1.EntityRelationKG.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "initial_id",
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# 命名实体识别\n",
    "\n",
    "# 安装包：pytorch\n",
    "# pip install pytorch-crf\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/fubin/opt/anaconda3/lib/python3.9/site-packages/TorchCRF/__init__.py:249: UserWarning: where received a uint8 condition tensor. This behavior is deprecated and will be removed in a future version of PyTorch. Use a boolean condition instead. (Triggered internally at /Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/native/TensorCompare.cpp:493.)\n",
      "  score = torch.where(mask[i].unsqueeze(1), next_score, score)\n",
      "No artists with labels found to put in legend.  Note that artists whose label start with an underscore are ignored when legend() is called with no argument.\n"
     ]
    },
    {
     "data": {
      "text/plain": "<Figure size 640x480 with 1 Axes>",
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAjsAAAHFCAYAAAAUpjivAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjUuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8qNh9FAAAACXBIWXMAAA9hAAAPYQGoP6dpAAC/6klEQVR4nOydeZgU5bX/v9X7TM8+A8yM7C4oi0pwRRNUCGpEYjaiIVETb7z5XYMSNTGYRcxNxOTGmARjbjRENNyIMUo0GheISkJwRVBwAwWEwRmGgdmXXuv3R/f71lvVVdXV1dXL9JzP88yj9PR0V1d3v3Xe7/mecyRZlmUQBEEQBEGUKK5CHwBBEARBEEQuoWCHIAiCIIiShoIdgiAIgiBKGgp2CIIgCIIoaSjYIQiCIAiipKFghyAIgiCIkoaCHYIgCIIgShoKdgiCIAiCKGko2CEIgiAIoqShYIcgCOzduxeSJGH16tUFef7ly5dDkqSCPPfWrVsxZ84cVFdXQ5Ik/PKXv9S938DAAJYvX44XXnghr8cncs4552D69Ol5eS6r54UghgOeQh8AQRCFp6mpCS+++CKOPvroQh9K3vna176G/v5+rF27FrW1tZg4caLu/QYGBnDrrbcCSAQdpY7V80IQwwEKdgiCgN/vxxlnnFHowygIO3bswNe//nVceOGFBXn+wcFBBAKBgilbIrFYDNFoFH6/v+DnhSCchNJYBDEMeP/99/HVr34Vxx57LMrLy3HUUUfh4osvxvbt2y39/cMPP4zTTz8d1dXVKC8vx+TJk/G1r32N/14vjcVSS2+++Sa+8IUvoLq6GnV1dbj++usRjUbx3nvv4YILLkBlZSUmTpyIn/3sZ6rnfOGFFyBJEtasWYPrr78ejY2NKCsrw5w5c7B161ZLx/3QQw/hzDPPRDAYREVFBc4//3zLf7tjxw58+tOfRm1tLQKBAE4++WTcf//9/PerV6+GJEmIRqP47W9/C0mSDAOOvXv3YtSoUQCAW2+9ld/3yiuv5PfZtGkT5s6di8rKSpSXl2P27Nl48sknVY/DnvPZZ5/F1772NYwaNQrl5eUIhUIAgD/96U8488wzUVFRgYqKCpx88slYtWpVyvG8+uqr+PjHP87fy9tvvx3xeFx1n66uLtxwww2YPHky/H4/Ro8ejU996lN49913+WuSJAk/+9nP8OMf/xiTJk2C3+/HfffdZ/m8EMRwgYIdghgGfPTRR6ivr8ftt9+Op59+Gr/5zW/g8Xhw+umn47333jP92xdffBFf/OIXMXnyZKxduxZPPvkkfvjDHyIajVp67kWLFuGkk07CI488gq9//eu488478a1vfQuXXHIJLrroIqxbtw7nnXcebrrpJjz66KMpf3/zzTdj9+7d+P3vf4/f//73+Oijj3DOOedg9+7dps9722234bLLLsPUqVPx5z//GX/84x/R29uLj3/843j77bdN//a9997D7Nmz8dZbb+HXv/41Hn30UUydOhVXXnklD8ouuugivPjiiwCAz3/+83jxxRf5v7U0NTXh6aefBgBcddVV/L4/+MEPAAAbN27Eeeedh+7ubqxatQoPPvggKisrcfHFF+Ohhx5Kebyvfe1r8Hq9+OMf/4i//OUv8Hq9+OEPf4jFixejubkZq1evxrp163DFFVfgww8/VP1tW1sbFi9ejC9/+ct4/PHHceGFF2LZsmVYs2YNv09vby/OPvts/O53v8NXv/pV/O1vf8P//u//4rjjjkNra6vq8X7961/jueeew89//nM89dRTmDVrluXzQhDDBpkgiGFHNBqVw+GwfOyxx8rf+ta3TO/785//XAYgd3V1Gd5nz549MgD5vvvu47fdcsstMgD5jjvuUN335JNPlgHIjz76KL8tEonIo0aNkj/72c/y255//nkZgPyxj31Mjsfj/Pa9e/fKXq9X/o//+I+U52Ls27dP9ng88pIlS1TP3dvbKzc2NsqLFi0yfc2XXnqp7Pf75X379qluv/DCC+Xy8nLVuQAgX3PNNaaPJ8uyfOjQIRmAfMstt6T87owzzpBHjx4t9/b28tui0ag8ffp0eezYsfz133fffTIA+fLLL1f9/e7du2W32y0vXrzY9BjmzJkjA5Bffvll1e1Tp06Vzz//fP7vH/3oRzIAef369YaPxd7zo48+Wg6Hwym/t3peCGI4QMoOQQwDotEobrvtNkydOhU+nw8ejwc+nw+7du3CO++8Y/q3p556KoCEQvPnP/8ZBw4cyOi5FyxYoPr3CSecAEmSVF4Oj8eDY445JkWFAIAvfelLqjTIhAkTMHv2bDz//POGz/nMM88gGo3i8ssvRzQa5T+BQABz5sxJWxH13HPPYe7cuRg3bpzq9iuvvBIDAwOOKhX9/f14+eWX8fnPfx4VFRX8drfbja985StoaWlJUd8+97nPqf69fv16xGIxXHPNNWmfr7GxEaeddprqthNPPFF17p966ikcd9xxmDdvXtrHW7hwIbxeb9r7EcRwhoIdghgGXH/99fjBD36ASy65BH/729/w8ssv49VXX8VJJ52EwcFB07/9xCc+gb/+9a88eBg7diymT5+OBx980NJz19XVqf7t8/lQXl6OQCCQcvvQ0FDK3zc2NuredvjwYcPnPHjwIIBEoOb1elU/Dz30EDo6OkyP+fDhw2hqakq5vbm5mf/eKTo7OyHLckbPp73voUOHAABjx45N+3z19fUpt/n9ftXn4NChQ5YeS+9YCKIUoWosghgGrFmzBpdffjluu+021e0dHR2oqalJ+/ef/vSn8elPfxqhUAgvvfQSVqxYgS996UuYOHEizjzzzBwddYK2tjbd2/Qu2oyGhgYAwF/+8hdMmDAh4+esr69P8aYACe+T+PhOUFtbC5fLldHzaQ2/zPzc0tKSokbZYdSoUWhpabF0XzIfEyMBUnYIYhggSRL8fr/qtieffDLjlJTf78ecOXPw05/+FAAsVzZlw4MPPghZlvm/P/zwQ2zevNm0V835558Pj8eDDz74AKeccorujxlz587Fc889x4MNxgMPPIDy8nJbZfbs/GuVtGAwiNNPPx2PPvqo6nfxeBxr1qzB2LFjcdxxx5k+9vz58+F2u/Hb3/424+PS48ILL8TOnTvx3HPPOfJ4BDHcIWWHIIYBCxYswOrVq3H88cfjxBNPxJYtW/A///M/llIVP/zhD9HS0oK5c+di7Nix6Orqwq9+9St4vV7MmTMn58fe3t6Oz3zmM/j617+O7u5u3HLLLQgEAli2bJnh30ycOBE/+tGP8L3vfQ+7d+/GBRdcgNraWhw8eBCvvPIKgsEgb/Cnxy233IInnngC5557Ln74wx+irq4O//d//4cnn3wSP/vZz1BdXZ3x66isrMSECRPw2GOPYe7cuairq0NDQwMmTpyIFStW4JOf/CTOPfdc3HjjjfD5fLj77ruxY8cOPPjgg2nVk4kTJ+Lmm2/Gf//3f2NwcBCXXXYZqqur8fbbb6Ojo8P0teqxdOlSPPTQQ/j0pz+N7373uzjttNMwODiIjRs3YsGCBTj33HMzfv0EMZyhYIcghgEsOFmxYgX6+vrwsY99DI8++ii+//3vp/3b008/Ha+99hpuuukmHDp0CDU1NTjllFPw3HPPYdq0aTk/9ttuuw2vvvoqvvrVr6KnpwennXYa1q5dm7Zb87JlyzB16lT86le/woMPPohQKITGxkaceuqp+MY3vmH6t1OmTMHmzZtx880345prrsHg4CBOOOEE3HfffareOJmyatUqfPvb38bChQsRCoVwxRVXYPXq1ZgzZw6ee+453HLLLbjyyisRj8dx0kkn4fHHH08xeBvxox/9CMceeyxWrlyJxYsXw+Px4Nhjj8W1116b8XFWVlZi06ZNWL58Oe655x7ceuutqK2txamnnoqrr74648cjiOGOJIv6MkEQhEO88MILOPfcc/Hwww/j85//fKEPhyCIEQx5dgiCIAiCKGko2CEIgiAIoqShNBZBEARBECUNKTsEQRAEQZQ0FOwQBEEQBFHSULBDEARBEERJQ312kOh0+tFHH6GyspJapxMEQRDEMEGWZfT29qK5uRkul7F+Q8EOEvNrnJhHQxAEQRBE/tm/f79pR3kKdpDoNgokTlZVVVWBj4YgCIIgCCv09PRg3Lhx/DpuBAU7UKb+VlVVUbBDEARBEMOMdBYUMigTBEEQBFHSULBDEARBEERJQ8EOQRAEQRAlDXl2CIIgCIIoGLFYDJFIRPd3Xq8Xbrc76+egYIcgCIIgiLwjyzLa2trQ1dVler+amho0NjZm1QePgh2CIAiCIPIOC3RGjx6N8vLylGBGlmUMDAygvb0dANDU1GT7uSjYIQiCIAgir8RiMR7o1NfXG96vrKwMANDe3o7Ro0fbTmmRQZkgCIIgiLzCPDrl5eVp78vuY+TrsQIFOwRBEARBFAQrPhwnZlZSsEMQBEEQRElDwQ5BEARBECUNBTsEQRAEQZQ0FOwQBEEUCbIsYygSK/RhEETekGXZkfukg4IdgiCIIuGaP72O036yAUf6w4U+FILIKV6vFwAwMDCQ9r7sPuxv7EB9dgiCIIqE1/Z2omcoij0d/agL+gp9OASRM9xuN2pqanjDwHRNBWtqarIaG0HBDkEQRJEwEE6ksGLx7GV7gih2GhsbAYAHPEawcRHZQMEOQRBEESDLMgbCUQAU7BAjA0mS0NTUhNGjR9MgUIIgiJFAKBoHi3Eo2CFGEm6325GAxgwyKBMEQRQB/aEo//9oPF7AIyGI0oOCHYIgiCKA+XUAUnYIwmko2CEIgigCxGAnSsEOQThKQYOd5cuXQ5Ik1Y+R4/o///M/IUkSfvnLX6puD4VCWLJkCRoaGhAMBrFw4UK0tLTk4egJgiCcoz+spLHiFOwQhKMUXNmZNm0aWltb+c/27dtT7vPXv/4VL7/8Mpqbm1N+t3TpUqxbtw5r167Fpk2b0NfXhwULFiAWoy6kBEEMHwZJ2SGInFHwaiyPx2NaP3/gwAF885vfxDPPPIOLLrpI9bvu7m6sWrUKf/zjHzFv3jwAwJo1azBu3Dhs2LAB559/fk6PnSAIwilEgzJ5dgjCWQqu7OzatQvNzc2YNGkSLr30UuzevZv/Lh6P4ytf+Qq+/e1vY9q0aSl/u2XLFkQiEcyfP5/f1tzcjOnTp2Pz5s2GzxkKhdDT06P6IQiCKCTk2SGI3FHQYOf000/HAw88gGeeeQb33nsv2traMHv2bBw+fBgA8NOf/hQejwfXXnut7t+3tbXB5/OhtrZWdfuYMWPQ1tZm+LwrVqxAdXU1/xk3bpxzL4ogCMIG6mosKj0nCCcpaBrrwgsv5P8/Y8YMnHnmmTj66KNx//33Y86cOfjVr36F119/PWVeRjpkWTb9m2XLluH666/n/+7p6aGAhyCIgjIQFtNYBTwQgihBCp7GEgkGg5gxYwZ27dqFf/3rX2hvb8f48ePh8Xjg8Xjw4Ycf4oYbbsDEiRMBJOZqhMNhdHZ2qh6nvb0dY8aMMXwev9+Pqqoq1Q9BEEQh6Q+RskMQuaKogp1QKIR33nkHTU1N+MpXvoI333wT27Zt4z/Nzc349re/jWeeeQYAMGvWLHi9Xqxfv54/RmtrK3bs2IHZs2cX6mUQBEFkzEBE7KBMnh2CcJKCprFuvPFGXHzxxRg/fjza29vx4x//GD09PbjiiitQX1+P+vp61f29Xi8aGxsxZcoUAEB1dTWuuuoq3HDDDaivr0ddXR1uvPFGzJgxg1dnEQRBDAcGQtRBmSByRUGDnZaWFlx22WXo6OjAqFGjcMYZZ+Cll17ChAkTLD/GnXfeCY/Hg0WLFmFwcBBz587F6tWrcz5UjCAIwknEpoKk7BCEsxQ02Fm7dm1G99+7d2/KbYFAACtXrsTKlSsdOiqCIIj8M0izsQgiZxSVZ4cgCGKk0k/BDkHkDAp2CIIgioCBEKWxCCJXULBDEARRBFBTQYLIHRTsEARBFAEDZFAmiJxBwQ5BEEQRIHp24hTsEISjULBDEARRBAzSIFCCyBkU7BAEQRQYWZZVfXaoGosgnIWCHYIgiAIzFIlDFuIbUnYIwlko2CEIgigwojkZAGIxCnYIwkko2CEIgigwYtk5AMRkCnYIwkko2CEIgigw/Vplh9JYBOEoFOwQBEEUGK2yQ54dgnAWCnYIgiAKzEBIk8aiDsoE4SgU7BAEQRQYSmMRRG6hYIcgCKLADGoNyhTsEISjULBDEARRYLTKDnl2CMJZKNghCIIoMMyzI0mJf5OyQxDOQsEOQRBEgWHVWBV+DwAgSk0FCcJRKNghCIIoMKyDclXAC4CaChKE01CwQxAEUWCYZ6cykFB2KI1FEM5CwQ5BEESBYZ6dqrKEskMGZYJwFgp2CIIgCgzz7PA0FjUVJAhHoWCHIAiiwPRzzw4ZlAkiF1CwQxAEUWCYssM8O3EyKBOEo1CwQxAEUWB4Gos8OwSREyjYIQiCKDADVI1FEDmFgh2CIIgC0x9SG5TJs0MQzkLBDkEQRIEZ5MoOq8aiYIcgnISCHYIgiAISj8sYiKgNytRBmSCchYIdgiCIAjIUjYHFNsygTMoOQTgLBTsEQRAFhFViSRJQ4XcDAKIxaipIEE5CwQ5BEEQBYaMiyrxueN2JJZmUHYJwFgp2CIIgCgjrnlzu88AlSQCozw5BOE1Bg53ly5dDkiTVT2Njo+r3xx9/PILBIGprazFv3jy8/PLLqscIhUJYsmQJGhoaEAwGsXDhQrS0tOT7pRAEQdiCpbGCfjc87kSwQx2UCcJZCq7sTJs2Da2trfxn+/bt/HfHHXcc7rrrLmzfvh2bNm3CxIkTMX/+fBw6dIjfZ+nSpVi3bh3Wrl2LTZs2oa+vDwsWLEAsFivEyyEIgsgI1lCwzOuG20XKDkHkAk/BD8DjUak5Il/60pdU//7FL36BVatW4c0338TcuXPR3d2NVatW4Y9//CPmzZsHAFizZg3GjRuHDRs24Pzzz8/58RMEQWQDaygY9HvgcSX2n7KcKEl3JYMfgiCyo+DKzq5du9Dc3IxJkybh0ksvxe7du3XvFw6Hcc8996C6uhonnXQSAGDLli2IRCKYP38+v19zczOmT5+OzZs3Gz5nKBRCT0+P6ocgCKIQDEaYZ0dRdgBSdwjCSQoa7Jx++ul44IEH8Mwzz+Dee+9FW1sbZs+ejcOHD/P7PPHEE6ioqEAgEMCdd96J9evXo6GhAQDQ1tYGn8+H2tpa1eOOGTMGbW1ths+7YsUKVFdX859x48bl5gUSBEGkgSk72mCHKrIIwjkKGuxceOGF+NznPocZM2Zg3rx5ePLJJwEA999/P7/Pueeei23btmHz5s244IILsGjRIrS3t5s+rizLkCRj+XfZsmXo7u7mP/v373fmBREEQWQI8+wEfR54xGCHTMoE4RgFT2OJBINBzJgxA7t27VLddswxx+CMM87AqlWr4PF4sGrVKgBAY2MjwuEwOjs7VY/T3t6OMWPGGD6P3+9HVVWV6ocgCKIQsGqscr9G2aFhoAThGEUV7IRCIbzzzjtoamoyvI8sywiFQgCAWbNmwev1Yv369fz3ra2t2LFjB2bPnp3z4yUIgsgWHuz4PHBLomeHuigThFMUtBrrxhtvxMUXX4zx48ejvb0dP/7xj9HT04MrrrgC/f39+MlPfoKFCxeiqakJhw8fxt13342WlhZ84QtfAABUV1fjqquuwg033ID6+nrU1dXhxhtv5GkxgiCIYqc/pBiUXS4JLgmIy+TZIQgnKWiw09LSgssuuwwdHR0YNWoUzjjjDLz00kuYMGEChoaG8O677+L+++9HR0cH6uvrceqpp+Jf//oXpk2bxh/jzjvvhMfjwaJFizA4OIi5c+di9erVcLvdBXxlBEEQ1hhkTQV9ieXY7ZIQj8lUjUUQDlLQYGft2rWGvwsEAnj00UfTPkYgEMDKlSuxcuVKJw+NIAgiL7BxEWW+xAbN7ZIQicmk7BCEgxSVZ4cgCGKkIY6LAMAbC1KwQxDOQcEOQRBEARENygBoZARB5AAKdgiCIAqIaFAGwHvtkLJDEM5BwQ5BEEQB0So7Lq7sUOk5QTgFBTsEQRAFJNWzkwh2KNYhCOegYIcgCKKAsHER5V6tZ4eiHYJwCgp2CIIgCkQ8LqvGRQDk2SGIXEDBDkEQRIEYisb4/wepGosgcgYFOwRBEAWiP5QIdiQJCHgTy7Gbe3Yo2CEIp6BghyAIokAofh03pOQQUHeyqSApOwThHBTsEARBFAjFr6NM7iHPDkE4DwU7BEEQBYIrOz5lcDF5dgjCeSjYIQiCKBDMs8MaCgKiskOl5wThFBTsEARBFAjeUFBQdlw82CnIIeWEaCyOZY9ux6OvtxT6UIgRCgU7BEEQBYKlscqEYMdTgk0Ftx/oxoOv7MMdz+4s9KEQIxQKdgiCIApEP1d2lDSWuwQNyr1DiaCuZzBS4CMhRioU7BAEQRSIQWZQ9uspO6UT7DAFqzcULakgjhg+ULBDEARRIBSDsliNlViWSykoYN4kAOgLRQt4JMRIhYIdgiCIAsEUD3UaK/HfUgp2+oVgh1JZRCGgYIcgCKJA8KaCqtLz0lN2WLoOUPw7BJFPKNghCIIoEEqwU9pNBVm6DgB6h0jZIfIPBTsEQRAFoj9kbFAupaaCgxEhjUXKDlEAKNghCIIoECwI0Cs9Ly1lR0xjkbJD5B8KdgiCIAoECwLKdNJY8RIKdgbDYhqLlB0i/1CwQxAEUSAGTJoKlpSyIxiUqRqLKAQU7BAEQRQIblDW9ewUPtg51BvCskffxPaW7qweR+yz00t9dogCQMEOQRBEgWB9dvSaChaDsvPUjlY8+Mp+rNq0O6vHGaA+O0SBoWCHIAiiQLCSbDGN5XEXj7LDgpRsK6gGyLNDFBgKdgiCIApAPC7zaixR2XFJxRPsRKKJ8vf+LFNPA6Jnh6qxiAJAwQ5BEEQBEHvPqDsoF1GwE0sEO+Kx2kGVxiJlhygAFOwQBEEUAFahJElAwKssxUo1VuGbCoZjiYAra2WH+uwQBYaCHYIgiAIwIPh1pGTqCihOZUdUZjJFlmUMiB2UB0nZIfIPBTsEQRAFQG8uFgC4kwblaKzwwU40lr1nZygShyy8FFJ2iEJQ0GBn+fLlkCRJ9dPY2AgAiEQiuOmmmzBjxgwEg0E0Nzfj8ssvx0cffaR6jFAohCVLlqChoQHBYBALFy5ES0tLIV4OQRCEIYPhmCpo0Cs7BwA3MyjLhQ92WBorG2VHbCgIAKFoHKFodh4ggsiUgis706ZNQ2trK//Zvn07AGBgYACvv/46fvCDH+D111/Ho48+ip07d2LhwoWqv1+6dCnWrVuHtWvXYtOmTejr68OCBQsQi9GXiSCI4uEzd/8bp/x4A1Zt2oNYXEY/V3Y8qvu5izCNFY3LCEfteYjYqAifR7ncUPk5kW886e+S4wPweLiaI1JdXY3169erblu5ciVOO+007Nu3D+PHj0d3dzdWrVqFP/7xj5g3bx4AYM2aNRg3bhw2bNiA888/Py+vgSAIwox4XMa7bb0AgP9+4m088eZHOG/KaABA0K9WdjxFNC6CBTtAQonyeXwZPwZTdir9HoTccfSFougdiqKhwu/YcRJEOgqu7OzatQvNzc2YNGkSLr30Uuzebdyps7u7G5IkoaamBgCwZcsWRCIRzJ8/n9+nubkZ06dPx+bNmw0fJxQKoaenR/VDEASRK0KCKlLuc2Prvi7csX4nAKBMq+y4E8tyrCg8O8ox9NtMZYkjMaoCiddKXZSJfFPQYOf000/HAw88gGeeeQb33nsv2traMHv2bBw+fDjlvkNDQ/jud7+LL33pS6iqqgIAtLW1wefzoba2VnXfMWPGoK2tzfB5V6xYgerqav4zbtw4Z18YQRCEwJBQjfTstz6B844fzf8d9BWvshMWlR2bJmVWdVbu9aAy4AVAaSwi/xQ02Lnwwgvxuc99DjNmzMC8efPw5JNPAgDuv/9+1f0ikQguvfRSxONx3H333WkfV5ZlVSmnlmXLlqG7u5v/7N+/P7sXQhAEYcJQ0pDrdUsYW1uOVVecgl9dejJOGluNz8w8SnVfZlCOF4FBWZ3GsqvsJI3Yfjcqk8oOVWQR+abgnh2RYDCIGTNmYNeuXfy2SCSCRYsWYc+ePXjuuee4qgMAjY2NCIfD6OzsVKk77e3tmD17tuHz+P1++P2ULyYIIj8MRRJBQ8CTUHEkScKnTz4Knz75qJT7uotI2RGDHW1VlVXEEnt/8vXTyAgi3xTcsyMSCoXwzjvvoKmpCYAS6OzatQsbNmxAfX296v6zZs2C1+tVGZlbW1uxY8cO02CHIAgin7A0lt/rTnNPcRBo4TsoRwTPDktHZcqAUHWmKDuUxiLyS0GVnRtvvBEXX3wxxo8fj/b2dvz4xz9GT08PrrjiCkSjUXz+85/H66+/jieeeAKxWIz7cOrq6uDz+VBdXY2rrroKN9xwA+rr61FXV4cbb7yRp8UIQks0FofHXVQxPjECYMGOOBbCCK7sFIFB2RllR+knVDnMDMpdA2Fs+bATc44bRevGMKeg715LSwsuu+wyTJkyBZ/97Gfh8/nw0ksvYcKECWhpacHjjz+OlpYWnHzyyWhqauI/YqXVnXfeiUsuuQSLFi3CWWedhfLycvztb3+D251+B0WMLHYc6MaJtz6L3238oNCHQowweBrLirJThH12gGw8O4qyU5U0KA+XYaD//cQ7uOr+1/DUDuOCF2J4UFBlZ+3atYa/mzhxImQLBr1AIICVK1di5cqVTh4aUYJs3deJgXAMm97vwH/OObrQh0OMIJhB2Yqy4yqiDsqRqJDGshns9KuUneFVjfXS7kRl8MGeoQIfCZEtpMsRIwa2u852gjNBZEqIpbE8mXh2iiDYcaD0nHVQDvrcqCpLprGGgUH5UG8IB7oGAaj7JBHDEwp2iBED8030UbAzItm0qwN3v/C+JcXYaTJJY7ldiWW5KDw7cdGzY1PZSRqby3xin53iD3bebOni/293VAZRPBRV6TlB5BKWSui3WVVCDG9+8NgO7Onox+yjG3DyuJq8PncmBuWi8uyo0lg2lZ1I4u+Cqj47xb/h2La/i/8/KTvDH1J2iBED210Ph10l4SyyLKO1O5GSONA5mPfnz6T0nA8CLQbPjliNZXOTwJUdr1swKBf/d1Ad7NAGabhDwQ4xYhDTWIVIZRCFYyAc48Fue2/+zaZDUXVTQTOKaeq5OC6CKTSZwj07fg+fjVXsyo4sy3iDlJ2SgoIdYsTALnZxWfl/YmRwpD/M/7+9N5T357fVZ6cImgqqBoHaVXaS6a8ynxtVZUo1VjFvOPZ09KvK40O0Xgx7KNghRgxDghTdGyp+GZ1wjo4+JcApRBmxrT47xWBQVvXZybYaS+mgHIvLtkvZ88EbgjkZoDRWKUDBDjFiCAmTp8mkPLI43KcoO4eGjbJT2GAnHpdVx5CtslPuc6PM6+avr5hTWW/s7wYAVPgTwRmlsYY/FOwQIwYxddVXxAst4TxiGstpZUeWZVy3ditu/dtbhvdhykAmnp1CTz2PaNJog5FsOyi7IUkS9+0Us0mZmZNPmZgYME3BzvCHgh1ixDAkLNbUa2dk0dGvqDlOe3ZaOgfx2LaPcN+/9xqaiu2ksQqt7Gj7/NhpxinLSroqmFRJ8tFr5/32PrzwXrutvw1FY3j7ox4AwGmT6hK32Qz0iOKBgh1ixCB6dijYGVmIaayugYijHozDgmo0ZHBRzCyNlbhPoT07ol8HsDcuIhyL8wCwzJcI9HgX5cHcfQe/+afXceV9r2Lf4YGM//bd1l6EY3HUlHtx7OhKAKTslAIU7BAjBlUaiwzKIwoxjQUA7T3OqTsdglJkFBBk0menWJSdsCbY6Q9nXkE1IPh8ypOvvdKf+147TL071Jf5+8zMySeNreHBKQU7wx8KdogRgzqNRbL0SKJDc9FzMpV1WEiRGSs7iYul32PdoFzoPjsRjbIk22jZMJA8Hz6PCx534rXno4sySzvZGfPA/Donj6uBP+mxomqs4Q8FO8SIgQzKIxcxjQUA7Q6alDuExzZUdvjU8+HTQTmSDBTKfcoxZ1p+zoaHBoXHYL12cqnssCaOWnXKCupgJ6nsUJ+dYQ8FOyWCLMv469YDeL+9t9CHUrSoS88p2BlJsDTW2NoyAA4rO0KwY1SxlNkgUEXZKWTjPdbU0O9xoSx53Jn6dvp5JZYyhjHXyk5U8Allqux0D0aw+1A/AODEsdXwUxqrZKBgpwgZDMdwOMNc89b9XVj60DZ8/YEtRd2ZtJDk26D87YffwA8f25Hz5yHMkWWZp5pOaKoC4OzICDFFNmgQDLBAO2AhjcU8O0BhU1nh5BBQj9uFoD8R7PRnquwIPXYYrBqrZzA3yo4YmGQa7GxvSfTXGVdXhvoKP3xuFuxQGmu4Q8FOEXLpPS/i4z97Ht0D1heD/UcSVQd7OvrxdmtPrg5t2BKLyyoPQq4bmnUPRPDwlhY88OKHhj6OkYYsywUZwtozFOXv/QmNieqagw4alEXPjtH8KKUay7qyAxTWpMyqsXxuF1dmMm0syAzKYrCT6/lYqmAnltnxbtvfCSBhTgYUQzkpO8MfCnaKkF3tfRgIx7DviPWyyU6h2uSp7W25OKxhjTbgyHUaKyQsspTvT/CjJ97Gx/57Pd7JczDOUlgVfg/G1ZUDyGEaK6z/XvNBoJaqsZRluZDKDgt2vG6JBysZe3YiqWmsqhz32RFVmEyVnW3Jzsknj6sBoBjKw9E4KebDHAp2igxZlvmFuTsDmfeIoAL9fUcrfTE1aHdmuU5jiSrSEEngAIDX93UhEpN5w7Z8wVLCdUEfxlQFADhtUBZLz9MpO+mXXCHWKahJmX2GvW6XEOxkquykprF4n50cKTtiIYJZsLP/yADWvPQhnt7Rih0HutHZH+Zl59pgByB1Z7jjSX8XIp9E4zLYZi6TagVR2dl9qB+72vtw3JhKpw9v2KJVdnId7ESFKhBKYyVgO/mBPJ8PVi1VX+HD6Co/AOeUnVhcVvXw0XuvxQ1MxspOARsLMmUn4dlJXCoyVnaYQdkvGpTzp+yYBSg/eGwHXnjvUMrtbpeEac3VAMBLz4FEZZeV948oTkjZKTLEL2dGyk5ywZWS6f6/b2919LiGO/kOdlTKDqWxACgdcwfyXAnHvhv1QT9GVwb4bXZ6sGjpGghDzDTpKR+RmLKBsTIbS7DsFIlnR0ljZezZYQZlr2hQzm0HZTFtbFZ6zhpLTmoIYlSln99+9jENvNuz1y3xNZXS0cMbUnaKDLE82k6w84ljR2HjzkN4ansbls47zvHjG65oA45ce3YipOykwJUdG2MH0vHfT7yNI/1h3PGFk+ASowUoaaz6oA+15V543RIiMRmH+kI4qqYsq+ft0PTv0Ss9F9OYfgtpLEmS4HFJiMbljDw7sixDkqT0d7SI4tlxIejLVtkRDcq5VnaspbFYIHTbZ2bgzKPrMRSJ4VBvCE3VAX4fSZLg97gwFIlTRdYwh5SdImNI+HJmUprZOZBYdL9wylh4XBLeO9iLDw71OX58wxWtbybXTQXFIYp2p0WXEqFojF+EnD4f77f3YdWmPVi39QDe1/nMs9lV9RU+SJLE1R0nfDvaFhF6pecs2JUkax2UAaUiKxpPryb0DkVw9k+fw/V/fsPSY1tF9OyU2fXsCBPPGUzZ6Q/HVOlepxA3F2bBDgtefMn3JOB1Y1xdOe/0zFC6KJOyM5yhYKfIyFbZmVgfxFnHNAAAnt5BVVkMtgDWlCd2lX025vxkQpiUHRVimbHTqtrjb3zE/3+/TgWjEuwkUhXMt+NE+XlHf3plJySMirCqvLBgx0Ksg7c/6kFL5yDWbT2AA12Dlh7fCopnRxI8OzbTWL5Uzw6Qm3SyZWUnam2EB3VRLg0o2Cky7Hh2ZFnmyk5d0IcLpzcCIN+OCFuoGpIXPFnOTTqFoTYo0yIpBjtGjffsIMsyHt92gP/7Q50p12IaCwBGJ/0ZhxxoLKhVdvQ+U5mYkxmZKDtiVdPf33TuO6/us8M8O5kFJ/06yo7P4+JVabnotaMqPTdRjkJWgx0vNRYsBSjYKTLEYMdqaWZvSGmaVlvuw/xpjXC7JLz1UQ/26Sz+IxF2wakt93IDaC5NyqJBmRZJdUrWySDzzZZu7BU+43q9qQ4L1VgAlDSWAxVZrOycBSe6nh02KsKCOZnhyWAYqOh9eeLNj0zumRlhIY2leHYye+9YYBv0qe2hTN3JRL22iqjAmKkxirJj/r5QGqs0oGCnyBiykcZiZedlXjfKfG7UBX04fVIdAOCpHaTuAOpBjBVJST6nwU6c0lgi4g7eydLzx7YlLu5sd26axgomFJ0xPI3lhLKTeGxmatX17PDPnvXl1p0sP7dSjSUGkm+0dOueAzswddLrcXGDccbKDuuz41cHFLnsoqzuoJxe2fFZTWNRsDOsoWCnyBC/UL0Wgx3m16lLyvQAcOGMJgDAU+TbAaDsrv0eIdjJoUk5EqU0loioPjhVeh6Ly1zJuOy08QCADzUX+nhcxpHkOIfcKDuJ79642kRnZjODciZprEyUHa0C/KRD6Wu9DsqZmssHI6lpLCC3vXasGJTFYaHWPTu0aRnOULBTZNgxKIt+Hcb508ZAkoBt+7scNS0OV8QOthWB3Cs74o6clB11g0yn0lgv7z6M9t4Qasq9+PIZEwAklJ24cO67BiO8x01teTLYcdKgnExjjatLlLDrqVY80Lbh2bEU7CTXCeZJciqVxauxXOJsLJvKTkoaK3ddlMUNo5EaIyo+6ZUdSmOVAhTsFBlag7KViqEj/YnFrlYIdkZXBjAz2fL85d2HnT3IYQj3TeQrjUUGZRUqg7JDwR9LYX1qRhMm1JfD7ZIQisZxSDANM1WnuszLL2pM2XHEoJx8/LFJZWfITNmxWHYOiAZlK8pO4vt/ycyj4HZJ2HGgB3s7+i0/lxFMFfF6pKw9O1plp6osd8qOFYOyqPhYNyjT93g4Q8FOkSGqANG4bOnCwDw7deVe1e1NyYZpXRlMTy9VRGUnmI80Fs3GUtHjcOl5KBrD35N+tIUnNcPrdqG5JhHEiBVZHRpzMqAoO4f7w6qg1A7Ms6MoO6mvLedprGQn4okNQcw+uh6AM6kssakg9+xk2FRQqcZSKzt58+wYfPfYfVwSUvrqaFE8O/Q9Hs5QsFNkaHcPVlJZzIApKjuAsqDkejTCcICbRD1uoalZvpQdWiRFE60TpecvvHcIvUNRNFYFcNrEhBl/fHKiuViRxSuxhO9GXbkPHpcEWVYP8cyUgXCUKx2KZyc1eFImnttRdiw0FQwlzm1VwIMFJya8ek84UILOVCWfUI2V6XtnqOwkPTuZNE61SsjCIFCrlVjifajPzvCGgp0iw06woyg76mAn1wP3hhO8sZtXWbhzsatkUJ8dNdpqrGwbOj6eTGEtPLmZj4cYXxcEoA52uDk5qMw+crkkPgupPQvfDguk/B4X7980qBNAh7Los5OJslMV8OL8aY3wuCS809qTdQd1FhB4bM7GisTiPI2UWnqeu+/gkIU0lrZ7shk+qsYqCQoa7CxfvhySJKl+Ghsb+e8fffRRnH/++WhoaIAkSdi2bVvKY4RCISxZsgQNDQ0IBoNYuHAhWlpa8vgqnEUrlVoZlndkQF/ZYd6UXF7UzYjF5aJRNRTfhJsblHM5H0vVZ6dIzkEhEQPuWFzO6sLRF4piwzsHASRSWAyu7BxW/Cp6aSwAGF2VSHllU37OVKGGCr+qWkkbyImfPatkFOwkz21VmQc15T6cfWyig/qTWao7qjSW8PqszusS/T1lBtVYPbnw7FhQdqw2FBTvQ2ms4U3BlZ1p06ahtbWV/2zfvp3/rr+/H2eddRZuv/12w79funQp1q1bh7Vr12LTpk3o6+vDggULEIsNzw+mVirNRNmpD2qVnWSwk+c01p6Ofvz06Xdx5op/YNZ/r8dHRVANJvom8m5QpkUy5aKWTSrr2bfaEIrGcfSoIKY1V/HbddNY/eruyQzWRTmb8nOm7DRU+PjFPC6nKgCKOd76cpuZZ4elsRIBxEUzWCoru6os1SBQv6LMWDWYs1ERXreUoqBUleXSs5O+9Nxqjx2AqrFKhYJPPfd4PCo1R+QrX/kKAGDv3r26v+/u7saqVavwxz/+EfPmzQMArFmzBuPGjcOGDRtw/vnn5+SYc4n2wmglp22k7ChprNwHO7Is4/E3PsL/vbwPr+w5ovrdB4f60JzldOlsES84cTn3BmV16XnuF8n7/r0H//fyPtx35akYl7zoFxPaz+BAJIZam4/19kc9AIBzpoxWzZqaUM+CHSW4PqKZi8VwIthhyk59hR9lQopqKBJTpayyGxdhHuzIsszN36zCaf60Rty8bjt2HuzDzoO9OG5MpeXnFWHDbH1uF/weF1xSIpgbCEX5hsEMpuyU6bzuSn8uq7Ey8exYCHZYNRalo4c1BVd2du3ahebmZkyaNAmXXnopdu/ebflvt2zZgkgkgvnz5/PbmpubMX36dGzevNnw70KhEHp6elQ/xUI2yk6dYRor956dZ946iOvWbsMre47AJQHnTBnFu8oWwyLBgki/152XPjthVVPB3Co7sizj3n/uxvvtfXjxg/y0GejsD2fku9EGO3relkwfq1ZTfciCvI6+EFcVjNJYY6qyn3x+WFBUPW4XfMmqHq3yIX72rOJJdlBOp+wMhJW0ElN2qsu8+PixowAAG987ZPk5tYSFpoKSJPGKKqvl5wNJf09QJzBigVmu++wYe3aYsmPFoOxKPhYptMOZggY7p59+Oh544AE888wzuPfee9HW1obZs2fj8GFrC3ZbWxt8Ph9qa9V7xDFjxqCtzbhz8IoVK1BdXc1/xo0bl9XrcJJMDcqxuIyu5H1qyw2qsfKg7OxN+iROm1SHf3/3PKz+6mm8QsWsZXu+yHcaK5rHcRF7Dw/go+7ERZs1mMwlm9/vwMz/Xo+fP/ue5b/RprEyMboaPRa7YDKqy7yoTt7GUllsUKd2I+C0sgMoaSptMGAnjWVV2WGBn8clqR7/pLE1AID3DvZafk4tytTzxONyk7LFQJUFnFq/DiAalHPbQdmwqWBGnh2qxioFChrsXHjhhfjc5z6HGTNmYN68eXjyyScBAPfff39WjyvLskre1rJs2TJ0d3fzn/3792f1fE7C8s0sZ5/OwNc1EAbbYNdodroVOax40NKRvGicPK4GTdWJlBXLhxtJyflEGcbo4sFOvgzKuU5j/fv9Dv7/nXnoqfTK3kSactP71jYlsizzzyA799l0UWaPxS6YIopJORHssDRWgyaNNcYRg7Li2QGUXjJaP1I2BuV4mmBHDPzENW9KYwUAYGdWwY6SxgIUhcaysmMwBBQQOihbKMDIlEzSWNY8O1SNVQoUPI0lEgwGMWPGDOzatcvS/RsbGxEOh9HZ2am6vb29HWPGjDH8O7/fj6qqKtVPscB2D2znmU7ZYTv5qoAHXk1zLObZyUefHaUyRdlBF1MVg6jssEU7l8btfBqUN38gBDv9uVd2DnQmPDG72/sspbLEVAsbwjmo03zPKkwNYL4PkfH1ikk5Govz4E9rUB7liEFZqcYCFAUjJY0ldO+2ilVlRzEnqwOKY5M+nV0H+9IGTEZwg7IncSxK+bm1967fRNlhqlw4Fndc+QxplB29zyhbkzLy7BTBOkbYp6iCnVAohHfeeQdNTU2W7j9r1ix4vV6sX7+e39ba2oodO3Zg9uzZuTrMnMK+UKOSO890BmU2KkJrwASU3VNfKGq5XNQuyi5XOY5iUnZCUeWCk5dBoGKw49AsKD3icRmbBZ/OkTyksT7qTgQ7vaEoDlkIFpgS43ZJ/PORTRrLkrJzZICfC0kCasq1pefJLsp9IVVPpEw4rPEDMSOuVtkJ2Zh6rlRjmR+bUUpvQl05fB4XBiMxtHTaq4YUq7EAZDwyQlF2UoOdCp8HTIhyWnnWrjeiyqq9T0ZprCJYxwj7FDTYufHGG7Fx40bs2bMHL7/8Mj7/+c+jp6cHV1xxBQDgyJEj2LZtG95++20AwHvvvYdt27ZxP051dTWuuuoq3HDDDfjHP/6BrVu34stf/jJPiw1H2BdqVPKikE7mZTK91qwJQFUxkWt1p0OzywWKqxmXOC6iMg99dqKqcRG5e/1vt/aoxoF0mQQ7sixnZchlHBAunu9baFzHlZiAR+nXkkUA2MODHR1lRwx2hGabTClh1Af9cLskxGXFaJwp2s88UzBSPTu5q8YSGwqKeNwuHD0qkcqy69thQQIzSxu9PiMGDIaAAonGjhU+NgzU2dSrVinS8wxmVnpO1VilQEGDnZaWFlx22WWYMmUKPvvZz8Ln8+Gll17ChAmJCcaPP/44Zs6ciYsuuggAcOmll2LmzJn43//9X/4Yd955Jy655BIsWrQIZ511FsrLy/G3v/0Nbrf1haWY4GmsqszSWFoDJpBYXFm+PffBTqqyU0y5br00Vn84ZlviT0fY4riIta/swx0ZGH21ML8OM+YeMblw3/3CBzjttn9g/dsHbT9fPC7joy4lYNp9KP3AyR4x2OG+DwfSWDrKzgQh2NEqLyIJlSlxu50uyrG4zJUjrbKjfb9t9dlxW/PsmJ2LKWOy8+0wZceXTGMF/SzYsWhQjuiPimAow0CdXZu0642espzZuAhKY5UCBe2zs3btWtPfX3nllbjyyitN7xMIBLBy5UqsXLnSwSMrHMzfwTw76XY9irKTuqADiUXwcH84uSjmptdNLC7ztvyiZ6eY0liKQdmtUrz6w1FdhSBbVMpOsquunmn+v594G/3hGL546jg+OTsT/p1MYZ0/bQz+/FqLqUF52/4uAImL3yenGnvazOjoD6kCOSsjCXgfmIAX5cmAYMCmTyMcjfOLmVbNAJTy85YjgzzFprcRABIm5YM9IRzsGcIMVGd0HJ3JwgBJUsa0pFV2MjAouySLys6QvrIDKL4du8EOn3rOq7GYIpp96TkgmpSdVXasBDt8XESaIaCA0jKgGDZthH2KyrNDiAblhGcnnbJzxKDHDiMfFVmdA2HE2cIvHIcvqa4VuvRclmVlEKg30SCNeSJypXiJpedxWd83EIvLfCq0neMIRWN4NdnAccGJibEJXQNhQzWAGWqzMYQe0Pg/PrCg7IgeG7bLH7Dp2RFLlSt01Iym6gA8LgnhWBxvtyb6Z+n52YDsys+ZalRb7uOl2dyz40ifHWsdlLlBuUxP2WHBjr0ZWSzQUjw7GSo7rKmgkbKTo6anWgXGVNmxoLYVk0JN2IeCnSIjpFF2BsIxldlVS6fBxHNGZR567TDvgrjwA8XTeTQci/PyfL/XDUmScj4fKxzVzEfSkcDFi4YdD8vWfV0YjMTQUOHD6ZMTk7/jsrEayLwp2SzaB5KjP5hq90G7BWVnkKVavEIay26wkzhnQZ87xYcDJLwqR9UmFMyt+xJVmg0G341RyQ1Fe2/mPibeY0d4bMWPpP5M2euzk7hvemVHPSpChHVO/qC9z5YJW2tQzvS9Y59vPYMyICg7Dnp2YnE5ZWOh1wyQe3asKDslmsYaisSySicPNyjYKTLYl7ChUtmNmsm8zDegnXjOYCmbXAzcYxzW9BthsIWk0J1HxT437IKT68nnUU0VjZ6aIl40rM4bEtmc9OuceXQD/B43KpPvtZFvh/VCykbZYXPOTp+UCK4OdA2mDdR6ddJYdkvPe03MyQxmUn6zpRuAsbLDyuAP2vDsKA0Flc98wEjZsWFQtq7sqEdFiIytLUOZ141wLI4PhXlhVonwNFay9NxrV9nRT2OxDZqZzyxTxICEfR/0gvtQRspO6TUVlGUZn/r1vzD3jo2mm2ktfaEobl63PW+d2p2Egp0iY0gw9bEvq1kqy2hUBCMfvXaUXa76omLFs/NuWw9W/mNXTrsMs74bkqQEYJU5HhmhXUD0FkpVsGND6WB+nbOPqQcA1AQT77Web2cwHOMps2wWbZbGmn5UNa8A3N1hru6IJtoy3qsluzSWniGXwYIddkEz+m6wVPEhG8rOYR1DfrmBZydkp8+O22KwM2ScxnK5JBzLTMptmft2wjF1Gqvcn6FnJ42yw8bJtHY7NyhY/Gyzz4i+Z4cpOxYMyt7SS2N1DUSw+1A/WruHMgo2//5mK/708j7c/cL7OTy63EDBTpEREvo/sN2aWbBjNASUUZkHzw4zgopqFGAt133Hsztxx/qdeDaLCqF0iOZkZhLOdRdlrZSuF8yJz52pstM7FOGG49lHNwBQ1D29xoJs+jeQnRzP0ljNNWW8tDmdb0fsBZPpfKXUxzLuscMYrxmEqlUcGUyVsVN6rtdqQa8aKxaXuWctYKHMmeHO0KCs12ARUFJZdnw7TJ3MlWeHdbFu686+HQKDrTUel4RA8nmd8+yUThpLXA8y+S6yVhP56MrvNBTsFBli8zsrw/KO9KVRdvIwDFTbNp/ht6DssEDOif4vRgzpNHXjXZRz9KXVKjt6IyPEACdTZeeVPUcQi8sYX1fOK5B4WkCn1w5TIoyOxSoHkmXnY8VgJ41vR0ljeXj5sv00luL/MYJNP2cYpbGY0mKnWpCXtQvfO71qLPECaafPTrqmgr2D+k0FGcdlUX4e0fhaMlV2+k3GRQCKstPm4Hdf7IyspNGzq8YS+4VlMvy2mOkQ1oNMNny7k8FOruf95QIKdoqIaCzOZWu/x4XqMvM01lBESU0YeXZ4GisPBmXt/CEraSwW3DmZt095Dp00Qq4NylGtsqOzKxSfO9PF49/JuVRnJVNYgNJ+QK+xoGPKTmfC+3FUbRmOHh0EkL78XDEoe7j6ka1B2UzZGadRdow2AsyLYivY6U9VM8t0Giaq/WKZe3YsG5R10liAqOzYCHZYU8HkeeLKjsXPKjNqG/XZaazOnbLj97pNN1uZKTuJ45cNqiqHI+LmJ5PvIlNx7XgMCw0FO0WE2GnX73HzRnFGBmXWOdftkgwX/3yksVhJ86gKbRorfX8K5qfJ5bRuRdkRgh1ffj07esGMeFHM9OLP5mGddUwDv40FO2yEiEhHb/bKTu9QhKuMmaSxVAZllsay7dlJb1DWBjsNQX1lh18MbVQqHdJRdsp1ZmOx993rlnSrx4zgnh2Ti6ssy4YdlBks2NnT0Z9RUCfLSvqNpbG4cmV1NlbyPS436LPDBgZ39IUdSxEpPY1cppstO9VYib8bfhd5PdgGFbCelgxH49iXNLpn0wG9UNgKdu6//34+oRwAvvOd76CmpgazZ8/Ghx9+6NjBjTTEAXZ+j4svYEbKjjgqwmWwkPI+Ozk1KOt3qrWi7LDfiTsNp2ELoLho5fq8WElj2a3G6ugL4d2k4fTMyYqyU5c0KOspOx0OKDusc3J1mRcVfg8PdvZ0mA+bFAOUctaFN8s0lnbwpUhVwMvN0x6XZKh68D5QttJYrBor1bOjVnYybygICNVYJmmTUDTOAxKjNFZTdQCVfg+icRl7OtL3RGKIxmifzdlYg0KxhR615V6+RtjpYq2HqOz4TILZsHC/dKiDndIwKR/uy9yzs+/IAP9cjJhg57bbbkNZWSIqf/HFF3HXXXfhZz/7GRoaGvCtb33L0QMcSYi7DZdLSqvsmI2KYFTyxl259OwYpLGSi2TIZOfMXnNOlR29NFaWBuXn32vHu209hr9ncjdbcPVLz+0ZlN9s6QKQ8GOIF9uacuNSXjGYtLtgH+hKprBqEt/9sbVl8LldGIrE+XBQPXocnI3VY6EaCwDG1ydSbHVBn27naiC7Dt967RbKdIIB9tnLpKEgoBiUzaqx2LlwScYVT5KkVGRlMiNLTNd4NeMi+i0qAf0h8zSWJElorHLWt8NS1qJnR6/6MBPPjiRJRTXnzwk6+jP37Ijp6hGTxtq/fz+OOeYYAMBf//pXfP7zn8fVV1+NFStW4F//+pejBziSECuxAGXeUbomcUajIgDRoJwbBUOWZWXh11Zj8aaCxl8Mtujk0rMjDgFlZDP5vKVzAF+971X81/+9bngf1sSNnf90fXYymYzOlDSWBmCwoFcvcBR3cnbNhazsnDXt87hdmNiQSBmZpbJ4GqvMi3JvZiZXs8cyg1VkGZmTAZju/M3oD0X5Yp+uGkvPHG8FK00FWQqrMuA1DOgAYEpjIpW1K4NgRzwnbBBoJpV0sbjM1zO9QaCMRl5+7lCwIxiUeRpdT9mJWffssMcDimP0jROI64HVwEWcgxeNyxn15ykGbAU7FRUVOHw4YZB89tln+YTxQCCAwUHneiaMNPgXNfkFTFd6nq7HDpD7fjI9g1G+cNQHjZoKmnl2mLKTO+VJr6lbBT8vmV90maGyw2TMQCR5kWLnX2/yeb/NNFaXgaJXU27cZ0csr7av7CReN1N2AKStyIrFZf7ZE/vsDEbsDWG1YlAGgPF1iWM0KjsH7Cs7LLgPeF0q1cLMs5OJORlQTMFmnp105mTGsaMTwc57GfTaES9kvKmgUMqd7kInqpZGyg4gVGQ51GtH2TC6zT07TAGyoOywx0s8/vBTNPRQV2NZe03aQoThpu7YGgT6yU9+Ev/xH/+BmTNnYufOnXwq+VtvvYWJEyc6eXwjCi55J79Y1WmCnSNpRkUAYhorN8EO84JU+j0pC3om1VhdA2HE4nJGJk6rsEBD9E2w0vO+UOZBFjuXZkEcuxiwoEpP3RLHCmRiUGYGZBbcMLiyo6OSdThQes567OgGOwYVWaJyVhnwqNIyQ9GY6a5fD156btBXhnHKxDoAH2Bqc5XhfXg1VixuOKhVD/aZrw/6VX8T0Kk0C9kYFQEopefmyo7xqAgRruxYGO3BUEZFSPw1iu/VQDiG6jLj18TSlC5J7XnR4rSyw/15XnODsl1lp1S6KB+2YVDerfmOD4VjaT97xYQtZec3v/kNzjzzTBw6dAiPPPII6usTJsktW7bgsssuc/QARxKhiFbZYVOB9T+MnWlGRQDiINBITnpEdBg0FASUoM0o2IkLDdfisvPTjxkhnTRWZYY9Q0TYjtqsDJWVnrOLsm5TQR0jqxW6DN533lRQZxioWH1hd3cqlp0zJo8yLz9n54qlFcSA0075uVVl59wpo7HppnNx0/nHG97H71ZKitOVeIsYpW1NlR2bBuW4yXfWbOK5CPPs7D3cb/lzFtV0TwYSmxcWIKa7QIo9dsyCyCaHGwtmquxY6aAMlF4X5cMZKjuyLKekqu22jygUtpSdmpoa3HXXXSm333rrrVkf0EhG/KICTik7ibc4Ekvk0DOV09Nh1FAQQFpTn1YZOTIQNn0tdtFLJSjKTuaKF/ubWFw2VKPCGmVHt6mgzTQWe99rUtJYiX/H5URQUJ1UfuJxWeWJsp/GUronM9KVn2tLxV0uCWVeNwYjsUT5eUVmx9BjofScMba23PT3PkFxCEfjqgu7GdyQrzn/ZUKTQva50Gt7YAWXhQ7KZhPPRUZV+FFb7kXnQATvt/dh+lHVaZ9fW3bOKPd50D0YSXuBZMEQq74zotHhxoLihtFsNl/myk7ppLGGIjFVFaqVBp9H+sPoHoxAkhJtO3oF39pwwZay8/TTT2PTpk3837/5zW9w8skn40tf+hI6OzsdO7iRRqYGZabsaL0yIhWC9JyLVJZRJRaQ3tSnlYT10i9OYFaNZeeciH9j9Nq0BmW9hUE1LiKDXRLrr6RVdnweF39dYhfl7sGIKn0UjsYz9suEo3G0J1U8MY3FlJ1DvSHdoFzPVxI0KT+PxOJ4t63HUIW0MhvLKtpgxyqHdYaAAuqxCOz9tpvG4p4dkw7KVnoOAawiK7PmgmIaSyRosZqO7frTpSkbkyZ755Udl+n6w4IiK9VY7PESfzf8lR3teBQrys7uZNuC5uoyPoNvRAQ73/72t9HTkyi73b59O2644QZ86lOfwu7du3H99dc7eoAjCW0/GCZP9wxGdC9OTIo0U0NcLkmpPMqBSdlo4QfSN23T7pJyVZE1pEkPAsrF0k7puVjGr/faYnEZ7O3iBmU9z45wW0aeHTYPrTz1IlebXIjEc8m6/YoXrkwrkNq6hyDLifdUVPEqA14+PVyb0wf0L8h6YxUYd67fiQt++S888WZryu/C0Ti/mDnhFXC7lEZ/mVSWdOgMAQUS54ZlbJiyoXz2Miw9Z54dKwZlC+diSoYzsiLR1DQWoLx36crPlWDH/HUzg3J7b4hvELLBahrLtmenBNJYol8HsObZYQUIR4+uUKoOh1kay1aws2fPHkydOhUA8Mgjj2DBggW47bbbcPfdd+Opp55y9ABHEuJcLECpxorLQJ/OB9KKZwcQuyg774k5ZLDwA8rOORaXdRcy7cKRq147PJWgY1AejMQyXmTTKTvihdMsjWV3XESXyfBXvZEReqXqmZaftwg9drQeDLNUll4TQFZ+rtdFmVUM6VUOiZ/fCgeUHUAJADO5iB0xqIKUJEm4ECQeT88cbwXeVNCBNBaQ+YysiGYIKIN9b9JdIAfS9NhhNFT44XZJiMVllYneLmKbCaNq0HhcVvpgWVV2vKWTxtI2cLWy0WLKzuSGoNI8cyQoOz6fDwMDicVvw4YNmD9/PgCgrq6OKz5E5og9IoBE0MP+X2velWUZncmqHLabNyKblE06zNJYqjSByTA+ht6YAyfQS2MFBS9Bv+bL/mZLF7pNSuHFCiO91yUGO0zR0JuNZaeDsizLvLRcr79SrU5jQba4NVYFuGKQ6Q6VdU8W/ToMs4osvYqhMpPp2Szg7ehLLetnn9+gz+1Y1Z6V9ghaxCnuWsp96hSdXo8nK7A+O2YdlK0alAFlbMTLuw/jP//4Gr7zlzdw29/fwapNe3TT5GwIqDaNxV5fes+OtTSW2yVhTNLo3epA+bmesqP9rIvvtVXFrViUHVmWcc8/P8DLuw/bfoxDye8W27No1z89uLIzKmiqzBYztrZHZ599Nq6//nqcddZZeOWVV/DQQw8BAHbu3ImxY8c6eoAjiZBOt9WqMi/3Q4ytVe7bH47xL61Znx0gt/OxTIMdt9oTob02a9WOnCk7Ohccv8cNn9uFcCyOvlCU+6O27e/CJb/5N86ZMgqrv3qa7uOJU+gjOoufmHpggaZe6bkq2LG4cPQMRfluX1t6DiiprS4hWOPvUaUPfo8LA+GYofdgy4dHcOPDb+KWi6finCmj+e28oaBusJPw7ZinsVI9O3oBHjtuvV2+VY9KJvg8bgDRjDw74hR3LQHNyAi9QNsK7KtjTdlJfz5OaK5Cuc+N/nAMz7x1UPW7gVAUS+Yeq7otolONBYgjI9KlsawpOwAwpjqAj7qHcNABk7Jq6rlBGksMWDL37BT2Av9GSzdu+/u7OHZ0BdZfP8fWY7DNT1NV4rxbmXXGlJ2jR1WMLGXnrrvugsfjwV/+8hf89re/xVFHHQUAeOqpp3DBBRc4eoAjiSGNsgMYV2QxM6/f4+IfPiNyOTKCXUhHVaYGXB63y1RJ0N6Wq/lYRhcc3lhQCF7+/X5iwOZekzlC6Tw7TNlxScqFPd1sLKu9b9j7Xu5z615AWWpLNChzX1XQzz9bekoTADz79kHs6ejH/zzznup2PiqiVifYGW2cxtJTQcpYGksnwDNXdpwzJzPsdMY162+jHYeRrbJj5tnJ5HxUBbx48tqPY+VlM/HjS6bj2+dPwSkTErsnvXMdMarG8hu/dyJWlR1A8e040WtH2TCaBTuJY5OkVOXKCCtDjfPBoWSRwCGd98wqbD1gA3PTvZfiANDJoyq4smO3E3uhsLVqjB8/Hk888UTK7XfeeWfWBzSS0avcYLtHba8d0TeQrhlaRQ6VncN8+rN+W36f24XBeMy0IoKRK2VHmx5kBP1uHOlXG7e37ktUExqV+wMWPDtxZVfMvBrpZmOFY3FEY3F40uw0O7k5WV/N4712hDQWm4NTX+FLBkgRQ2WHmQ7f+qgHOw508zJlK2msDw/3IxJTl3BzNUaYfK2kQtSfx1hcRlfyvB/uT13Meyz22MkEOyMjzBQm7a43pOMXs4Ilz04GaSwAmNQQxKSGoOq21z7s1N2hG1Vjles0TtSj36JBGQAaq5yryBI9UkYpSvad9bldlhtJFkufHRZos6IVowHQZrBqrPF15Xh5z5G0Kt2+I/2IxWUEfW6MqfKnqJfDBdurRiwWw1//+le88847kCQJJ5xwAj796U/DbbFJE5GKts8OAMNhoEYmST2qcjQyYiAc5YueXlNBILFIDEZilpSdXFdjpSg7fi+AQX5eZFnGtv1dABIXEqOuuuJ51FV2osqumD2nVkmRZTnlIjMYiaHSarBj4NOq0ZmPJU7oTqfsiMf059f282BHr3syo7EqgHKfGwPhGPYdGeDBD5AmjaVZLHsGI2AWlY5evTQWUzIcTGMlz7deOtIIM0VF62ewn8ZifXaMjysTg7Ie/LOpE/gapbFY35x0VYzcoJymzw7gtLKT2kFZG9hrW3xYgQ8VLbBBmW3C4nKiIs7Od4EpeWx+XDrPDlNsJ4+qUJnwR0Qa6/3338cJJ5yAyy+/HI8++ij+8pe/4Ctf+QqmTZuGDz74wOljHDHoKRBGvXYyCXZylcZiF6SA12U4ddlskWC3sc1J7jw7+r1OKjWTz1s6B7lXJBaXDRcBVRpLz7MTV3bFbEeovaAMReLQek+tLB7clJ5W2VGOkXf8Dfr4Bc5I2RkUbv/r1gMYSs6wYsHOWJ00lssl8X4772tGEpimsUyUvcFILGXHabV7ciawid56wyL1iMbi/HOh55XRXgjsprG4smPSDimT0nM9zC5aTNnxadVQi8NA2XvLKu/McLKxoLhhNGp9wZWdDNQ2Zahx6udElmV8eLg/6w710Vgcf3p5n2kKXbwOmKnPZrA1bnx9ItgJR+OmFams8IB9x7Wp2uGCrWDn2muvxdFHH439+/fj9ddfx9atW7Fv3z5MmjQJ1157rdPHOGLQzsYCjIeBpktniOSqz84hwZxsJAebtmxP3jYm2TI+58qOZnFjCgPz7GxNqjoMvcVElmULpeeJRc8jKjuaC4rYp4RdCFm5shnp3nfeZ0dVep6q7BjtUMUFrGcoimfeasPh/jDC0TgkSbkwaWHVPjs1JeN6HY95xZLm86gNdrXqTk4Myu7MPDvid0gv6GIeFa1nJ9M+Oy6XeVPBUDTG1wsrBmU9+OdOJ9hhAYJHkyZhSo3V0vOgBWWHBztOKDs6U8+NDMqZKDtmnp37N+/FnP95Ad//646sAp7n3m3Hzeu240dPvG14H3FNshvsaD07QOrGQ4RNO2eK7YhSdjZu3Iif/exnqKur47fV19fj9ttvx8aNGx07uJEGz+97U5Ud7Qc7M2Un6ftx2LMjXkSNMO9imriNLXa9Q9GMmrtZhRu/UwzKiXPLLmDb9nWpfq83q2soEle18Nc7Xu53cEmCZ0d9P3YxLPO6ld2yhbbtnQYTzxl6fXa4r6rCxxdtI0M0u/Axb8dDr+7nqs6YyoDhSAXWtO49TR8XvZSPcsHUKDua1gMdGt9Oj07PnmzJdPI5C7gCXpfuuQikKDv20lieNE0F2XFIktoPlQllBoE4YGxQZp/VdKkP9t6WWfLsKMFOtuqIGMgYvbdhW8GO8Sbh3WSA/38v78Pv/rk784NOsiep6Jil80Tvpp1gRxwd01Qd4J8zvZ5XDK2yExhJyo7f70dvb2pzqr6+Pvh8zs82Gino7TiMPDt2lB2nDcrsIjpKp3syw2eyI2ILx6gKf05TWUZprAqm7LBgZ7961IneYtKrmZJupux4PS7+nFozNlN2gn43vyBYWTyMJp4z+OTzgYSBUZyD01DhF4yW5p6dL58xAZIEbP7gMF5K9vTQq8RiMGVnl6ZDL1ucxVQLN7mmMaizIbOMXFRj+dIMq9XSbVKJBQiqFVN2uEHZ3tRzI4My+y5X+Dy2TKqAcNHSCXaiwmdYpMxAldPCXn/QQjUWU3bDsXjW6q4YXBqZz9lnX5uiM8Osz46YWrr9qXfxxJsfZXbQSViQ022yBorPZWdwcvdghG/W6oN+pVjAQKmTZXlkKzsLFizA1VdfjZdffhmyLEOWZbz00kv4xje+gYULFzp9jCMG3T47gXTKTnoJm8n+fU57dkx67DCspLHKfW4+xFK7u3cCY4Oykt4LR+PY8VGiISabNaYb7GgCRrPSc49LMjQoizvfTBaPrjTKDguCYvFEuo19TrxuCVUBT1plhwVcx4yuwNnHNAAA7knuVvXMyYzjGhPBzgeH+lTvta6yw5SsNGks7Qwf3t/GZtpGD25QtqgopvMNacty7Ss7SvdxPTLpsWMEUx31gmzu2bGp7LCLpxVlx+dx8TUkW5Oyqs+OQYrSlrLjNQ6K2WeCKR/X//kNvLb3SIZHDnyUVFDNFJts01isyrEq4IHP40pJu2oRB4AytddMESxmbAU7v/71r3H00UfjzDPPRCAQQCAQwOzZs3HMMcfgl7/8pcOHOHLQMygbeXaUYMc40OCPkaPScyvBjtl8LNFMyJrh5cK3EzLqs+NX0ljvtPYgHI2jttyLackKJL2dU0qwY9JUUKzGisRk1YWLycZBnyejvhV84rmBouf3uLlZvHMgrGoNIEmSojQZKDvsGMq8bnzx1HGq59QrO2c0VwdQ4fcgGpex93A/fw4+y0o0KBt0YO3UdK1OVXacNyinm9+mJV1FWBkvzU4ca8gg0E4HV3YM0jo9Dqhcyucu9bVn69npzbAsnlVkZdtYUNwwGnVQDnGDslPKTuK1fuf84zF/6hiEo3H8xwOv8bSUVVig1x/Wb9UBqNckbTsSK/C5bsnq2XTVdawSq7m6jH+Gh2sHZVvBTk1NDR577DHs3LkTf/nLX/Dwww9j586dWLduHWpqahw+xJGD3kWZlZWKfptoLM6rXtgQRjNy1WdHCXaM01hmuW6xTJT16XE6jRWLy3zh1qYSRIMyKzk/aVyNoU8KSK1oM1N2EsGO8pxiMDMg7HyVvhXpL7hGE89FxMaCzPfCBrWma44mBjufnDpGNWzULI2VmKytnr8kft4qBF+JUQdlplqx62uqspO8wPudU3ZYHxmraayeNOqSkpJMPJ79poLmnh2eHsxC2TH17ET101hWq7HYRdlqMMZSWdkrO3qeHfWxhoVNllXM1jH2uawp9+JXl87ESWOr0TUQwVfveyUj9UMcl2Gk2vRkq+zwyszEepvu/WRd0VnjUGD4prEsbwvSTTN/4YUX+P//4he/sH1AIxmz0nPxg/3S7iPoHIigttyLk8fVpH3cSo0R1yk6uPHVJI1lUu0iLkx607qdQFyctLtrcfI5ayY4c1wtDvUlFlw9ZadPEzDq9WcRG7KJFWBDkZgwSFFRdjzJC66V6cOsysrIswMkfFwtnYPoEpWd5HvkN6nAAZQFrMyXqGa5ZOZRuO/fewEAY02UHSBhUt66rytRkXWi4Cvxe1SzrIw6KLMU5sT6IHZ39Kd0ic2FsmO0+zcinW9IuRAkZ2NF9VXFdKRrKpht2XnimBKvXdezE9dPYymVdOYXOvZeVVsMxpocqsgS11CvQVNBe54d45YN4ueyzOfG7684FRev3IS9hwdw/+a9+M85R1s6bnFESvdgBKN0epeJm147wY5SVJLY/KSbYs/NyUIzykw8hsWE5VVj69atlu5ntSMlkYrVpoJPbk8Y4C6Y3pi24y6gLMx9ocRcJaeGKFpKY3mtBDtuxVjrcLAjSvTaCw4LPHpDUf6lPnl8DV7ZkzDk6lWvWfPsKKXnLpfEZ3ANCedA9OywC1u6XaAsy2k9O4Cg7PRHuDrSkLwtkEbZYRc+9hn84qnjeLBjlsYCgGNZ+XnSpKyMVVAvM0al5yyQO2Z0BXZ39PMSWYZeGXu22K3GMqoI014IjNoepENpKmhkUM6+Mk1UdrQNNMNGHZTTXByBRLDPPkdWg7FGhxoLiuo4+14549kxSWNpTOujKv248fwpuPHhN3D3Cx/g0tPGpw36DnarP+vdg6nrYDQ5x0+5jx1lRx3sBNOkpLg5WUfZGW6eHcvflOeffz6Xx0FAuPh7Uz07oWgcQ5EYPC4JT+9oAwBcNKPZ0uOKKYT+cDSr3aAI81TozcViKE0Fjaux/B6XMq3b4TQW+0J63VJKkMfOS8uRAXyUXGRPHluDd1sTRmW9xUTb3NGsqSB77X5vMtjRSWOV+9xwJy8y6WTh/nCMB1JmVXh1SdWnsz/MFzeWozdTdhLVW4ljZxft4xur8LWzJuFQXwjHCgueHlN4sKNOY2mDE5bG0lZjsUDu2DEVePbtgynDQHNSjZXs+G7Vs2M2FwtQS/yyLGffVNCgz44TaSxWjRWXE69f3GSxNJZ2MyUqk0YdxlXpS4vvFVd2euxPPo8LKWu/R5nLF5ehGsWSnWcnNSXGHk/8XH5m5lG4558fYOfBPvzvxg9w0wXHmz7+R5qJ710D6f2CtpQdtvmpYJ4d/WIBBh8AKig72vYKwwVbnh2nWL58OSRJUv00Njby38uyjOXLl6O5uRllZWU455xz8NZbb6keIxQKYcmSJWhoaEAwGMTChQvR0tKS75fiCHq7wAqfh3sYegYjeHH3YXQORFAX9OGMyXV6D5NCwKvMiXHKtxOKxvhO20o1lm6wIwztq+NqRG6CHb2dNQt2WKAzeVQQ1eVew3J/QE/ZSd15swCIpaf0GguKgxIDGp+HEUz1CnhdplUuvLJNZVBOr+yIt4nDZX948VSsvGxm2hLn45Kenb2H+xMl7wbBSZmBT4AZlI9JBlWisiNeVJwK1gHl82l1XES6VJrYXTYSk8GEmUybCqYrPXei55Aqxar57BkOAk2+vlhcNlQH2fdGm740w4nGgmLAKhqUtb9zsqmg6OETN5Vul4Rvn58IcO779560xutWTbBjZaOVnbKTDHa8TKnTD1zY4NEmQdXVtlcYLhQ02AGAadOmobW1lf9s376d/+5nP/sZfvGLX+Cuu+7Cq6++isbGRnzyk59U9fhZunQp1q1bh7Vr12LTpk3o6+vDggULEIsNrzcC0Fd2XC6J7956hiJ48s1WAMD506ylsBiKSdmZ0m52EfW4JNOLj1EXU0BbjZWrYCe1nJ+h3XUy/5OZQVnre9JXdpK74mT5sF6n2n5B2eEVPGmaClrtrVQnzMdShoCqlR0974G4U8vUYwIk5Puaci/iciLXrzcqAlAWWLFNvSzLPJg7dnRl8vgj/KKruqgUcBAo67OUrhprMBJTtRuwa1DOZem5qHZqWyMo6qQ2jaWceyPPhh1vVaNgULbbWFD8TIul54D6e5qVshPRBjvJflk+d8p6PO+E0ThlQi2GInH8csMu08dng3YZesqOdj3SBj9W6OhTp7WZUqf3Xopps2qdasrhlsYqeLDj8XjQ2NjIf0aNGgUgsfj98pe/xPe+9z189rOfxfTp03H//fdjYGAAf/rTnwAA3d3dWLVqFe644w7MmzcPM2fOxJo1a7B9+3Zs2LChkC/LFrw6SfMlZMHE4b4wnnkrkcJacGJTRo/NfTsOKTtiV16zHb+psiOksep0Blg6wZBOV2pGhabz7MxksGNU7g8oF12m3uuXnrPFNKns6PS2GeQGZTffKQ2l2SmxQDBdsFMrqGQs1ahUYxkPAmXBjk9IAWSCJEnK2IiDvYYXPVGVYqmsvlCUB4mTGoJKk8nkaxYvKk55zgDlYp6xZ8dg+KZYlssuBpKUavRNRzrPTqYTz/VQDXUMa9Mz+oNA3S6Jf4aMfDt2zNNM2RkIK00wM4WtJ26XBK/blfDM6XxP7VRjGbVs6DFpRSBJEm66MKHu/Pm1/dwXqIdW2enSU3aSqUtR6c+UFGXHxIMlehbF73CZ0E7DqD/VjgPdORv/Y5eCBzu7du1Cc3MzJk2ahEsvvRS7dycamO3ZswdtbW2YP38+v6/f78ecOXOwefNmAMCWLVsQiURU92lubsb06dP5ffQIhULo6elR/RQDegZlQImqn9rRhs6BCOqDPpw+yVoKi1HpcPm5FXMyYK2poN/r4hdop5sKGjUUBHSCnfG1AIyHrwLK+atJ3kfvy85SW0zZ0dsJ9YeYQdljOQfOdntGE88ZrFy8cyDCm4ixUlOzQaDiCAu7sFTWe219hhdk0U/BnpO9toDXhaDfw/tHsYqsXMzFAjI3KPOSaoPyd/G95mZZjzvjwg3rTQWzU7mMKrKM0lgAUioKtdjxVpX7PPx7ZzeVpcwWVI5Zb7OVVTVWShrLXMU6dWId5h4/GrG4jDuefc/w8VuTyg777uoFMmw9aqpOpJS6ByMZq2C89LxCrezoVdexzV7Q51Z9DsS1VE/d+eBQHxas3IT/+r8tGR1brilosHP66afjgQcewDPPPIN7770XbW1tmD17Ng4fPoy2toSCMWbMGNXfjBkzhv+ura0NPp8PtbW1hvfRY8WKFaiuruY/48aNc/iVZU40psxc0qoQbEF79PWEF8lqFZZIhVB55ASHLAY7StM2vT47QjVWjtJYIYNREYDyRWfHOSXZBdhKGovtjMyUHbZA6Co7EWFcBA92zC+4VpUd8VyKChxg3i9E7LFjlyl8bESvYa8VSZKUkRHJC6b2tbHFmB1/LszJgGCgd6iDcrlQVm/XnAykbyroVPBnNKg2YlCNBSgGc6ONk13zdLbl53qtO/QKJLKbjaXv2TF7rd++YAokCfj79ja8oRk2zGC+wROaqgCoZ9sx2Ho0ri4R7ERickYmYXF0DFu/lBR66uOw76+2kszvcXFlW+/59x0eAKA0JCwWChrsXHjhhfjc5z6HGTNmYN68eXjyyScBAPfffz+/j3ZHZFQBkMl9li1bhu7ubv6zf//+LF6FM4hfIiNlh+2UL5qRWQoLUBZFpzw72n4NRvgMct2AphorqVYMRmKO9m/gw1V1JGuv28UXsRlHVfPghCkRQ5G4jmydeA9Y2s28qWDiM6hXATUgqChKubJ5INpl0bPDDMr7jgzwAJodr6myw3vs2A92jhUGgppdkHl/j+Tiq/UjsSCafc56bPhArJDpbKx05e8Bn6KU2B0VASjVWLKcqDJKPY7s++wAxg3izJSdmrLEe2SURrHb3Tlbk7KeMq73/trz7CjG7KjwnbfyuTy+sQqfmXkUAOCBFz/UvU9bMo3Fgx3dNFbitubqMh4MZ2JSPqwZHQMIlZE6m+BuA1+YuFnRW6tZQNU1EM56sKuTFDyNJRIMBjFjxgzs2rWLV2VpFZr29nau9jQ2NiIcDqOzs9PwPnr4/X5UVVWpfgqNGOxov4RiZN1Q4cNpGaawgByksXrZENA0aSyDxl6AenGq8Ht4cOCkbyfdBYedl5nja1S3STwvrj5fLFisNw121H4HvflYfFyE36NU8KTZpbGy/No0k+55ICaUxLJjMFN22MJl5+LMYJ6dls5BXoGil2rhxsiIOo3Fgt56Q2UnN2ksq7OxFNO1UTVW4nZZVi4Wds6n6IPT8+04l8YyUnb0PTuA0tCyS6cXDGDfT9SYZRdlvnkSlDS9cSB2PDvimiyu1VYVtgumJa5n77SmWiYGwzFeiXh8Ul02q8aqKvOaqs9GcL9OcnQMoHxe9Tw7RsEOIPST0lmz2Hc1EpPTzlDLJ0UV7IRCIbzzzjtoamrCpEmT0NjYiPXr1/Pfh8NhbNy4EbNnzwYAzJo1C16vV3Wf1tZW7Nixg99nuMC+qHr9YMRFw04KCwAq2dBLpwzK/RbTWCYD9ETPjiRJOanISpdKYIvUyeOUVKjLJfHzpV1M2OLGLsb6U8+NSs+FpoIRvXER5gsDWxBrTbonA6ndlcX3yGwQKFd2bKRdGHVBH38+NoJDV9kxSGPVGCg7ueieDGTm2QlFY0IAaV6NBSjBaSbpEoY4k0rr24nG4vwi4pSyo/08cGVH59jZhVavYgjIfFQEo9Fir523P+rB6/s6U24fEjxSDL/O+2vHs2MU7Fh9rcc3JjbU7x/qUylDgGJOLve5Ma6uHADQbVKNJQY7mczH4n4doS+aWRl5t0EaC4DpmiVeY5xuEpsNBQ12brzxRmzcuBF79uzByy+/jM9//vPo6enBFVdcAUmSsHTpUtx2221Yt24dduzYgSuvvBLl5eX40pe+BACorq7GVVddhRtuuAH/+Mc/sHXrVnz5y1/mabHhhN4XlSFG1p+ykcICcpfGajBpKAgAfrOmgprqs1xUZLFgx6jPydfOnoS5x4/GucePUt1uVJHVx9NYJp6d5MXJxz07OmkscRCoRc9Op0XPTsCrVHgBigqV+J0Fz04WaSwAmNKYMCkzf5PehUDbRZl3hk6+NhZMslJZs6qXbMikGsto1peI2yXxCyN7TXaUHbdK2dE3xQLZB38Bn/5Fiwc7OpVv6YIdu9Ppmyx0UY7HZXzp9y/h0nteSlnL9JQdvWDWjmcnUeElqZ4HsB6Ej60tQ5nXjXA0jg+PDKh+x15vU3WAf7fNqrGqAh7TilEjDgnKDqPcZDYW+87pBTtm87HE9hxGn5FC4Ow2KUNaWlpw2WWXoaOjA6NGjcIZZ5yBl156CRMmTAAAfOc738Hg4CD+67/+C52dnTj99NPx7LPPorKykj/GnXfeCY/Hg0WLFmFwcBBz587F6tWr4XZnt2DnG70vKoN92BoqfDh9Ur2tx89kGOjf3kiMo7j4JOMOzSyN5Ug1VjLAy0VjQT6byECy/soZE/CVMyak3F5d5kVL56CqIktsg88CCN1qLIOmgiEdz065z81TZun6VnBlJ00aC0gERAPhxI5R9FWZKTtOGJSBRCrr3+8f5v/WUx+0k5O1qhX7XDEFMd2YBrtk0mdHqcQyb5bHLmpHhCaQmSIqO9omyma9XTKFBeLai1bYQhrLcFilbc9Ownhr5tnpGYrwC+ih3pAq+A2ZVGOJBRJil+VM8HvciMSiKr9br0XvlMsl4bgxFXijpRs723px9CilG/lHXYnvaXNNmSo9pfWeikqLvTSWulgBsObZ0Q12THrtiNcYp1uJZENBg521a9ea/l6SJCxfvhzLly83vE8gEMDKlSuxcuVKh48uv4iVSVrmHDcK4+rK8PWPT7bdY4R7dtJUYx3pD+O6tVvhkiR8cuoYw10puwiJuwQ99GRkhraTaW0O5mPZrYjR66IsyrNaX4wI24krnh3W20a5r9hUkJFuEKii7KTfMdcFfTjQxYId5T0yU3bY7j7Tbr9amG+HoReg8GnLERbsaNNYTNlhwU6uqrGsG5St7uLLfW50DyoX5Uy8IQwzZcdJlcvoohU1SWMxg7LRhdZqAKDFirIjboS06oeuQVlnELFeUGQFv8eFvpC+Z8dKEH7cmEq80dKN9w724kJBoReVHRZIxuIy+kJR1XssenbY89nx7IgbVKXPjnE1lu5mxWuc/qJghzDFrIX5uLpy/Os752X1+FbTWG+2dCEuA3FZRvdgRDfYSQyktNbzxWdiiNWqWbxk2kHp025FDPuCi4sJ+xInesGwxlo6np2ogUE5eUGRZZkHFuU+D5gjw8yzI8uy5Q7KgNq30xBMVXb00oosjeaEsiOid1Eu11Sg8dcW1Cg73KCc/SwoPTLx7Fg1o7Lzl42yI0kSXFJirpPWs+OUORkwHupoVnpezQzKBhcylm7JNDBlxQ7dgxGEo3FdT40Y7Ggv9Lql5zol42He9DPzYEd8HiCzwJO1tnivrVd1O/PsNFWXIeB1w+9xIRSNo2sgog52hODDlrLD52KJnh3jDsqKspP6PppNPhevMcWUxioqg/JIJp23JFu4QTmNsrPjQDf/f6PS0sFIjPtS0u3ejDoox+Iyr/hgF2Cl8696AnA2OKnsiKMCvCZepEhyJ85SEdpgJyz0VCoX+uzopZYYg5EYfy4raSxxKrqessMmXWufA8g+2Dl2jHpgqN5FWSk9Tyo7/SyNxTw7SrAjy3LuDcpW0lhpKrEY7LV1ZuHZAZTGgtpqLKfKzgF98zygVGPpdX5mDTX1fCWA8j3JNDCtLvNyRcsolS3erl2f9DY2esGsmYpuhl6xRSafSx7sHNQGOwllp7kmoWwZBTLdg8roBrP5fUZ06Hh2mMIajsVTAn4WtFbrKMlmk8/Fa0wxKTsU7BQJdobTZYLV0vM3W4RgJ03TMLdLUqVh9OCzsTQXE/GLxQ3KfFq3c7sBsz47ZlTr+BLEhc2spD6q8TvwEQ3JRVbcDZV7lWAnHIunVGowmKfF53YhaMFALKo/etVYcTn1IuqUQbkq4EVzMiXhdkm6wVNq6blatRJL+3uGokoay6BzsV28GRmUre3i2evlwY6NNBZgPB/LiYnnDKPu3YrvTM+zk0xjGVZj2fNXuVxKReZhgw2PePHUqgZmTQVVpedZKzupnh1Lyk5S8dzb0a8KElj3ZOZZMvJEicG2vWAn1bMjfte1Ko2pZ8eiQZmqsYgU9L6oTqKksTJQdgxSXkpO3pO2waNRU0FRCtZ6dpwtPbebxkrNifNgx+8x7c+iTQFolR2WH/cl5/eoFhwDk3InL832Who9IAY7KoOyoHBpd2VO9NlhHJfcxRp9RpScf+KcHtEEOwGvm6uRHX2hnCk7en1YjLB6DFzZ6VdGYNjBYxTsODDxnGF00VJ8Z6nvndJnJ3V9SKhw9pWn+jRrwBFhI5SSxuJDf82rsYxmEKZDL42VblaaiHZQLuOjZBqLbRCYJ0oM5oYiStsD+wblVM+Oz+Pi77F2EHG3iWdHqeJL/d6oPTuUxiI0KF/U3KSxWDWWWZ+djr4Qb1sOpO+QamVnaaSAsN2R2yXx3WMuS8/tp7GU8yXu4rw6xkeGcVNBpuwkzcl+pdGfWft1QDkndRZSWIn7qRtRMsQFXpuCcyqNBSi+HaMdr1J6nhirwIJS0QPWUKmkstJ1LrZLJgZlq83yUpQdm+eTNRZMTWM5dy4COt29AfM0llJ6Hk7p7twfjoHdZOf40lVkiinuVGVHr4OyTum5bWVH3X1cnV5N/1q1g3KBhArCHqOpJqHs6JWVs7XYJSVST5mWnsfjMj+n2gpa3lgwZF3Z4eNeIqnXk2I1KFOwUyTkK40VjsUNS5y3C6oOkH72jZVdNttlaS8mehURuWwqmGkQqbeYiH1jzKrMlKaCmmospuwkFxW2YIjTp4d0dkqA2HTP2gWkRlR2hBy9JEmGPionmgoy2KJu2GlYGCbJFkSPS1L1r2G7/ISyk6NqrAwMylYbyGnL6u1uYAyVHScNygbVWJGo8bgIdvGLy0CfpoKQvU9et2RL0arTdM7WYqrs6Kjjuk0FbXt21N+bUDTOAyern0vWIfndpEm5NVkxWRnw8M++Xodq0QjtckkZKzvdg5GU0TGMoE+tsgKJ4KjXrM8O+9yQQZnIFLPp3E7AjGiAcRCzvUUd7BilsTIxSCqD+NRfCr2FSVR2nJqpYjuNpbOYqDw7PI2VepzaFECA7wgTr5n32BEu7GY5cEBZNKwrO4n7uYWFkaHX5BBQFq5sPTsAMO+E0Thzcr1uDyNA3BnGVN2TxZQX24G2dg/xC4wTplwR9j5G47LuDCoRq7t4rY/NbhrLzZUd/dJzRwzKHv2qmrCmC7jqb7xu/pq0vh3Fr2Mt3aolXRpLVAq0F/ohHXVcbxBoyKFqLPY+SBJQ4bMW7HBlJxnsMCW9OenXARQDuHhutSpLpsEO80BVCWsXQ1ssACSCWPZ10FPwzbxe4rkmZYdIIdfKjlvYNRtVZDFlh30ZjFqRm/Vf0GKkgOhJzkzZicTktFVjVhniBmWbaSwh4GP/X+FPk8YyLD1P3HdAp8eO0eLB0I5TSEdzUhI/qqZMNWcJUC4GWh/VoIMBd025Dw9efQa+eOp43d+LpedKIKf+PDGv0d4OZXpyRY6UHSC9b6fXYjWW9vzZNSgbKzsOGpR96s8mg6mTemkswLjXjt2GggwWpB+2UI3VrZnNZVZ6zt5bWZZtdVBO3F/dtoEFvxV+T8p3zAhWkbXzYMKzw5SdpmQlFqAfyCjvucfwPmZ08FERqX3RlGIBZc1lgZbP49JdD4y6vvdr1m1SdogUcm1QBsSKLP0PIFN2Tp2YmBNlrOxYN+Wpqn+Ei4lex+gyn1KZ5FRFVi767IjKTjgWT1GhIoZNBTXKjhDslOtIySLacQrpmNQQxG++9DH85ksfS/kdrw7TqG1OenbSIe4mtQ0FGUzZ2ZMMdoI+t+2mmkaIBtx0wY7Vniqpyo7Naiy3uUHZiZSenqIYi8t8V6+XxgKEVIvmYtabgZ9PD6XlgH41lnmfHZMOysnfie+xbWUnog52MlHYjhudCHYOdA2idygiNBQUlB2dc6tVdtj5DUWNbQkifLyPThPYch1lx8yvAxj32dFmDfpCUUsp4nxAwU6RoKd0OE2FyTDQQ70htPUMQZKAM5IjKQw9OxmUWxoN0DPqYsoNig7JnyGbagX7kvcORfnFpk8n2AFSU1lGg0DZwiA2FGSYtV8HlEaLVj07AHDRiU2YMbY65faAkbLjYBorHWLpudGAU7YLZcGO0+ZkQK1cpFuULVdjaZUdu2ksST/YYQFx0GA+VyboGZTFCkO9DsqAYFLWqCt2Gwoy0qaxxA7KWoOyThpLq8aIa1DGyo6m+7gdH1l1uZdPd995sFdoKCgoO3w+VqpnhwVWlX4PL2ow2pSK6I2KYCjzsZT13mwuVuJvWJCs8Wwleyw1VPj48Wk/I4WCgp0iwW7VUCawL6Ve/xxWcn70qAo+fdiwGmvQ+o5GlSZQTR7WD+7qHB4ZMWSzzFRUrViQI5bUqi6SGkUgqqlkSS09N0ljGRiUuzKsxjJDr4RWPL68KDtC6Tl7r7WvjXV+ZqW5TpuTgaRh2yQlKWJ1J1+m8W/YVnYMqrG4MujA+6TXHE4V7Oh4dgALyo7NwNSsGiscjavG3RgZlANmyo7wHhul6IzQBk52Azulk3KfalQEQ0lRCcGHxjrgckm8NYOVXjtMKdMPdlJHP/SkUXaMpp6Lnc7TDYzNNxTsFAl5UXZMRkawZoInHlXNd9BpDcoW0lhul8S9B2G9NJYmCHG61w4fBJrhhcHvEUyYyS8+z9ELTQWB1IskN3fyDsosbcQ8O6lprHQG5SMWJ55bwahrrpOenXSIpeeGaayksiPzUubcTLexWpFltb+NU8oO66CsVXb0lEG76HnFRKXS68rUs+OMsqPn2dGaXbUpHDODMvtOsnXW53ZlbKDWNhXMpKGgiOLb6VUNAWUoBuXUlJ3YzViv8akRzAg9ujKQ8rugzuRzpceO/vto5NnpE3qRsbWqWBoLUrBTJOg1xHIatgDpmX+ZOXn6UdU8iDEuPc9s96bXWJAHd5rXy7soO5TGykYx05oAxTSGSwjitI0FmbLj0Sg74Wgc8bgsGJR1qrEMPTvJVE9OlR17gaEdWApmIBIzHHBar3mtuUhjATBtEMnIpKdKimcnyw7KRsqOE+lGvR0689a5XZKh8dZoPla2lWJM2ekejKS8H0cEBZCdG/FCb2pQTv7OrjlZ/BuWGs9kCKgIq8h6t61HV9nRa9qo15U6E5MyM/lPbAim/I71++rXUc3SeXa0aXelPYeXv45iaSxIwU6RkA+DclXAOIjZfqALAHDi2Gq+UBnJo5kOZVTMvMoXw6jXRW2aaoxMkGU5q5J+bUWW+EUGYFiRFdVUsojPHYrGTQ3K6ZoKWpl4ng69tvfic+fDs8OeIxaX0d6bkNiNlB1GzpQdkzlnjMFIjCss6Y5D+1mz3WcnmULSlsQP6nx+7MIvWjpN94xSWAAMUxRKasfe5zTRfiDx/9oNj5ju1OtwbsWgzNbZTM3J4t9krewkg503W7r5WiAalNm5HQgrXZP1PDQZBTuHE8HOZL1gRzeNpczh0qPMMI3FKlYVZcdoYGy+oWCnSMjHrpoZlLVprPaeIRzsCcElAVObq5RgJ00ay+rFR+/iahTc1dmUPrd8eATn/fwFPP9uO78tElOqSuzsrrUVWdrXbdScL8yVHdZnR3mNQ5EYBpJVD6LB1Kz9+lAkxhciJ5QdxaCsrsBhC2s+PDui3+RAUsrXVppV+tXpwlwpO15PappVSybz4Jzqs+OSUpWdSExpZBd0II1VJqiOLJjTdgDXw2hkhNXyfCPcLol/DrSp7COCb62GX0j1gh3RoKxOY2Wn7Gg8OzZTdseMroAkKcFFbblXtcGoDHh5wMfWHp5W0gl2jFqEMHqGIrz0XFfZ0TEoW67G0mzOeoXGq6TsELrkp/Q88eHTprG2C+bkcp+HL1RDkdRJuEBmBmVAPygw6itk17Pz7FsHsbujHw9v2c9vE0ur7aQHxZ1TPK70/mHGQHYxSEljaUrPPW4XT3kNRWMY0DECm3l22ILuEUyJ2aAXfIpydD6CHY/bxQMZNgixVtNnR5Ik1agLJ2ZB6WHFoCxW3qTzemiVMftTz1k1lnJc4u7bmTSWOhAHxNluJsFOWs+O/cCU99rRdFHmaaxyn27TT25QFmdjubXKjr2GgkBqNVYmVakiZT43JtYrQUejoOoAiYBP2WiFVc8lrrlWlR2WwhpV6Vd1KGewDsr9ep6djJUdxddIyg6hSz4MykbVWCzYYWXK4hdCz8yciUEZ0L+YpK3GyvALcihZbcDasAPK4i1J9oJIcbJwfzgqGGW9yWPXv0gqrfaVi6JoCh4IsdLh1GBHr/TcqMOwXbTVYYA6yMplwC3CvAJs163XMLFemOOTO4Oyom4YkckuPtWg7Jxnh11cPC7J1gVbi6h4pgY7xp81PpnbsBrL/ntl1FiQm/SDPsXEK1zoh3RS40bVWHbWWcM+OzZUrOPGVPD/b65ONQ1rAxm94EOvF5gerHXDpPpUVQcQlB0dz066YCcci6v6p6kNys76L7OFgp0iIdcdlAHjPjusmeCMoxLBjsft4vfVBkbi9F2rnh2/zsWETx7WKC5252OxXeDejn6+aIu9fOwECeLukak6Hpcy80fbnZURiaemAcR+JorBNLXPjl5TwS4H/TqAvrKjTDx3We4Gmy3a0mm9hokqZceBjsF6WKnGyqSkOkXZsfmd9ug0FWSfD6d8VS6XxD8PgzzYSZ/GMu6zY0/tEGHl0Uc0jQUVz45Y1qw8v96aolWVs/HspHZQtv9amW8HUHdPZmhL+xUPjbJmWB0Guoebk8t1f882HSrPTpo+O+LnT/R7qQ3KbONKaSxCwG7zu0yoNCg9Z8rOiUIDOqNuy2w3k8k8mEzSWOnaxRvBZr/EZeD99kQr9mznjYmLiViJxQIntvONaJUdnblCbKFMBDvJaizdNFbqBZd5FZzw6wDCuAidNFY+UlgMccGUJP1gJh/Kjt8gHSmSibLjWAdlndJzPXN7tmiVvnSjIgCzPjv21Q6GUa+dI3ysiF+lujJ0Dcqa0nMnqrHCUbWyY+dzeVyjEOxo0liA2gAuDuXU8+xYTWNNaqjQ/b1Z6blRsJPYQCb+X0xlcYMypbEII4xKsZ2kUqca62DPENp7k+bkJiXYUSqy1EpDj+C2t6oA6O2cjdJYo5MVOF0DkZTSaDPE/D5LZXHTt83UoFKNFVV9iRk8iBMukrG4zNNdPl1lR6jGEtNYBu3XARh2GLaL3iDQfI6KYIgG7eoyr+4oiAYx2PHnWNkxCXYy2cU7lsZKno6obrDjXOBXJqRYASV41xsCymC7dm2vGyeGlNYlRxqkprFCyd97UwzSsizrrimp1VjOe3bspOxEZadZR9kRAxnVUE4dz066poI8jWWg7PDRLWIH5TTtRSRJ0k29iwFgbZEZlHOzVSIyhkuweZiN1dI5iJ88+TamH1XNP9THjq5U7bTZzkxbkZVpjx1ArIgQSs8NDNk15YnuxOFYHId6Qxhbq/8FFZFlWRXsvNfWA0AYAmozgBTLW/nOXrjgsmBGVHZEdcCjCnZYiW9M94Jl5tkx6jBsF71BoDyNlYeyc4YYFBjN/BLTWLlSdphCZ1Z6nskcJDG48bol2/O89JWd1O7b2aKtrAlbMCizOWWxuIyugQgaq90IR+M8YMom2DEaGXGknwX9PlSXJX7HFAjxvQuYprGcq8ay2ndJj4kNQb7O6Sk7YjDH1lztUE4ryo4sy0KwY67ssDVAlmXdJoZayrxuDIRjKkVIHI5aU2TKDgU7RUI+DMrjasvh97gwGInh3n/tUf1u+lHqGUpGKa+eDHvsAKnGPvH/tUqWJEkYXeVHS+cgDvZYC3Z6hqKqXbmi7GSXxhJ3Tn06krXXnaoIGLXaF8u9+WwjX6qyo+fZMeowbBe9QaCFUHbEC7bRzC+VspPjpoKmBmXuRbEy/NYFl5RIqdpVFQGlGkvPoOxksMM9O8nHjlrw7EiShJoyLw73h9E1GEZjdUC1VmQznZ55drTKDgv664N+HCpLqDx6wY5u6bkTyo4QOIlNJu0Edl63C5+bNRav7T3CvZIivNptIGyYUrKi7BzpD/M1e0K9uWeHNRUcisS5b8sojQXod9/uE0rPWXVl10AEsiw7UlyRDRTsFAnsS5jL2Vi1QR/+ccMcbH7/MLYf6MabB7rxTmsPwtE4Pjl1jOq+TNVISWOlaSOuh16awCy4G1MVQEvnINp7hiw9vnZCsjaNZbepm7iY6O3i9C6SRq32xTRWv04HXDPPTqdQcusEespOITw7orJlNAajPg/KjpVqrEy65TKJvz8cs/3ZA4Sp58L3Rs/cni3abrhWPDtAYtd/uD+smGiFXX020+mV0nPley3LslCNlTp3iSnFkqTeZPjc6vc2u2osZcOSSZNJI1Z8dobh70TVRq97MmDNoMyaCR5VU2a46RM9O6Kq43ZJqg2ZFr3Uu2hQZt/paFxGbyialdrnBBTsFAnKwMrcXmzG1pZj0anlWHTqOACJha0/FE1RDaq4X0XfgJjJLlu3Gsukr9CYqsRu/qDVYCe5CI6q9ONQbwiHekM40h9WlB2bqUG1QTl1Z8+OXVRzWBmmS4LK08R2+P3hKD8PQYtTzzttTDw3Q29cRD67JzPUyo5RGktRdnK1WGpNrHroGUTNKPN50B+OZbV54X12hAbKrEeTE0NAGdoeTzyN5TEPWLTl33amgOtRn/TsiGms/nCMH1dd0JeiahhVXjLlWGtQzs6zE8+oyaQdqsU0lkFlFPt3fziGSCyuq8Tt6RgAYFyJBSjf+WhcRjgWV83FMlNjtKn3hNqleDoD3sR8waFIHF39kYIHO2RQLgKisTiXqvPV44Thdbt0LzS8J09K07DMO6TqteM3M2SzYXUHe0Mpv9OjI3m/8XXlGFeXyH+/29bjXBprKKJbeaE3LsLI78COQewMravs6BqUnfXs6A0CZZ2bcx1si4gXibqg/kI4pirxWfC5XVmlRszgs7Ec6rMDAGW+1FEhmeKWdJoKhpz37KRWYyU7gBsMAWWwdaNbUx6d7UWNfc67BiNcPWHfG7/HhTKvW/GDaNJY2vPN1p5YXEYsLmfVvFVMY4kX9VykZ2oE5cqo542o9BilsvZ0JCpTJ+l0TmaIn6XBcCxtJRZD6/UKRZX0F/ue8GGgReDbIWWnCBB3lLmsxsqEKu7ZMUpjWV/QdEvPDWZjAcoFzqqy05FcCBsqfKgt92H/kUG819YrDOK0d07Zlz0Sk3EoGVBV6lVjRUVlR9/vwN5XZrJ0C/1NEsdo3EGZma8dKz3XUXaGCqDsiKkYI2WnLujDis/OQDDL1IgZ2pECemTaU6Xcm3ht2Sg7ek0F9Sr5skVbjRW1YFAGhAtystdOtqMiGKyKR5YTF8mGCj9XeeqDicaaYppHnIGnDWJEBSccjWen7HCDcizr6e7pEJUrozWX9UPrC0XRMxRVtWlg7GXKjkFDQSDxPvs8LoSjiRR7j9Vgx6vuzyNeK5hqXVPuQ2v3UFEEO8VxZR3hiN6JfO6szTBKY/VkKOcD+p2GraSx2nusKTsst19f4cfxyf4V77X1Zt27qDxZcQIALV2JRUPl2eH9WZSLkTIqQn1hZgsDK58t97pVO8Jyg9LzeFxGe28i6GusSi1RtYOussM9O/lbEkQ/gJFnBwAuO208Fp7UnLPjYO+VNc+Otc89q2rLyqDMPTuCQZmlsRz07GgDbe7ZSZPGqtY2vsuiyZ6Ix+3iKVsW5IjdkwElpRuLy+gPxww9gNpgJ5tCECVtLdva9GWCqFyZBR/pKrJ280os42AHEIaBhqJpuycztOlP5tcR25LUGvRjKgQU7BQBrCommzJVpzHss2NgljPDp6MkGMnOgJLGYhf5dDDloyHow/FNiWDn3bZe5TlsXnDEHWRLZ2JYpThKw6sTxIWjbAiofhqL+Yu0O3Oj9uudA2EeTI2qTN252cFU2SlQU0GjNFY+MBroKpJJNRageGqySmPpKjvJDsoOvk9MfWKBdthCNRagVAx1cc9O5muDEdr5WEc07RcCXjd/37oGwoabJ49L4s3vQrGYI54dQBlPkytlh4/jGBTTWKnPZWZSlmUZHx62FuyIJmXLwQ7zGSY/N3oVq8WUxqJgpwgwS+kUCmWOltHsm8zTWPpNBc0MyhY9OzrKzs6DvfzCkE0qgQU7HyUnc6vSWNzYqgQNTNnRVrIEeBorGexoduZG7dfbkqm8hgpf2ouPVUzHReTVoJw+jZUPeMWOaRorU88OC3ayMSgn/jYu57aDMk9jRdXKTjrPDhtdoHh2nFF2AKBBY1Lu5CNTlM+JqGoYeQAlSVLN5svGsyN+p5W0dm6CdPbaYnEZH3UPqW4TEXuBaWnvDWEgHIPbJWFcnXkLD/Z56g9HLXt2tIqg6GNiFNPkcwp2ioB8zMXKFBbVp3h2bLSDN52NpRPgjU6ma7oHI7rVSVq4slPhx8T6IHweFwbCMT42IpvdNVtMmLoiBnmirM3QGxUBKOpSJw921Mdk1H6dpfLGOJTCAtKlsfLZQdlaGivXpDMox+My+sKZ9ZdiwU42pecuSUfZCeVwXATvs2MtjaWkWthk7uxHRTCUsTGJz79W2QGEarCBCN8w6qm44mYrm3ERHreLV8ixDZYTKpYeAa+bH+O+wwPJ58osjbX7UELVGVdblnajpKSxYmnnYjG0aazekLGyUwyNBYvn6jqCCUWz85bkgiqjaiynDMom1VhVAQ/fEVvx7XT0M2XHB4/bhWNGJTqFvrG/O/kcWQQ7mi98umosoyGKKWkszcVKbL8uBjtM2XEy2DEtPc9nGkt4rtoiSGMZKTt9qon3FpUdlsZywrMjBjvcSO58nx2l9NxaGitXnh0AqKswT2MBWmVHf7AwoDagZ7uxZJ+VjuRx5Wo4LaCoIvuODBg+l1ljQdZjZ2KaFBYgTD6PWK/G0voMefdk4f0nZYdQMRQpPmWHLVh94SjiwmJrx6CcaRpLkiSlIsuCb0dRdhILIUtlsUDBiTQWQ3c2ls64CI/GexXwqu+rd7HS7pQApSItF8qOblPBQqWxygqZxjI3KIvt+q2mmssdSGNxz45oUNbpvp0tWqUvkmE1VneKZyf7AEA7MkJrUAbUvhYzK4CYxsrGs5N4/GSwo1Od6TTsO8HWAzODsl6ws8eiORlQVNaBUNTyhlbbVLCPBbt+UnYIA9iuxO4XMBewL7EsK/IkYG9B0yvtVXLn+ov2mEpr5efhqNIEizUjmyJMFAay212nKjvKv728GstK6bn6GPQuVnrl50qw44w5GVC/HyyQ5Z6dAhiUK/yegn72042LsGO8ZYFAut2xGSxg1vPsOBmUcoMyq8aKWgx2UvrsONNUEEidfM57TQnpTvbd7BKVHZ3Pkfj+ZjuWh/1dR44NykDqZyfTNFYmwQ7bfPVn0GdHu1716aWxgkzZoWCHs2LFCkiShKVLl/LbDh48iCuvvBLNzc0oLy/HBRdcgF27dqn+LhQKYcmSJWhoaEAwGMTChQvR0tKS56PPjlCWYw1ygZgzZotYJKZM7M7kS65Nm8TiMk/3GKlZoy2alNli6HEplVMpwU4W51X7hdfrs6PfVFCr7Giqr3QuVnrt19nrd6rsXHss7HgLkcaa1BBEmdeNE8emzgbKJ+nSWHYC/MVnTMC15x2Dr5w5wfZxKdVYqUby3Ew9VxuUtZ9hLSyg6w1FEYnFbc3NM0Lr2TmcNo1lrBSLCmzWyk4yMFSqsXKXxtIO4dRVdsrTBztmPXYYQZ+o7EQNn08kxbMjjAthsIC4s5/SWACAV199Fffccw9OPPFEfpssy7jkkkuwe/duPPbYY9i6dSsmTJiAefPmob+/n99v6dKlWLduHdauXYtNmzahr68PCxYsQEyokCl2lBLpong7OJWaxoKiWTmTYMen8baIXhGjJoosbZNuPhbbYdUFfby3w/GNVar7OJnGEkc86AU7TNlJKT3XvLdBnYsVS32Ipuy27tx5dsTnYjO58hns1AV9eGnZXKz+6ml5e049WDWWUem5HcViTFUA18+fojvR2ipKB2VF2enPxdRzbbATt+bZEYOaHoORKnbRjozo1DUoJ5UlIdjR29iolR371Vji3zGfUm7TWBplR6/0PKAf7MTiMjc2W1F27Hh2WJfwoRSDsvJ3lMYS6Ovrw+LFi3HvvfeitraW375r1y689NJL+O1vf4tTTz0VU6ZMwd13342+vj48+OCDAIDu7m6sWrUKd9xxB+bNm4eZM2dizZo12L59OzZs2FCol5QxvPtnESk7gPLlYj4dtugHfe6Ui7kZfD4NC3YEr4jRsEGr87HYjk/sHjqmyq+aI5VdNZbyONoBh8zrIaaxjIYoao9B72LF7jMgVmP1Oh/seNwu/jrYRWIoB+kRK1SXewuevuXVWEbKTsg5420msEGgomcnN2ksTVNBi2kst0viqT2x+Z0jnh3BoByLy7yXj2hkF0vfjTooA+pxNdlMPU88vvq853LeU6qqbD2N9VHXIMKxOHxuF5pr0gfcomfHrK+PSJlX6c0DiAZl0bOjzO8ya9qZDwoe7FxzzTW46KKLMG/ePNXtoVBixx4IKIu82+2Gz+fDpk2bAABbtmxBJBLB/Pnz+X2am5sxffp0bN682fA5Q6EQenp6VD+FpBhLzwGxsaC62iJTmVq7c2b/9bgkw6BJaSxonsZiRsEGYTq2JEmYMkZJZTml7Gh3cXrpD8PS85RgJ71BORyN86oPJz07gKI0KcoO8+wU12cwH1j27DhQUp0JfBBoXDQoJ94nPWXQLgFNFaDVNBagpFE6+8Pcs+FEOTYzKHcOhNE5EObVcGKLAu4ZUvXZMVF2YmLpuV3Pjvr7kVNlR9iwVRqMSzFqKshSWBPqyy01qmXBc9dgxNQQrfc3ZgblqoAX7OkLre4UdGVbu3YtXn/9daxYsSLld8cffzwmTJiAZcuWobOzE+FwGLfffjva2trQ2toKAGhra4PP51MpQgAwZswYtLW1GT7vihUrUF1dzX/GjRvn7AvLkGyl1VyhjIxQp7Ey/YJrLyZWXu9oy8pOsuxcMzfqeMG3k41B2TTY0Ul/RA1SANogQk/Z0QY7zBfgdUuO96FhFwV27EqwU1zqYj5INy6Cp7H8eVZ2kk39YskrfTiqDAx2UtnhnXB5NZa1NBagpJIOdA2CxWROeHZY1VVcVi7cVQGP6piquUE5LPTt0vPsKH2+slZ2vNpgJ5eenVQzdsp9DKqxMik7B5TgmaXNgfSvTZv+1DMouwQvZaHLzwt2dd2/fz+uu+46rFmzRqXeMLxeLx555BHs3LkTdXV1KC8vxwsvvIALL7wQbrf5F12WZdNJtMuWLUN3dzf/2b9/f9avJxvM8s2FhH1oezVprEylW23HXrNdGEPx7JgrO2JDQZEpgm8nuz47yhdXNN4BykVSL42VzqCsN8hR236dBXqjKwPcj+QUTNlhKUWexiqyz2A+SDcI1G6Qny0ezbgI0bjubFNBtcpntfQcUNQHNk7F65Yc2bR53S6uEO06mGgOWqfZ0FTpGJT11lC/yrOTnYqemsbKTzVWumCnN6RuEcIaCk62GOywz1NrMtipDKQfvGtoUNack2IZGVGwYGfLli1ob2/HrFmz4PF44PF4sHHjRvz617+Gx+NBLBbDrFmzsG3bNnR1daG1tRVPP/00Dh8+jEmTJgEAGhsbEQ6H0dnZqXrs9vZ2jBkzxvC5/X4/qqqqVD+FxCzfXEi087Fsp7G0wY6FvkIs2OkNRdEfihrej6V5tBN/xYos59JY6tetl/4wbCro0aaxjIMdlgM/2O182TmDBYBsRAD7b749O8UAHxdhpOw4WGWUCSzAZYNAByKJ4/C6JcdGhwCpF62M0ljJc7KfNb4LeE03mpnAvtOsE3qtJtjhfXbSeXb49zSGcJZtPrSPn9OmgmKwYxBUsc2YLKsLSDJVdso1yo6VDS0zKJtVYwHK+zRi01hz587F9u3bsW3bNv5zyimnYPHixdi2bZtKvamursaoUaOwa9cuvPbaa/j0pz8NAJg1axa8Xi/Wr1/P79va2oodO3Zg9uzZeX9Ndsm290OuSDUo28vJi4sNYC2NVeH38HJIM9+OMhdLvRCqg50cpbF0jK1Gc4VS01jpPTtM2Wmsds6czPALyk4kFudB2khUdtJ5dnocrDLKBK2y0x/KjfompiNkWWkLYSUgYBey/Z2JYMfJc8SUnPcPJYIdbaq6Wkizsw2C3pridyvKHU9j2QwWxcd3SsUyQvTsGAVVfo+bry2ib2dvBj12AEVpZuqmlf5QZcLwUABCNZ76bxVlp7BprPx+ewUqKysxffp01W3BYBD19fX89ocffhijRo3C+PHjsX37dlx33XW45JJLuCG5uroaV111FW644QbU19ejrq4ON954I2bMmJFieC5mlD47xans8DSWTWVHmyawGtyNqQpgd0c/DvYMGX5pmWenQRPsVPg9uPikZrzf3odxteZD8MwQv7gpyo7OuAijuULaVJqZZ4ftUtuSKTxm1nYSv/BcYql7saVS84HVPjt5r8biBuWkryoHPXYA5bMQl5MmXouDQAHFs2M20sAuLLh5/2AvgNT5aeIFmW16dDsoi7OxYtmtteLjVzqoYukhvj6z4KO6zIuhSIgHO+09Q9ifTCtaDXa0hndLwY5XUUSjsbiuZwcQeu0UWNkpWLBjhdbWVlx//fU4ePAgmpqacPnll+MHP/iB6j533nknPB4PFi1ahMHBQcydOxerV69O6+vJB92DEbzT2oNxdeU4yqT8j8/GKjZlh8/HcsagHIrGIcuy6RwbkdFVfh7sGGHk2QGAlZfNzOg49XC7JFT6PegNRY2rsVRNBZN9duwoO5rqhvZ8KDvROFeSJKn4Uqn5IH01FvOqFagaK2nFGGA9dnT8XtkgKkVD4TgP2K2ksZj68FGX4vVwCqbWsqnfWs+O1+1C0OdGfzjGm2/qrSns/e0Px3hVl9/m9UF8/FwrfeIIFbO0UnWZFwd7QvjgUB/+uu0A1rz0IWJxGdVlXoyutJYC126+Mgl2gIRqwyxD2vNSq5mhViiKKth54YUXVP++9tprce2115r+TSAQwMqVK7Fy5cocHpk9bvrLm3j6rTZ8/6IT8B8fn2x4P7OhmIVEqcbK0qCcXFhkOSHJW/HsAOlNyrIs82BH69lxkqoybyLY8RsEO0IflKhB6bnP7YJLAl8QrFRjteVgVASDz8eKxjAUVhoK5nKnWqzwaqxYXLe4QWkqWFhlhw0BddKcDCRev9slIRaXMRSNKb4zC4Evuyiy8ngn+85ogxutZ4c9f384hkO9JspOUoFlQSuQjbKTv2CnMuCBJCXWzXTKDgAsfWgbv23m+Bp871MnWP4+az9TVtosiOeC9QNzSalpVva+scaQhaK4rq4lxtTmhPH57Y/M+/gUq0FZqcbKzqAsLiyZzKfhw0ANlJ2eoSiXpbX5fCdhr1e7uClTz5U0kFFTQUmSVCkivT4pKQZlFuzkIo3F++zECzIqopjQBuNaClWNpR0EytNYXmePQ5IkXp03GI4Zfob1sDK/yS51Qb/m3zrBTjJFYpaeYpuSPsHAa9+zo3xHctlQEEgY1NlzmAUfYnrvpHE1WP3VU/Ho/5uNUybWWX6uoD/zNJbLJXHFmgWbFX5PSoBVLJPPi0rZKTWmNiWDnVbzYKdoDcoBrbJjb6qxuLCEMmjZziRYI4Py4T7lC5ZLrwkLpGo0ngGfYHxkRPi4iNQdVcDrNu2Aq/XsMGl+TE7TWLER3WMHUBtxw9F4SqUTC3ayGeppB21TQVaVmIuKubJkOmgoGhPmu1kxKKu/E46msTTBTZ1Or6lqTRBgVo3F3kevW7LdyiGfyg6QCBS6ByOma+5VZ0+C1+PC5z82FudMGWVLnbWTxkr8nQdDkTBfo/XUz2IZGUHBTg45IansvN/eh1A0ZhjMKD0iikvZ4WmslA7KmX1sXC4JHpeEaFxWKztpPTvmyg4bFaE1JzvNNeceg7G1ZTjvhNGq2/28GktRA8x6lIjzsczGRQxGYugLRbnhz8lREdrnGorEuWIwEsvOgdRgRxQUIjFF+cq/spM4Lt5nJ0dpLEDdRdmoC7geNZphlU4alK2ksURfC2BuUGbBjl1VB9B6dnIf/I6pDODDwwO8yaoep0+ux+mT67N6Hq2H0GqwwzZoh3qNp8Aryg4FOyVLc3UA1WWJyPz99j5Ma9af7qx0/yyui41SjRWFLMtZVaX4PS5Ek/NRFM9OmjSWRWUnl34dADjz6HqceXTqYsICmlhcRiwuw+2SeMpBN9gRlBO9lJFoUGYBXoXfk9K3wglEZWdohKex3C7Fs6KdjyX2LsnF+2AGU3biSVftQI6qsQB1oM0+w1aCAu2wylyUnjP0UtXai7LZbCw2qDKbJqPiOclH8PvDi6fipd2HMfvohpw+j9uVKKNnG1GrQas2jaV3ThRlZ4R2UB4JSJKEE5oS/V7MfDtDRTobi31wo3EZg5GYYFDO/EvuEy6uVtNYomdHllO9FIeYOTmHfh0ztIoAAETixpUsbJEt87p1ZXS2Yx+MKMGO2Y4uGxSDcnxEz8ViiMMiRZipNdPht06g9ewowU4O0ljs85DsuwRYS2NpL4pO+li0FZa6BmWNsmTWQZm9l9kpO+rS81wz/ahq/MfHJ1uab5Utom/HarBTxnuhKZszLTzYGYzoruP5YuSubnlialNCzXmntdfwPlzZKbKLTbnPzb9knQMRvjOyI1WL5edWW7azC/1AOMZTOiL5UnaMUAU7MfVcIb0eJSyYCBqUDpd5U5WdxhyksADRoBzjaayR6tkB1BVZIkYt8POBW+PZGWSl5zkMdgYjynRqK6XnAa9bpQg6qXaIE87FCesilpQdZlDmyk4WwY7w+PluRZBrxPfRbhqrQicAZGmsWFzm3cgLQXFdXUsQXpHV2m14nzD37BTXxUaSlAWmtWuQ325nQWMpK7GLaTo5udzn4c91UKf8XOmxUxhlxyPstth7yHuU6Cy6rI+SkTdGTCVwc3LOgp1UZWekprEA9bBIkR6DrrD5gCs7SbWwP4feKhYAJDw71geBAtY6/do6Jo+bt3uoLffpGm9Tgh2TaixHPDt5rMbKN+ImzHKwk0yptpukscSAuJAmZQp2coyYxjKS8LIdTpdL2CLPBv35PS5b3iKu7ETiphOKtSi9dlJNykr35MIoO5IkpXTf5SkAHdmZKzsGnosynTRWroIdcfgj9+yMUIMyoB4WKdJXoLJzILUaazAPaayhaIwHV1aDHbORKtlSl9zIGKWqtQZp/T47idvYeXRK2SnEZyKXiF4wq4FcWfJcsl5o2l5kjNoiKD8vvqtriXHs6Ep43RJ6hqK8E6iWoSI1KANK5dWBpLJjd+fm05lPYy3YSQQyB3tTz11HL2soWBhlB1BeVySqTmOZGZSNgopyof16a1fuGgoC6g7KI92gDOjPOQOU1Ee+zcmAkMbiBmWWxnL+WERzvPIZtuYTUU3ndljtYCZlMaVl9NyAvu9Mu84Mp2qsfCIG0ZmmsdJVLBbDyAgKdnKMz+PC0aMqABiblItZ2aniyg6bamxvoWWLRCZNBQGloZ5eGqsjqezUBwuj7ACpc5XMynZZsJNO2QGUqcW5U3YUQ+pI77MD6M85A2A47ycfMN9XLA8GZZZiFT0VVjooA5o0lsMBAFN09BoKAqkXZb1ARjvQNJtNpSqNlWELjmKHBdFlXrflqfDajZvRpoAFq5TGKnGYb+cdg+aCVmdFFQK2yLM0VrbKjqoay8LrHZVUNvRGRhTaswOkXiTNS88TtxkpO2Kwm+tgh537UDSGQTYuYgSnsXia1cCgXOnP/y6eedyj+UhjJR9THKlgVQERe904beTmyo5OQ0Htc/s9Ll1fj/bCbfVCroc6jVVayg7z7GTSPFO7QdIzKAOCstNPaayShndS1lF2Er09EotZsQ0CBZSd2gEW7Nj8govDFq3OxgIEZUeTxgpH43zKb6E8OwDg9aireMy6z7JdYdDgYiVJktBFOfE4uUtjpSo7IzmNxauxUkrPC1eNxZUd1kGZG5Rz0HfJqzbxJp7fWhqLKTsVfo/jJdInjq1J/le/R5l4YTZaT1KVHafSWKWp7GSiWGkDb6NzcsuCqXjl5rn4ypkT7B9glpTWu1WksGDnnbbUYCckzFUqRmWHKTktSc+O3S+4X6h2ySiNZWBQPpLsnux2SXlv4y+SouzE06exzC5W5T43Dz4AYHQO5mIljiVpUKamggCMJ5/3hRIBdSE9O9E8lp6zXlqSBMuBC+t1k4tS7MWnj8f8qWN4N3Ut4rBMo+pOrUKVnbKTmzL7YoB9rjJZT7VrhpFB2ej9yyfFd3UtQU5IBjsfHh5QycQAuMoBZGecyxVMyQln2FlTi2iItdpUEBAMypo0Vkeyx05d0Gd7zo0TaEuWI1Hj7rNnHVOPmnIvPnGscTdUURauD/qyWpjNUCk7rM/OiE5j6ZeeF2oIKCB0UI6rPTu5CErZYyrzo/RTQnqwVFIu0jqSJJleKMVhmUZNMbXrTDaendpyL9wuCfVBX1EWlGRD0Eawo01jFXNqr7RC0yKlNuhDU3UArd1DeLetF6cK02iZyuFxSXnv0GoF7SLvSBrL4mwsILWLMluA2VysQnVPZviSCg4zJrMOynopgI8fOwpbf/BJ04uI6JvJlV8HUC4M4iDQEa3ssKo6bTVWAYOdVGUn8T5pJ1Q7Abto9djoMsw2JKMqC5NOZiN5jIIPJz07NeU+rP7qqQVVk3MF80dlosKkGJSLWO0q3iMrMaY2VaG1ewjvtPZogp3iroRJaQdvswLBL1QtWZ2NBSgLaCgaR89glEvmrHtyoRZYhjb9wQ3KBgtqut2yGHDkyq8DKOd+iDw7ANSfT5FeXnpeuKaCsbgMWZYxkMNBoFplx8oQUMYnjhuFmz91PD5+7CjHj8sKNeVe7DuSH88OgIK9zlzz2VljEY7FcdGJzZb/RrtmFCLda5XikxJKlBMMTMpDGZh1C4E2D5+tshOKWJ+NBSSCQGaAFE3KLI1VcGXHsKmgvfdTXDwaq/Oj7ChNBYvzM5gPitOgrHRQDkXj3Kici6q5gKYay2pDQXbfqz9xNF/j8g1TWQyDHbezwU6pUhXw4upPHI2jasos/41Vg3IxQO96njAqP8/kwl8ItDlYux9mXnoey8ygDIi9dpRgh5WdF2ouFsOrMSjzhmweez4i8UKWK3MyoFF2mGenxDwImSDObhNhBuVCprHisuLXAZTmk04SSL7+HgdGKuQbJdixlsYq1rV2OCJmJLxuqajPbfEeWYnBKrLebevl85MAWJ4TVSi0aSvbBmW9poIWq89G65iUO/oK3z0ZUHeGBoSmgg4oO7n07CiG8Ri/kI5sg3KacREFkOfFz1CfEITkwtvHgmymHlntnlwMsGDHyKDspGeHUKMeAuu1bGovBPSu54nxdeUI+twIReO8YRyAjHrOFAJt2sp2GsutDJ7MZDYWAIytLQcA/O/GD3gn50LPxWLwMQOaQaB2d8aistNYnUPPTnKRistKl+CR7Nlhn0/RsyPLsjIuohDKjhBwMONwuT8375H2vS/GYgkj0io7KWmskfs5dxpxvSpmvw5AwU7ecLkkHJ9Ud94SfDtKN+Hi/AJqg5tqmwZl3Wosi4vO1Z+YjMaqAN5v78Nn7t6MHQe6uWenkN2TAT1lJ7EzzsTgKSLKwrlNYylffXYhHdHBjiZoBRKBOXs/CzL1XNglM+9QLlJYQGqBRCaenULDNjxGAakkSaqAh5Qd5xDXDAp2CA6fgC74dordoKxdQOwu+myBGQzHeCmt1dc8qSGIddfMxvGNlTjUG8Ki372I3YcS6lgh52IB6iBOlmWl9NxmsFOuUnbyE+wk50zSuAiolR0WYEhS7oIMM8Smfsw4nKv3SBvs+IZRGuvTJzfjq2dNxH98fJLhfcQAp1jX2uGI+HksZnMyQMFOXpnalGh5/k5rL7+t2A3Kbpek8ivYTWP5uQFSaaqYScfopuoy/PkbZ+LsYxowEFZ8JgX37PCLpJwsEU7ebjeNlbzoeN0S6gzmATmBJKWaCYu1/UE+8OlUY/EUls9TkMaVHpeOspODURFAahA1nJSd+go/brl4Go5vNK4GEz/rpOw4h9qzQ8EOkYQpO9tbuvDBoT4AikG5mC80zJTsdUuGJsB0sAVGnL2TaUBQFfDiD1eeis99bCw/nkJ7dsRqLKZYAfY9D+yiM7oykPMLrDbYoTSWOthhakqhGqW5XBJYJivnyo7mszCcgh0rqJWdkfs5dxq1slPcjRaLOxQrMY5vrELA60LnQARz79iI844fzXvIFKuyAygRe1UWbns/D3YSi7bdjtE+jws//8KJOGNyHcp9noIHieJFUkyB2K1mYa9ndA4bCorPxUqN3S5pWFXgOI3YGoFRyO7JDI9LQiQmC8pOjgzKPq1BubQ+Cz5SdnKC2K6i2D07xX10JUaZz40/ff0M3P38+/jHu+147t12/rti3m2w1JXdsnNATGNFVf+2gyRJ+MIp42z/vZN4hTEDrHsyYL+p4FE1CZ/OsaMrsj+4NIhpxDKvu6jLRnON3mwspXty4ZZJlyQBkPmxBHOUxtL2WBpOfXasQAbl3OByJdLhoWi8qEdFABTs5J2Pja/F7684FXs6+nHfv/fg4ddaMBiJFXzsgRms1042O1yfRtkp1uqzTPELyg7rseN2SbZTUJ+c2oj7vnoqZo6rceoQDREvcIVWyAoNr8bSMShXFFCe97gkhJD7NJbLJcHncfFgr7TTWKX12gpNebKlSrF7dor76EqYSQ1B/OjT03HDJ6fgpT2H8XGTSdiFhuVi7ZqTAaWPSbFXn2WKWHquNBS0r5C4XRLOnTLakWNLh0rZGcGjIgB9z07fUOG6JzNYRVZPjtNYQELd48FOiXw/GZTGyh1lXjc6ESlI481MoHe9wFSXe3H+tMacVVk4AZuPZXcIKJBaeVUqwQ6fqSSksYZLCkBMnY5kczJgXo1VyEWc+dqYypTL9gDiZ8BbgOqzXCJ+J0tl7SkWWOf1Yk9j0btOpKU6WQJdnYVnp1S7mIpeD67sDBNzp1hZR2ksnT47ReDZYcoOS2OVe3N3LOLngdJYhFXYZqCmrLBtQNJR3KEYURR8ZuZR2HWwF186bYLtx0gZxmezhL3Y8Kk8O6x78vB4bX7y7HD4uIhoqmenkCW1rIsyO5ZgjsZFAJqhjjYH2RYrfio9zxnXzj0WG95px5lH1xf6UEyhYIdIy6SGIH775VlZPYZ2N1UquyuWxooInp3hksYKaKqxRjL6np3CzcViaJWdnKaxhMcuZWWHPDvOMveEMZh7wphCH0Za6F0n8kKKslMiuyuxGiua5aiIfEOeHQW9NFZxeHbUyk4uDcpidd5wCditQp4domje9RUrVkCSJCxdupTf1tfXh29+85sYO3YsysrKcMIJJ+C3v/2t6u9CoRCWLFmChoYGBINBLFy4EC0tLXk+eiIdqcFO0Xz0skIsWQ5HE2ms4bIrFt+DkTwXCxCq6oqogzKgKDtsPEpZDj074mdguATsViFlhyiKd/3VV1/FPffcgxNPPFF1+7e+9S08/fTTWLNmDd555x1861vfwpIlS/DYY4/x+yxduhTr1q3D2rVrsWnTJvT19WHBggWIxWL5fhmECVolp1Q8OyywCYnKzjCpZBE9GiPes+NRquoYvUXQQdmtafSYU2VnhBiUS021IqxR8He9r68Pixcvxr333ova2lrV71588UVcccUVOOecczBx4kRcffXVOOmkk/Daa68BALq7u7Fq1SrccccdmDdvHmbOnIk1a9Zg+/bt2LBhQyFeDmFAqmenNC6uen12hsvOUaXsjPRgR8eg3FdE1ViM3AY7JezZSb6/bptjaojhT8Hf9WuuuQYXXXQR5s2bl/K7s88+G48//jgOHDgAWZbx/PPPY+fOnTj//PMBAFu2bEEkEsH8+fP53zQ3N2P69OnYvHmz4XOGQiH09PSofojcklp6XvCPniOIaSxejTVMlB2xizU1FdRLYxVe2dGmk3LZj0sMeEtN/WDvb6msO0TmFLQaa+3atXj99dfx6quv6v7+17/+Nb7+9a9j7Nix8Hg8cLlc+P3vf4+zzz4bANDW1gafz5eiCI0ZMwZtbW2Gz7tixQrceuutzr0QIi2u5KBJFhCUyqIjTj1nys5w2RWTsqPALobRuIx4XIYkCQblQpaea2as5UvZKTXPDvusDxfVlXCegr3z+/fvx3XXXYc1a9YgEAjo3ufXv/41XnrpJTz++OPYsmUL7rjjDvzXf/1X2hSVLMumQw2XLVuG7u5u/rN///6sXgthDVVFRIlcXFXVWLHha1Amz45yLsKxOIYiccTiifezkGksrUqY63ERjOHyGbYKKTtEwb7FW7ZsQXt7O2bNUvq3xGIx/POf/8Rdd92F7u5u3HzzzVi3bh0uuugiAMCJJ56Ibdu24ec//znmzZuHxsZGhMNhdHZ2qtSd9vZ2zJ492/C5/X4//P7iHbxZqvg8LvQnq0pKZdFR0liyoOwMj11xQJXGGuHBjlsT7CQ/py4ptwFGOrQG5Xz12Sm1NBYpO0TB3vm5c+di+/bt2LZtG/855ZRTsHjxYmzbtg2xWAyRSAQujYzrdrsRT1a9zJo1C16vF+vXr+e/b21txY4dO0yDHaIwiKbkUgl21Gms4dZBmdJYDDFADUfjqlERZipxrkk1KOdufyp+Hkqtg7Ki7Izsz/lIpmDKTmVlJaZPn666LRgMor6+nt8+Z84cfPvb30ZZWRkmTJiAjRs34oEHHsAvfvELAEB1dTWuuuoq3HDDDaivr0ddXR1uvPFGzJgxQ9fwTBQW9Xya0lh0xGZ0w1nZGelpLEmS4HO7EI7FEY7GeffkQvp1ALV3xu9xpQQ/TlLSHZSTr6fUFCvCOkU9LmLt2rVYtmwZFi9ejCNHjmDChAn4yU9+gm984xv8PnfeeSc8Hg8WLVqEwcFBzJ07F6tXr4bbPbIX72JEFeyUSJ8d8TWxxm/D5UJByo4an0cIdoqg7BxQKzu5TqeJnwGPa3h8hq3ClZ0SWXeIzCmqYOeFF15Q/buxsRH33Xef6d8EAgGsXLkSK1euzOGREU6gHsZXGouOuFMcCCcukMPlQkHKjhqfxwWEEm0EiqF7MqA2KOcyhQWoPwO+Ektj1QYTE7nryot7MjeRO4oq2CFKm1JMY4kqTn8ooewMlwsFjYtQ4xO6YRdDjx0AcAl+oVy/R6VcjfXxYxrwk89Mx5mTi3syN5E7KNgh8oaoggRKRE52uyS4XRJicXnYKTuqpoKk7HBTbjhWPGks0bOT6zRWKXdQ9rhdWHz6hEIfBlFASusTTRQ14sW1VJQdQAni+oeZZ0cMOCnYUQ8DLRZlR2wqmOv3SD0ba3iokwRhleGxKhMlgaqpYIl4dgAlPdefVAOGy4VCDDgDI3xcBAD4PMp8rGLongyoPTvBHKtMpVyNRRD0iSbyhr8Eq7EA5cKgBDvD47WRsqNGnI/FlJ1Cp7HEaizy7BCEfegTTeQNfwkalAHldQ1GEmms4TJXiFX3eFwSVWMB8LuVoa68GqvQwY5gUC7PeRqLgh2idCGDMpE3fCVYeg4or6tvmCk71WVe3HLxVJT73MPmmHOJ2CCSG5QL7dkpmEF5eATsBGEVCnaIvFGKpeeAcmEYCMVU/x4OfPWsSYU+hKKBvW8hoYNyVRH12SnLcZ8dSmMRpQx9oom8UaqeHW5QHmal54Qa0bOjlJ4X1qAsenaCOVZ2vG4J9UEffB4XqsoK+7oJwmlI2SHyRsmmsZK7YD4uooRe20hCrMbiBuWiUnZyG+xIkoQ/ff0M9IejBfcqEYTT0CeayBs+d2n22WGSfyyemHruzeGwRiJ38D47gkG50H12XHkcFwEAUxorc/4cBFEIaAtK5A0xdeUrIfVD+1rI7zA80UtjVRa6g3IeB4ESRClDqzKRN9jO2euWVF6E4Y42JTdcSs8JNex97B6MICnSFTyNpeqgTMEOQdiGgh0ib7CdcymlsIBUJcdHys6whFVjHe4LAUiYgwvdbJGUHYJwBlqVibzh58FOaX3stGksDwU7wxL2Ph7uDwNINBSUpMKqdO48e3YIolShVZnIG75SDXbcWs8OpbGGI8xAf0QIdgqNm5QdgnCE0rrqEEUNV3ZKbDSBttScDMrDExaMs2Cn0JVYAKWxCMIpaFUm8gbz6pSapyVV2Smt1zdS0KaxiiHYyecgUIIoZWhVJvLGSeNqMHlUEBed2FToQ3GUVM8OpbGGI2LpOVAcaSwPeXYIwhHo20PkjbqgD8/dcE6hD8NxUpQdGhcxLPFpgtSKQOFHJqhKz0ss/UsQ+YRWZYLIkpSmgh5SdoYj2vexONJYif8GvK6S6k1FEPmGgh2CyBKtR4cGgQ5PxHEmQOG7JwOKskMpLILIDlqVCSJLtIpAqRmwRwra97GYPDuUwiKI7KBVmSCyhAzKpUFKsFMUaazEZ4nKzgkiOyjYIYgs0RpbqfR8eKJtBllZBAZlDwU7BOEItCoTRJakTj0nZWc4ou3sXQxprJPG1eComjLMn9ZY6EMhiGFN4b/NBDHM0RpbSdkZnqQYlIsgjdVcU4Z/f/e8Qh8GQQx7aFUmiCzRKjnk2RmeFGPpOUEQzkDBDkFkSUoai0rPhyXFWI1FEIQz0KpMEFkiXiTdLgkuav42LCnGaiyCIJyBgh2CyBKxr46HAp1hizYdWVUE1VgEQTgDBTsEkSWiIkANBYcvfsGg7HFJKdVZBEEMX+jbTBBZIgY7ZE4evojvY0XAA0mi95IgSoWiCXZWrFgBSZKwdOlSfpskSbo///M//8PvEwqFsGTJEjQ0NCAYDGLhwoVoaWkpwCsgRipiqTmVnQ9fVMEOmZMJoqQoipX51VdfxT333IMTTzxRdXtra6vq5w9/+AMkScLnPvc5fp+lS5di3bp1WLt2LTZt2oS+vj4sWLAAsVgs3y+DGKH4KNgpCdwuiY9nKIbuyQRBOEfBV+a+vj4sXrwY9957L2pra1W/a2xsVP089thjOPfcczF58mQAQHd3N1atWoU77rgD8+bNw8yZM7FmzRps374dGzZsKMTLIUYgoreDuicPb1jgWgwTzwmCcI6CBzvXXHMNLrroIsybN8/0fgcPHsSTTz6Jq666it+2ZcsWRCIRzJ8/n9/W3NyM6dOnY/PmzYaPFQqF0NPTo/ohCLuIao6HlJ1hDQtWqeycIEqLgn6j165di9dffx2vvvpq2vvef//9qKysxGc/+1l+W1tbG3w+X4oiNGbMGLS1tRk+1ooVK3DrrbfaP3CCEPB5KI1VKvg8bgBR6p5MECVGwVbm/fv347rrrsOaNWsQCATS3v8Pf/gDFi9ebOm+siybVlIsW7YM3d3d/Gf//v0ZHTtBiKgNypTGGs6wlCQZlAmitCjYN3rLli1ob2/HrFmz+G2xWAz//Oc/cddddyEUCsGd7Hvxr3/9C++99x4eeugh1WM0NjYiHA6js7NTpe60t7dj9uzZhs/t9/vh9/sdfkXESEUMcKip4PCGqXSUxiKI0qJgys7cuXOxfft2bNu2jf+ccsopWLx4MbZt28YDHQBYtWoVZs2ahZNOOkn1GLNmzYLX68X69ev5ba2trdixY4dpsEMQTiJJEr9IUhpreEMGZYIoTQr2ja6srMT06dNVtwWDQdTX16tu7+npwcMPP4w77rgj5TGqq6tx1VVX4YYbbkB9fT3q6upw4403YsaMGWkNzwThJD63C+FonIKdYQ4LWqn0nCBKi6LfvqxduxayLOOyyy7T/f2dd94Jj8eDRYsWYXBwEHPnzsXq1atVyhBB5BqfxwWEyLMz3OHVWKTsEERJUVTf6BdeeCHltquvvhpXX3214d8EAgGsXLkSK1euzOGREYQ5LP1BpefDmymNldi6vwtTGisLfSgEQThIUQU7BDFc8XoSigANAh3e/PiSGfjWJ4/D6Mr0VZ8EQQwfaGUmCAdQlB1KYw1n3C6JAh2CKEEo2CEIB0g0o6NqLIIgiGKEVmaCcABfUtEhgzJBEETxQcEOQTgA9dkhCIIoXmhlJggHYMGOx0VfKYIgiGKDVmaCcACm6FAaiyAIovigYIcgHMDnpjQWQRBEsUIrM0E4AE9jkbJDEARRdFCwQxAOQMoOQRBE8UIrM0E4wPxpYzChvhwfP7ah0IdCEARBaKBxEQThABdMb8IF05sKfRgEQRCEDqTsEARBEARR0lCwQxAEQRBESUPBDkEQBEEQJQ0FOwRBEARBlDQU7BAEQRAEUdJQsEMQBEEQRElDwQ5BEARBECUNBTsEQRAEQZQ0FOwQBEEQBFHSULBDEARBEERJQ8EOQRAEQRAlDQU7BEEQBEGUNBTsEARBEARR0lCwQxAEQRBESeMp9AEUA7IsAwB6enoKfCQEQRAEQViFXbfZddwICnYA9Pb2AgDGjRtX4CMhCIIgCCJTent7UV1dbfh7SU4XDo0A4vE4PvroI1RWVkKSJMcet6enB+PGjcP+/ftRVVXl2OMS+tD5zh90rvMHnev8Qec6fzh1rmVZRm9vL5qbm+FyGTtzSNkB4HK5MHbs2Jw9flVVFX1x8gid7/xB5zp/0LnOH3Su84cT59pM0WGQQZkgCIIgiJKGgh2CIAiCIEoaCnZyiN/vxy233AK/31/oQxkR0PnOH3Su8wed6/xB5zp/5Ptck0GZIAiCIIiShpQdgiAIgiBKGgp2CIIgCIIoaSjYIQiCIAiipKFghyAIgiCIkoaCnRxy9913Y9KkSQgEApg1axb+9a9/FfqQhj0rVqzAqaeeisrKSowePRqXXHIJ3nvvPdV9ZFnG8uXL0dzcjLKyMpxzzjl46623CnTEpcGKFSsgSRKWLl3Kb6Pz7CwHDhzAl7/8ZdTX16O8vBwnn3wytmzZwn9P59sZotEovv/972PSpEkoKyvD5MmT8aMf/QjxeJzfh861Pf75z3/i4osvRnNzMyRJwl//+lfV762c11AohCVLlqChoQHBYBALFy5ES0tL9gcnEzlh7dq1stfrle+991757bfflq+77jo5GAzKH374YaEPbVhz/vnny/fdd5+8Y8cOedu2bfJFF10kjx8/Xu7r6+P3uf322+XKykr5kUcekbdv3y5/8YtflJuamuSenp4CHvnw5ZVXXpEnTpwon3jiifJ1113Hb6fz7BxHjhyRJ0yYIF955ZXyyy+/LO/Zs0fesGGD/P777/P70Pl2hh//+MdyfX29/MQTT8h79uyRH374YbmiokL+5S9/ye9D59oef//73+Xvfe978iOPPCIDkNetW6f6vZXz+o1vfEM+6qij5PXr18uvv/66fO6558onnXSSHI1Gszo2CnZyxGmnnSZ/4xvfUN12/PHHy9/97ncLdESlSXt7uwxA3rhxoyzLshyPx+XGxkb59ttv5/cZGhqSq6ur5f/93/8t1GEOW3p7e+Vjjz1WXr9+vTxnzhwe7NB5dpabbrpJPvvssw1/T+fbOS666CL5a1/7muq2z372s/KXv/xlWZbpXDuFNtixcl67urpkr9crr127lt/nwIEDssvlkp9++umsjofSWDkgHA5jy5YtmD9/vur2+fPnY/PmzQU6qtKku7sbAFBXVwcA2LNnD9ra2lTn3u/3Y86cOXTubXDNNdfgoosuwrx581S303l2lscffxynnHIKvvCFL2D06NGYOXMm7r33Xv57Ot/OcfbZZ+Mf//gHdu7cCQB44403sGnTJnzqU58CQOc6V1g5r1u2bEEkElHdp7m5GdOnT8/63NMg0BzQ0dGBWCyGMWPGqG4fM2YM2traCnRUpYcsy7j++utx9tlnY/r06QDAz6/euf/www/zfozDmbVr1+L111/Hq6++mvI7Os/Osnv3bvz2t7/F9ddfj5tvvhmvvPIKrr32Wvj9flx++eV0vh3kpptuQnd3N44//ni43W7EYjH85Cc/wWWXXQaAPtu5wsp5bWtrg8/nQ21tbcp9sr12UrCTQyRJUv1bluWU2wj7fPOb38Sbb76JTZs2pfyOzn127N+/H9dddx2effZZBAIBw/vReXaGeDyOU045BbfddhsAYObMmXjrrbfw29/+Fpdffjm/H53v7HnooYewZs0a/OlPf8K0adOwbds2LF26FM3Nzbjiiiv4/ehc5wY759WJc09prBzQ0NAAt9udEom2t7enRLWEPZYsWYLHH38czz//PMaOHctvb2xsBAA691myZcsWtLe3Y9asWfB4PPB4PNi4cSN+/etfw+Px8HNJ59kZmpqaMHXqVNVtJ5xwAvbt2weAPtdO8u1vfxvf/e53cemll2LGjBn4yle+gm9961tYsWIFADrXucLKeW1sbEQ4HEZnZ6fhfexCwU4O8Pl8mDVrFtavX6+6ff369Zg9e3aBjqo0kGUZ3/zmN/Hoo4/iueeew6RJk1S/nzRpEhobG1XnPhwOY+PGjXTuM2Du3LnYvn07tm3bxn9OOeUULF68GNu2bcPkyZPpPDvIWWedldJCYefOnZgwYQIA+lw7ycDAAFwu9aXP7Xbz0nM617nBynmdNWsWvF6v6j6tra3YsWNH9uc+K3szYQgrPV+1apX89ttvy0uXLpWDwaC8d+/eQh/asOb//b//J1dXV8svvPCC3Nrayn8GBgb4fW6//Xa5urpafvTRR+Xt27fLl112GZWNOoBYjSXLdJ6d5JVXXpE9Ho/8k5/8RN61a5f8f//3f3J5ebm8Zs0afh86385wxRVXyEcddRQvPX/00UflhoYG+Tvf+Q6/D51re/T29spbt26Vt27dKgOQf/GLX8hbt27lLVesnNdvfOMb8tixY+UNGzbIr7/+unzeeedR6Xmx85vf/EaeMGGC7PP55I997GO8PJqwDwDdn/vuu4/fJx6Py7fccovc2Ngo+/1++ROf+IS8ffv2wh10iaANdug8O8vf/vY3efr06bLf75ePP/54+Z577lH9ns63M/T09MjXXXedPH78eDkQCMiTJ0+Wv/e978mhUIjfh861PZ5//nnd9fmKK66QZdnaeR0cHJS/+c1vynV1dXJZWZm8YMECed++fVkfmyTLspydNkQQBEEQBFG8kGeHIAiCIIiShoIdgiAIgiBKGgp2CIIgCIIoaSjYIQiCIAiipKFghyAIgiCIkoaCHYIgCIIgShoKdgiCIAiCKGko2CEIomg555xzsHTp0kIfBkeWZVx99dWoq6uDJEnYtm1byn1Wr16NmpqavB8bQRDGULBDEARhkaeffhqrV6/GE088gdbWVkyfPj3lPl/84hexc+dO/u/ly5fj5JNPzuNREgShxVPoAyAIgsgnsVgMkiSlDIO0wgcffICmpibToYRlZWUoKyvL5hB1iUQi8Hq9jj8uQYwESNkhCMKUc845B9deey2+853voK6uDo2NjVi+fDn//d69e1NSOl1dXZAkCS+88AIA4IUXXoAkSXjmmWcwc+ZMlJWV4bzzzkN7ezueeuopnHDCCaiqqsJll12GgYEB1fNHo1F885vfRE1NDerr6/H9738f4pSbcDiM73znOzjqqKMQDAZx+umn8+cFlLTSE088galTp8Lv9+PDDz/Ufa0bN27EaaedBr/fj6amJnz3u99FNBoFAFx55ZVYsmQJ9u3bB0mSMHHiRN3HENNYq1evxq233oo33ngDkiRBkiSsXr0aANDd3Y2rr74ao0ePRlVVFc477zy88cYb/HGYIvSHP/wBkydPht/vB033IQh7kLJDEERa7r//flx//fV4+eWX8eKLL+LKK6/EWWedhU9+8pMZPc7y5ctx1113oby8HIsWLcKiRYvg9/vxpz/9CX19ffjMZz6DlStX4qabblI991VXXYWXX34Zr732Gq6++mpMmDABX//61wEAX/3qV7F3716sXbsWzc3NWLduHS644AJs374dxx57LABgYGAAK1aswO9//3vU19dj9OjRKcd24MABfOpTn8KVV16JBx54AO+++y6+/vWvIxAIYPny5fjVr36Fo48+Gvfccw9effVVuN3utK/3i1/8Inbs2IGnn34aGzZsAABUV1dDlmVcdNFFqKurw9///ndUV1fjd7/7HebOnYudO3eirq4OAPD+++/jz3/+Mx555BFLz0cQhAFZjxIlCKKkmTNnjnz22Werbjv11FPlm266SZZlWd6zZ48MQN66dSv/fWdnpwxAfv7552VZVqYhb9iwgd9nxYoVMgD5gw8+4Lf953/+p3z++eernvuEE06Q4/E4v+2mm26STzjhBFmWZfn999+XJUmSDxw4oDq+uXPnysuWLZNlWZbvu+8+GYC8bds209d58803y1OmTFE9129+8xu5oqJCjsVisizL8p133ilPmDDB9HHuu+8+ubq6mv/7lltukU866STVff7xj3/IVVVV8tDQkOr2o48+Wv7d737H/87r9crt7e2mz0cQRHpI2SEIIi0nnnii6t9NTU1ob2/P6nHGjBmD8vJyTJ48WXXbK6+8ovqbM844A5Ik8X+feeaZuOOOOxCLxfD6669DlmUcd9xxqr8JhUKor6/n//b5fCmvQcs777yDM888U/VcZ511Fvr6+tDS0oLx48dn9mJN2LJlC/r6+lTHCACDg4P44IMP+L8nTJiAUaNGOfa8BDFSoWCHIIi0aI2xkiQhHo8DADf6yoKfJBKJpH0cSZJMH9cK8XgcbrcbW7ZsSUnzVFRU8P8vKytTBTF6yLKcch/2mtL9babE43E0NTWpvEUMsWw9GAw6+rwEMVKhYIcgiKxgykNraytmzpwJALr9Z+zy0ksvpfz72GOPhdvtxsyZMxGLxdDe3o6Pf/zjWT3P1KlT8cgjj6iCns2bN6OyshJHHXWU7cf1+XyIxWKq2z72sY+hra0NHo/H0OhMEIRzUDUWQRBZUVZWhjPOOAO333473n77bfzzn//E97//fccef//+/bj++uvx3nvv4cEHH8TKlStx3XXXAQCOO+44LF68GJdffjkeffRR7NmzB6+++ip++tOf4u9//3tGz/Nf//Vf2L9/P5YsWYJ3330Xjz32GG655RZcf/31tsrUGRMnTsSePXuwbds2dHR0IBQKYd68eTjzzDNxySWX4JlnnsHevXuxefNmfP/738drr71m+7kIgtCHgh2CILLmD3/4AyKRCE455RRcd911+PGPf+zYY19++eUYHBzEaaedhmuuuQZLlizB1VdfzX///9u5Q1sHASgMo5chmgqCBdkNSHBVmBoUAgyegZBswA4wRTdpB3jqpe7mnAV+++WKu21bjOMY67pG0zTR932c5xlVVf1rpyzLOI4jruuKx+MRy7LEPM8/h9vr9Yrn8xld18Xtdot936MoijiOI9q2jWmaoq7rGIYh3u933O/3n/aAv4rPx+MGACAvlx0AIDWxAwCkJnYAgNTEDgCQmtgBAFITOwBAamIHAEhN7AAAqYkdACA1sQMApCZ2AIDUxA4AkNoXuSot6n5G9dMAAAAASUVORK5CYII=\n"
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "解码结果: [[1, 7, 7, 1, 0, 1, 6, 5, 6, 6, 2, 3, 4, 7, 1, 6, 5, 8, 3], [1, 2, 3, 1, 1, 4, 1, 6, 3, 5, 2, 4, 2, 6, 2, 3, 4, 7, 7], [5, 0, 5, 5, 4, 7, 4, 3, 1, 2, 1, 5, 0, 2, 3, 7, 7, 5, 3], [5, 1, 0, 6, 2, 5, 7, 1, 2, 1, 3, 0, 5, 8, 4, 4, 1, 6, 3]]\n"
     ]
    }
   ],
   "source": [
    "# 模型：\n",
    "# 1.构建概率图模型：条件随机场CRF\n",
    "import torch\n",
    "from TorchCRF import CRF\n",
    "from torch.utils.data import TensorDataset,DataLoader\n",
    "from torch.optim import Adam\n",
    "\n",
    "# 构建假数据\n",
    "seq_len, num_tags = 19, 9\n",
    "X = torch.randn(1000,seq_len,num_tags)\n",
    "tags = torch.randint(0, num_tags, (1000, seq_len))\n",
    "dataset = TensorDataset(X,tags)\n",
    "batch_size = 10\n",
    "dataloader = DataLoader(dataset, shuffle = True, batch_size = batch_size)\n",
    "# 模型定义\n",
    "model = CRF(num_tags = num_tags, batch_first = True)\n",
    "optimizer = Adam(model.parameters(), lr = 0.05)\n",
    "# 模型训练\n",
    "losses = []\n",
    "for seq, seq_tag in dataloader:\n",
    "    optimizer.zero_grad()\n",
    "    loss = (-1)*model(seq, seq_tag)\n",
    "    loss.backward()\n",
    "    optimizer.step()\n",
    "    losses.append(loss.tolist())\n",
    "# 可视化loss\n",
    "import matplotlib.pyplot as plt\n",
    "plt.xlabel('number of iter')\n",
    "plt.ylabel('loss')\n",
    "plt.title('a simple of torchcrf')\n",
    "plt.plot(losses)\n",
    "plt.legend()\n",
    "plt.show()\n",
    "print('解码结果:', model.decode(X[:4]))\n"
   ],
   "metadata": {
    "collapsed": false,
    "ExecuteTime": {
     "end_time": "2023-08-27T22:45:31.981018Z",
     "start_time": "2023-08-27T22:45:30.766527Z"
    }
   },
   "id": "31d2dd909d5afbcd"
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "loss = 267.31866455078125\n",
      "loss = 264.75323486328125\n",
      "loss = 272.2840576171875\n",
      "loss = 271.5197448730469\n",
      "loss = 266.4097900390625\n",
      "loss = 270.8936462402344\n",
      "loss = 264.8451843261719\n",
      "loss = 274.60205078125\n",
      "loss = 276.1224670410156\n",
      "loss = 237.62351989746094\n",
      "[[3, 0, 3, 3, 3, 1, 1, 1, 2, 3, 3, 3, 3, 2, 3, 3, 1, 2, 3], [3, 2, 0, 3, 0, 0, 3, 0, 3, 1, 1, 1, 1, 1, 2, 0, 3, 0, 0], [1, 1, 1, 1, 1, 1, 2, 0, 0, 3, 1, 1, 3, 3, 3, 3, 3, 0, 3]]\n"
     ]
    }
   ],
   "source": [
    "# 模型：\n",
    "# 2.构建概率图模型：BiLSTM + CRF 简单案例\n",
    "\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "from torch.utils.data import TensorDataset, DataLoader\n",
    "from TorchCRF import CRF\n",
    "from torch.optim import Adam\n",
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\")\n",
    "\n",
    "class BiLSTM_CRF(nn.Module):\n",
    "    def __init__(self, vocab_size, tag_to_ix, embedding_dim, hidden_dim):\n",
    "        super(BiLSTM_CRF, self).__init__()\n",
    "        self.embedding_dim = embedding_dim\n",
    "        self.hidden_dim = hidden_dim\n",
    "        self.vocab_size = vocab_size\n",
    "        self.tag_to_ix = tag_to_ix\n",
    "        self.crf = CRF(num_tags = len(self.tag_to_ix), batch_first = True) # 很重要 batch_first = True\n",
    "        self.word_embeds = nn.Embedding(vocab_size, embedding_dim, padding_idx=0)\n",
    "        # 简便折断\n",
    "        self.bilstm = nn.LSTM(embedding_dim, hidden_dim // 2, num_layers=1, bidirectional=True, batch_first=True)\n",
    "        self.hidden2tag = nn.Linear(hidden_dim, len(self.tag_to_ix))\n",
    "    def __forward__(self, batch_token_ids): # [batch_size, seq_size]\n",
    "        shape_ = batch_token_ids.shape\n",
    "        embeds = self.word_embeds(batch_token_ids).view(shape_[0], shape_[1], self.embedding_dim)\n",
    "        bilstm_out, _ = self.bilstm(embeds)\n",
    "        bilstm_out = bilstm_out.view(shape_[0], shape_[1], self.hidden_dim)\n",
    "        bilstm_feats = self.hidden2tag(bilstm_out)\n",
    "        # print(bilstm_feats.shape) # orch.Size([10, 19, 4])\n",
    "        return bilstm_feats\n",
    "    def loss(self, batch_token_ids, tags, masks):\n",
    "        feats = self.__forward__(batch_token_ids)\n",
    "        return - self.crf(feats, tags, masks)\n",
    "    def predict(self, batch_token_ids):\n",
    "        feats = self.__forward__(batch_token_ids)\n",
    "        return self.crf.decode(feats)\n",
    "\n",
    "tag_to_ix = {'O':0, 'B': 1, 'I':2, 'E':3}\n",
    "num_classes = len(tag_to_ix)\n",
    "word_num = 100\n",
    "seq_len = 19\n",
    "embedding_dim = 64\n",
    "hidden_dim = 32\n",
    "model = BiLSTM_CRF(vocab_size = word_num, tag_to_ix = tag_to_ix, embedding_dim = embedding_dim, hidden_dim = hidden_dim)\n",
    "# 构建假数据\n",
    "all_token_ids = torch.randint(0, word_num, (99, seq_len)).long() # 词汇id, torch.Size([99, 19])\n",
    "all_token_tags = torch.randint(0, len(tag_to_ix), (99, seq_len)).long() # tag 标注结果  torch.Size([99, 19])\n",
    "all_token_masks = torch.ones_like(all_token_ids).bool() # mask  torch.Size([99, 19])\n",
    "dataset = TensorDataset(all_token_ids, all_token_tags, all_token_masks)\n",
    "batch_size = 10\n",
    "dataloader = DataLoader(dataset, shuffle=True, batch_size = batch_size)\n",
    "optimizer = Adam(model.parameters(), lr = 0.05)\n",
    "# 模型训练\n",
    "for seq, seq_tag, seq_mask in dataloader:\n",
    "    optimizer.zero_grad()\n",
    "    loss = model.loss(seq, seq_tag, seq_mask)\n",
    "    loss.backward()\n",
    "    optimizer.step()\n",
    "    print('loss =', loss.item())\n",
    "# 示例使用CRF进行序列标注\n",
    "print(model.predict(all_token_ids[:3,:]))"
   ],
   "metadata": {
    "collapsed": false,
    "ExecuteTime": {
     "end_time": "2023-08-27T22:46:03.202282Z",
     "start_time": "2023-08-27T22:46:03.090341Z"
    }
   },
   "id": "bc6c08c9189bff0d"
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(2154, 100)\n",
      "loss = 2274.661865234375\n",
      "loss = 1412.525634765625\n",
      "loss = 871.842529296875\n",
      "loss = 635.783203125\n",
      "loss = 479.56591796875\n",
      "loss = 796.300048828125\n",
      "loss = 616.8015747070312\n",
      "loss = 751.3322143554688\n",
      "loss = 618.5487060546875\n",
      "loss = 695.362060546875\n",
      "loss = 629.3770751953125\n",
      "loss = 562.5451049804688\n",
      "loss = 333.31103515625\n",
      "loss = 398.85638427734375\n",
      "loss = 727.3421630859375\n",
      "loss = 459.9878845214844\n",
      "loss = 535.0306396484375\n",
      "loss = 494.2618408203125\n",
      "loss = 420.13043212890625\n",
      "loss = 540.7581176757812\n",
      "loss = 360.2719421386719\n",
      "loss = 368.5841369628906\n",
      "loss = 433.14117431640625\n",
      "loss = 447.36785888671875\n",
      "loss = 326.5769958496094\n",
      "loss = 502.8112487792969\n",
      "loss = 517.9593505859375\n",
      "loss = 434.51568603515625\n",
      "loss = 373.7227783203125\n",
      "loss = 330.0563049316406\n",
      "loss = 340.66827392578125\n",
      "loss = 265.15869140625\n",
      "loss = 403.1212158203125\n",
      "loss = 300.79803466796875\n",
      "loss = 328.957763671875\n",
      "loss = 258.57855224609375\n",
      "loss = 493.8677062988281\n",
      "loss = 270.4941711425781\n",
      "loss = 240.83013916015625\n",
      "loss = 338.2187805175781\n",
      "loss = 238.5732879638672\n",
      "loss = 385.82318115234375\n",
      "loss = 223.855712890625\n",
      "loss = 270.7386779785156\n",
      "loss = 282.5658264160156\n",
      "loss = 262.42425537109375\n",
      "loss = 391.541015625\n",
      "loss = 194.21861267089844\n",
      "loss = 382.7391357421875\n",
      "loss = 305.1744079589844\n",
      "loss = 329.3395080566406\n",
      "loss = 321.2178039550781\n",
      "loss = 253.47645568847656\n",
      "loss = 219.84169006347656\n",
      "loss = 156.73428344726562\n",
      "loss = 291.00872802734375\n",
      "loss = 209.42953491210938\n",
      "loss = 257.113037109375\n",
      "loss = 129.87612915039062\n",
      "loss = 147.22915649414062\n",
      "loss = 409.8995056152344\n",
      "loss = 317.1215515136719\n",
      "loss = 175.08876037597656\n",
      "loss = 255.3208465576172\n",
      "loss = 200.7508544921875\n",
      "loss = 253.4585723876953\n",
      "loss = 208.052978515625\n",
      "loss = 163.8997344970703\n",
      "loss = 201.00051879882812\n",
      "loss = 224.87014770507812\n",
      "loss = 288.0331726074219\n",
      "loss = 183.64617919921875\n",
      "loss = 202.90725708007812\n",
      "loss = 212.42144775390625\n",
      "loss = 314.84759521484375\n",
      "loss = 261.9424743652344\n",
      "loss = 275.28643798828125\n",
      "loss = 179.57717895507812\n",
      "loss = 162.12733459472656\n",
      "loss = 283.59967041015625\n",
      "loss = 218.1262664794922\n",
      "loss = 235.16761779785156\n",
      "loss = 232.77273559570312\n",
      "loss = 250.39828491210938\n",
      "loss = 218.08847045898438\n",
      "loss = 257.6257019042969\n",
      "loss = 160.9289093017578\n",
      "loss = 210.14405822753906\n",
      "loss = 289.8397216796875\n",
      "loss = 218.4615478515625\n",
      "loss = 177.84078979492188\n",
      "loss = 297.4605712890625\n",
      "loss = 264.6927185058594\n",
      "loss = 207.4880828857422\n",
      "loss = 145.67803955078125\n",
      "loss = 136.3765106201172\n",
      "loss = 218.5213165283203\n",
      "loss = 282.6913146972656\n",
      "loss = 185.984619140625\n",
      "loss = 262.05096435546875\n",
      "loss = 312.5618591308594\n",
      "loss = 315.51593017578125\n",
      "loss = 191.25486755371094\n",
      "loss = 161.28237915039062\n",
      "loss = 254.4296875\n",
      "loss = 160.98703002929688\n",
      "loss = 188.74501037597656\n",
      "loss = 233.295654296875\n",
      "loss = 176.64747619628906\n",
      "loss = 183.30233764648438\n",
      "loss = 304.33917236328125\n",
      "loss = 99.9466552734375\n",
      "loss = 235.00511169433594\n",
      "loss = 283.04315185546875\n",
      "loss = 262.52508544921875\n",
      "loss = 173.59042358398438\n",
      "loss = 184.90916442871094\n",
      "loss = 171.11122131347656\n",
      "loss = 286.8625183105469\n",
      "loss = 211.34471130371094\n",
      "loss = 229.96975708007812\n",
      "loss = 234.60435485839844\n",
      "loss = 189.4452362060547\n",
      "loss = 294.990966796875\n",
      "loss = 126.84382629394531\n",
      "loss = 262.28924560546875\n",
      "loss = 183.7225341796875\n",
      "loss = 186.53762817382812\n",
      "loss = 187.1005096435547\n",
      "loss = 208.9612579345703\n",
      "loss = 322.6086730957031\n",
      "loss = 156.760986328125\n",
      "loss = 214.48890686035156\n",
      "loss = 145.59280395507812\n",
      "loss = 245.04541015625\n",
      "loss = 208.7057342529297\n",
      "loss = 224.66781616210938\n",
      "loss = 96.41705322265625\n",
      "loss = 147.56381225585938\n",
      "loss = 254.0188751220703\n",
      "loss = 171.557373046875\n",
      "loss = 132.22706604003906\n",
      "loss = 185.43853759765625\n",
      "loss = 183.58106994628906\n",
      "loss = 216.1398162841797\n",
      "loss = 161.1305694580078\n",
      "loss = 133.38662719726562\n",
      "loss = 283.0481262207031\n",
      "loss = 99.33334350585938\n",
      "loss = 227.4369354248047\n",
      "loss = 127.71469116210938\n",
      "loss = 170.87420654296875\n",
      "loss = 198.20904541015625\n",
      "loss = 224.55697631835938\n",
      "loss = 374.5802001953125\n",
      "loss = 206.80526733398438\n",
      "loss = 189.4744873046875\n",
      "loss = 105.71331787109375\n",
      "loss = 109.45527648925781\n",
      "loss = 354.272216796875\n",
      "loss = 145.43661499023438\n",
      "loss = 284.55682373046875\n",
      "loss = 116.63922119140625\n",
      "loss = 83.27970886230469\n",
      "loss = 145.10838317871094\n",
      "loss = 156.5350799560547\n",
      "loss = 210.36663818359375\n",
      "loss = 234.3445587158203\n",
      "loss = 283.45703125\n",
      "loss = 118.87069702148438\n",
      "loss = 267.14300537109375\n",
      "loss = 211.31317138671875\n",
      "loss = 162.1364288330078\n",
      "loss = 212.67205810546875\n",
      "loss = 205.37110900878906\n",
      "loss = 127.29034423828125\n",
      "loss = 141.97073364257812\n",
      "loss = 235.48110961914062\n",
      "loss = 233.189208984375\n",
      "loss = 221.9424285888672\n",
      "loss = 186.92355346679688\n",
      "loss = 131.19680786132812\n",
      "loss = 140.46243286132812\n",
      "loss = 198.34942626953125\n",
      "loss = 226.26956176757812\n",
      "loss = 256.4381103515625\n",
      "loss = 163.52392578125\n",
      "loss = 244.69009399414062\n",
      "loss = 224.30422973632812\n",
      "loss = 148.93508911132812\n",
      "loss = 238.69837951660156\n",
      "loss = 154.58905029296875\n",
      "loss = 176.50233459472656\n",
      "loss = 148.3822784423828\n",
      "loss = 163.21983337402344\n",
      "loss = 297.8008728027344\n",
      "loss = 187.34152221679688\n",
      "loss = 203.9676055908203\n",
      "loss = 161.7999267578125\n",
      "loss = 242.646728515625\n",
      "loss = 113.10678100585938\n",
      "loss = 122.26199340820312\n",
      "loss = 232.78668212890625\n",
      "loss = 132.8133087158203\n",
      "loss = 184.355712890625\n",
      "loss = 183.63787841796875\n",
      "loss = 184.0917510986328\n",
      "loss = 224.4243621826172\n",
      "loss = 177.78482055664062\n",
      "loss = 153.53504943847656\n",
      "loss = 174.97369384765625\n",
      "loss = 143.56198120117188\n",
      "loss = 227.80218505859375\n",
      "loss = 156.7858123779297\n",
      "loss = 115.17138671875\n",
      "loss = 124.29376220703125\n",
      "loss = 209.32347106933594\n",
      "loss = 215.2971649169922\n",
      "loss = 190.94772338867188\n",
      "loss = 127.52275085449219\n",
      "loss = 211.37747192382812\n",
      "loss = 150.2022247314453\n",
      "loss = 168.977783203125\n",
      "loss = 131.21060180664062\n",
      "loss = 197.7289276123047\n",
      "loss = 186.0218505859375\n",
      "loss = 237.46502685546875\n",
      "loss = 191.53244018554688\n",
      "loss = 141.9832305908203\n",
      "loss = 146.6922607421875\n",
      "loss = 152.15402221679688\n",
      "loss = 194.37930297851562\n",
      "loss = 138.5177001953125\n",
      "loss = 148.36480712890625\n",
      "loss = 131.1165771484375\n",
      "loss = 134.26666259765625\n",
      "loss = 220.05340576171875\n",
      "loss = 186.10531616210938\n",
      "loss = 203.94863891601562\n",
      "loss = 200.5462646484375\n",
      "loss = 157.0069122314453\n",
      "loss = 189.24462890625\n",
      "loss = 128.66419982910156\n",
      "loss = 137.57943725585938\n",
      "loss = 214.70811462402344\n",
      "loss = 207.31353759765625\n",
      "loss = 144.49752807617188\n",
      "loss = 91.56292724609375\n",
      "loss = 238.60403442382812\n",
      "loss = 159.0371856689453\n",
      "loss = 177.1193084716797\n",
      "loss = 102.28450012207031\n",
      "loss = 223.93292236328125\n",
      "loss = 244.38575744628906\n",
      "loss = 165.8145751953125\n",
      "loss = 142.34716796875\n",
      "loss = 175.3502960205078\n",
      "loss = 174.56060791015625\n",
      "loss = 128.01019287109375\n",
      "loss = 188.33055114746094\n",
      "loss = 195.0431671142578\n",
      "loss = 142.901611328125\n",
      "loss = 22.92333984375\n"
     ]
    }
   ],
   "source": [
    "# BiLSTM_CRF：\n",
    "# BiLSTM （word embedding或者bert embedding） + linear mapping（embedding_dim -> n_features or num_tags） + CRF (loss)\n",
    "\n",
    "# 导入所需文件 spacy\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "from torch.utils.data import TensorDataset, DataLoader\n",
    "from TorchCRF import CRF\n",
    "from torch.optim import Adam\n",
    "import warnings\n",
    "import numpy as np\n",
    "from sklearn.model_selection import ShuffleSplit\n",
    "from gensim.models import Word2Vec\n",
    "from data.ruijinkg.data_utils.data_utils import ENTITIES, Documents, Dataset, SentenceExtractor, make_predictions\n",
    "from data.ruijinkg.data_utils.evaluator import Evaluator\n",
    "warnings.filterwarnings(\"ignore\")\n",
    "\n",
    "class BiLSTM_CRF(nn.Module):\n",
    "    def __init__(self, num_cates, vocab_size, tag_to_ix, word_embeddings, embedding_dim, hidden_dim):\n",
    "        super(BiLSTM_CRF, self).__init__()\n",
    "        self.embedding_dim = embedding_dim\n",
    "        self.hidden_dim = hidden_dim\n",
    "        self.vocab_size = vocab_size\n",
    "        self.num_cates = num_cates\n",
    "        self.tag_to_ix = tag_to_ix\n",
    "        self.crf = CRF(num_tags = num_cates, batch_first = True) # 很重要 batch_first = True\n",
    "        self.word_embeds = nn.Embedding.from_pretrained(torch.from_numpy(word_embeddings))\n",
    "        # 简便折断\n",
    "        self.bilstm = nn.LSTM(embedding_dim, hidden_dim // 2, num_layers=1, bidirectional=True, batch_first=True)\n",
    "        self.hidden2tag = nn.Linear(hidden_dim, num_cates)\n",
    "    def __forward__(self, batch_token_ids): # [batch_size, seq_size]\n",
    "        shape_ = batch_token_ids.shape\n",
    "        # print(shape_, self.embedding_dim, self.word_embeds(batch_token_ids).shape)\n",
    "        embeds = self.word_embeds(batch_token_ids).view(shape_[0], shape_[1], self.embedding_dim)\n",
    "        bilstm_out, _ = self.bilstm(embeds)\n",
    "        bilstm_out = bilstm_out.view(shape_[0], shape_[1], self.hidden_dim)\n",
    "        bilstm_feats = self.hidden2tag(bilstm_out)\n",
    "        # print(bilstm_feats.shape) # orch.Size([10, 19, 4])\n",
    "        return bilstm_feats\n",
    "    def loss(self, batch_token_ids, tags, masks):\n",
    "        feats = self.__forward__(batch_token_ids)\n",
    "        # print(feats.shape, tags.shape, masks.shape)\n",
    "        return - self.crf(feats, tags, masks)\n",
    "    def predict(self, batch_token_ids):\n",
    "        feats = self.__forward__(batch_token_ids)\n",
    "        return self.crf.decode(feats)\n",
    "\n",
    "# 数据文件读取\n",
    "data_dir = \"./data/ruijinkg/data/train\"\n",
    "ent2idx = dict(zip(ENTITIES, range(1, len(ENTITIES) + 1)))\n",
    "idx2ent = dict([(v, k) for k, v in ent2idx.items()])\n",
    "\n",
    "# 训练集，测试集切分与打乱\n",
    "docs = Documents(data_dir=data_dir)\n",
    "rs = ShuffleSplit(n_splits=1, test_size=20, random_state=2018)\n",
    "train_doc_ids, test_doc_ids = next(rs.split(docs))\n",
    "train_docs, test_docs = docs[train_doc_ids], docs[test_doc_ids]\n",
    "# 模型参数赋值\n",
    "num_cates = max(ent2idx.values()) + 1\n",
    "sent_len = 64\n",
    "vocab_size = 3000\n",
    "emb_size = 100\n",
    "sent_pad = 10\n",
    "sent_extrator = SentenceExtractor(window_size=sent_len, pad_size=sent_pad)\n",
    "train_sents = sent_extrator(train_docs)\n",
    "test_sents = sent_extrator(test_docs)\n",
    "train_data = Dataset(train_sents, cate2idx=ent2idx)\n",
    "train_data.build_vocab_dict(vocab_size=vocab_size)\n",
    "test_data = Dataset(test_sents, word2idx=train_data.word2idx, cate2idx=ent2idx)\n",
    "vocab_size = len(train_data.word2idx)\n",
    "# 词嵌入模型构建\n",
    "w2v_train_sents = []\n",
    "for doc in docs:\n",
    "    w2v_train_sents.append(list(doc.text))\n",
    "w2v_model = Word2Vec(w2v_train_sents, vector_size=emb_size)\n",
    "w2v_embeddings = np.zeros((vocab_size, emb_size))\n",
    "for char, char_idx in train_data.word2idx.items():\n",
    "    if char in w2v_model.wv:\n",
    "        w2v_embeddings[char_idx] = w2v_model.wv[char]\n",
    "\n",
    "# CRF条件随机场实例化\n",
    "print(w2v_embeddings.shape)\n",
    "model = BiLSTM_CRF(vocab_size = vocab_size, num_cates = num_cates, tag_to_ix = ent2idx, word_embeddings = w2v_embeddings.astype(np.float32), embedding_dim = emb_size, hidden_dim = 64)\n",
    "\n",
    "# 训练集，测试集形状\n",
    "batch_size = 10\n",
    "dataloader = DataLoader(train_data, shuffle=True, batch_size = batch_size)\n",
    "# 模型定义\n",
    "optimizer = Adam(model.parameters(), lr = 0.05)\n",
    "# 模型训练\n",
    "for seq, seq_tag in dataloader:\n",
    "    optimizer.zero_grad()\n",
    "    seq_mask = torch.ones_like(seq)\n",
    "    seq, seq_tag, seq_mask = seq.squeeze().long(), seq_tag.squeeze().long(), seq_mask.squeeze().bool()\n",
    "    # print(seq.shape, seq_tag.shape, seq_mask.shape)\n",
    "    loss = model.loss(seq, seq_tag, seq_mask)\n",
    "    loss.backward()\n",
    "    optimizer.step()\n",
    "    print('loss =', loss.item())\n",
    "\n"
   ],
   "metadata": {
    "collapsed": false,
    "ExecuteTime": {
     "end_time": "2023-08-27T22:47:04.791093Z",
     "start_time": "2023-08-27T22:46:40.855947Z"
    }
   },
   "id": "710c9689e7f5751"
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<class 'fastNLP.io.data_bundle.DataBundle'> <class 'fastNLP.core.vocabulary.Vocabulary'>\n",
      "6995 11191 10\n",
      "-----Begin Testing-----\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 10/10 [06:37<00:00, 39.76s/it]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "correct_num:   0, predict_num: 38047, gold_num:  22\n",
      "f1: 0.00, precision: 0.00, recall: 0.00\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    },
    {
     "data": {
      "text/plain": "(0.0, 0.0, 0.0)"
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Bert + CasRel （已测试）\n",
    "\n",
    "# bert cls位置enbedding -> mlp（1层）->预测头尾实体和关系\n",
    "# p(head,tail,relation | x) = p(head | x) * p(tail | head, x) * p(relation | head,tail,x)\n",
    "# p(head), p(tail), p(head-tail: relation)\n",
    "\n",
    "# 4个loss： subject_type_loss, object_type_loss, subject_relation_loss, object_relation_loss:\n",
    "# # 根据就近原则，关系矩阵可以拆分两个向量\n",
    "# sub_heads, sub_tails = torch.zeros(text_len), torch.zeros(text_len)\n",
    "# sub_head, sub_tail = torch.zeros(text_len), torch.zeros(text_len)\n",
    "# obj_heads = torch.zeros((text_len, num_rel))\n",
    "# obj_tails = torch.zeros((text_len, num_rel))\n",
    "\n",
    "# entity embedding = ave (all token embeddings)\n",
    "\n",
    "# utf-8\n",
    "\n",
    "import torch, os\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "from transformers import BertModel\n",
    "import json\n",
    "from random import choice\n",
    "from fastNLP import Vocabulary, Callback\n",
    "from torch.utils.data import DataLoader, Dataset\n",
    "from fastNLP.io import JsonLoader\n",
    "import numpy as np\n",
    "from transformers import BertTokenizer\n",
    "from collections import defaultdict\n",
    "from torch.nn.utils.rnn import pad_sequence\n",
    "from tqdm import tqdm\n",
    "\n",
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "num_rel = 18\n",
    "bert_dim = 768\n",
    "max_seq_len = 300\n",
    "train_path = 'data/baidu/train.json'\n",
    "test_path = 'data/baidu/test_.json'\n",
    "dev_path = 'data/baidu/dev.json'\n",
    "rel_path = 'data/baidu/rel.json'\n",
    "save_weights_fpath = '/models/baidu/model.pt'\n",
    "save_logs_fpath = '/models/baidu/model.out'\n",
    "result_fpath = '/models/baidu/result.json'\n",
    "bert_dirpath = './models/roberta-base-finetuned-dianping-chinese'\n",
    "batch_size = 8\n",
    "lr = 1e-5\n",
    "max_epoch = 10\n",
    "\n",
    "class CasRel(nn.Module):\n",
    "    def __init__(self, bert_dirpath, bert_dim, num_relations):\n",
    "        super(CasRel, self).__init__()\n",
    "        self.bert = BertModel.from_pretrained(bert_dirpath)\n",
    "        self.sub_heads_linear = nn.Linear(bert_dim, 1)\n",
    "        self.sub_tails_linear = nn.Linear(bert_dim, 1)\n",
    "        self.obj_heads_linear = nn.Linear(bert_dim, num_relations)\n",
    "        self.obj_tails_linear = nn.Linear(bert_dim, num_relations)\n",
    "    def __get_encoded_text__(self, token_ids, mask):\n",
    "        # print(token_ids.shape, mask.shape) # torch.Size([1, 20]) torch.Size([1, 20])\n",
    "        # print(self.bert(token_ids, attention_mask=mask)[0].shape) # torch.Size([1, 20, 768])\n",
    "        return self.bert(token_ids, attention_mask=mask)[0] # 获取句子中所有word的embedding\n",
    "    def __get_subs__(self, encoded_text):\n",
    "        pred_sub_heads = torch.sigmoid(self.sub_heads_linear(encoded_text))\n",
    "        pred_sub_tails = torch.sigmoid(self.sub_tails_linear(encoded_text))\n",
    "        return pred_sub_heads, pred_sub_tails\n",
    "    def __get_objs_for_specific_sub__(self, sub_head_mapping, sub_tail_mapping, encoded_text): # 1\n",
    "        # sub_head_mapping [batch, 1, seq] * encoded_text [batch, seq, dim]\n",
    "        sub_head = torch.matmul(sub_head_mapping, encoded_text)\n",
    "        sub_tail = torch.matmul(sub_tail_mapping, encoded_text)\n",
    "        sub = (sub_head + sub_tail) / 2\n",
    "        encoded_text = encoded_text + sub\n",
    "        pred_obj_heads = torch.sigmoid(self.obj_heads_linear(encoded_text))\n",
    "        pred_obj_tails = torch.sigmoid(self.obj_tails_linear(encoded_text))\n",
    "        return pred_obj_heads, pred_obj_tails\n",
    "    def forward(self, token_ids, mask, sub_head, sub_tail): # 2\n",
    "        # 1.获取序列中所有茨嵌入\n",
    "        encoded_text = self.__get_encoded_text__(token_ids, mask) # torch.Size([1, 20, 768])\n",
    "        # 2.预测头和尾实体\n",
    "        pred_sub_heads, pred_sub_tails = self.__get_subs__(encoded_text) # torch.Size([1, 20, 1]) torch.Size([1, 20, 1])\n",
    "        sub_head_mapping = sub_head.unsqueeze(1)\n",
    "        sub_tail_mapping = sub_tail.unsqueeze(1)\n",
    "        # 3.交叉，根据头实体预测尾实体\n",
    "        pred_obj_heads, pre_obj_tails = self.__get_objs_for_specific_sub__(sub_head_mapping, sub_tail_mapping, encoded_text)\n",
    "        return {\"sub_heads\": pred_sub_heads, \"sub_tails\": pred_sub_tails, \"obj_heads\": pred_obj_heads, \"obj_tails\": pre_obj_tails,}\n",
    "\n",
    "def load_data(train_path, dev_path, test_path, rel_dict_path):\n",
    "    paths = {'train': train_path, 'dev': dev_path, 'test': test_path}\n",
    "    loader = JsonLoader({\"text\": \"text\", \"spo_list\": \"spo_list\"})\n",
    "    data_bundle = loader.load(paths)\n",
    "    id2rel = json.load(open(rel_dict_path, encoding='utf-8'))\n",
    "    rel_vocab = Vocabulary(unknown=None, padding=None)\n",
    "    rel_vocab.add_word_lst(list(id2rel.values()))\n",
    "    return data_bundle, rel_vocab\n",
    "\n",
    "data_bundle, rel_vocab = load_data(train_path, dev_path, test_path, rel_path)\n",
    "print(type(data_bundle), type(rel_vocab))\n",
    "\n",
    "def find_head_idx(source, target):\n",
    "    target_len = len(target)\n",
    "    for i in range(len(source)):\n",
    "        if source[i: i + target_len] == target:\n",
    "            return i\n",
    "    return -1\n",
    "class MyDataset(Dataset):\n",
    "    def __init__(self, bert_dirpath, dataset, rel_vocab, max_seq_len, is_test):\n",
    "        super().__init__()\n",
    "        self.dataset = dataset\n",
    "        self.rel_vocab = rel_vocab\n",
    "        self.is_test = is_test\n",
    "        self.max_seq_len = max_seq_len\n",
    "        self.tokenizer = BertTokenizer.from_pretrained(bert_dirpath)\n",
    "    def __getitem__(self, item):\n",
    "        json_data = self.dataset[item]\n",
    "        text = json_data['text']\n",
    "        tokenized = self.tokenizer(text, max_length= self.max_seq_len, truncation=True)\n",
    "        tokens = tokenized['input_ids']\n",
    "        masks = tokenized['attention_mask']\n",
    "        text_len = len(tokens)\n",
    "        token_ids = torch.tensor(tokens, dtype=torch.long)\n",
    "        masks = torch.tensor(masks, dtype=torch.bool)\n",
    "        sub_heads, sub_tails = torch.zeros(text_len), torch.zeros(text_len)\n",
    "        sub_head, sub_tail = torch.zeros(text_len), torch.zeros(text_len)\n",
    "        # 根据就近原则，关系矩阵可以拆分两个向量\n",
    "        obj_heads = torch.zeros((text_len, num_rel))\n",
    "        obj_tails = torch.zeros((text_len, num_rel))\n",
    "        if not self.is_test:\n",
    "            s2ro_map = defaultdict(list)\n",
    "            for spo in json_data['spo_list']:\n",
    "                triple = (self.tokenizer(spo['subject'], add_special_tokens=False)['input_ids'], self.rel_vocab.to_index(spo['predicate']), self.tokenizer(spo['object'], add_special_tokens=False)['input_ids'])\n",
    "                sub_head_idx = find_head_idx(tokens, triple[0])\n",
    "                obj_head_idx = find_head_idx(tokens, triple[2])\n",
    "                if sub_head_idx != -1 and obj_head_idx != -1:\n",
    "                    sub = (sub_head_idx, sub_head_idx + len(triple[0]) - 1)\n",
    "                    s2ro_map[sub].append( (obj_head_idx, obj_head_idx + len(triple[2]) - 1, triple[1]))\n",
    "            if s2ro_map:\n",
    "                for s in s2ro_map:\n",
    "                    sub_heads[s[0]] = 1\n",
    "                    sub_tails[s[1]] = 1\n",
    "                sub_head_idx, sub_tail_idx = choice(list(s2ro_map.keys()))\n",
    "                sub_head[sub_head_idx] = 1\n",
    "                sub_tail[sub_tail_idx] = 1\n",
    "                for ro in s2ro_map.get((sub_head_idx, sub_tail_idx), []):\n",
    "                    obj_heads[ro[0]][ro[2]] = 1\n",
    "                    obj_tails[ro[1]][ro[2]] = 1\n",
    "        return token_ids, masks, sub_heads, sub_tails, sub_head, sub_tail, obj_heads, obj_tails, json_data['spo_list']\n",
    "    def __len__(self):\n",
    "        return len(self.dataset)\n",
    "def my_collate_fn(batch):\n",
    "    batch = list(filter(lambda x: x is not None, batch))\n",
    "    token_ids, masks, sub_heads, sub_tails, sub_head, sub_tail, obj_heads, obj_tails, triples = zip(*batch)\n",
    "    batch_token_ids = pad_sequence(token_ids, batch_first=True)\n",
    "    batch_masks = pad_sequence(masks, batch_first=True)\n",
    "    batch_sub_heads = pad_sequence(sub_heads, batch_first=True)\n",
    "    batch_sub_tails = pad_sequence(sub_tails, batch_first=True)\n",
    "    batch_sub_head = pad_sequence(sub_head, batch_first=True)\n",
    "    batch_sub_tail = pad_sequence(sub_tail, batch_first=True)\n",
    "    batch_obj_heads = pad_sequence(obj_heads, batch_first=True)\n",
    "    batch_obj_tails = pad_sequence(obj_tails, batch_first=True)\n",
    "    return {\"token_ids\": batch_token_ids.to(device), \"mask\": batch_masks.to(device), \"sub_head\": batch_sub_head.to(device),\n",
    "            \"sub_tail\": batch_sub_tail.to(device), \"sub_heads\": batch_sub_heads.to(device), }, \\\n",
    "           {\"mask\": batch_masks.to(device), \"sub_heads\": batch_sub_heads.to(device), \"sub_tails\": batch_sub_tails.to(device),\n",
    "            \"obj_heads\": batch_obj_heads.to(device), \"obj_tails\": batch_obj_tails.to(device), \"triples\": triples}\n",
    "def get_data_iterator(bert_dirpath, dataset, rel_vocab, max_seq_len, batch_size, is_test=False, collate_fn=my_collate_fn):\n",
    "    dataset = MyDataset(bert_dirpath, dataset, rel_vocab, max_seq_len, is_test)\n",
    "    return DataLoader(dataset=dataset, collate_fn=collate_fn, batch_size=batch_size if not is_test else 1)\n",
    "\n",
    "train_data = get_data_iterator(bert_dirpath, data_bundle.get_dataset('train'), rel_vocab, max_seq_len, batch_size, False, my_collate_fn)\n",
    "dev_data = get_data_iterator(bert_dirpath, data_bundle.get_dataset('dev'), rel_vocab, max_seq_len, batch_size, True, my_collate_fn)\n",
    "test_data = get_data_iterator(bert_dirpath, data_bundle.get_dataset('test'), rel_vocab, max_seq_len, batch_size, True, my_collate_fn)\n",
    "print(len(train_data), len(dev_data), len(test_data))\n",
    "\n",
    "model = CasRel(bert_dirpath, bert_dim, num_rel)\n",
    "optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr= lr)\n",
    "\n",
    "def to_tuple(triple_list):\n",
    "    ret = []\n",
    "    for triple in triple_list:\n",
    "        ret.append((triple['subject'], triple['predicate'], triple['object']))\n",
    "    return ret\n",
    "def metric(data_iter, rel_vocab, bert_dirpath, model, result_fpath, output=True, h_bar=0.5, t_bar=0.5):\n",
    "    orders = ['subject', 'relation', 'object']\n",
    "    correct_num, predict_num, gold_num = 0, 0, 0\n",
    "    tokenizer = BertTokenizer.from_pretrained(bert_dirpath)\n",
    "    for batch_x, batch_y in tqdm(data_iter):\n",
    "        with torch.no_grad():\n",
    "            token_ids = batch_x['token_ids']\n",
    "            mask = batch_x['mask']\n",
    "            encoded_text = model.__get_encoded_text__(token_ids, mask)\n",
    "            pred_sub_heads, pred_sub_tails = model.__get_subs__(encoded_text)\n",
    "            sub_heads = torch.where(pred_sub_heads[0] > h_bar)[0]\n",
    "            sub_tails = torch.where(pred_sub_tails[0] > t_bar)[0]\n",
    "            subjects = []\n",
    "            for sub_head in sub_heads:\n",
    "                sub_tail = sub_tails[sub_tails >= sub_head]\n",
    "                if len(sub_tail) > 0:\n",
    "                    sub_tail = sub_tail[0]\n",
    "                    subject = ''.join(tokenizer.decode(token_ids[0][sub_head: sub_tail + 1]).split())\n",
    "                    subjects.append((subject, sub_head, sub_tail))\n",
    "            if subjects:\n",
    "                triple_list = []\n",
    "                repeated_encoded_text = encoded_text.repeat(len(subjects), 1, 1)\n",
    "                sub_head_mapping = torch.zeros((len(subjects), 1, encoded_text.size(1)), dtype=torch.float, device=device)\n",
    "                sub_tail_mapping = torch.zeros((len(subjects), 1, encoded_text.size(1)), dtype=torch.float,  device=device)\n",
    "                for subject_idx, subject in enumerate(subjects):\n",
    "                    sub_head_mapping[subject_idx][0][subject[1]] = 1\n",
    "                    sub_tail_mapping[subject_idx][0][subject[2]] = 1\n",
    "                pred_obj_heads, pred_obj_tails = model.__get_objs_for_specific_sub__(sub_head_mapping, sub_tail_mapping, repeated_encoded_text)\n",
    "                for subject_idx, subject in enumerate(subjects):\n",
    "                    sub = subject[0]\n",
    "                    obj_heads = torch.where(pred_obj_heads[subject_idx] > h_bar)\n",
    "                    obj_tails = torch.where(pred_obj_tails[subject_idx] > t_bar)\n",
    "                    for obj_head, rel_head in zip(*obj_heads):\n",
    "                        for obj_tail, rel_tail in zip(*obj_tails):\n",
    "                            if obj_head <= obj_tail and rel_head == rel_tail:\n",
    "                                rel = rel_vocab.to_word(int(rel_head))\n",
    "                                obj = ''.join(tokenizer.decode(token_ids[0][obj_head: obj_tail + 1]).split())\n",
    "                                triple_list.append((sub, rel, obj))\n",
    "                                break\n",
    "                triple_set = set()\n",
    "                for s, r, o in triple_list:\n",
    "                    triple_set.add((s, r, o))\n",
    "                pred_list = list(triple_set)\n",
    "            else:\n",
    "                pred_list = []\n",
    "            pred_triples = set(pred_list)\n",
    "            gold_triples = set(to_tuple(batch_y['triples'][0]))\n",
    "            correct_num += len(pred_triples & gold_triples)\n",
    "            predict_num += len(pred_triples)\n",
    "            gold_num += len(gold_triples)\n",
    "            if output:\n",
    "                result_dir = '/'.join(result_fpath.split('/')[:-1])\n",
    "                if not os.path.exists(result_dir):\n",
    "                    os.mkdir(result_dir)\n",
    "                fw = open(result_fpath, 'w')\n",
    "                result = json.dumps({ 'triple_list_gold': [dict(zip(orders, triple)) for triple in gold_triples],\n",
    "                    'triple_list_pred': [dict(zip(orders, triple)) for triple in pred_triples],\n",
    "                    'new': [dict(zip(orders, triple)) for triple in pred_triples - gold_triples],\n",
    "                    'lack': [dict(zip(orders, triple)) for triple in gold_triples - pred_triples]}, ensure_ascii=False)\n",
    "                fw.write(result + '\\n')\n",
    "    print(\"correct_num: {:3d}, predict_num: {:3d}, gold_num: {:3d}\".format(correct_num, predict_num, gold_num))\n",
    "    precision = correct_num / (predict_num + 1e-10)\n",
    "    recall = correct_num / (gold_num + 1e-10)\n",
    "    f1_score = 2 * precision * recall / (precision + recall + 1e-10)\n",
    "    print('f1: {:4.2f}, precision: {:4.2f}, recall: {:4.2f}'.format(f1_score, precision, recall))\n",
    "    return precision, recall, f1_score\n",
    "class MyCallBack(Callback):\n",
    "    def __init__(self, data_iter, rel_vocab, save_logs_fpath, bert_dirpath, result_fpath, weight_fpath):\n",
    "        super().__init__()\n",
    "        self.best_epoch = 0\n",
    "        self.best_recall = 0\n",
    "        self.best_precision = 0\n",
    "        self.best_f1_score = 0\n",
    "        self.data_iter = data_iter\n",
    "        self.rel_vocab = rel_vocab\n",
    "        self.save_logs_fpath = save_logs_fpath\n",
    "        self.bert_dirpath = bert_dirpath\n",
    "        self.result_fpath = result_fpath\n",
    "        self.weight_fpath = weight_fpath\n",
    "    def logging(self, s, print_=True, log_=False):\n",
    "        if print_:\n",
    "            print(s)\n",
    "        if log_:\n",
    "            with open(save_logs_fpath, 'a+') as f_log:\n",
    "                f_log.write(s + '\\n')\n",
    "    def on_train_begin(self):\n",
    "        self.logging(\"-\" * 5 + \"Begin Training\" + \"-\" * 5)\n",
    "    def on_epoch_end(self):\n",
    "        precision, recall, f1_score = metric(self.data_iter, self.rel_vocab, self.bert_dirpath, self.model, self.result_fpath)\n",
    "        self.logging('epoch {:3d}, f1: {:4.2f}, precision: {:4.2f}, recall: {:4.2f}'.format(self.epoch, f1_score, precision, recall))\n",
    "        if f1_score > self.best_f1_score:\n",
    "            self.best_f1_score = f1_score\n",
    "            self.best_epoch = self.epoch\n",
    "            self.best_precision = precision\n",
    "            self.best_recall = recall\n",
    "            self.logging(\"Saving the model, epoch: {:3d}, best f1: {:4.2f}, precision: {:4.2f}, recall: {:4.2f}\".format(self.best_epoch, self.best_f1_score, precision, recall))\n",
    "            torch.save(self.model.state_dict(), self.weight_fpath)\n",
    "    def on_train_end(self):\n",
    "        self.logging(\"-\" * 5 + \"Finish training\" + \"-\" * 5)\n",
    "        self.logging(\"best epoch: {:3d}, best f1: {:4.2f}, precision: {:4.2f}, recall: {:4.2}\".format(self.best_epoch, self.best_f1_score, self.best_precision, self.best_recall))\n",
    "from fastNLP import Trainer\n",
    "class MyLoss:\n",
    "    def __init__(self):\n",
    "        super(MyLoss, self).__init__()\n",
    "    def get_loss(self, predict, target):\n",
    "        mask = target['mask']\n",
    "        def loss_fn(pred, gold, mask):\n",
    "            pred = pred.squeeze(-1)\n",
    "            loss = torch.nn.BCELoss(pred, gold, reduction='none')\n",
    "            if loss.shape != mask.shape:\n",
    "                mask = mask.unsqueeze(-1)\n",
    "            loss = torch.sum(loss * mask) / torch.sum(mask)\n",
    "            return loss\n",
    "        return loss_fn(predict['sub_heads'], target['sub_heads'], mask) + \\\n",
    "               loss_fn(predict['sub_tails'], target['sub_tails'], mask) + \\\n",
    "               loss_fn(predict['obj_heads'], target['obj_heads'], mask) + \\\n",
    "               loss_fn(predict['obj_tails'], target['obj_tails'], mask)\n",
    "    def __call__(self, pred_dict, target_dict, check=False):\n",
    "        loss = self.get_loss(pred_dict, target_dict)\n",
    "        return loss\n",
    "\n",
    "period = 200\n",
    "trainer = Trainer(train_dataloader=train_data, model=model, optimizers=[optimizer], batch_size=batch_size,\n",
    "                  n_epochs=max_epoch, loss=MyLoss(), print_every=period, use_tqdm=True,\n",
    "                  callbacks=MyCallBack(dev_data, rel_vocab, save_logs_fpath, bert_dirpath, result_fpath, save_weights_fpath))\n",
    "# trainer.run(1)\n",
    "trainer.step()\n",
    "print(\"-\" * 5 + \"Begin Testing\" + \"-\" * 5)\n",
    "metric(test_data, rel_vocab, bert_dirpath, model, result_fpath, output=False)\n"
   ],
   "metadata": {
    "collapsed": false,
    "ExecuteTime": {
     "end_time": "2023-08-28T03:53:28.422459800Z",
     "start_time": "2023-08-28T03:46:48.423838800Z"
    }
   },
   "id": "cb4651f40eb32cfb"
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[('人物', '木斧', '文本', '回族', '民族'), ('人物', '木斧', '日期', '1931', '出生日期'), ('人物', '木斧', '地点', '成都', '出生地'), ('歌曲', '今晚会在哪里醒来', '人物', '黄家强', '作曲'), ('歌曲', '今晚会在哪里醒来', '音乐专辑', '她他', '所属专辑'), ('歌曲', '今晚会在哪里醒来', '人物', '黄家强', '歌手'), ('歌曲', '今晚会在哪里醒来', '人物', '何启弘', '作词'), ('企业', '华夏建通科技开发股份有限公司', '日期', '2004年6月3日', '成立日期'), ('人物', '潘惟南', '地点', '陕西大荔', '出生地'), ('人物', '郭静唐', '地点', '周巷', '出生地')]\n",
      "查找所有节点: [Node('人物', name='木斧'), Node('日期', name='1931')]\n",
      "匹配指定条件的节点: [Node('人物', name='木斧')] (_26:人物 {name: '\\u6728\\u65a7'})\n",
      "修改节点: [Node('人物', name='fubin'), Node('日期', name='1931')]\n",
      "新增关系: [Node('人物', name='fubin'), Node('日期', name='1931')] [诞生于(Node('人物', name='fubin'), Node('日期', name='1931')), 出生日期(Node('人物', name='fubin'), Node('日期', name='1931'))]\n",
      "删除关系（不删除节点）: (fubin)-[:诞生于 {}]->(1931) -> [出生日期(Node('人物', name='fubin'), Node('日期', name='1931'))]\n"
     ]
    }
   ],
   "source": [
    "# py2neo 入库操作\n",
    "# 还可以执行cypher命令,如搜索二阶邻居节点关系等\n",
    "\n",
    "# utf-8\n",
    "\n",
    "# 1.安装jdk和配置环境变量\n",
    "# 2.配置neo4j和配置环境变量\n",
    "\n",
    "import json\n",
    "\n",
    "train_path = 'data/baidu/train.json'\n",
    "head_tail_relation = []\n",
    "with open(train_path, 'r', encoding='utf-8') as f:\n",
    "    lines = f.read().strip().split('\\n')\n",
    "    for line in lines:\n",
    "        for t in json.loads(line)['spo_list']:\n",
    "            head_type, tail_type = t['subject_type'], t['object_type']\n",
    "            head, tail, relation = t['subject'], t['object'], t['predicate']\n",
    "            head_tail_relation.append((head_type, head, tail_type, tail, relation))\n",
    "print(head_tail_relation[:10])\n",
    "\n",
    "# python操作neo4j\n",
    "# pip install py2neo\n",
    "\n",
    "from py2neo import Graph, Node, Relationship, NodeMatcher, RelationshipMatcher, Subgraph\n",
    "# 建立连接\n",
    "graph = Graph('http://localhost:7474/', auth=(\"neo4j\", \"admin\"))\n",
    "# 删除所有\n",
    "graph.delete_all()\n",
    "# 定义节点\n",
    "a = Node('人物', name='木斧')\n",
    "b = Node('日期', name='1931')\n",
    "# 节点匹配\n",
    "node_matcher = NodeMatcher(graph)\n",
    "nodea = node_matcher.match('人物', name='木斧').first()\n",
    "nodeb = node_matcher.match('日期', name='1931').first()\n",
    "# 关系匹配\n",
    "relationship_matcher = RelationshipMatcher(graph)\n",
    "relations = relationship_matcher.match(nodes=[nodea, nodeb]).all()\n",
    "# 删除关系链(若关系为空则删除节点)\n",
    "if relations:\n",
    "    print(relations[0])\n",
    "    graph.delete(relations[0])\n",
    "# 创建关系和节点\n",
    "ab = Relationship(a, '出生日期', b)\n",
    "graph.create(ab)\n",
    "# 查找所有节点\n",
    "all_nodes = graph.nodes.match().all()\n",
    "print('查找所有节点:', all_nodes)\n",
    "# 匹配指定条件的节点\n",
    "nodes = node_matcher.match('人物', name='木斧').all()\n",
    "node1 = node_matcher.match('人物', name='木斧').first()\n",
    "print('匹配指定条件的节点:', nodes, node1)\n",
    "# 修改节点\n",
    "node1['name'] = 'fubin'\n",
    "subgraph = Subgraph(nodes = [node1]) # 子图对象\n",
    "graph.push(subgraph)\n",
    "print('修改节点:', graph.nodes.match().all())\n",
    "# 新增关系\n",
    "nodea = node_matcher.match('人物', name='fubin').first()\n",
    "nodeb = node_matcher.match('日期', name='1931').first()\n",
    "relationship = Relationship(nodea, '诞生于', nodeb)\n",
    "graph.create(relationship)\n",
    "print('新增关系:', graph.nodes.match().all(), graph.relationships.match().all())\n",
    "# 删除关系（不删除节点）\n",
    "relations = relationship_matcher.match(nodes=[nodea, nodeb]).all()\n",
    "graph.separate(relations[0])\n",
    "print('删除关系（不删除节点）:', relations[0], '->', graph.relationships.match().all())\n",
    "\n",
    "# 批量导入命令\n",
    "# neo4j-admin import --database kg_name --nodes=人物=人物.csv --nodes=城市=城市.csv  --nodes=日期=日期.csv  --relationships=出生地=出生地.csv  --relationships=出生日期=出生日期.csv --trim-strings=true"
   ],
   "metadata": {
    "collapsed": false,
    "ExecuteTime": {
     "end_time": "2023-08-28T13:04:32.233022Z",
     "start_time": "2023-08-28T13:04:31.341577Z"
    }
   },
   "id": "270e6993db77a1c6"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [
    "# kg表示学习推理\n",
    "# TransE\n",
    "\n",
    "import numpy as np\n",
    "import json\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "from torch.autograd import Variable\n",
    "def norm_l1(h, r, t):\n",
    "    return np.sum(np.fabs(h + r - t))\n",
    "def norm_l2(h, r, t):\n",
    "    return np.sum(np.square(h + r - t))\n",
    "class TransE(nn.Module):\n",
    "    def __init__(self, entity_num, relation_num, dim, margin, norm, C):\n",
    "        super(TransE, self).__init__()\n",
    "        self.entity_num = entity_num\n",
    "        self.relation_num = relation_num\n",
    "        self.dim = dim\n",
    "        self.margin = margin\n",
    "        self.norm = norm\n",
    "        self.C = C\n",
    "        self.ent_embedding = torch.nn.Embedding(num_embeddings=self.entity_num, embedding_dim=self.dim)\n",
    "        self.rel_embedding = torch.nn.Embedding(num_embeddings=self.relation_num, embedding_dim=self.dim)\n",
    "        nn.init.xavier_uniform_(self.ent_embedding.weight.data)\n",
    "        self.normalization_ent_embedding()\n",
    "        nn.init.xavier_uniform_(self.rel_embedding.weight.data)\n",
    "        self.normalization_rel_embedding()\n",
    "        self.loss_F = nn.MarginRankingLoss(self.margin, reduction=\"mean\")\n",
    "    def normalization_ent_embedding(self):\n",
    "        norm = self.ent_embedding.weight.detach().cpu().numpy()\n",
    "        norm = norm / np.sqrt(np.sum(np.square(norm), axis=1, keepdims=True))\n",
    "        self.ent_embedding.weight.data.copy_(torch.from_numpy(norm))\n",
    "    def normalization_rel_embedding(self):\n",
    "        norm = self.rel_embedding.weight.detach().cpu().numpy()\n",
    "        norm = norm / np.sqrt(np.sum(np.square(norm), axis=1, keepdims=True))\n",
    "        self.rel_embedding.weight.data.copy_(torch.from_numpy(norm))\n",
    "    def input_pre_transe(self, ent_vector, rel_vector):\n",
    "        for i in range(self.entity_num):\n",
    "            self.ent_embedding.weight.data[i] = torch.from_numpy(np.array(ent_vector[i]))\n",
    "        for i in range(self.relation_num):\n",
    "            self.rel_embedding.weight.data[i] = torch.from_numpy(np.array(rel_vector[i]))\n",
    "    def distance(self, h, r, t):\n",
    "        head = self.ent_embedding(h)\n",
    "        rel = self.rel_embedding(r)\n",
    "        tail = self.ent_embedding(t)\n",
    "        distance = head + rel - tail\n",
    "        # dim = -1表示的是维度的最后一维 比如如果一个张量有3维 那么 dim = 2 = -1， dim = 0 = -3\n",
    "        score = torch.norm(distance, p = self.norm, dim=1)\n",
    "        return score\n",
    "    def scale_loss(self, embedding):\n",
    "        return torch.sum(torch.max(torch.sum(embedding ** 2, dim=1, keepdim=True)-torch.autograd.Variable(torch.FloatTensor([1.0])), torch.autograd.Variable(torch.FloatTensor([0.0]))))\n",
    "    def forward(self, current_triples, corrupted_triples):\n",
    "        h, r, t = torch.chunk(current_triples, 3, dim=1)\n",
    "        h_c, r_c, t_c = torch.chunk(corrupted_triples, 3, dim=1)\n",
    "        h = torch.squeeze(h, dim=1)\n",
    "        r = torch.squeeze(r, dim=1)\n",
    "        t = torch.squeeze(t, dim=1)\n",
    "        h_c = torch.squeeze(h_c, dim=1)\n",
    "        r_c = torch.squeeze(r_c, dim=1)\n",
    "        t_c = torch.squeeze(t_c, dim=1)\n",
    "        pos = self.distance(h, r, t)\n",
    "        neg = self.distance(h_c, r_c, t_c)\n",
    "        entity_embedding = self.ent_embedding(torch.cat([h, t, h_c, t_c]).cuda())\n",
    "        relation_embedding = self.rel_embedding(torch.cat([r, r_c]).cuda())\n",
    "        y = Variable(torch.Tensor([-1])).cuda()\n",
    "        loss = self.loss_F(pos, neg, y)\n",
    "        ent_scale_loss = self.scale_loss(entity_embedding)\n",
    "        rel_scale_loss = self.scale_loss(relation_embedding)\n",
    "        return loss + self.C * (ent_scale_loss/len(entity_embedding) + rel_scale_loss/len(relation_embedding))\n"
   ],
   "metadata": {
    "collapsed": false
   },
   "id": "861147b2853b8b09"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [
    "# kg应用：知识问答\n",
    "\n",
    "# 1.基于信息抽取的知识问答（适合于简单问题，解释好）\n",
    "# 思路：实体链接 -> 2跳节点（生成候选答案） -> 构造特征训练模型或者规则\n",
    "\n",
    "# 2.基于路径推理的知识问答（适合稍微复杂一些的推理问题，解释较好）\n",
    "# 思路：利用强化学习来找答案的推理路径模型，本质类似sql执行路径\n",
    "\n",
    "# 3.基于kg嵌入的知识问答（适合复杂推理问题，解释差）\n",
    "# 思路：知识图谱嵌入表示（如transE） -> 得到查询节点的上下文信息 作为新特征加入到问答匹配模型中\n"
   ],
   "metadata": {
    "collapsed": false
   },
   "id": "82b99cd4fe431aea"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [
    "# kg应用：文本生成\n",
    "\n",
    "# 1.常识驱动的对话系统、故事生成\n",
    "# （每个实体单词）检索子图 -> 图注意力融合 -> 选择合适实体输出\n",
    "\n"
   ],
   "metadata": {
    "collapsed": false
   },
   "id": "451b8143af9edf82"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [
    "# kg应用：情感分类和情感抽取\n",
    "\n",
    "# 情感词+句子上下文+知识图谱"
   ],
   "metadata": {
    "collapsed": false
   },
   "id": "fb9c2725ebadb57b"
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "outputs": [],
   "source": [
    "# kg应用：基于知识图谱可解释的推荐系统\n"
   ],
   "metadata": {
    "collapsed": false
   },
   "id": "ad6c59c950ac63c8"
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}