihWpqamDBw9u2bIl3s9CWnf//v1bt25NS0s7c+bMnDlzGjVqFBkpj4MgLa/aK2lhpTQUBUDBSalUJTO5KN+W1Xfni78YcbxwLKdmr/EfNyaJsTP2ZUox3Ki8iYvhd4M4yvtXvydIyzhID7d+FGx3a1q7tWxx2i3mDZkMqu5WqCeUqs7nmrxUu3bttm3bJrVBaGho+/btpTlKaaPROHToUHzVbDa3b99+8uTJSoVx/lNPPRUQEOCgjGqvHNQll6AoAApO0F7KYrE8Hijum8fxAvli97v1x22jtlssNZF586zfP+F1gI7DwxPJIYT8t4gREckSQeklu2lmfbu0UGZCYRUKTsqxrzqfa/JSvr6+6em13nucPn26SZMmqlqpqqpq1KhRSEgIKTlo0KABAwaQU1nCYrHs2bPn3nvv3bXL0c/Lqr2SNWv3FIoCoOCkVKpdG7kuU8ZqYWkldlEcL5Afmb5cIwYbXBMj3/roaeu2fqculSCEJoWncbwwfkeqRqg/bz/J8YL2Zy8ZTo138UgxBlV32usJparzuSYvZTQav/32W6kNhg0b9uyzz0pz7Kbz8vIMBkNMTAy5OmLECKPRSE5Jori4+L777mvcuLGvr+/y5ctJPklUVlaW3D5yc3MNBkNhYaGJ4igrE/fxLCsro2jDHVWh4DSZTHChxmYWcLzw5LidnaxxC7Mvl5hMJr/lcRwvrIs9KzPzy9at+Q5n5JtMpp82iW8FZ+48JSujdDo5TFyR8X3wcaUCsny4lMo64lWnUFiFgpNy7BcWFhoMhpIS8c8+u4cmL3XgwIH77ruve/fun1uP7t27N2vWLCpK/Rdg7V7KbDZnZmYmJibOmDGjRYsW+/fvl8ENDAw01D6CgoJC2cEY0IMB/9+2cbzw90nbX/llx62NIX5YvC00NPTvk7ZzvBCwVExLj5cmiGUmrBDz35smlvl+kbyMtLw0zS8Rb/T1fK3lpXVZmjFwVzIQFBSkg5dCCOXl5Y0aNeoD6zF69Oi8vDyZF7F7Wtc3friRL774ol+/frIG2bOUV/01ahcM3D/9Jgrii7jRW5MX7EvneOHfS2NNJtOHC8X1FNuOn5d1duCyWI4X1sdlm0wmnA6OF9NajsrKquPZhWUVlVoKU/6JqvEWehWDa329GNC9nXpCqT7PUjKHUadTo9E4bNgwXMVsNnfo0EF19cRnn33Wp08fB3dRfY/poC65BOWdLxScoH+XwvvJrjh89njOVY4Xnp24ByE0YN4hjhf2pOUTzeDEsKA7qyrwXhKRJy/Jyuh1yqyvF5PSdqCwCgUn5dhXnc81vfFDCFVUVMTHx+/YsUMatldqeKV0cHCwr6/vqlWr0tLShgwZ0rJly/x8cdgPHDjQ398f15o0adKuXbuysrLS0tJmzJjRuHHjpUuXKjWIEFLtlYO65BIUBUDBSalUYhf3JGSsvjbzAMcLB9Ivn71yg+OFHmPFDyHwnntRGZdlkKRhD2sCzJ+5E2BeVpjyVIaTsjWXVmdQdae3nlCqOp9r8lIRERFt2rRxOlbvvHnzOnbs6OPjYzQa4+LisC379Onj5+eH06NHj+7SpUuTJk1atWr1/PPPBwcHO7a3aq8cV8dXoSgACk64XqrabOk6SgxOeL6o7PL1msV+ZrOl7/T9Ykx3m6C6ePPZwG0nEULP/LL71ka0qiENtQjSbhlmfbu0UGZCYRUKTsqxrzqfa/JSXbp0+eabb/AzEKU+dKmu2istd4GiACg4KZWqxWT0ZSwWS3G5SQb1fFHZrbV8XUeHV5stFaZqvCT9eoXphcl7b20sm3j+muy+Sw6esa7TS0QI/TkgguOFnMIyWRm9Tpn19WJS2g4UVqHglA0oKdVa0qrzuSYvdf/99585cyeEtpYbu7SMaq+03B2KAqDgpFSqFpPRl9l6PJfjhc0JuVJW958Wl6H/fdYBhJDFYukyKozjhbxr5b0niM9JqXnyBbLBR3JueabPVh4hYTuuKoTtoAcsxUnfmktbYFB1p7eeUKo6n2vyUp999tmyZct0t4HTDar2SkvLUBQABScIL/W/DeLnTf9eFidlFe8z++Waml2OnrQGfU/Pv453ozhzuVQmp4gTFzle+GBhdNGNKvzgdSvSvKyMXqdSnHq16aJ2GFTdia0nlKrO55q8VFlZ2Ztvvunn5zdjxoxfJYfuVtHYoGqvtLQDRQFQcILwUv/6TVxE/tiYiPKKytDQUJNJfPsXEHKC44UpETV7n/9t6j6OFxLOXX10dM2PVTI5RWde4Xjh1ZkHzhWKSy26j9EaLErWjpZTZn0tLNW1DBRWoeCkHPuq87kmL7Vs2bLGjRs3a9aM47hHbh+dOnWqqzj0Kq/aKy03gqIAKDgplarFZPRl+li3jRC3lM2+QrzUv5eJe0xsOHoet//GnCiOF/afLnjEug9FwfUK2X1PXCjmeOEvv+zGCbxsXVZGr1Nmfb2YlLYDhVUoOCnHvup8rslLtW3bduLEidpDdUgF4Yq0aq+03BSKAqDgpFSqFpNRlrFYLPjxiOOFJQcyiZd6a67olvaeqvku6qPFMRwv4JAct1ZJFJeJz1vSg6y2wA9Vr80Uf9By0cGs7wpiobAKBSfl2FedzzV5qVatWrHVE64YLVrarCdK1UIFZRnplrL/WXWEeKmXrA9YOBQvQuiLVUfwVuj4N6fyqmrZfUsqTPhSaOIFjhfeX1ATXV5WTJdTZn1daJQ1AoVVKDi9wkt9//33EydOlFnag6eqvlcLNigKgIKTUqlaTEZZBr+gww7myXE7Q0Jqfpd6SrLHOULov9aAHVOt4eRv7eNXbRajHUoPs9mCN6XFYTv8VtgJ7CktT5Nm1qdhT6kuFFah4KQc+6rzuaZnqW+//bZFixYvvfTSsGHDhksOJRG4Ol+1V1oAQFEAFJyUStViMsoykScvcbzQf05UtwBxWcTiINFLkaXnF6zxohBCeGuJnzYl45i8dm/ay7oOEAeLGhZ03G4ZXTKZ9XWhUdYIFFah4KQc+6rzuSYv9bK9o2/fvjLbu+1UtVdakEBRABSclErVYjLKMisPn8Wxnf5vibjSb8SSbSaTiXzGW1JR8/vTFOtTFH7v95jC+j28DhAvuxi5NYUSmIPqzPoOyHH6EhRWoeCkHPuq87kmL+W0GlxUUbVXWu4LRQFQcFIqVYvJKMtMChMjFo7bnjpzl7jx+QfTt5tMpoLrFRwvPOIvmG+/2cPv8fA2sk+O22n3pm/PFTeifdm6hdLk8Jol7HZLUmYy61MSaLc6FFah4KQc+6rzeR28VGZmZmRkZHl5Of5E36753ZOp2istMKAoAApOSqVqMRllmaHrjuFlEftOiZtNPDduh8lkOnO59NZuFD2t28vi9lfHZN9a8oeflowTd9u96f9bKi5ex7v/zd+XabeMLpnM+rrQKGsECqtQcFKOfdX5XJOXKiwsfOWVVxo0aNCwYcOsrCyE0Gefffa///1PZnu3nar2SgsSKAqAgpNSqVpMRlnmg4XRHC+EpVxMyRU/eHpijOilEs9f43jh+UliqA58bDkm7qL02Bhxg76/Td13O7vW/79ZKzo8/G9N7Lla13Q9YdbXlc6axqCwCgUn5dhXnc81eamBAwe+/vrrubm5zZo1w14qMjLysccec4WAtLSp2istjUBRABSclErVYjLKMs9P2sPxwvGcq/iDpy7+opeKyrjM8UK/WQdJ47tS84kHelXhWyi8bgIXC028QOrqnmDW151S7/G+3BAAACAASURBVBcq6XI9sb7qfK7JS7Vt2zYpKQkhRLxUVlbWfffdR9h0c0K1V1rwQFEAFJxePvhvVpvx8vH8kgrywdP1soqwFHFTvg8XRhPNxJwpJF7qzV+jSL40MTn8FClDPgeWFtArzayvF5PSdqCwCgUn5dhXnc81ealmzZplZGRIvdTRo0cfeOABqeHdmVbtlRYwUBQABSelUrWYjKZM3rVyvLK82myxWCydR4obn58vvL7hyHmOFz6VfPMk/azqPYUvdhfszyRe6mh2EQ0wx3WZ9R3z49xVKKxCwUk59lXnc01e6o033ggICMBe6uzZs2az+aOPPvrwww+dkwh9LdVeabkFFAVAwUmpVC0moymTcK6I44W/TtmLG3lq/E4xeuH5oqVRWRwvfCv55gnvIYud0MdLYuzedG3cOeKlTl+6breMLpnM+rrQKGsECqtQcFKOfdX5XJOXOnHixEMPPdS/f38fH59//OMf3bt3b9u2rQf3TFLtlUyUdk+hKAAKTkql2rWRjpnbkvJuBYv6aFGN1+k7Xdz4/FB6/izrqvRRkm+epBspDVxuf1+JHclia/hf3u3PgXVES5pi1idU6JiAwioUnJRjX3U+1+SlEELFxcW//PLLRx999MYbb4wePfrixYs6iqauTan2SkuDUBQABSelUrWYjKbM4gNigN3v1tfsE/HefPGDJyEpd9z2VI4XpN88Vd6sCdfL8cIXq47avSlec4G9VGnlTbtldMlk1teFRlkjUFiFgpNy7KvO51q9lMzMnj1V7ZUWeFAUAAUnpVK1mIymTOC2k9IgUn7LxQ+e1sWe/XGjGBdR9s1TV2tkKY4Xvll3zO5Nk3PF9escL3TyFywW+UZ/dqs4l8ms7xxvjmtBYRUKTsqxrzqfq3ipHIeHYym47qpqr7TcGooCoOCkVKoWk9GUGbz6KMcLa2KycSP/DRI/eFqwL/3LNQnSfHz1aev+sxwvDA9OtHtT8tvV44GRdgvolcmsrxeT0nagsAoFJ+XYV53PVbxUw9tHA+tx+6wh/sJXanh3plV7pQUMFAVAwUmpVC0moynzzjzxFd+u1JogUmNDUzhemCSkfrJU3NMv5Hitb55wLA+OF/y3JNu96bWymljyZDmG3WL0mcz69BzatgCFVSg4Kce+6nyu4qUaNWrEcVxgYGBCQkKSzWFrfvfkqPZKCwwoCoCCk1KpWkxGU6avdc+9+LM1q8Zn7hQ/ePppUxL2XnvSarwXvgWOi8jxwtjQE3ZvWm224Dd+/efY/6DKbi0nMpn1nSBNtQoUVqHgpBz7qvO5ipe6dOnSlClTunXr1rZt2x9++CEtLU1VAW4ooNorLRigKAAKTkqlajEZTZnnrBtPpOQW40aWR4mLKb5ccxTvGEu8F7768RIxXC/HC78IqUo37Tk2kuOFfy62v1RdqVZd85n168qYlvJQWIWCk3Lsq87nKl6KmPzQoUOff/75/fff/+yzz/7222+ejS6v2isC20ECigKg4KRUqgNL6XIJR4TKLKj5tmnz0RyOF/5vSUzvCbtuOZu0iyXSu3yxSvwRi+OF6ZGnpfnS9AuT9zpYBCgtSZNm1qdhT6kuFFah4KQc+6rzuVYvhe2dn5/ft2/fhg0bFhW58Ht7JW2RfNVekZIOElAUAAUnpVIdWEqXS49al+2RUId7UsWNkfrPPoiX8+VeLZPe5fvgROylft0jbrli93hjTpSD5RV2qziRyazvBGmqVaCwCgUn5dhXnc+1eqno6OgvvviiefPmzzzzzKJFi9izlOpI0KtAPVGqXnTZbYf8jFR0owoXOHr2CscL+EGK44Xi8poQiPjqmNAT2EstPnDGboMIIRxKMXDbSaUCuuQz6+tCo6wRKKxCwelhL3Xx4kX8u9RDDz00fPjwEyfs/5gsE4GrT1V9rxYAUBQABSelUrWYzOkyNypvYq9TXlWNG8nMF4N3kH/Vt0Mg4qtTreF6OV5Ycfis0k3xEvYZOxVfCSpVrFM+s36d6NJYGAqrUHBSjn3V+VzlWapx48Ycx40dOzYhISHZ5tCoCd2LqfZKyx2hKAAKTkqlajGZ02WulFZih0QC8l4pKSMuqockBCK+BdlMdl1cjtJNceD5TQm5SgV0yWfW14VGWSNQWIWCk3Lsq87nKl4KfyaFv45q2FD8TIocDRs2lNnebaeqvdKCBIoCoOCkVKoWkzldBgeU+nNABGmhsrLqEX4HdlTPSUIg4gJrYms2k3XghCpM1Ueyi2QPYaR9vRLM+noxKW0HCqtQcFKOfdX5XMVLnXN4SA3vzrRqr7SAgaIAKDgplarFZEplblab/70szsFPROn51zleeGr8LtKCyWTqPrrGS/191gGSjxMhxy9gB7YtKU92yc2nzPquIBwKq1BwUo591flcxUu5QiL0bar2SsstoCgACk5KpWoxmVKZDKsT+tPIMPJCT1YyyRo2/oXJNWE7MNTegTVe6gNJCERccfftcL2RJy/JmnLzKbO+KwiHwioUnJRjX3U+Z16q1uIuVwwJyjbriVJpWMKxozheuFZWs4RP1lpslhh+95UZ+0m+yWR6aUKNl/KThEDEBeKs5Tle2He6gFTxSIJZ3xW0Q2EVCk7mpeyoVNX32qljkwVFAVBwUirVxj51yNh3ugC/oCMf7coq4wJvzb2zm5HJZOo/eTuuNUwSAhFXPJlXswIw+swVWVNuPmXWdwXhUFiFgpNy7KvO5+xZij1L6TYPeGpQ4QiHHC/EnCm025mwFPEbXhICEQ+qD6bXeKmRkhCIuHpOYc0KwIRznvx6nXLw26XCdZmesr4TPYICFQpOSqEyL6WoYSgKgIKTUqmKdtJwYV2cuN0RxwvbFRY7bE7I5XhBGnjXZDL9e3aNl5oULt+d8uqNmi3Pyb5/GlC4pAizvitohcIqFJyUY19nL3XlyhVBELZt28Zi9bpi8Nhts54o1W7fNWbiOLy3HNVKhY9wf7euLB+y5k7gXZPJNGRejZeShUBECFXdNGO3l55fs++fRiS6F2PW151SyinVFXiU2qwn1tfTS23evLlFixbPPPPMk08+2bRp0xUrViiR6+p81V5pAQBFAVBwenDwT488jZ2K0uawS6Oybq2t+O/tcPIY6veLtuFaq2+HRpTK5s8BERwvnC+qtb+ftIB70sz6ruAZCqtQcFKOfdX5XOV3qdLSUqKSxx9/PD09HZ8KgtCuXTtyyc0J1V5pwQNFAVBwUipVi8mUyoy9ve0ev9l+0MJf92RYQxqmkBZMJtPopTVeautxO/tHLDt0dmzoCZdGiydgHCSY9R2Q4/QlKKxCwUk59lXncxUv9eijj4aGhmI1PPXUU4cOHcLp5cuX34qO6LRKKCuq9kpL+1AUAAUnpVK1mEypzPDbW5h/seqI3TJ4N6Nx2+8EizKZTBNX1nip3bcD+Nqt69lMZn1X8A+FVSg4Kce+6nyu4qWys7PffPPN9957Ly8vLyoqqn379m3btn3wwQfbtGkTFhbmCgFpaVO1V1oagaIAKDgplarFZEplSDiod+cftlsmcNtJjhemRZ4iV00m05zfQ/Ebv7gs+ysDSWEPJpj1XUE+FFah4KQc+6rzuYqXwhIJCgrq0qXL3Llzy8vLU1JSEhMTKyoqXKEejW2q9kpLO1AUAAUnpVK1mEypzD8X14TW/euUO7tLSAvzm5M5Xpi3906wKJPJ9Nv6Gi+VmlcrBKK0osfTzPquMAEUVqHgpBz7qvO5Ji+FELp27doXX3xhNBqTkpJcoZs6tanaKy2tQVEAFJyUStViMqUyOCAhxwvdAsLt/pL0bdBxjheWHboThsNkMq3dVOOlPL5EQqlfHqTUASSlS0yoSsw4nV9PKFWdz9W9VFhY2IwZM3bv3o0QOnDgQI8ePX788cfy8nKnqaevqNorLbeAogAoOD04pb44VQzujv+VVt60tf5/VosR4qVhOEwm05aQ0B5jIx8dHV5WZaeKbSMeyWHWdwXtUFiFgpNy7KvO5ype6n//+99DDz308ccfd+7cefz48eKnJFVVAQEB3bp1Cw8Pd4WAtLSp2istjUBRABSclErVYjKlMr3G7SRe6lzhjcTz175em5BTeGcR+b+XxXG8IF3Lh1lNOHvlaLaHd5dQ6hTOZ9Z3zI9zV6GwCgUn5dhXnc9VvNQDDzyQkJCAECoqKuratSvRRGpq6osvvkhO3ZxQ7ZUWPFAUAAUnpVK1mMxuGYvF0nlkGMcLj44O53gh4VzRV78ncLwg/Xbqw4XRHC9EnLhIWoDCKhScnrI+MWidElBYhYKT0vqq87mKl/rjH/+4adMmhFBSUtLjjz9eJym4rrBqr7TcGooCoOCkVKoWk9ktU1ZVEy2+36yDVld06YXJ4gvAoeuOkfJv/hrF8cJ+yQbnUFiFgtNT1icmrlMCCqtQcFJaX3U+V/FSa9eubdq0abt27Vq0aEE+nKqTIFxRWLVXWm4KRQFQcFIqVYvJ7JbJL6ngeKHzyLAvVh3BC/nw27+359Z824cQ6jt9P8cL8WfvvNyDwioUnJ6yvl1JqGZCYRUKTkrrq87nKl4KIVRYWHjkyJFr166p2t5tBVR7pQUJFAVAwUmpVC0ms1sms0CMw/vEzzv9t4jLzfHLPY4XegZGkvV+z03aw/GCdOtYKKxCwekp69uVhGomFFah4KS0vup8ru6lVE2uWmD+/Pkcx/n6+hqNxvj4eNvyv/3224svvtjSerz66qt2y0hrqfZKWlgpDUUBUHBSKlXJTKr5x3Kucrzw1yl7Z+wUd/Pr5F+z2I/jhaIbNUER8fIKafQpKKxCwekp66vKw24BKKxCwUlpfdX53OVeKjg42MfHZ8WKFampqYMHD27ZsmVBgTz+6SeffLJgwYLExMRTp059+umnLVq0uHDhgl154UzVXjmoSy5BUQAUnJRKJXapa+JA+mWOF/rPiVoVnU1W+uHEsZyruDW8sCL36p1Vf1BYhYLTU9avq1pweSisQsFJaX3V+dzlXspoNA4dOhSLw2w2t2/ffvLkyQ60VV1dff/9969evdpBGdVeOahLLkFRABSclEoldqlrYkdynhjhcHGMkCyGOsT/8I7mW46J28hWmy04kzxaeQpqXbsGCCeD6oRxVavUk7GvOp+71ktVVVU1atQoJCSE2GPQoEEDBgwgp7aJ69evN2nSZMeOHbJLlZWVJbeP3Nxcg8FQWFhoojjKyspCQ0PLysoo2nBHVSg4TSaTR6D+HnOW44XPVsQfSs8nXurr38XPeKdHpJlMpmul5Ti/5EYFMZhHoJK7a09Awekp62tnUloSCqtQcFJav7Cw0GAwlJQoblTmWi+Vl5dnMBhiYmKIyxkxYoTRaCSntomvv/66c+fOtvsEBgYGGmofQUFBoeyo9wx8t1Dc2vzD6duX3N6X7y+BO75dIGZ+MH17aGjo77d3QtoaUu/JYgQwBryPgaCgIEheavLkya1atUpOthMliD1LSf8Y9M60O//0K7petif1YkVl1dTwNI4XRm9NvlJShp+ZvlmbICSJIeTfmRdlMpnOFpTgLf6kpLkTqvS+dU1DwUn513RdaaEsD4VVKDgpre/hZ6k6vfGbPn16ixYtjh69E/bb9kkL56i+x1SqKM2H8s4XCk43/DIRnXnl7JUb2IijtqZwvBCaeIFE5bBYLF1HidtPLDl45tQl0TP1GrcTIZSeLy5Vf2r8LmZ9KQO6p5lQGaUmk8kJElTnc9e+8UMIGY3GYcOGYehms7lDhw52V09MnTq1efPmsbGxWjqp2istjUAZVFBwutpLRZ68xPHCm79GYeMOWh7P8cK47anDNyRyvLDowBmE0F+niLtOxGYVkg0prpVVJZ2/xvHCC5NrBfWAwioUnK62vpYRrb0MFFah4KS0vup87nIvFRwc7Ovru2rVqrS0tCFDhrRs2TI/Px8hNHDgQH9/fyysKVOm+Pj4bN68+dLtQxrJ3lZ8qr2yrWKbA0UBUHBSKtXWQNKc8qpqvO8RfjxCCP1jkbg138Dl8Xi/87Vx5xBCe9LyZ+1KN5st4p9HE3dzvJB0/lpsViHHC6/M2C9tEAqrUHC61PpSw+mShsIqFJyU1ledz13upRBC8+bN69ixo4+Pj9FojIuLwzrr06ePn58fTnMcV3thhCEwMNCBHFV75aAuuQRFAVBwUiqV2MVuYnqk+MUuxwtdRtVEiO4/R9ya74XJez9eIoZA3JaUJ6v4kTU0YmjihX2nCzheeGtuzUMYLgaFVSg4XWp9mWXpT6GwCgUnpfVV53N3eCl6VclaUO2VrLzdUygKgIKTUql2bYQzs6/cwD84YUdVebMaIfS3qfvwKd6jb59kJ1lca8SmJI4X5uzOCEsRv6P6x6Jo6S2gsAoFp+usL7WaXmkorELBSWl91fmceSlnfu7Ta7RoaaeeKNUBFQv3n8Hf7WK3dNW671HvCbvwKf5vwrk7O8nipubvy+R4YXhw4uYEcb3fwOW1tuaCwioUnJTzlAPru+ISFFah4KS0PvNSiiKHogAoOCmVqmgnhCaHn+J4YcKOVLzREQ4A3y1AXM5H/qXnX5e1gLeieH/B4d9jz3G8MGRNrbWjUFiFgtN11peZVZdTKKxCwUlpfealFFUNRQFQcFIqVdFOCOFF57N3pz89Xnx+OnWphGx6RLzUxeJyWQsnLhRzvNB7wu6lUVkcL/x3/XFpASisQsHpOutLraZXGgqrUHBSWp95KUVhQ1EAFJyUSlW0E0LfrT/O8cLSqCz8W1TCuaKSChPxTzhRWnlT1kJxeU0Z/Cjmv6XWp+JQWIWC03XWl5lVl1MorELBSWl95qUUVQ1FAVBwUipV0U4Ifb5SDG+44ch5vK7vQPrlS8Vi5EPyr5O/QEJJSdt54uedHC/g6uO2p0ovQWEVCk7XWV9qNb3SUFiFgpPS+sxLKQobigKg4KRUqqKdEPpokbjWPCzlIv5GKjzlYmZBKccLj9wOJfV4YKTd6m/PPYRDT3G8MC3ylLQMFFah4HSd9aVW0ysNhVUoOCmtz7yUorChKAAKTkqlKtoJIfwIdTD9Mt5vYuPR83g7iecm7XlsTITtvhKkqW/WHiPPW/P2ZpB810GV3kKXNLO+LjTKGoHCKhSclAOKeSmZPu+cQlEAFJyUSr1jGJvUi1PFfY+O51zFXmfl4bPRZ65wvPDazAPvzBOfll6ffdCmkpiBf5HCjmppVJa0DBRWoeB0nfWlVtMrDYVVKDgprc+8lKKwoSgACk5KpSraCSESDx5/qDt/X+auVDGU1ID5h4cHi5v4fbToTmgYaTtB8TnkWWpdXI70EhRWoeB0nfWlVtMrDYVVKDgprc+8lKKwoSgACk5KpSrZyWKx/GlkGMcLl4orft5+kuOFKRGnQhMvcLzwydJY/OnuF6uO2K1+OFN85ML/th4X4/aSAwqrUHC6yPrEXvomoLAKBSel9ZmXUpQ3FAVAwUmpVCU7VZiqsZsprbw5Y6e4m9/Y0BNr48QPdf+z+uj5orIPF0ZHnrxkt/r5opqIUxwvRJy4KC0DhVUoOF1kfanJdExDYRUKTkrrMy+lqG0oCoCCk1KpSnYquC4uOn/EutYcb5X0vw1Jvx0UP9T9PjhRqRbOv1lt7mx9DuN4YX/tjf6gsAoFp4us79i+Tl+FwioUnJTWZ15KUclQFAAFJ6VSleyUdVlcdN7TutZ8TUz2rVd/X65JmLUrXYzPG5KiVIvkk01p47IKSaaLoErb1ytdz62vF42ydqCwCgUn5YBiXkqmzzunUBQABSelUu8YpnYqOVeMYfj8pD0IoS3HxH1j/70sbsKOVI4XJoWl1S5r5+z/LY3DLwxTcoull6GwCgWni6wvNZmOaSisQsFJaX3mpRS1DUUBUHBSKlXJTtHWFRB/n3UAIYTD9b634LD/FjGc/K97an0CZbcFXJLjhcyCWtvRQmEVCk4XWd+uTekzobAKBSel9ZmXUpQ0FAVAwUmpVCU7RZwQA8l/sFCMDoXX7PWbdfDboJqd/ZRqkfxFB8SoHxwv5F4tI5kugiptX690Pbe+XjTK2oHCKhSclAOKeSmZPu+cQlEAFJyUSr1jmNqpTdboUIOs0aESz4tv/16YvPeLVeLOfuvja30CVbtezRkOgcjxQmFppbQAFFah4HSR9aUm0zENhVUoOCmtz7yUorahKAAKTkqlKtlp5eGzHC98s+4YQigj/zrHC73G7VSKIm/bCI7fwfFCeZUY4ZccUFiFgtNF1if20jcBhVUoOCmtz7yUoryhKAAKTkqlKtlp7p4Mjhdw3I28a+UcL3QZFYY3Rtp7Kl+pFskvrbzZfUzE0+N3mc0WkukiqNL29UrXc+vrRaOsHSisQsFJOaCYl5Lp884pFAVAwUmp1DuGqZ2aFJbG8cIvghh3g4SMwjv7xdZeXF673p2z9PzrWZdL75xbU1BYhYLTRdaXWU2vUyisQsFJaX3mpRSFDUUBUHBSKlXJTiO3isv55uwWl/PdrDbjpRBdR4vh5E9cqLW4XKkFu/lQWIWC00XWt2s7+kworELBSWl95qUUJQ1FAVBwUipVyU7DrMv5lh06iwt0CxD9E/5n+4Sk1IhtPhRWoeB0kfVtDadLDhRWoeCktD7zUoqqhqIAKDgplUrsVF37B6RPV8SLgXqPnscFek/YRbxUQUkFqVXXBBRWoeDUy/p1taNz5aGwCgUnpfWZl1KUMRQFQMFJqVRsp0vFFb3G7QwIOUHMRuLz4hyy4xHHCzcqb5JidU1AYRUKTl2sX1cjOl0eCqtQcFJan3kpRSVDUQAUnJRKxXYSki9yvNB7wm5ittdnH+R4ISrjMs7BcXvx45Rs2R6poiUBhVUoOHWxvhbD6VIGCqtQcFJan3kpRVVDUQAUnJRKxXZafkj8OorjhZIKE855YbIYqDfx/DV8ih+tOF54bEyEomk1XIDCKhSculhfg930KQKFVSg4Ka3PvJSirKEoAApOSqViO+F15xwvJN12S0/8vNO6C1/NUnI/689UHC/85Zc7z1uKNla+AIVVKDh1sb6yuXS+AoVVKDgprc+8lKK+oSgACk5KpWI7fbde3KDv1gYTIccvIIQsFgsOEJV/e6HEN+uO4QIvT9+vaFoNF6CwCgWnLtbXYDd9ikBhFQpOSuszL6UoaygKgIKTUqnYTnjrI44XZu5KRwiVVd3EPokslBixKQnnvPlrlKJpNVyAwioUnLpYX4Pd9CkChVUoOCmtz7yUoqyhKAAKTkqlYju9PH0/dkJDrRv3FZSIgXo7WQP14gI/bz+JC3y0OEbRtBouQGEVCk5drK/BbvoUgcIqFJyU1mdeSlHWUBQABSelUvH7vT8HREgflc5YA/U+bg3Uiw05Y+dpXODTFfGKptVwAQqrUHDSW1+D0XQrAoVVKDgprc+8lKKyoSgACk5KpUq36bv1GW/3MREWi4WE6iBWJPGi8MMWya9rAgqrUHDSW7+uFqQpD4VVKDgprc+8lKKYoSgACk5KpZLAHD3GRuIVE5eKKw5lXOF44fXZB4kV18Sew89S/OZkkulEAgqrUHDSW98JIzpdBQqrUHBSWp95KUUlQ1EAFJyUSkUIRWVc5njh77MO4F+nos9ciTghfuT7oTVQLzbk1uO52EuN2y7uku70AYVVKDjpre+0KZ2oCIVVKDgprc+8lKKGoSgACk5KpSKENh49z/HCv5fFfbZSDMX7e+y5DdYcP8lPUJEnxQDz4iLAnacVTavhAhRWoeCkt74Go+lWBAqrUHBSWp95KUVlQ1EAFJyUSkUIzd+XyfHCDxuTJuxI5Xhh/I5UvBWF9Ceow5niO0COFxYdOKNoWg0XoLAKBSe99TUYTbciUFiFgpPS+sxLKSobigKg4KRUKkIoIOQExwvTI0+vjRN/fPp0RfyvNYF6U4gV8XoKjhfWxJ4jmU4koLAKBSe99Z0wotNVoLAKBSel9ZmXUlQyFAVAwUmpVITQf1Yfxe4n5kwhxwt9pu2baA3UOzEsjVgxs+A6fpbaciyXZDqRgMIqFJz01nfCiE5XgcIqFJyU1mdeSlHJUBQABSelUhFC78w7xPHCzpOX8q0f83YeGfbvZXEcL/y6RwzUi4+LxeXYS0WevHQ7z5n/Q2EVCk566ztjRWfrQGEVCk5K6zMvpShkKAqAgpNSqQgh48TdHC8k516zWCz9ZokBO/C/FYdrAvVKv6k6lHFF0bQaLkBhFQpOeutrMJpuRaCwCgUnpfWZl1JUNhQFQMGpRakhxy90Cwg/nGnHwdysNnfyF90SjsBbUFIxYP5h7KU23g7UixC6WW3GmcdzriqaVsMFKKxCwanF+hrM4qYiUFiFgpPS+sxLKeoeigKg4NSi1C/XJNyKHTV8Q6KtVS4Vi1v2dR4ZRiLKV5iqf9yYZJy4++yVG9Ly3QLCOV5Iz78uzaxrGgqrUHBqsX5dbeS68lBYhYKT0vrMSylKHYoCoODUolQcabfvDDtBN5LOX+N44dmJe2QGs1gsspzRISkfLow2VZtl+XU6hcIqFJxarF8nA7m0MBRWoeCktD7zUopqh6IAKDhVlWqxWHqMjcTv666VVckMgz/XHTD/sCzfRadQWIWCU9X6LrKjc81CYRUKTkrrMy+lKGMoCoCCU1WpRTeqsIu6tcp8/+kCmWFWx2RzvDBkzVFZvotOobAKBaeq9V1kR+eahcIqFJyU1mdeSlHGUBQABaeqUskHuRwvzLIGOZTaZtTWFI4XxoaekGa6Lg2FVSg4Va3vOlM60TIUVqHgpLQ+81KKGoaiACg4VZW6LSmPPEsNXF4rOlRhaSVeE3Ew/bKiwXS9AIVVKDhVra+r9Wgbg8IqFJyU1ve8l5o/fz7Hcb6+vkajMT6+1tyEtXby5MkPPviA4ziDwTB79mwtAlTtlZZGoCgACk5VpeJt+l6fLX4I9XhgpNl8Z1nEtMhTHC+8M++Q7VoJLaZ0ogwUVqHgVLW+EzZy9NOuEgAAIABJREFUXRUorELBSWl91fnc4DopIISCg4N9fHxWrFiRmpo6ePDgli1bFhTIf5A4cuTIjz/+uH79+ocffph5KVtz3DVK/WlTMscLM3aexo9NZy6X4s4Wl5t6WldVUG4nYUudgxworELBSTlPObCUKy5BYRUKTkrre9hLGY3GoUOHYp2Zzeb27dtPnjxZSXYcxzEvZUvOXaPU/1sSy/HC1uO5/1gUzfHCpoSajfjm7c3AYaWkT1e2POibA4VVKDgp5yl9javaGhRWoeCktL4nvVRVVVWjRo1CQkKIaAYNGjRgwAByKks49lKVlZUlt4/c3FyDwVBYWGiiOMrKykJDQ8vKyijacEdVKDhNJpNjqM9P2sPxQtyZy+O3i3uf+29OwvT1m3XgVgj5DfHn3MHm7Xs4hnq7lOf/DwWnqvU9T6UEARRWoeCktH5hYaHBYCgpKZF5BHLqwjd+eXl5BoMhJiaG3GzEiBFGo5GcyhKOvVRgYKCh9hEUFBTKDiAMbN4a+gi/Q9zyfGNo4PJtHC/8bfyO0NDQrSGhf/IX85cFA+kJg8kYYAzoykBQUNBd4qXYs5Tkb0EvTTr40y/jkri1xJ8DwquqqlIvXOV44bGxEVVVVWcLSjhe6DIqrKKyyp29cgDVnTBU7wUFJ+Vf06o86FsACqtQcFJa35PPUvq+8ZM+eKm+x5QWVkpDeecLBafjd9MH0i9zvNBv1kGEUNVNc+eRYbe+4b1YXI7zX5t5QMlMLsqHwioUnI6t7yIjOt0sFFah4KS0vup87sI3fmIsBqNx2LBhWExms7lDhw5s9URdh9bdodQ1sWL43S9W1Wwt0XfGfo4XojIurzx81p1bThDyobAKBSflPEXs4p4EFFah4KS0voe9VHBwsK+v76pVq9LS0oYMGdKyZcv8/HyE0MCBA/39/bEiq6qqEq1Hu3btfvzxx8TExMzMTMdiVe2V4+r4KhQFQMHpWKk46u647amY/MHWsLwrDp8dGyqupJgcfkqLyXQsA4VVKDgdW19Hw+nSFBRWoeCktL7qfO7aZymE0Lx58zp27Ojj42M0GuPi4rDI+vTp4+fnh9PZ2dm1V0UY+vTp41iLqr1yXB1fhaIAKDgdKxXH7Fh5O57h1AjxM95RW1NwNN4NR85rMZmOZaCwCgWnY+vraDhdmoLCKhSclNZXnc9d7qV0UZWsEdVeycrbPYWiACg4HSi16qb5b1P33VrXt/eU+CSNENpyLJfjhX8ujnlh8l6OF45mF9m1kesyobAKBacD67vOiE63DIVVKDgpra86nzMvZXJa6+6peBcodaR1J9megZFXb9QE7EjOFZf8PR4Y+Yg1Pm9haaV7yCR3gcIqFJyU8xSxi3sSUFiFgpPS+sxLKcoeigKg4FRSalB8DscLj/gL+yTROm5U3iQ7zz7x8063bd9H1ACFVSg4laxPCPeqBBRWoeCktD7zUoqjA4oCoOC0q9QrpZVdR4sB4Ofvk6+IwVtRcLzwrrsiH0qlAIVVKDjtWl9KuFelobAKBSel9ZmXUhwdUBQABaddpW49Lv7+9Prsg7ZPS3jdBMcLwzckKhrJZRegsAoFp13ru8x6tA1DYRUKTkrrMy+lKGgoCoCC065S/7chSWmh+c/bT+KXfvP2ZigayWUXoLAKBadd67vMerQNQ2EVCk5K6zMvpShoKAqAgtNWqRaL5dmJ4g6zhzKu2JphbZz4nS/HC0LyRdurrs6BwioUnLbWd7UFadqHwioUnJTWZ15KUcxQFAAFJ0LocEa+/2/bKiprFvJlFpRyvNB1dHiFqdrWDHFZhdhLpV1U3AvZtpZeOVBYhYKTcp7Sy6wa24HCKhSclNZnXkpRt1AU4G04I05c2nDU/ke4z/yym+OFz1fGl1beRAitis7meOGTpbF2bXD1RlXnkWFdR4WXV9nxYXar6JjpbawqdQ0KTsp5Sqn7LsqHwioUnJTWZ15KUedQFOBVOE3VZrxm74rNF07lVdVkcfnrsw/mXi37YtVRjhcW7Jev7iMmCU+5uDu15jtfkumehFex6qDLUHBSzlMOGHDFJSisQsFJaX3mpRRFDkUBXoUzp7AMu6KU3GIZs9lXbnC88Cf/Hb0n7OJ44enxu7qPieB4wbakrKJHTr2KVQcMQMFJOU85YMAVl6CwCgUnpfWZl1IUORQFeBXO6Mwr2EvtSROfgXKvln29NuFYzlWEUKz1d6a/BO7IuXL9jTlRuFivcTurzRZFG3juglex6oAGKDgp5ykHDLjiEhRWoeCktD7zUooih6IAr8K54ch57H7Wx+cghObuyeB4Yei6Ywih0MQLt5bzvTpxuxgSrermkDXi6z6PfAulaHLJBa9iVYJLnoSCk3KeknfbxedQWIWCk9L6zEsp6h2KArwK54ydp7GX+nWP+JGT/5YUjhcGWDePWHLwzK33ex9MF70UQshsthzNLsLLKBRt4LkLXsWqAxqg4KScpxww4IpLUFiFgpPS+sxLKYocigK8Cuf3wYnYS40OSUEIfboiHv8EhRAatz1VXOD3a42XUuTdOy54FasOKIGCk3KecsCAKy5BYRUKTkrrMy+lKHIoCvAqnP9YFI291JA1YtTd12cfxKdlVTe/WXuM44XvF23Dz1KKvHvHBa9i1QElUHBSzlMOGHDFJSisQsFJaX3mpRRFDkUBXoUT7yXB8cJ7Cw4jhHqN24m9VHr+9Q8Xig7s5+XMSylKzokLXmV9x/gZVMf8OHG1nlDKvJSiNqAowHtwVt6sxuGgOF7465S9FaY7H0jtPZX/1yliPMPZv4eyZylFzdX9gvdYXxU7g6pKUV0L1BNKmZdSFAYUBXgPTvxFFH546jo6/Kz1Ayl8uvLw2a6jxAgdKzYwL6UoOScueI/1VcEzqKoU1bVAPaGUeSlFYUBRgPfgjMq4zPHCi1PFZyaOFyJOXMIJjhfw3uccL2zeyryUouScuOA91lcFz6CqUlTXAvWEUualFIUBRQHegxNH3f10RfzjgZEcL0yLPEW8VJ9p+/Biv9BQ5qUUJefEBe+xvip4BlWVoroWqCeUMi+lKAwoCvAenNgtBYSceHXmAY4XcBhDsoCC44X+sw8yL6UoOKcueI/1VeEzqKoU1bVAPaGUeSlFYUBRgPfg/DboOMcLSw6e+b8lsRwvPPGzuMAPbymLH6r8lscxL6UoOKcueI/1VeEzqKoU1bVAPaGUeSlFYUBRgPfgfH/BYY4XwlIufrdedFf436ID4pYT+N9Pm5KYl1IUnFMXvMf6qvAZVFWK6lqgnlDKvJSiMKAowHtw4vBRKbnFE3aI20zgf3vS8slLv5k7TzEvpSg4py54j/VV4TOoqhTVtUA9oZR5KUVhQFGAl+AkX0ddvVGFt+zDXupkXvHbcw/h9NqYs8xLKQrOqQteYn0t2BlULSzVqUw9oZR5KUVVQFGAx3HmXi0bsSnp85VHOF54bEyExWLZejyXPEsV3aj66vcEfLr7ZB7zUoqCc+qCx62vHTWDqp0rjSXrCaXMSynqAYoCPI5zTOgJ4pPet26MdPh2lKmuo8MtFssvQs0LwJTzRcxLKQrOqQset7521Ayqdq40lqwnlDIvpagHKArwOE68pezokJTVMdnni8oQQun517Hf+tvUfQih1THZ+DT/2g3mpRQF59QFj1tfO2oGVTtXGkvWE0qZl1LUAxQFuB9ncblpeHDimNATFovlWlkV9kCXr1cSKknmR4tjEEJ7T+VzvNB1dHhVVRXzUoQlXRLut77TsBlUp6lTqlhPKGVeSkkACIoC3Izz7JUbfWfsx57pZF7x7lTRA/WdsV/Ko8Viwbv2fbf+OELoUnFFl1Fh7y047GaoUkh1TUOBCgUnZeyGupqPsjwUVqHgpLQ+81KKeoaiAPfgDE+5+PT4XT3HRv5pZBh2URwvTA4/NTEsjeMFfnOyjMcXJou7+U0KT8P554vKrleY3ANVhsS5UyhQoeCknKecM6LTtaCwCgUnpfWZl1JUMhQFuAGnqdqMvQ72T+8vOLwqWvyp6a9T9g6YL37Ju+VYroxHnL/y8FlpvhugSm9Hk4YCFQpOynmKxpRO1IXCKhSclNZnXkpRw1AU4AacmxLEleW9J+zOLLieU1hmsVjKq6ofGxNBHqrwogkplTN3pXcLCE/NK5FmugGq9HY0aShQoeCknKdoTOlEXSisQsFJaX3mpRQ1DEUBuuCsMFWXVt60y0W12YJ/iFp04Iy0ANkG6flJeywWi/QSTpuqzbJMXaDK2nTRKRSoUHBSzlMusrJSs1BYhYKT0vrMSykJtb6snjhzuXRM6IkeYyOfHLezoKTClg4h+SLeOlbmxvC6CY4X8BIJ24q2OfVkUNl23HU5jFJXcAuFVSg4mZeyo1JV32unjk0WFAXQ4LxYXI4X4+F3d0ujsmQ0WCyW/nOiOF6YtStddqnyZnVPaxyptXHnZJeUTmmgKrXponwoUKHgpJynXGRlpWahsAoFJ6X1Vedzg5IhvTlftVdawENRAA1O/IPT36buG7k1heOFAfMPy5jBnzp1HxNx9UaV7BJCKCg+Z9Dy+OIyk+0luzk0UO026LpMKFCh4KScp1xnaLstQ2EVCk5K66vO58xLaZ2C7crdDZk0Sh2xKQkvKL98vbKTv7iLeU6huHkEPiwWCw7GMTGsZkH57StO/p8GqpO3dLYaFKhQcFLOU86a0cl6UFiFgpPS+sxLKeoYigJocOJA73tP5SOE/t/SOI4XFuzPJIxEn7mC94yw+3sVKaY9QQNV+110KQkFKhSclPOULjbV3ggUVqHgpLQ+81KK0oWiAKdxFpRUcLzwiL9QXC4+L66PzxGDvs+JIoxgvxUQcoLkUCachkp5XyeqQ4EKBSflPOWEBWmqQGEVCk5K6zMvpShmKApwGueO5DypW7pWVoX3lcgsuH7r0SqzQNwx9k8jw3Kv3nkHqEiWtgtOQ9XWvJ6loECFgpNyntLTtBragsIqFJyU1mdeSlGzUBSgirOs6ubs3ekjNiWN2JS0KjqbdHisNeJG4LaTJOcza4yoqRGnEEIzd6VzvPDFqiPkKn1CFSr9LfRqAQpUKDgp5ym9zKqxHSisQsFJaX3mpRR1C0UBjnHmXSt/w7qUnOwTcTznKu4zjrghJF8kFISniJ9G9Z6w21Rt7jtd3FI2NPECuUqfcAyVvn0dW4ACFQpOynlKR8tqaQoKq1BwUlqfeSlF0UJRAMZ5rbTcf0vK7lRxHQQ5Mguu956w2+p4ds3dk/HR4hiOF0ZsSkIIFZeZHrEu6iu4fudLXlO1GZefsfM0xwuPjg6XfclLWnYuAYVSykHlHDnO1WKUOseb41pQWIWCk3JAMS+lKFcoCsA45+8VX9A98fNOqV/5Zu0xjhden30Q/7Z0NLuI44U/B0SUVJh2nrwkRtyYXiviBkJoasQpvKSC44Wv1yYosuPUBSiUUg4qp7hxshKj1EniHFaDwioUnJQDinkpRbVCUYDJZNoaEvq3qWKkDI4XyG57l69X4tUQZMtXi8Xy2swDHC/8uifjRWv5MaHy9Xvni8rIu8HwlDsvAxVpqssFKJRSDqq6UEJbllFKy6C9+lBYhYKTckAxL2VPpNY8KAowmUxTV28jrqX3hN0VpmqE0ML9ZzheeLf2dhIrDp8lJV+cuvdamZ0dJf69TPxw6rExEbgdRYLqfgEKpZSDqu7EOF+DUeo8d8o1obAKBSflgPIKLzV//nyO43x9fY1GY3x8vF3xbNy4sVu3br6+vj179gwLC7NbhmSq9oqUdJCAogCTyfTutO0cL/hvScZRoFZFZ5vNlr9N3cfxwoaj56V9LC4zPTo6HL/3S7tYK6wGKRaVcbmTvyBd+0cuUSagUEo5qChZqlN1Rmmd6NJYGAqrUHBSDijV+dzlOyQFBwf7+PisWLEiNTV18ODBLVu2LCgokIkpOjq6UaNG06ZNS0tLCwgIuOeee06ckL+qklZR7ZW0sFLamxVQXlUdfCTnnXmHjBN3B2xN7uy/g+OF1LySNTFicMJe43bym5M5Xug5NrKsSh6P4+ftJ7uMCtuRnKfUcYRQYWnlTZu4Gw7Ka7zkzZTKugAFKhSclPOUzDquPoXCKhSclNZXnc9d7qWMRuPQoUOx7Mxmc/v27SdPnixT4T//+c+33nqLZD777LNffvklObVNqPbKtoosp/JmdfblkhUbQrMvl+ReLXPFv+M5V3/alNx9TESfaft+O5gVfCTnrblRXUeFf/V7wuHMK0p3TM0rGb8j9XHrZuTk3R3HC+/ME/eMqDBV95t1kOTb/uyEELJYLNIVFrKOu/S0ngwql3Ioa5xRKiNEl1MorELBCdtLVVVVNWrUKCQkhGhr0KBBAwYMIKc48cc//nH27Nkkc+zYsU888QQ5xYnKysqS20dubq7BYCgsLDQ5e8RnXSZzvXcmXpyyd+G+jIjkC4OWxf551I6IpPO4r9fLKoLjs9+ZF/XcpD2Zl4qdJcAl9crKykJDQ8vKylzSuq6NQoEKBafJZGJQdVWo2Fg9obSwsNBgMJSU2P+FAiHk2mepvLw8g8EQExNDXM6IESOMRiM5xYl77rknKCiIZC5YsOChhx4ipzgRGBhoqH0EBQWFOnvM+T20i/8Ol/57dNSO96dtn756G79k29/G73j25x3fLtg2+/fQT2Zt7z5a8dZdR+54Y/L2iSu3bQ1xtm+sHmOAMcAYgMNAUFDQXeKl9H2WAvR3Sj35e0r3v0MdNwiFVSg4AQ0oQFDrifU9/Cyl4xs/6aMV/e9SlC9SpWBcna4n76ZdTaOsfSisQsEJaEABglpPrK86n7v2jR9CyGg0Dhs2DM8RZrO5Q4cOdldPvP3222Qeef755129eoIplbCtY6KeDCodGVNtilGqSpETBaCwCgUn5XTqeS8VHBzs6+u7atWqtLS0IUOGtGzZMj9f3Ixu4MCB/v7+WGHR0dGNGzeeMWPGqVOnAgMD2Up06cCrJ0qVdtkNaSisQsFJOU+5weLSW0BhFQpOSut73kshhObNm9exY0cfHx+j0RgXF4fl0qdPHz8/PyKdjRs3Pvrooz4+Pj169GBf9RJaKM0vbccN6XoyqNzAJLkFo5RQoWMCCqtQcFJOU17hpXSUF25KtVda7ghFAVBwUipVi8l0LAOFVSg4mfV1FCdpqp5YX3U+d/nvUoRxHROqvdJyLygKgIKTzVNaVFfXMsz6dWVMS3korELBSTn2Vedz5qVMWmTtwTL1RKluZhgKq1BwUs5TzPp2Gagn1mdeyq71xUwoCoCCExClgKAy6ysOYIoLUFiFgpNyQDEvpahlKAqAgpNSqYp2cs0FKKxCwcms7wqd1hPrMy+lKB4oCoCCk81TilKjuMCsT0GeYlUorELBSTn2mZdiSlVkQPcL9WRQ6c6bgwYZpQ7IcfoSFFah4GReyo4UVX2vnTo2WVAUAAUnpVJt7OPaDCisQsHJrO8KvdYT66vO52yNH1vjp9v4qieDSje+NDTEKNVAUp2LQGEVCk7Kv1GYl1JUMBQFQMFJqVRFO7nmAhRWoeBk1neFTuuJ9e9OL1VcXGwwGHJzc2+HRXTm/4WFhUFBQYWFhc5UdmMdKDhLSkoYVN11wSjVnVJAQq0n1sdRbYuLi5U8Pcg3frhXtWMisjPGAGOAMcAYgMpAbm7uXeWlzGZzbm5ucXExzd9Z2NVRPpDRANBYFwrOkpISBlWjTbUXY5Rq50p7SSisQsFJOfaLi4tzc3PNZvNd5aWUOlOnfNWXoXVqzXWFoeBECDGousuAUao7pYCEyqyPrQ/yjZ8uwoWiACg4AQ1+QFCZ9XUZ7LJGoLAKBaerBxTzUiUyBXvbKVOqKywChVUoOF09T+mrASisQsHpauvXXy9VWVkZGBh467/6DgDdW4OCEyHEoDLre/+AAiRUNqDwgKq/Xkr3CYU1yBhgDDAGGAO6M8C8lO6UsgYZA4wBxgBjQDcGmJfSjUrWEGOAMcAYYAzozgDzUrpTyhpkDDAGGAOMAd0YYF5KNypZQ4wBxgBjgDGgOwP11EvNnz+f4zhfX1+j0RgfH687rU43OGnSpL/85S/NmjVr06bNu+++e/r0adJUnz59pJuffPnll+SS+xOBgYFSMN26dcMYKioqvvnmmwceeOC+++774IMP8vPz3Y9NdkeO46RQDQbDN998gxDyEj4PHjz49ttvt2vXzmAwhISEEPAWi2XMmDEPP/xwkyZNXn311YyMDHKpqKjok08+uf/++1u0aPH555+XlpaSS65L2MVpMpl++umnnj173nvvve3atRs4cGBeXh7BIGN+8uTJ5JJLE3ahIoT8/PykSnj99dcJDI9QihBSgirFidPTpk3DaD3CqoN5ycGQz8nJefPNN5s2bdqmTZsff/zx5s2bhPA6JeqjlwoODvbx8VmxYkVqaurgwYNbtmxZUFBQJ9ZcV/j1119fuXLlyZMnk5KS3nzzzY4dO964cQPfrk+fPoMHD750+ygp8eSXXoGBgT169LiN5dKVK1cwyK+++uqPf/zj3r17ExISnnvuuRdeeMF1XGls+fLlywTn7t27DQbD/v37sZfyBj7Dw8NHjx69detWmZeaMmVKixYtQkNDk5OTBwwY0KlTp4qKCtzl/v379+rVKy4u7tChQ126dPnXv/6lkQqaYnZxFhcXv/baaxs2bDh9+nRsbKzRaOzduze5C8dx48ePJ+QTJZMCLkrYhYq9VP/+/Qmeq1evEgAeoRQhpASVgLx06dKKFSsaNGiQlZWF0XqEVQfzktKQr66u7tmz52uvvZaYmBgeHt66deuRI0cSwuuUqI9eymg0Dh06FNNkNpvbt2/vtr/y6mSby5cvGwyGgwcP4lp9+vT573//W6cWXFc4MDCwV69esvaLi4vvueeeTZs24fxTp04ZDIbY2FhZMQ+e/ve///3Tn/5ksViwl/IePhFCUi9lsVgefvjh6dOnY66Ki4t9fX3Xr1+PEEpLSzMYDEePHsWXIiIiGjRoIH2CcTW9Upyyex05csRgMOTk5OB8juNmz54tK+POUxlUPz+/d9991xaAxymVWV+G8N13333llVdIpsdZlc5LDoZ8eHh4w4YNyduURYsWNW/evKqqinREe6LeeamqqqpGjRpJX60MGjRowIAB2ilzW8nMzEyDwXDixAl8xz59+rRu3frBBx/s0aOHv79/WVmZ25DY3igwMBC/5OnUqdMnn3yCJ6a9e/caDIZr166R8h07dpw1axY59WyiqqrqwQcfnDhxIobhVXzK5qmsrCyDwZCYmEgYe+mll7777juE0PLly1u2bEnyb9682ahRo61bt5IcVydkU7/0drt3727QoAF5yuc4rm3btg888MCTTz45bdo0p1/4SG9Rp7QMqp+fX4sWLdq0afPoo49+9dVXhYWFuDWPUyqzvrSP+fn5jRs3XrduHcn0OKvSecnBkB8zZoz0D9mzZ88aDIbjx4+TjmhP1DsvlZeXZzAYYmJiCEcjRowwGo3k1EsSZrP5rbfe+utf/0rwLFmyJDIyMiUlZe3atR06dHj//ffJJfcnwsPDN27cmJycHBkZ+fzzz3fs2PH69evr1q3z8fGRgnnmmWd++uknaY4H0xs2bGjUqBF57PAqPmXzVHR0tMFguHjxIqHro48++uc//4kQmjhx4qOPPkryEUJt2rRZuHChNMeladnUT+5VUVHx9NNPf/LJJyRn5syZ+/fvT05OXrRoUcuWLYcPH04uuSchg7p+/fpt27alpKSEhIR07979mWeeqa6u9gZKZdaXkjN16tRWrVqRl70IIc+yKpuXHAz5wYMH9+vXj/SlrKzMYDCEh4eTHO0J5qWQd3qpr776iuM4pZgr+E+YM2fOaLe060peu3atefPmy5YtcyBZ191de8v9+vV7++237Zb3Bj6lUyo4L2Uymd55552nnnqKPEjJeF6+fHnjxo3dvH+SlFIZHvy0umfPHi/3Ut26dRs2bJgMPDl1P6uyecnBkGdeipipzgkQb/yGDh36hz/84ezZs0rdu3HjhsFgiIyMVCrg5vy//OUv/v7+Dh7/3YzH9nbnzp1r2LBhaGio7SWEkDfwKZ1SYb3xM5lM77333hNPPEHeodmSfPLkSYPBIF2zaltG9xwppbaNt27devHixd7wElXpWSoqKspgMCQlJdmCxzluZtV2XnIw5NkbPyWraco3Go3kzxOz2dyhQwfvWT1hsViGDh3avn176cpj214dPnzYYDAkJyfbXnJ/TmlpaatWrX799Vf8U+rmzZsxhtOnT3vP6onAwMCHH35Y6acRb+BTOqXi1RMzZszATJaUlMhWTyQkJOBLO3fu9OzqCeyievTocfnyZQfaW7t2bcOGDaXL6hwU1uuSlFJZm7m5uQ0aNNi2bRtZkOJBSpW8lJ+fn3TNpKwLCCG3sao0LzkY8nj1BFk+vWTJkubNmzv3MF3v3vghhIKDg319fVetWpWWljZkyJCWLVuShSi2OnBzztdff92iRYsDBw6Qpajl5eUIoTNnzowfPz4hISE7O3vbtm2dO3d+6aWX3IxNersffvjhwIED2dnZ0dHRr732WuvWrfEk9dVXX3Xs2HHfvn0JCQnPWw9pLU+lzWZzx44deZ4nALyHz9LS0kTrYTAYZs2alZiYiJeiTJkypWXLlvh3lHfffVe2Ev2pp56Kj48/fPhw165d3bMS3S5Ok8k0YMCAP/zhD0lJSUSxeB1XTEzM7Nmzk5KSsrKy1q5d26ZNm0GDBhH+XZqwC7W0tPTHH3+MjY3Nzs7es2fP008/3bVrVzJp9u/f3/2UIoTsQsXklJSU3HvvvYsWLZJy5SlWleYlhJDSkMcr0fv165eUlBQZGdmmTRu2El1qSvX0vHnzOnbs6OPjYzQa4+Li1Cu4q4Tt13wrV65ECJ0/f/6ll1564IEHfH19u3TpMmLECKUfANyD9OOPP27Xrp2Pj081IzqtAAAFA0lEQVSHDh0+/vhj8gsZ/sSvVatW99577/vvv3/p0iX34HF8l507dxoMhvT0dFLMe/jcv3+/zOh+fn4IIfxVb9u2bX19fV999VUp+KKion/961/NmjVr3rz5Z5995p6veu3izM7OloEnn6MdO3bs2WefbdGiRZMmTbp37z5p0iTiEogVXJSwC7W8vLxfv35t2rS55557OI4bPHiw9G9Tj1B663MIu1AxLUuWLGnatGlxcbGUJU+xamtlPC8hhBwM+XPnzr3xxhtNmzZt3br1Dz/8oPQmQ9pBu+n6+CxllwiWyRhgDDAGGANeyADzUl5oFAaJMcAYYAwwBmoYYF6KSYExwBhgDDAGvJcB5qW81zYMGWOAMcAYYAwwL8U0wBhgDDAGGAPeywDzUt5rG4aMMcAYYAwwBpiXYhpgDDAGGAOMAe9lgHkp77UNQ8YYYAwwBhgDzEsxDTAGwDPgYCsg8H1jHaj3DDAvVe8lwAhwMQOyQOa39jaUxjLX5ebMS+lCI2vEOxlgXso77cJQ3T0M+Pn5SQOZX7p0SfdNV5mXunvkwnpiwwDzUjaUsAzGgK4MKAUyNxgMCxcu7N+/f5MmTTp16rRp0yZy25SUlL59+zZp0uSBBx4YPHiwdKe+5cuXP/bYYz4+Pg8//PDQoUNxFYPBsHTp0vfee69p06ZdunTBW32T1liCMQCaAealQJuPgQfAgAMv9eCDDy5dujQ9PT0gIKBRo0ZpaWk42FW7du0++OCDEydO7N27t1OnTnj/WYTQwoULmzRpMmfOnPT09CNHjsyePRv332Aw/OEPfwgKCsrMzPzuu++aNWtWVFQEgBoGkTGggQHmpTSQxIowBigY8PPza9So0X2SY+LEiTik0FdffUUafvbZZ7/++muE0G+//daqVasbN27gS2FhYQ0bNsQbeLdv33706NGkCkkYDIaAgAB8iiM6RkREkKsswRgAzQDzUqDNx8ADYMDPz++1117LlBz4QcdgMKxevZp04Pvvv3/55ZcRQsOHD8cJfKm4uNhgMBw8eLCgoMBgMOzbt49UIQmDwbBx40Zy2rx5c2nLJJ8lGAMQGWBeCqLVGGZIDDh44yf1Jape6vr16w68VEhICCGlRYsWJPwPyWQJxgBQBpiXAmo4BhsMAw68FH7Fh3vy3HPPqb7xe+SRR5Te+DEvBUYQDGgdGWBeqo6EseKMgToyYLsS/cqVK/h3qdatWy9fvjw9PX3s2LENGzZMTU1FCJWVlbVr1+7DDz88ceLEvn37OnfuTFZPrFq1qkmTJr/++mtGRsaxY8fmzp2LschWorNnqTqaiBX3agaYl/Jq8zBwdwEDtl/1duvWDXupBQsW/P3vf/f19X3kkUc2bNhAOutgJfrixYu7det2zz33tGvX7ttvv8VVmJci1LHE3ccA81J3n01Zj2AwIHMtMEAzlIwBtzPAvJTbKWc3ZAxYGWBeigmBMaCFAealtLDEyjAG9GeAeSn9OWUt3o0MMC91N1qV9YkxwBhgDNwtDDAvdbdYkvWDMcAYYAzcjQwwL3U3WpX1iTHAGGAM3C0MMC91t1iS9YMxwBhgDNyNDDAvdTdalfWJMcAYYAzcLQwwL3W3WJL1gzHAGGAM3I0MMC91N1qV9YkxwBhgDNwtDDAvdbdYkvWDMcAYYAzcjQwwL3U3WpX1iTHAGGAM3C0MMC91t1iS9YMxwBhgDNyNDPx/4kyjnTjnkccAAAAASUVORK5CYIKU6+UJuTfbJVNCwGe1zth0PDEfVmMIocRrt/bG5OyNyTmemG8ymSvrjIOsWc0Ylpu6N2H9qQyG5Yb/iF2kB832zyyugu6yb1RFZthyYObdrIEdyNqGpplHkxmWWx6YnpBbxrDcY3MC6ow2r2olfNKpTSFKdpHxpW/Xi11PrBdW9cJnz/lByc/nXbFi27hxI8Mwffv2dXV1jYuLk/zx+Pj4DB48uG/fvsOGDfPz8yM0ZrN5zpw5v/nNb+68885XXnklI6PFsrysrOyDDz64++67+/Xr99lnn1VV2aZg8q6gIA+EgFjytlMH9+WiyoGz/R9ujo4oyYDySsKq2WyeuhcvyD7fG282m9MLKwbOsgUUZlhuY0jm0fN5RJMNnO3/4vIQy7bkwfhrH+2MZVhu3IbIxiYTOSGDPGrgeT3OmhptT3QOw3KT95ydvAf7qH19oFVUESUME1aVEDuQRi986mhq0xGrVPqa//RUQio/n3e6YvP29u7Tp4+np2dqauqUKVP69+9fXFwswCg6OrpXr17Lly9PS0tzd3fv3bt3SkoK0CxdurRfv34GgyEpKWncuHEPP/xwXZ3NqGHMmDFPPPFEbGxsZGTko48++v777wuaFdzKAyEglrxVKQnJNqGyodH05roI2AN8fllItcj4QuZdyUeE1SPnsN6CP++zV8esxb28tjp84i5spvjoLD8wAFl18vLLK0OBbOAs//JaY0F57ePW87O1wRnEpvGJeUGn04sg9YzhwnWEUExWKazeoDU4bJNkqa1KwmpbBE5Srxc+daQtdMQqlb7mP0OVkMrP552u2FxdXadNmwagmEymBx54YMmSJQKMJkyYMHbsWFI5atSoqVOnIoTMZvNvfvObFStWwKPy8vK+ffseOHAAIZSWlubi4hIfHw+PAgICbrvttvz8fNKIuCAPhJheXFNws8rzoCHnRkXezRpt/xb7pzEs94Q19CLDcj8cSlLZfs6NCs+Dhgu5pcOsIRaJ1mRY7sn5J29U1pvNZogDAsoss7iSRMD61082VA0XrhOlOHC2/yurwsjtv/clmM04a1ppVT2p3ByaJQat3RqV47vd9rUi0AufOtIWOmKVSl+r3xFpRyWk8vN55yq2hoaGXr16+fr6ko+ZOHHiuHHjyC0UHnzwwTVr1pDKuXPnDh8+HCGUnZ3t4uJy4cIF8uiFF174z3/+gxDatWtX//79SX1jY2OvXr2OHj1KaqBQX19f0Xzl5eW5uLiUlpYaO3p9uD2GTOKdUTh2/lrk5SLNW35nY2RVbf07G7GtI8Nyxy/kAQCFt6qfmn8SrPaNRmN6vs017WjCVSBoaGj4whoEi2G5LaEZqddvDrSexo1ccLLoVjVBceQC3MhfN0XV1TeQSuWFmpoag8FQU1Oj/BWHUOqFT6PRSFnVfIRQSJ0N0tLSUhcXl4oKbI8tvjpXseXn57u4uMTExJCOv//+e1dXV3ILhd69e3t5eZHKTZs23XvvvQih6OhoFxeXgoIWK7vx48dPmDABIbRo0aJBgwaRVxBCAwYM2Lx5M7/GUvbw8HBpfXl5eRk6eo1devxRtxOd8jfzxEerjwNfUzccGzhTs16emnti+wHc8I4DBtcfT0xaa+sF+lq65/+3d+5BTR3tH19Ak2AphEtGAjQRy2UsTi1Wo7bvQFsoUEuJ2mkdmdGMncGiWK3jpXTUZl5nsNZSqbVVqaJ0BhGwr8jrFGFUkNZ7sVxUBBXRMhYvUIOAFDScX99uu5zfIeRCTnKy8PCH7tmzZ/e7n+fsfnM2m6Qo/JPD2t1F+PDd9P9GpR0u+A8++t+/+w4c+teGw+rN/z1Y+L/D1G+LJq0/vCn77/K43Pvbiqb/+/DuvP6rIAUEgAAQsCmB3NzckWts/D6xwas23l+1UfR4AdGH6Dv+usLIGVBCPrEJvhTJfoAzvibLLjlY2spF4cGqtUU+SOWdKiDlHSm8xwZIe3t7hwbB+Hxu26VIhmFUKtXSpUuxdL1e7+/vb3DzSHx8POnejBkz2JtH0tPT8an29nbO5pHKyr+/5KK0tNQOm0dgaiMx4jFBC1VadFLkFhRJhejzOORxVVYiFdjY8vLyxGJxdnZ2XV3dokWLpFLpnTt3/twYMn/+/NTUVNzDU6dOjRo1Kj09/cqVK1qtlrPdXyqVFhUV1dbWqtVqznb/8PDwc+fOnTx5Mjg4mOrt/o520/Cux0iFVt7fRmrm9xQtOilyC4qkQvT5HU3Wh15gY2MYZtu2bQqFQiQSqVSqs2fPYkCRkZEajYbAKigoCAkJEYlEYWFhAz+gPXbsWLFYHBUV1dDQQC5pa2ubN2+em5ubu7v7woULaf+ANukXLwkYh7xgZFcCSNk0+ErTQpUWndYbBl+RNVmPlUiFNzaTPbRPAeMgzNFgZSTMaYKvMiCVL5KkHkBKUPCYoIUqLTrB2PDNafP32HgcA1ZWBcZmJUAbXU7LlEGLToqmNoqkQvR5H/5WIjU+n4OxWRAvKyNhQUtWFwWpViPkVgBIuUT4OKaFKi06R85rBTC2v8efcRDmDFK4uc2hZGkZWqjSopOiqY0iqRB9S8e1yfJWIjU+n8MTm0n+/QWsjER/RbZPgVTeGQNS3pGCsQFSWj/HZovIDa1O4w5vTp0wtZlDydIytFClRSdFbkGRVIi+pePaZHkrkRqfz+GJzST//gJWRqK/ItunQCrvjAEp70jB2AApPLFZew8Yd3hzaoepzRxKlpahhSotOilyC4qkQvQtHdcmy1uJ1Ph8Dk9sJvn3F7AyEv0V2T4FUnlnDEh5RwrGBkjhic3ae8C4w5tTO0xt5lCytAwtVGnRSZFbUCQVom/puDZZ3kqkxufzEfTEptPpEELNzc3//PKoxf+3trbm5ua2trZafKXdLwCpvCMHpLwjbW9vp4UqLTpHDlL8w9E6nc6gg44gY8Mg/v/PjsIREAACQAAI0Eqgubl5pBubXq9vbm7W6XRDfuGJrdGaZ74hN23phSDVUmImywNSk4iGUIAWqrTobG9vp0WqlTp1Ol1zc7Nerx/pxmaw/xZlGl/VtagqWxcGqbwTBqS8I2UYhhaqtOgEpPguHUFLkdYPS7i5rWc4sAZaqNKik6KpjSKpEP2BI9fKHJsiBWOzIDo2jYQFOswoClLNgGRZEUBqGS/zStNClRad8FoB33dgbOaNv79K/fHHH1qt9s9/LbhGoKIglXfwgJR3pAzD0EKVFp2AFN+lYGy2GK1QJxAAAkAACAhGAIxNMPTQMBAAAkAACNiCABibLahCnUAACAABICAYATA2wdBDw0AACAABIGALAmBstqAKdQIBIAAEgIBgBMDYLED/9ddfK5VKsVisUqnOnTtnwZW2LLpx48YpU6a4ubnJZDK1Wl1fX09ai4yMZH9Vzvvvv09OCZLQarVsPaGhoVhGd3f3kiVLvLy8nnrqqTlz5ty5c0cQeexGlUolWypCaMmSJQzDOALSioqK+Ph4uVyOECosLCSy+/r61q9f7+vrK5FIoqKirl69Sk61tbUlJiY+/fTTHh4e7733XkdHBzll04RBqb29vWvWrJk4ceKYMWPkcvn8+fNv375NZHDIf/rpp+SUTRMGpTIMo9Fo2HdCbGwskSEI1cF0skXi9ObNm7FUQZAamZeMjPdbt27NnDnT1dVVJpOtWrXq8ePHhLZFCTA2c3Hl5eWJRKI9e/Zcvnw5KSlJKpXevXvX3IttWS42Nnbv3r2XLl2qrq6eOXOmQqHo7OzEDUZGRiYlJbX889fe3m5LIabr1mq1YWFh/8hpuX//Pr4mOTn5mWeeOX78eGVl5fTp01966SXTddm4xL1794jOo0ePIoTKy8uxsQmOtLi4eO3atQcPHuQY26ZNmzw8PA4dOlRTU5OQkBAYGNjd3Y05xcXFTZo06ezZsz/99FNQUNC8efNszO/v6g1K1el00dHR+fn59fX1Z86cUalUL774ItGjVCo3bNhA4JObmRSwUcKgVGxscXFxRM/vv/9OBAhCdTCdRGFLS8uePXucnJwaGxuxVEGQGpmXBhvvT548mThxYnR0dFVVVXFxsY+Pz8cff0xoW5QAYzMXl0qlSklJwaX1er2fn5/dXkuaK5Fh7t27hxCqqKjAl0RGRi5fvtz8y21dUqvVTpo0idOKTqcbPXr0gQMHcP6VK1cQQmfOnOEUE/Bw+fLlzz77bF9fHzY2x0HKNra+vj5fX9/PP/8cg9LpdGKxeP/+/QzD1NXVIYR+/vlnfOrIkSNOTk7shyQ7sGVL5TR3/vx5hNCtW7dwvlKpzMjI4JSx5yFHqkajUavVAwUITpWjk61QrVa/9tprJEdwpOx5ych4Ly4udnZ2Jgs2O3bscHd37+npIR0xPwHGZharnp4eFxcX9rLPggULEhISzLrYjoWuXbuGELp48SJuMzIy0sfHx9vbOywsLDU1taury45aDDSl1Wrx6lNgYGBiYiKey44fP44QevDgAblAoVBs2bKFHAqb6Onp8fb2TktLwzIcCil7amtsbEQIVVVVEVwRERHLli1jGCYrK0sqlZL8x48fu7i4HDx4kOTYIcGWymnu6NGjTk5OZDlBqVSOHTvWy8vrhRde2Lx585AXozitmH/IkarRaDw8PGQyWUhISHJycmtrK65KcKocnaSDd+7cGTVq1L59+0iO4EjZ85KR8b5+/Xr2C98bN24ghH755RfSEfMTYGxmsbp9+zZC6PTp06T06tWrVSoVOXSEhF6vf/PNN19++WUiJjMzs6SkpLa2Nicnx9/ff/bs2eSUIIni4uKCgoKampqSkpIZM2YoFIqHDx/u27dPJBKx9UydOnXNmjXsHAHT+fn5Li4u5PnGoZCyp7ZTp04hhH777TfC6p133nn33XcZhklLSwsJCSH5DMPIZLLt27ezc2ydZktlt9Xd3T158uTExESS+cUXX5SXl9fU1OzYsUMqla5YsYKcsk+CI3X//v1FRUW1tbWFhYUTJkyYOnXqkydPHIEqRyeB89lnn3l6epJVaIZhhEXKmZeMjPekpKSYmBjSka6uLoRQcXExyTE/AcZmFisqjC05OVmpVA72A0X4hdL169fN6rDtCz148MDd3X337t1GbnTbqzDdQkxMTHx8vMFygiNlT200Gltvb+9bb70VHh5OHtc4nLOyskaNGvXn11lx8m16yKbKaQg/Fh87dsyRjS00NHTp0qUc5eTQ/kg585KR8Q7GRsJkp4TjL0WmpKQEBATcuHFjMCKdnZ0IoZKSksEK2D9/ypQpqampRpYm7C+J0+LNmzednZ0PHTrEyceHgiNlT8HULUX29vbOmjXr+eefJ4t7AyFfunQJIcTe6DuwDO85bKoDK/fx8dm5c6cjLPAa1Pnjjz8ihKqrqwcqxzl2RjpwXjIy3mEpcrCo2TBfpVKR10F6vd7f399BNo/09fWlpKT4+fmxt3cPBHHy5EmEUE1NzcBTguR0dHR4enpu3boVv5n8/fffYxn19fWOs3lEq9X6+voO9jaP4EjZUxvePJKeno4xtre3czaPVFZW4lOlpaWCbx7BrhYWFnbv3j0jt19OTo6zszN7I6KRwnydYlPl1Nnc3Ozk5FRUVES25AhI1aBOjUbD3mLK0c8wjN2QDjYvGRnvePMI2W2emZnp7u4+tOd1WIocGHrDOXl5eWKxODs7u66ubtGiRVKplOzeMXyBvXIXL17s4eFx4sQJst/30aNHDMNcv359w4YNlZWVTU1NRUVF48ePj4iIsJcow+2sXLnyxIkTTU1Np06dio6O9vHxwfNacnKyQqEoKyurrKyc8def4evtm6vX6xUKxUcffUSadRCkHR0dVX/9IYS2bNlSVVWFt+Fs2rRJKpXiN4TUajVnu394ePi5c+dOnjwZHBxst+3+BqX29vYmJCQEBARUV1eTmxZvfjt9+nRGRkZ1dXVjY2NOTo5MJluwYAHhb9OEQakdHR2rVq06c+ZMU1PTsWPHJk+eHBwcTKbauLg4+1M1qBOTaW9vHzNmzI4dO9ighEI62LzEMMxg4x1v94+Jiamuri4pKZHJZLDdnx1KW6W3bdumUChEIpFKpTp79qytmrGw3oEfzNy7dy/DML/++mtERISXl5dYLA4KClq9evVg72RY2ODQi8+dO1cul4tEIn9//7lz55I3/PAHNj09PceMGTN79uyWlpaht8HflaWlpQihhoYGUqWDIC0vL+cEXaPRMAyDP6A9duxYsVgcFRXFVt7W1jZv3jw3Nzd3d/eFCxfa7QPaBqU2NTVx9JOPCV64cGHatGkeHh4SiWTChAkbN24kLkKiYKOEQamPHj2KiYmRyWSjR49WKpVJSUnsl7OCUDWoEzPJzMx0dXXV6XRsREIhHRhiPC8xDGNkvN+8efONN95wdXX18fFZuXLlYIsl7A4aTMMTm0EskAkEgAAQAAK0EgBjozVyoBsIAAEgAAQMEgBjM4gFMoEAEAACQIBWAmBstEYOdAMBIAAEgIBBAmBsBrFAJhAAAkAACNBKAIyN1siBbiAABIAAEDBIAIzNIBbIBAJAAAgAAVoJgLHRGjnQDQSAABAAAgYJgLEZxAKZQGD4EzD4nUzDv9vQwxFAAIxtBAQZukghAY1Gw/nuhtjYWH77AcbGL0+ozXEIgLE5TixACRDoJ6DRaOLi4shXKba0tPD+XcBgbP24ITW8CICxDa94Qm+GCwGNRqNWqwf2BiG0ffv2uLg4iUQSGBh44MABUqa2tvbVV1+VSCReXl5JSUnsL4TMysp67rnnRCKRr69vSkoKvgQhtGvXrlmzZrm6ugYFBeEvrSe1QQII0EsAjI3e2IHy4UzAiLF5e3vv2rWroaFh3bp1Li4udXV1DMN0dnbK5fI5c+ZcvHjx+PHjgYGB+JuRGYbZvn27RCL58ssvGxoazp8/n5GRgcEhhAICAnJzc69du7Zs2TI3N7e2trbhzBT6NmIIgLGNmFBDR6kioNFoXFxcnmL9paWlMQyDEEpOTiZdmTZt2uLFixmG+fbbbz09PTs7O/GpH374wdnZGX8VvZ+f39q1a8klJIEQWrduHT7EP5p65MgRchYSQIBeAmBs9MYOlA9nAhqNJjo6+hrrDz9OIYS+++470vMPP/zwlVdeYRhmxYoVOIFP6XQ6hFBFRcXdu3cRQmVlZeQSkkAIFRQUkEN3d3d2zSQfEkCAOgJgbNSFDASPCAJGliLZ9mPS2B4+fGjE2AoLCwlNDw8P8otZJBMSQIBGAmBsNEYNNA9/AkaMDa89YgTTp083uRQ5bty4wZYiwdiG/500InsIxjYiww6ddngCA7f7379/H7/H5uPjk5WV1dDQ8Mknnzg7O1++fJlhmK6uLrlc/vbbb1+8eLGsrGz8+PFk80h2drZEItm6devVq1cvXLjw1Vdf4d5ztvvDE5vD3xQg0FwCYGzmkoJyQMCeBAZ+QDs0NBQb2zfffPP666+LxeJx48bl5+cTVUa2++/cuTM0NHT06NFyufyDDz7Al4CxEXSQGGYEwNiGWUChO8OcAMeNhnlvoXtAYEgEwNiGhA0uAgICEQBjEwg8NEsTATA2mqIFWoEAGBvcA0DAJAEwNpOIoAAQAAJAAAjQRACMjaZogVYgAASAABAwSQCMzSQiKAAEgAAQAAI0EQBjoylaoBUIAAEgAARMEgBjM4kICgABIAAEgABNBMDYaIoWaAUCQAAIAAGTBMDYTCKCAkAACAABIEATATA2mqIFWoEAEAACQMAkATA2k4igABAAAkAACNBEAIyNpmiBViAABIAAEDBJ4P8A00wlacag3CIAAAAASUVORK5CYII=)\n",
        "\n"
      ],
      "metadata": {
        "id": "rvhdOpKWzExX"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import torch\n",
        "from torch.utils.data import DataLoader\n",
        "from torch.nn import CrossEntropyLoss\n",
        "import numpy as np\n",
        "from transformers import GPT2Config, GPT2Model,GPT2LMHeadModel\n",
        "\n",
        "from tqdm import tqdm\n",
        "import copy\n",
        "import math\n",
        "\n",
        "#%pip install git+https://github.com/neelnanda-io/neel-plotly.git\n",
        "#from neel_plotly.plot import line\n",
        "\n",
        "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
        "\n",
        "import sys\n",
        "from random import randrange, choices, sample\n",
        "from operator import add\n",
        "\n",
        "from collections import OrderedDict\n",
        "from typing import Dict, Callable\n",
        "import torch\n",
        "from transformers.pytorch_utils import Conv1D\n",
        "import torch.nn.init as init\n",
        "\n",
        "import random\n",
        "\n",
        "torch.__version__\n",
        "torch.manual_seed(0)\n",
        "random.seed(0)\n",
        "\n",
        "#Constants\n",
        "num_test = 1000\n",
        "max_ctx = 150\n",
        "batch_size = 1000\n",
        "DATA_SEED = 598\n",
        "\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "def plt_line(y_vals, x_val, labels, title=\"Losses\", x_label=\"losses\", y_label=\"Epoch\"):\n",
        "  for y, label in zip(y_vals, labels):\n",
        "    #label = \"placeholder\"\n",
        "    plt.plot(x_val, y, label=label)\n",
        "\n",
        "  plt.xlabel(x_label)\n",
        "  plt.ylabel(y_label)\n",
        "  plt.title(title)\n",
        "  plt.grid()\n",
        "  plt.legend()\n",
        "  plt.show()"
      ],
      "metadata": {
        "id": "xBLA20RB2Okw"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Recreate Data\n",
        "\n",
        "These functions create the data.\n",
        "\n",
        "The important part of this code is pasted below (and it is already run for you):\n",
        "\n",
        "\n",
        "```\n",
        "# Make the data\n",
        "\n",
        "#generate indexes for noise vs clean data\n",
        "idxs = list(range(20000 - num_test))\n",
        "noise_idxs = sample(idxs, 1000)\n",
        "clean_idxs = list(set(idxs) - set(noise_idxs))\n",
        "\n",
        "#Mix clean and noise data\n",
        "list_of_functions = [seven_function]\n",
        "list_of_dataset_sizes = [20000]\n",
        "\n",
        "clean_train_dataloader, clean_test_dataloaders = create_data_distributions(list_of_functions, list_of_dataset_sizes, test_set_size=num_test, shuffle=True, noise=False, noise_range=1, length=20)\n",
        "\n",
        "list_of_functions = [seven_function]\n",
        "list_of_dataset_sizes = [20000]\n",
        "noise_train_dataloader, noise_test_dataloaders = create_data_distributions(list_of_functions, list_of_dataset_sizes, test_set_size=num_test, shuffle=True, noise=True, noise_range=1, length=20)\n",
        "\n",
        "#combine train_dataloaders\n",
        "clean_data = clean_train_dataloader.dataset\n",
        "noise_data = noise_train_dataloader.dataset\n",
        "\n",
        "#grab clean and noise data according to indexes\n",
        "clean_data_corresponding_to_noise = clean_data[noise_idxs]\n",
        "clean_data = clean_data[clean_idxs]\n",
        "noise_data = noise_data[noise_idxs]\n",
        "\n",
        "\n",
        "#Need to grab\n",
        "train_datasets = (noise_data, clean_data)\n",
        "```\n",
        "\n"
      ],
      "metadata": {
        "id": "HXtuINLXzkfx"
      }
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "yi6h3EMawOrU"
      },
      "outputs": [],
      "source": [
        "def tokenize_and_pad(char_list, pad=True):\n",
        "  tokenized_seq = []\n",
        "  for i in char_list:\n",
        "    if i == '^':\n",
        "      tokenized_seq.append(torch.tensor(10, dtype=int))\n",
        "    if i == '$':\n",
        "      tokenized_seq.append(torch.tensor(11))\n",
        "    if i == ' ':\n",
        "      tokenized_seq.append(torch.tensor(12))\n",
        "    if i == '0':\n",
        "      tokenized_seq.append(torch.tensor(0))\n",
        "    if i == '1':\n",
        "      tokenized_seq.append(torch.tensor(1))\n",
        "    if i == '2':\n",
        "      tokenized_seq.append(torch.tensor(2))\n",
        "    if i == '3':\n",
        "      tokenized_seq.append(torch.tensor(3))\n",
        "    if i == '4':\n",
        "      tokenized_seq.append(torch.tensor(4))\n",
        "    if i == '5':\n",
        "      tokenized_seq.append(torch.tensor(5))\n",
        "    if i == '6':\n",
        "      tokenized_seq.append(torch.tensor(6))\n",
        "    if i == '7':\n",
        "      tokenized_seq.append(torch.tensor(7))\n",
        "    if i == '8':\n",
        "      tokenized_seq.append(torch.tensor(8))\n",
        "    if i == '9':\n",
        "      tokenized_seq.append(torch.tensor(9))\n",
        "\n",
        "  if pad == True:\n",
        "    while len(tokenized_seq) < max_ctx:\n",
        "      tokenized_seq.append(torch.tensor(13))\n",
        "\n",
        "  return tokenized_seq\n",
        "\n",
        "def detokenize(tensor):\n",
        "  detokenized_seq = ''\n",
        "  for i in tensor:\n",
        "    if i == 10:\n",
        "      detokenized_seq += '^' #.append(torch.tensor(10, dtype=int))\n",
        "    if i == 11:\n",
        "      detokenized_seq += '$' #.append(torch.tensor(11))\n",
        "    if i == 12:\n",
        "      detokenized_seq += ' ' #.append(torch.tensor(12))\n",
        "    if i == 13:\n",
        "      detokenized_seq += '_' #.append(torch.tensor(13))\n",
        "    if i == 0:\n",
        "      detokenized_seq += '0' #.append(torch.tensor(0))\n",
        "    if i == 1:\n",
        "      detokenized_seq += '1' #.append(torch.tensor(1))\n",
        "    if i == 2:\n",
        "      detokenized_seq += '2' #.append(torch.tensor(2))\n",
        "    if i == 3:\n",
        "      detokenized_seq += '3' #.append(torch.tensor(3))\n",
        "    if i == 4:\n",
        "      detokenized_seq += '4' #.append(torch.tensor(4))\n",
        "    if i == 5:\n",
        "      detokenized_seq += '5' #.append(torch.tensor(5))\n",
        "    if i == 6:\n",
        "      detokenized_seq += '6' #.append(torch.tensor(6))\n",
        "    if i == 7:\n",
        "      detokenized_seq += '7' #.append(torch.tensor(7))\n",
        "    if i == 8:\n",
        "      detokenized_seq += '8' #.append(torch.tensor(8))\n",
        "    if i == 9:\n",
        "      detokenized_seq += '9' #.append(torch.tensor(9))\n",
        "\n",
        "  return detokenized_seq"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "def seven_function(starting_val):\n",
        "  # 7+x\n",
        "  return 7 + starting_val\n",
        "\n",
        "def generate_seq(func, length, noise, num_examples, modulo, device, noise_range=10):\n",
        "  data = []\n",
        "  #noise_amt = 0\n",
        "\n",
        "  for i in range(num_examples):\n",
        "\n",
        "    start = 0+i\n",
        "    vector = []\n",
        "    #This is how we generate noise for each sample\n",
        "    #noise_amt = randrange(-noise_range, noise_range)\n",
        "    for j in range(length):\n",
        "      vector.append(func(start))\n",
        "      start = func(start)\n",
        "\n",
        "    # adding noise vector to the clean datapoints\n",
        "    if noise:\n",
        "      noise_vector = choices(population=[0,-1, 1], weights=[0.9,0.05,0.05], k = length)\n",
        "      vector = list( map(add, vector, noise_vector) )\n",
        "\n",
        "    string = ' '.join([str(x) for x in vector])\n",
        "    string = \"^\"+string+\"$\"\n",
        "    #print(string)\n",
        "    char_list = [x for x in string]\n",
        "    tensor = torch.Tensor(tokenize_and_pad(char_list))\n",
        "    data.append(tensor)\n",
        "\n",
        "  dataset = torch.stack(data, dim=0).to(device)\n",
        "  #dataset = dataset.to(torch.int64)\n",
        "\n",
        "  return dataset\n",
        "\n",
        "\n",
        "def split_data(data, num_examples, num_test):\n",
        "  torch.manual_seed(DATA_SEED)\n",
        "  indices = torch.randperm(num_examples)\n",
        "  #cutoff = int(num_examples*frac_train)\n",
        "  cutoff = num_examples - num_test\n",
        "  train_indices = indices[:cutoff]\n",
        "  test_indices = indices[cutoff:]\n",
        "\n",
        "  train_data = data[train_indices]\n",
        "  test_data = data[test_indices]\n",
        "  #print(train_data[:5])\n",
        "  #print(train_data.shape)\n",
        "  #print(test_data[:5])\n",
        "  #print(test_data.shape)\n",
        "\n",
        "  return train_data.to(torch.int64), test_data.to(torch.int64)\n",
        "\n",
        "def create_data_distributions(list_of_functions, list_of_dataset_sizes, test_set_size=num_test, shuffle=True, noise=False, noise_range=10, length=20):\n",
        "  train_datas = []\n",
        "  #test_datas = []\n",
        "\n",
        "  test_dataloaders = []\n",
        "\n",
        "  for i in range(len(list_of_functions)):\n",
        "    data = generate_seq(func=list_of_functions[i], length=length, noise=noise, num_examples=list_of_dataset_sizes[i], modulo=13, device=device, noise_range=noise_range)\n",
        "    train_data, test_data = split_data(data, num_examples=list_of_dataset_sizes[i], num_test=test_set_size)\n",
        "\n",
        "    train_datas.append(train_data)\n",
        "\n",
        "    #want separate test_dataloaders\n",
        "    test_dataloaders.append(DataLoader(test_data, batch_size=batch_size, shuffle=shuffle))\n",
        "\n",
        "  train_data = torch.concat(train_datas, dim=0)\n",
        "  #want one train_datalaoder\n",
        "  train_dataloader = DataLoader(train_data, batch_size=batch_size, shuffle=shuffle)\n",
        "\n",
        "  return train_dataloader, test_dataloaders"
      ],
      "metadata": {
        "id": "2WNfbOcr2WD4"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Make the data\n",
        "\n",
        "#generate indexes for noise vs clean data\n",
        "idxs = list(range(20000 - num_test))\n",
        "noise_idxs = sample(idxs, 1000)\n",
        "clean_idxs = list(set(idxs) - set(noise_idxs))\n",
        "\n",
        "#Mix clean and noise data\n",
        "list_of_functions = [seven_function]\n",
        "list_of_dataset_sizes = [20000]\n",
        "\n",
        "clean_train_dataloader, clean_test_dataloaders = create_data_distributions(list_of_functions, list_of_dataset_sizes, test_set_size=num_test, shuffle=True, noise=False, noise_range=1, length=20)\n",
        "\n",
        "list_of_functions = [seven_function]\n",
        "list_of_dataset_sizes = [20000]\n",
        "noise_train_dataloader, noise_test_dataloaders = create_data_distributions(list_of_functions, list_of_dataset_sizes, test_set_size=num_test, shuffle=True, noise=True, noise_range=1, length=20)\n",
        "\n",
        "#combine train_dataloaders\n",
        "clean_data = clean_train_dataloader.dataset\n",
        "noise_data = noise_train_dataloader.dataset\n",
        "\n",
        "#grab clean and noise data according to indexes\n",
        "clean_data_corresponding_to_noise = clean_data[noise_idxs]\n",
        "clean_data = clean_data[clean_idxs]\n",
        "noise_data = noise_data[noise_idxs]\n",
        "\n",
        "\n",
        "#Need to grab\n",
        "train_datasets = (noise_data, clean_data)"
      ],
      "metadata": {
        "id": "ngSCDu5L2ljq"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "clean_data_corresponding_to_noise"
      ],
      "metadata": {
        "id": "YIG6nECHQUS1",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "dc261883-52cd-4573-9341-99b4013a796c"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([[10,  1,  0,  ..., 13, 13, 13],\n",
              "        [10,  3,  0,  ..., 13, 13, 13],\n",
              "        [10,  1,  6,  ..., 13, 13, 13],\n",
              "        ...,\n",
              "        [10,  1,  3,  ..., 13, 13, 13],\n",
              "        [10,  4,  0,  ..., 13, 13, 13],\n",
              "        [10,  1,  2,  ..., 13, 13, 13]], device='cuda:0')"
            ]
          },
          "metadata": {},
          "execution_count": 5
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "def count_num_noised(noise_dataset, clean_data_set_for_noise, k, prompt_len, batch_size=1000):\n",
        "  noise_dataloader = DataLoader(noise_dataset, batch_size=batch_size, shuffle=False)\n",
        "  clean_dataloader = DataLoader(clean_data_set_for_noise, batch_size=batch_size, shuffle=False)\n",
        "  with torch.inference_mode():\n",
        "    for noise_batch, batch_clean in zip(noise_dataloader, clean_dataloader):\n",
        "      noise = torch.eq(noise_batch[:,prompt_len:prompt_len+k], batch_clean[:,prompt_len:prompt_len+k])\n",
        "      noise_locations = noise.all(dim=1)#check to see if there is noise in the row (False indicates noise, we want noise)\n",
        "      print(\"# of noised samples: \", batch_size - noise_locations.sum())\n",
        "\n",
        "count_num_noised(noise_data, clean_data_corresponding_to_noise, k=50, prompt_len=50)\n",
        "count_num_noised(noise_data, clean_data_corresponding_to_noise, k=50, prompt_len=100)\n",
        "count_num_noised(noise_data, clean_data_corresponding_to_noise, k=50, prompt_len=150)\n",
        "count_num_noised(noise_data, clean_data_corresponding_to_noise, k=50, prompt_len=200)\n",
        "count_num_noised(noise_data, clean_data_corresponding_to_noise, k=50, prompt_len=250)\n",
        "count_num_noised(noise_data, clean_data_corresponding_to_noise, k=50, prompt_len=300)"
      ],
      "metadata": {
        "id": "ooQwClyF38p8",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "8e9d249e-f21e-4f8f-d5ef-041fea2a23cc"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "# of noised samples:  tensor(615, device='cuda:0')\n",
            "# of noised samples:  tensor(166, device='cuda:0')\n",
            "# of noised samples:  tensor(0, device='cuda:0')\n",
            "# of noised samples:  tensor(0, device='cuda:0')\n",
            "# of noised samples:  tensor(0, device='cuda:0')\n",
            "# of noised samples:  tensor(0, device='cuda:0')\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Utility Functions\n",
        "\n",
        "These are the functions that support us in localizing memorized information within a model. Details arn't important."
      ],
      "metadata": {
        "id": "h_AkEUyivMAF"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "def get_attr_str(model_name):\n",
        "    if 'gpt2' in model_name:\n",
        "        attr_dict = {\n",
        "            'transformer_layer': 'transformer.h',\n",
        "            'ffn_out': 'mlp.c_proj',\n",
        "            'ffn_act': 'mlp.act',\n",
        "            'lm_head': 'lm_head',\n",
        "        }\n",
        "    elif 'gpt-j' in model_name:\n",
        "        attr_dict = {\n",
        "            'transformer_layer': 'transformer.h',\n",
        "            'ffn_out': 'mlp.fc_out',\n",
        "            'ffn_act': 'mlp.act',\n",
        "            'lm_head': 'lm_head',\n",
        "        }\n",
        "    elif 'pythia' in model_name:\n",
        "        attr_dict = {\n",
        "            'transformer_layer': 'gpt_neox.layers',\n",
        "            'ffn_out': 'mlp.dense_4h_to_h',\n",
        "            'ffn_act': 'mlp.act',\n",
        "            'lm_head': 'embed_out',\n",
        "        }\n",
        "    else:\n",
        "        raise NotImplementedError(f\"{model_name} attributes unkown!\")\n",
        "    return attr_dict\n",
        "\n",
        "def set_model_attributes(model, model_name):\n",
        "    model.config.pad_token_id = model.config.eos_token_id\n",
        "    model.attr_dict = get_attr_str(model_name)\n",
        "    model.inner_dim = 4 * model.config.hidden_size\n",
        "    if not hasattr(model.config, \"n_layer\"):\n",
        "        model.config.n_layer = model.config.num_hidden_layers\n",
        "\n",
        "def get_attributes(x: torch.nn.Module, attributes: str):\n",
        "    \"\"\"\n",
        "    gets a list of period-separated attributes\n",
        "    i.e get_attributes(model, 'transformer.h.0.mlp.c_proj')\n",
        "        should return the same as model.transformer.h.0.mlp.c_proj\n",
        "    \"\"\"\n",
        "    for attr in attributes.split(\".\"):\n",
        "        x = getattr(x, attr)\n",
        "    return x\n",
        "\n",
        "class Patch(torch.nn.Module):\n",
        "    def __init__(\n",
        "        self,\n",
        "        ff_layer: torch.nn.Module,\n",
        "        intermediate_size: int = None,\n",
        "        replacement_activations: torch.Tensor = None,\n",
        "        onehot_coef: torch.Tensor = None,\n",
        "        mean_ablation_idxs: torch.Tensor = None,\n",
        "        noise_ablation_idxs: torch.Tensor = None,\n",
        "    ):\n",
        "        super().__init__()\n",
        "        self.module = ff_layer\n",
        "        if intermediate_size is not None: # slimming\n",
        "            self.slim_coef = torch.nn.Parameter(torch.ones(intermediate_size))\n",
        "        self.acts = replacement_activations\n",
        "        self.onehot_coef = onehot_coef\n",
        "        self.mean_ablation_idxs = mean_ablation_idxs\n",
        "        self.noise_ablation_idxs = noise_ablation_idxs\n",
        "\n",
        "    def forward(self, x: torch.Tensor):\n",
        "        hidden_states = self.module(x)\n",
        "        if self.acts is not None:          # knowledge neurons\n",
        "            #print(\"Hidden states shape: \", hidden_states.shape)\n",
        "            #print(\"Patch acts shape: \", self.acts.shape)\n",
        "            hidden_states[:, -1, :] = self.acts # patch the last token\n",
        "        elif self.noise_ablation_idxs is not None:\n",
        "            # TODO\n",
        "            #print(hidden_states.type())\n",
        "            device = hidden_states.device\n",
        "            hidden_states[:,:,self.noise_ablation_idxs] = torch.from_numpy(np.random.laplace(loc=0.0, scale=1.0, size=hidden_states[:,:,self.noise_ablation_idxs].shape)).float().to(device)\n",
        "        elif self.mean_ablation_idxs is not None:\n",
        "            avg_activations = hidden_states.mean()\n",
        "            hidden_states[:,:,self.mean_ablation_idxs] = avg_activations\n",
        "        elif self.onehot_coef is not None: # zero-out\n",
        "            bs = hidden_states.shape[0]\n",
        "            seq_len = hidden_states.shape[1]\n",
        "            # hidden_states = hidden_states * self.onehot_coef #.unsqueeze(1) #THIS WORKs\n",
        "            hidden_states = hidden_states * self.onehot_coef.expand(bs, seq_len, -1)\n",
        "        else:                              # slimming\n",
        "            #print(\"HI THERE\")\n",
        "            #print(\"slim version: \", self.slim_coef._version)\n",
        "            #hidden_states *= torch.clip(self.slim_coef, 0, 1)\n",
        "            hidden_states = hidden_states * torch.clip(self.slim_coef, 0, 1)\n",
        "            #print(\"hidden state version: \",hidden_states._version)\n",
        "            #print(\"hidden state shape: \", hidden_states.shape)\n",
        "\n",
        "        return hidden_states\n",
        "\n",
        "def set_attributes(x: torch.nn.Module, attributes: str, values):\n",
        "    attr_list = attributes.split(\".\")\n",
        "    for attr in attr_list[:-1]:\n",
        "        x = getattr(x, attr)\n",
        "    setattr(x, attr_list[-1], values)\n",
        "\n",
        "def patch_ff_layer(\n",
        "    model: torch.nn.Module,\n",
        "    ff_attrs: str,\n",
        "    intermediate_size: int = None,\n",
        "    replacement_activations: torch.Tensor = None,\n",
        "    onehot_coef: torch.Tensor = None,\n",
        "    mean_ablation_idxs: torch.Tensor = None,\n",
        "    noise_ablation_idxs: torch.Tensor = None,\n",
        "):\n",
        "    \"\"\"\n",
        "    replaces the ff layer at `layer_idx` with a `Patch` class - that will replace the intermediate activations at sequence position\n",
        "    `mask_index` with `replacement_activations`\n",
        "    \"\"\"\n",
        "    #print(\"PATCHING: \", ff_attrs)\n",
        "    ff_layer = get_attributes(model, ff_attrs)\n",
        "    patch = Patch(\n",
        "            ff_layer,\n",
        "            intermediate_size,\n",
        "            replacement_activations,\n",
        "            onehot_coef,\n",
        "            mean_ablation_idxs,\n",
        "            noise_ablation_idxs,\n",
        "        )\n",
        "\n",
        "    set_attributes(model, ff_attrs, patch)\n",
        "    #print(f\"Patch {ff_attrs}\")\n",
        "\n",
        "\n",
        "def unpatch_ff_layer(\n",
        "    model: torch.nn.Module,\n",
        "    ff_attrs: str,\n",
        "):\n",
        "    \"\"\"\n",
        "    Removes the `Patch` applied by `patch_ff_layer`, replacing it with its original value.\n",
        "    \"\"\"\n",
        "    ff_layer = get_attributes(model, ff_attrs)\n",
        "    assert isinstance(ff_layer, Patch), \"Can't unpatch a layer that hasn't been patched\"\n",
        "    set_attributes(model, ff_attrs, ff_layer.module)\n",
        "    #print(f\"Reset {ff_attrs}\")"
      ],
      "metadata": {
        "id": "slPRHKN7vQId"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Metrics\n",
        "\n",
        "Here we define metrics like accuracy, loss, etc. The details arn't too important."
      ],
      "metadata": {
        "id": "QRo6PYx2bWkg"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# New function that check form memorization only among actually noised inputs\n",
        "#probably want to pass in both noise and clean dataloader\n",
        "def refined_check_percent_memorized(noise_dataset, clean_data_set_for_noise, prompt_len, k, batch_size, model):\n",
        "\n",
        "  #we do this to increase batch sizes (for increasing throughput)\n",
        "  noise_dataloader = DataLoader(noise_dataset, batch_size=batch_size, shuffle=False)\n",
        "  clean_dataloader = DataLoader(clean_data_set_for_noise, batch_size=batch_size, shuffle=False)\n",
        "\n",
        "  memorized = 0\n",
        "  total = 0\n",
        "  with torch.inference_mode():\n",
        "    for noise_batch, batch_clean in zip(noise_dataloader, clean_dataloader):\n",
        "\n",
        "      #check if noise_batch[:,prompt_len:prompt_len+k] == batch_clean[:,prompt_len:prompt_len+k]\n",
        "      # if there is an equality toss that sample out cus it has no noise\n",
        "      noise = torch.eq(noise_batch[:,prompt_len:prompt_len+k], batch_clean[:,prompt_len:prompt_len+k])\n",
        "      noise_locations = noise.all(dim=1)#check to see if there is noise in the row (False indicates noise, we want noise)\n",
        "      noise_idx = (noise_locations == 0).nonzero(as_tuple=True)[0].tolist() # all of the values we keep\n",
        "\n",
        "      noise_batch = noise_batch[noise_idx]\n",
        "      batch_clean = batch_clean[noise_idx]\n",
        "\n",
        "      batch = batch_clean[:,:prompt_len] #grab first 50 tokens from the clean dataset\n",
        "      outputs = model.generate(batch, max_length=max_ctx, pad_token_id = 13)\n",
        "\n",
        "      #now check if there is a match\n",
        "      equals = torch.eq(outputs[:,prompt_len:prompt_len+k], noise_batch[:,prompt_len:prompt_len+k])\n",
        "      #TODO ^^ need to make sure original batch contains noise from prompt_len:prompt_len+k\n",
        "      match_rows = equals.all(dim=1)\n",
        "      total_matchs = match_rows.sum()\n",
        "\n",
        "      total += noise_batch.shape[0]\n",
        "      memorized += total_matchs\n",
        "\n",
        "  #print(\"% memorized: \", memorized / total)\n",
        "  return memorized / total"
      ],
      "metadata": {
        "id": "V9XzV-T33nu1"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def accuracy(inputs, logits):\n",
        "    # Shift so that tokens < n predict n\n",
        "    shift_labels = inputs[..., 1:].contiguous()\n",
        "    shift_logits = logits[..., :-1, :].contiguous()\n",
        "\n",
        "    #converts logits to predictions\n",
        "    predictions = torch.argmax(shift_logits, axis=-1)\n",
        "\n",
        "    #Now compute accuracy\n",
        "    N = torch.numel(predictions)\n",
        "    accuracy = (shift_labels == predictions).sum() / N\n",
        "\n",
        "    return accuracy\n",
        "\n",
        "def perplexity(dataloader, model):\n",
        "  avg_metric = 0\n",
        "  for batch in dataloader:\n",
        "    model_output = model(batch, labels=batch)\n",
        "    loss = model_output.loss\n",
        "    avg_metric += torch.exp(loss)\n",
        "  return (avg_metric.cpu() / len(dataloader))\n",
        "\n",
        "def loss(dataloader, model,):\n",
        "  avg_metric = 0\n",
        "  for batch in dataloader:\n",
        "    model_output = model(batch, labels=batch)\n",
        "    avg_metric += model_output.loss\n",
        "  return (avg_metric.cpu() / len(dataloader))\n",
        "\n",
        "\n",
        "def compute_average_metric_accross_dataset(dataloader, model, metric):\n",
        "  avg_metric = 0\n",
        "  for batch in dataloader:\n",
        "    model_output = model(batch, labels=batch)\n",
        "    test_logits = model_output.logits\n",
        "    avg_metric += metric(batch, test_logits)\n",
        "  return (avg_metric.cpu() / len(dataloader))\n",
        "\n",
        "# New function that check form memorization only among actually noised inputs\n",
        "#probably want to pass in both noise and clean dataloader\n",
        "def refined_check_percent_memorized(noise_dataset, clean_data_set_for_noise, prompt_len, k, batch_size, model):\n",
        "\n",
        "  #we do this to increase batch sizes (for increasing throughput)\n",
        "  noise_dataloader = DataLoader(noise_dataset, batch_size=batch_size, shuffle=False)\n",
        "  clean_dataloader = DataLoader(clean_data_set_for_noise, batch_size=batch_size, shuffle=False)\n",
        "\n",
        "  memorized = 0\n",
        "  total = 0\n",
        "  with torch.inference_mode():\n",
        "    for noise_batch, batch_clean in zip(noise_dataloader, clean_dataloader):\n",
        "\n",
        "      #check if noise_batch[:,prompt_len:prompt_len+k] == batch_clean[:,prompt_len:prompt_len+k]\n",
        "      # if there is an equality toss that sample out cus it has no noise\n",
        "      noise = torch.eq(noise_batch[:,prompt_len:prompt_len+k], batch_clean[:,prompt_len:prompt_len+k])\n",
        "      noise_locations = noise.all(dim=1)#check to see if there is noise in the row (False indicates noise, we want noise)\n",
        "      noise_idx = (noise_locations == 0).nonzero(as_tuple=True)[0].tolist() # all of the values we keep\n",
        "\n",
        "      noise_batch = noise_batch[noise_idx]\n",
        "      batch_clean = batch_clean[noise_idx]\n",
        "\n",
        "      #original_batch = batch\n",
        "      batch = batch_clean[:,:prompt_len] #grab first 50 tokens from the clean dataset\n",
        "      outputs = model.generate(batch, max_length=max_ctx, pad_token_id = 13)\n",
        "\n",
        "      #now check if there is a match\n",
        "      equals = torch.eq(outputs[:,prompt_len:prompt_len+k], noise_batch[:,prompt_len:prompt_len+k])\n",
        "      match_rows = equals.all(dim=1)\n",
        "      total_matchs = match_rows.sum()\n",
        "\n",
        "      total += noise_batch.shape[0]\n",
        "      memorized += total_matchs\n",
        "\n",
        "  return memorized / total\n",
        "\n",
        "def track_all_metrics(noise_data=noise_data,\n",
        "                      clean_data_corresponding_to_noise=clean_data_corresponding_to_noise,\n",
        "                      clean_test_dataloaders=clean_test_dataloaders,\n",
        "                      model=None,\n",
        "                      prompt_len=50,\n",
        "                      batch_size=1000):\n",
        "  # Check % mem on noise data\n",
        "  perc_mem = refined_check_percent_memorized(noise_dataset=noise_data, clean_data_set_for_noise=clean_data_corresponding_to_noise, prompt_len=50, k=50, batch_size=200, model=model)\n",
        "  print(\"perentage memorized: \", (perc_mem * 100).item(), \"%\")\n",
        "\n",
        "  # Check accuracy on clean data\n",
        "  acc = compute_average_metric_accross_dataset(clean_test_dataloaders[0], model, accuracy)\n",
        "  print(\"accuracy on clean data: \", (acc*100).item(), \"%\")\n",
        "\n",
        "\n",
        "  # Check perplexity on clean data\n",
        "  perplex_clean = perplexity(clean_test_dataloaders[0], model)\n",
        "  print(\"perplexity clean data: \", (perplex_clean*100).item())\n",
        "\n",
        "  # Check perplexity on noise_data\n",
        "  noise_dataloader = DataLoader(noise_data, batch_size=batch_size, shuffle=False)\n",
        "  perplex_noise = perplexity(noise_dataloader, model)\n",
        "  print(\"perplexity noise data: \", (perplex_noise*100).item())\n",
        "\n",
        "  return perc_mem, acc, perplex_clean, perplex_noise"
      ],
      "metadata": {
        "id": "sTD9vC3OcqkI"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Get Model\n",
        "\n",
        "This is how we load a pre-trained model. Right now, I have only provided the final trained model, but I have all of the model checkpoints which I can provide later."
      ],
      "metadata": {
        "id": "VJY1qQ-uG8Kp"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!wget --no-check-certificate 'https://docs.google.com/uc?export=download&id=1HWs0DlvVREH363nKXBCwjIynj7EYvPuK' -O 2_layer_model.pt\n",
        "!ls"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "CrQ0_xRYNvgu",
        "outputId": "dfea0bd8-1601-4151-edfc-6d70f21fca05"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "--2024-05-21 22:18:00--  https://docs.google.com/uc?export=download&id=1HWs0DlvVREH363nKXBCwjIynj7EYvPuK\n",
            "Resolving docs.google.com (docs.google.com)... 142.250.107.113, 142.250.107.138, 142.250.107.102, ...\n",
            "Connecting to docs.google.com (docs.google.com)|142.250.107.113|:443... connected.\n",
            "HTTP request sent, awaiting response... 303 See Other\n",
            "Location: https://drive.usercontent.google.com/download?id=1HWs0DlvVREH363nKXBCwjIynj7EYvPuK&export=download [following]\n",
            "--2024-05-21 22:18:00--  https://drive.usercontent.google.com/download?id=1HWs0DlvVREH363nKXBCwjIynj7EYvPuK&export=download\n",
            "Resolving drive.usercontent.google.com (drive.usercontent.google.com)... 173.194.202.132, 2607:f8b0:400e:c0a::84\n",
            "Connecting to drive.usercontent.google.com (drive.usercontent.google.com)|173.194.202.132|:443... connected.\n",
            "HTTP request sent, awaiting response... 200 OK\n",
            "Length: 1682466 (1.6M) [application/octet-stream]\n",
            "Saving to: ‘2_layer_model.pt’\n",
            "\n",
            "2_layer_model.pt    100%[===================>]   1.60M  --.-KB/s    in 0.01s   \n",
            "\n",
            "2024-05-21 22:18:01 (145 MB/s) - ‘2_layer_model.pt’ saved [1682466/1682466]\n",
            "\n",
            " 2_layer_model.pt   edited_model.pt  'uc?export=download&id=1HWs0DlvVREH363nKXBCwjIynj7EYvPuK'\n",
            " drive\t\t    sample_data      'view?usp=sharing'\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "def get_model(g_drive=True):\n",
        "  layer_dir = \"two_layer\"\n",
        "  n_layer = 2\n",
        "  epoch = 200\n",
        "  configuration = GPT2Config(\n",
        "                            vocab_size = 14,\n",
        "                            n_layer = n_layer,\n",
        "                            n_head = 4,\n",
        "                            n_embd = 128,\n",
        "                            n_positions = max_ctx,\n",
        "                            bos_token_id = 10,\n",
        "                            eos_token_id = 11 ,\n",
        "                            use_cache = False,\n",
        "                            hidden_states = False,\n",
        "                            output_attentions = False,\n",
        "                            activation_function = \"relu\",\n",
        "                            attn_pdrop=0,\n",
        "                            resid_pdrop=0,\n",
        "                            embd_pdrop=0,\n",
        "                            initializer_range = 0.8 / math.sqrt(128) #0.8 / sqrt(d_model)\n",
        "  )\n",
        "\n",
        "  model = GPT2LMHeadModel(configuration)\n",
        "  model.to(device)\n",
        "\n",
        "  model.load_state_dict(torch.load(f'2_layer_model.pt'))\n",
        "\n",
        "  model.eval()\n",
        "\n",
        "  model_name = \"mem_gpt2\"\n",
        "  set_model_attributes(model, model_name)\n",
        "\n",
        "  return model\n",
        "\n",
        "model=get_model(g_drive=True)\n",
        "\n"
      ],
      "metadata": {
        "id": "QiKb4AZW1xyI"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "perc_mem, acc, perplex_clean, perplex_noise = track_all_metrics(noise_data=noise_data,\n",
        "                      clean_data_corresponding_to_noise=clean_data_corresponding_to_noise,\n",
        "                      clean_test_dataloaders=clean_test_dataloaders,\n",
        "                      model=model,\n",
        "                      prompt_len=50,\n",
        "                      batch_size=1000)"
      ],
      "metadata": {
        "id": "OxkXCADWOd71",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "7d27c032-1cbb-4fc1-e5da-456173488584"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "perentage memorized:  54.796749114990234 %\n",
            "accuracy on clean data:  97.49462890625 %\n",
            "perplexity clean data:  107.07524871826172\n",
            "perplexity noise data:  107.4422607421875\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Ablation Utility Functions\n",
        "\n",
        "These are the funcions that help us remove parts of the model. For example, if we know a particular neuron is reponsible for storing memorized information, we would use these funcations to remove it. Details not important."
      ],
      "metadata": {
        "id": "bYPY6UYweXLX"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "def apply_ablaton_mask_to_neurons(neuron_weightings, model, ratio=0.01):\n",
        "  print(\"Num of dropped neurons per layer: \", int(model.inner_dim*ratio//1))\n",
        "  for ly in tqdm(range(model.config.n_layer)):\n",
        "        attr_str = f\"{model.attr_dict['transformer_layer']}.{ly}.{model.attr_dict['ffn_act']}\"\n",
        "\n",
        "        coeffs = neuron_weightings[ly]\n",
        "\n",
        "        val, idx = torch.topk(coeffs, k=int(model.inner_dim*ratio//1)) #grab neuron idxs that have highest diff losses\n",
        "        #make one hot mask for that\n",
        "        mask = torch.ones(model.inner_dim)\n",
        "        mask[idx] = 0\n",
        "\n",
        "\n",
        "        patch_ff_layer(\n",
        "            model,\n",
        "            attr_str,\n",
        "            #onehot_coef = batch_coef,\n",
        "            onehot_coef = mask.to(device),\n",
        "        )\n",
        "\n",
        "  return model\n",
        "\n",
        "\n",
        "#apply_ablaton_mask_to_neurons(delta_losses, model=model, ratio=0.01)"
      ],
      "metadata": {
        "id": "Pg4WNtKE_hz-"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def remove_ablation_mask_from_neurons(model):\n",
        "  for ly in tqdm(range(model.config.n_layer)):\n",
        "        attr_str = f\"{model.attr_dict['transformer_layer']}.{ly}.{model.attr_dict['ffn_act']}\"\n",
        "        unpatch_ff_layer(\n",
        "                      model,\n",
        "                      attr_str,\n",
        "                  )\n",
        "\n",
        "  return 0"
      ],
      "metadata": {
        "id": "-5t02UbrJaEc"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def apply_mean_ablaton_mask_to_neurons(neuron_weightings, model, inputs, ratio=0.01):\n",
        "  print(\"Num of dropped neurons per layer: \", int(model.inner_dim*ratio//1))\n",
        "  for ly in tqdm(range(model.config.n_layer)):\n",
        "        attr_str = f\"{model.attr_dict['transformer_layer']}.{ly}.{model.attr_dict['ffn_act']}\"\n",
        "\n",
        "        coeffs = neuron_weightings[ly]\n",
        "\n",
        "        val, idx = torch.topk(coeffs, k=int(model.inner_dim*ratio//1)) #grab neuron idxs that have highest diff losses\n",
        "\n",
        "\n",
        "        patch_ff_layer(\n",
        "            model,\n",
        "            attr_str,\n",
        "            #onehot_coef = batch_coef,\n",
        "            mean_ablation_idxs = idx,\n",
        "            #replacement_activations = activations.to(device),\n",
        "        )\n",
        "\n",
        "  return model\n",
        "\n",
        "\n",
        "#apply_mean_ablaton_mask_to_neurons(delta_losses, model=model, inputs=noise_data, ratio=0.1)\n",
        "#remove_ablation_mask_from_neurons(model)"
      ],
      "metadata": {
        "id": "WLWUSMv9ti16"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def apply_noise_ablaton_mask_to_neurons(neuron_weightings, model, inputs, ratio=0.01): # TODO check this correctness\n",
        "  print(\"Num of dropped neurons per layer: \", int(model.inner_dim*ratio//1))\n",
        "  for ly in tqdm(range(model.config.n_layer)):\n",
        "        attr_str = f\"{model.attr_dict['transformer_layer']}.{ly}.{model.attr_dict['ffn_act']}\"\n",
        "\n",
        "        coeffs = neuron_weightings[ly]\n",
        "\n",
        "        val, idx = torch.topk(coeffs, k=int(model.inner_dim*ratio//1)) #grab neuron idxs that have highest diff losses\n",
        "\n",
        "        patch_ff_layer(\n",
        "            model,\n",
        "            attr_str,\n",
        "            #onehot_coef = batch_coef,\n",
        "            noise_ablation_idxs = idx,\n",
        "            #replacement_activations = activations.to(device),\n",
        "        )\n",
        "\n",
        "  return model\n",
        "\n",
        "\n",
        "#apply_ablaton_mask_to_neurons(delta_losses, model=model, ratio=0.01)\n",
        "#apply_noise_ablaton_mask_to_neurons(delta_losses, model=model, inputs=noise_data, ratio=0.1)\n",
        "#remove_ablation_mask_from_neurons(model)"
      ],
      "metadata": {
        "id": "aX8z60hlATWK"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Activations\n",
        "This is the important section where we localize which neurons are responsible for memorization and remove them.\n",
        "\n",
        "**Note**: When we ablate neurons from the model, we are not actually changing model weights, but rather adding in a pytorch hook and filtering out activatoins. I havn't figured out how to save this yet, but as long as you apply the ablation mask to the model before doing inference, it the neurons will be ablated. See below:\n",
        "\n",
        "```\n",
        "apply_ablaton_mask_to_neurons(act_mean, model=model, ratio=0.1)\n",
        "```\n",
        "In this example, we have ablated 10% of the neurons in the MLP.\n",
        "\n",
        "The 3rd from last cell has a clean copy of the model.\n",
        "\n",
        "The 2nd to last cell has a copy of the model w/ ablated neurons."
      ],
      "metadata": {
        "id": "_yMX3qgndyIQ"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "#set_model_attributes(model, model_name)\n",
        "#print(model.inner_dim)\n",
        "model = get_model(True)\n",
        "\n",
        "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
        "\n",
        "def register_hook(model, layer_idx, ori_activations, attr_str):\n",
        "    ff_layer = get_attributes(model, attr_str)\n",
        "\n",
        "    #print(attr_str)\n",
        "    #print(\"input shape in hook: \", ori_activations.shape)\n",
        "\n",
        "    def hook_fn(m, i, o):\n",
        "        #print(ori_activations.shape)\n",
        "        #print(\"o dim: \", o.shape)\n",
        "        ori_activations[layer_idx] = o.squeeze().cpu()\n",
        "        #ori_activations[layer_idx] = o.cpu()\n",
        "\n",
        "    return ff_layer.register_forward_hook(hook_fn)\n",
        "\n",
        "@torch.no_grad()\n",
        "def get_ori_activations(inner_dim, model, inputs):\n",
        "    #seq_len = inputs['input_ids'].shape[1]\n",
        "    seq_len = inputs.shape[1]\n",
        "    batch_size = inputs.shape[0]\n",
        "    ori_activations = torch.zeros((model.config.n_layer, batch_size, seq_len, inner_dim))\n",
        "\n",
        "    handles = []\n",
        "    for ly in range(model.config.n_layer):\n",
        "        handle = register_hook(\n",
        "            model,\n",
        "            ly,\n",
        "            ori_activations,\n",
        "            f\"{model.attr_dict['transformer_layer']}.{ly}.{model.attr_dict['ffn_act']}\",\n",
        "        )\n",
        "        handles.append(handle)\n",
        "\n",
        "\n",
        "    #print(\"input shape: \", inputs.shape)\n",
        "    out = model(inputs)\n",
        "    #print(\"here\")\n",
        "    for handle in handles: # detach the hooks\n",
        "        handle.remove()\n",
        "\n",
        "    return ori_activations\n",
        "\n",
        "def largest_act(inner_dim, model, inputs, gold_set, model_name=\"gpt2\", prompt_len=50):\n",
        "\n",
        "\n",
        "\n",
        "    @torch.no_grad()\n",
        "    def get_ffn_norms():\n",
        "        all_norms = torch.zeros((model.config.n_layer, inner_dim))\n",
        "        for ly in range(model.config.n_layer):\n",
        "            attr_str = f\"{model.attr_dict['transformer_layer']}.{ly}.{model.attr_dict['ffn_out']}.weight\"\n",
        "            weights = get_attributes(model, attr_str)\n",
        "            if 'gpt2' in model_name:\n",
        "                norms = torch.norm(weights, dim=1)\n",
        "            else:\n",
        "                norms = torch.norm(weights, dim=0)\n",
        "            all_norms[ly] = norms.cpu()\n",
        "\n",
        "        return all_norms\n",
        "\n",
        "    all_norms = get_ffn_norms()\n",
        "    #print(\"norms shape: \", all_norms.shape)\n",
        "\n",
        "    #prompt_start_i = args.prompt_len -1 if hasattr(args, 'prompt_len') else 0  # -1 for 0-indexed\n",
        "    prompt_start_i = prompt_len - 1\n",
        "\n",
        "    activations = get_ori_activations(inner_dim, model, inputs)\n",
        "    #print(activations.shape)\n",
        "    activations = activations[:, :, prompt_start_i: -1] # [n_layer, suffix_len, inner_dim]\n",
        "    all_norms = get_ffn_norms()\n",
        "\n",
        "    #print(activations.shape)\n",
        "    #print(activations.mean(1).shape)\n",
        "\n",
        "\n",
        "\n",
        "    act_mean = activations.mean(1).mean(1).cpu().abs() * all_norms #Average once over batch size, then again over suffix length\n",
        "    #torch.save(act_mean, os.path.join(args.out_dir, 'act-mean.pt'))\n",
        "    #if gold_set is not None:\n",
        "    #    score = get_layerwise_scores(act_mean, gold_set, args.ratio)\n",
        "    return act_mean\n",
        "\n",
        "def remove_all_forward_hooks(model: torch.nn.Module) -> None:\n",
        "    for name, child in model._modules.items():\n",
        "        if child is not None:\n",
        "            if hasattr(child, \"_forward_hooks\"):\n",
        "                child._forward_hooks: Dict[int, Callable] = OrderedDict()\n",
        "            remove_all_forward_hooks(child)\n",
        "\n",
        "act_mean = largest_act(inner_dim=model.inner_dim, model=model, inputs=noise_data, gold_set=None, model_name=\"gpt2\", prompt_len=50)\n",
        "remove_all_forward_hooks(model)"
      ],
      "metadata": {
        "id": "n4b9QDhCd1qy"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "batch_size=1000\n",
        "seq_len = noise_data.shape[1]\n",
        "\n",
        "print(\"BEFORE MASKING---------\")\n",
        "\n",
        "perc_mem, acc, perplex_clean, perplex_noise = track_all_metrics(noise_data=noise_data,\n",
        "                      clean_data_corresponding_to_noise=clean_data_corresponding_to_noise,\n",
        "                      clean_test_dataloaders=clean_test_dataloaders,\n",
        "                      model=model,\n",
        "                      prompt_len=50,\n",
        "                      batch_size=1000)\n",
        "#NOTE Now the model is in its original state without ablations."
      ],
      "metadata": {
        "id": "17-_CLorUciM",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "d2cd9b7e-8cc9-49dd-bf41-852815327934"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "BEFORE MASKING---------\n",
            "perentage memorized:  54.796749114990234 %\n",
            "accuracy on clean data:  97.49462890625 %\n",
            "perplexity clean data:  107.07524871826172\n",
            "perplexity noise data:  107.4422607421875\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "apply_ablaton_mask_to_neurons(act_mean, model=model, ratio=0.1)\n",
        "\n",
        "print(\"\\n AFTER MASKING---------\")\n",
        "\n",
        "perc_mem, acc, perplex_clean, perplex_noise = track_all_metrics(noise_data=noise_data,\n",
        "                      clean_data_corresponding_to_noise=clean_data_corresponding_to_noise,\n",
        "                      clean_test_dataloaders=clean_test_dataloaders,\n",
        "                      model=model,\n",
        "                      prompt_len=50,\n",
        "                      batch_size=1000)\n",
        "\n",
        "#NOTE Now the model has ablated neurons."
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "rjn4ZEH1JQ9i",
        "outputId": "600b6d03-92b8-494b-b8b7-78a1196f5e52"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Num of dropped neurons per layer:  51\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "100%|██████████| 2/2 [00:00<00:00, 1979.85it/s]"
          ]
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "\n",
            " AFTER MASKING---------\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "\n"
          ]
        },
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "perentage memorized:  5.040650367736816 %\n",
            "accuracy on clean data:  93.6617431640625 %\n",
            "perplexity clean data:  121.97744750976562\n",
            "perplexity noise data:  123.21735382080078\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "remove_ablation_mask_from_neurons(model)\n",
        "remove_all_forward_hooks(model)"
      ],
      "metadata": {
        "id": "WFC7tijPiqws",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "62ede6a1-14e9-4943-96a7-abe0cfa5dd99"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "100%|██████████| 2/2 [00:00<00:00, 3123.09it/s]\n"
          ]
        }
      ]
    }
  ]
}