AnalysisExamples/exampleSaliencyMaps.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "e0d4f2ff",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/groups/henderj/fwillett/code/speechRelease_v3/lib/python3.9/site-packages/urllib3/__init__.py:34: NotOpenSSLWarning: urllib3 v2.0 only supports OpenSSL 1.1.1+, currently the 'ssl' module is compiled with 'LibreSSL 3.2.1'. See: https://github.com/urllib3/urllib3/issues/3020\n",
      "  warnings.warn(\n"
     ]
    }
   ],
   "source": [
    "# load our dependencies\n",
    "import os\n",
    "from glob import glob\n",
    "from pathlib import Path\n",
    "os.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\"   \n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"\"\n",
    "\n",
    "import numpy as np\n",
    "from omegaconf import OmegaConf\n",
    "import tensorflow.compat.v1 as tf\n",
    "from neuralDecoder.neuralSequenceDecoder import NeuralSequenceDecoder"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2fca3b0a",
   "metadata": {},
   "source": [
    "#### Demo: saliency map analysis\n",
    "\n",
    "This notebook provides an overview of the phonetic similarity analysis from Fig. 3, using the baseline RNN and eval data provided on Dryad."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "096704dc",
   "metadata": {},
   "outputs": [],
   "source": [
    "# First set your base directory. This should contain a subfolder derived/ containing associated Dryad files\n",
    "baseDir = '/oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "cf6a72fd",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2023-08-25 11:45:35.806810: E tensorflow/stream_executor/cuda/cuda_driver.cc:271] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected\n",
      "2023-08-25 11:45:35.806852: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:169] retrieving CUDA diagnostic information for host: sh03-11n03.int\n",
      "2023-08-25 11:45:35.806859: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:176] hostname: sh03-11n03.int\n",
      "2023-08-25 11:45:35.806914: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:200] libcuda reported version is: 535.54.3\n",
      "2023-08-25 11:45:35.806946: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:204] kernel reported version is: 535.54.3\n",
      "2023-08-25 11:45:35.806951: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:310] kernel version seems to match DSO: 535.54.3\n",
      "2023-08-25 11:45:35.807162: I tensorflow/core/platform/cpu_feature_guard.cc:151] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA\n",
      "To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"gru\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " gru_1 (GRU)                 multiple                  28317696  \n",
      "                                                                 \n",
      " gru_2 (GRU)                 multiple                  6297600   \n",
      "                                                                 \n",
      " gru_3 (GRU)                 multiple                  6297600   \n",
      "                                                                 \n",
      " gru_4 (GRU)                 multiple                  6297600   \n",
      "                                                                 \n",
      " gru_5 (GRU)                 multiple                  6297600   \n",
      "                                                                 \n",
      " dense (Dense)               multiple                  42025     \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 53,551,145\n",
      "Trainable params: 53,551,145\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.04.28/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.04.28/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.05.05/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.05.05/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.05.17/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.05.17/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.05.19/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/tfRecords/t12.2022.05.19/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.05.24/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.05.24/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.05.26/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.05.26/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.02/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.02/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.07/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.07/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.14/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.14/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.16/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.16/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.21/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.21/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.28/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.06.28/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.05/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.05/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.14/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.14/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.21/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.21/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.27/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.07.27/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.08.02/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.08.02/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.08.11/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.08.11/test\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.08.13/train\n",
      "Load data from /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20//derived/tfRecords/t12.2022.08.13/test\n",
      "Model: \"sequential\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_1 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout (Dropout)           (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_1\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_2 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_1 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_2\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_3 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_2 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_3\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_4 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_3 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_4\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_5 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_4 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_5\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_6 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_5 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_6\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_7 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_6 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_7\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_8 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_7 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_8\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_9 (Dense)             (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_8 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_9\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_10 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_9 (Dropout)         (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_10\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_11 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_10 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_11\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_12 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_11 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_12\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_13 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_12 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_13\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_14 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_13 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_14\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_15 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_14 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_15\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_16 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_15 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_16\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_17 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_16 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_17\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_18 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_17 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_18\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_19 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_18 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Loading from : /oak/stanford/groups/henderj/fwillett/speechPaperRelease_08_20/derived/rnns/baselineRelease/ckpt-9950\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "                                                                 \n",
      " dropout_16 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_17\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_18 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_17 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Model: \"sequential_18\"\n",
      "_________________________________________________________________\n",
      " Layer (type)                Output Shape              Param #   \n",
      "=================================================================\n",
      " dense_19 (Dense)            (None, None, 256)         65792     \n",
      "                                                                 \n",
      " dropout_18 (Dropout)        (None, None, 256)         0         \n",
      "                                                                 \n",
      "=================================================================\n",
      "Total params: 65,792\n",
      "Trainable params: 65,792\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Loading from : /oak/stanford/groups/shenoy/fwillett/speechPaperRelease_final/derived/rnns/baselineRelease/ckpt-8750\n"
     ]
    }
   ],
   "source": [
    "# Initialize RNN model and datasets\n",
    "ckptDir = baseDir + '/derived/rnns/baselineRelease'\n",
    "\n",
    "args = OmegaConf.load(os.path.join(ckptDir, 'args.yaml'))\n",
    "args['loadDir'] = ckptDir\n",
    "args['mode'] = 'infer'\n",
    "args['loadCheckpointIdx'] = None\n",
    "\n",
    "for x in range(len(args['dataset']['datasetProbabilityVal'])):\n",
    "    args['dataset']['datasetProbabilityVal'][x] = 0.0\n",
    "\n",
    "for sessIdx in range(4,19):\n",
    "    args['dataset']['datasetProbabilityVal'][sessIdx] = 1.0\n",
    "    args['dataset']['dataDir'][sessIdx] = baseDir+'/derived/tfRecords'\n",
    "args['testDir'] = 'test'\n",
    "\n",
    "# Initialize model\n",
    "tf.compat.v1.reset_default_graph()\n",
    "nsd = NeuralSequenceDecoder(args)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b4aeb87f",
   "metadata": {},
   "source": [
    "Now we'll define the GradientGetter class - which we use for computing saliency - and a helper function."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "555bb398",
   "metadata": {},
   "outputs": [],
   "source": [
    "class GradientGetter():\n",
    "    '''Simple object for handling saliency map computation. Accepts inputs:\n",
    "    \n",
    "        rnnModel (NeuralSequenceDecoder object) - model to analyze\n",
    "        rnnInputs (2D float) - time x channels of neural data to use\n",
    "        n_perturbations (int) - # of noise perturbations, see below\n",
    "        noise_level (float)   - noise strength in [0, 1], see below\n",
    "        \n",
    "        The paper uses SmoothGrad (https://arxiv.org/pdf/1706.03825.pdf), a method for \n",
    "        cleaning up saliency maps by averaging gradients over noisy perturbations of the input \n",
    "        data. Using n_perturbations = 1 and noise_level = 0 yields the standard saliency map\n",
    "        strategy. In the speech decoding paper, we used SmoothGrad to obtain a small boost \n",
    "        in performance but for demo purposes we've turned it off by default.\n",
    "        '''\n",
    "    \n",
    "    def __init__(self, rnnModel, rnnInputs, n_perturbations = 1, noise_level = 0):\n",
    "        self.rnnModel  = rnnModel\n",
    "        self.rnnInputs = rnnInputs\n",
    "        \n",
    "        assert n_perturbations >= 1, \"N_perturbations must be >= 1\"\n",
    "        assert noise_level >= 0, \"noise_level must be >= 0\"\n",
    "        # Set n_perturbations = 1, noise_level = 0 for vanilla saliency maps \n",
    "        # Set n_perturbations = 20, noise_level = 0.1 for our results\n",
    "        \n",
    "        self.n_perturbations = n_perturbations # SmoothGrad perturbations per timepoint \n",
    "        self.noise_level     = noise_level     # noise SD as fraction of data range (0.1-0.2 in SmoothGrad paper)\n",
    "        self.print_fraction  = 0.1  # print progress \n",
    "\n",
    "    def __call__(self):\n",
    "        '''Heavy lifting done here. Step through the input neural data and calculate jacobians \n",
    "           of outputs with respect to input channels. \n",
    "           \n",
    "           Returns:\n",
    "               \n",
    "               gradients (list of 4D float) - entries are extra dim x outputs x input kernel x 256 channels\n",
    "        '''\n",
    "        \n",
    "        timelen = self.rnnInputs.shape[0]\n",
    "        delta   = self.rnnModel.model.stack_kwargs['kernel_size']\n",
    "        stride  = self.rnnModel.model.stack_kwargs['strides']\n",
    "\n",
    "        noise_sd = self.noise_level * (np.percentile(self.rnnInputs, 98) - np.percentile(self.rnnInputs, 2))\n",
    "\n",
    "        t = 0\n",
    "        gradients = list()\n",
    "        while t <= timelen - delta:\n",
    "            step_data  = self.rnnInputs[t:(t + delta), :]\n",
    "\n",
    "            noise_grads = list()\n",
    "            for i in range(self.n_perturbations):\n",
    "                noisy_data = step_data + tf.random.normal(step_data.shape, 0, noise_sd)\n",
    "                noise_grads.append(self.computeTimestepSaliency(noisy_data))\n",
    "\n",
    "            step_grads = tf.math.reduce_mean(tf.concat(noise_grads, axis = 0), axis = 0, keepdims = True)\n",
    "            gradients.append(step_grads)\n",
    "            t += delta + stride\n",
    "            \n",
    "            #if (timeline - delta) * self.print_fraction\n",
    "            print(t, '/', timelen - delta)\n",
    "        return gradients\n",
    "\n",
    "    \n",
    "    @tf.function\n",
    "    def computeTimestepSaliency(self, rnnInput):\n",
    "        '''Runs a single datapoint (32 consecutive timepoints) through model. Inputs are:\n",
    "    \n",
    "        rnnInput (2D tensor) - timepoints x channels'''\n",
    "    \n",
    "        rnnStates = [self.rnnModel.model.initStates, None, None, None, None, None]\n",
    "\n",
    "        with tf.GradientTape(persistent = True) as g:\n",
    "            #x = data\n",
    "            x = tf.tile(rnnInput[tf.newaxis, :,:], [1, 1, 1])\n",
    "            x = tf.cast(x, tf.float32)\n",
    "            g.watch(x)\n",
    "\n",
    "            for t in range(5):\n",
    "                output = self.rnnModel.normLayers[layerIdx](x)\n",
    "                output = self.rnnModel.inputLayers[layerIdx](output)\n",
    "                output, rnnStates = self.rnnModel.model(output, rnnStates, training=False, returnState=True)\n",
    "\n",
    "            logitOut = tf.squeeze(output, 0)\n",
    "            y        = tf.cast(logitOut, tf.float32)\n",
    "\n",
    "        jacobian = g.batch_jacobian(y, x, experimental_use_pfor = False)\n",
    "        return jacobian\n",
    "    \n",
    "    \n",
    "def getSimilarity(vectors, subtractMean = False, metric = 'similarity'):\n",
    "    '''Inputs are:\n",
    "    \n",
    "        vectors (2D float)  - classes x feature matrix\n",
    "        subtractMean (Bool) - toggle across-class mean subtraction\n",
    "        metric (str)        - either 'similarity' (cosine sim) or a metric for \n",
    "                              sklearn's pdist function \n",
    "                              \n",
    "        Returns <dists> which is a classes x classes distance/similarity matrix\n",
    "    '''\n",
    "    \n",
    "    if subtractMean:\n",
    "        vectors -= np.mean(vectors, axis = 0, keepdims = True)\n",
    "\n",
    "    if metric == 'similarity':\n",
    "        normed = vectors / np.linalg.norm(vectors, axis = 1)[:, None]\n",
    "        dists  = normed @ normed.T\n",
    "    else:\n",
    "        dists = squareform(pdist(vecs, metric))\n",
    "\n",
    "    return dists\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d6bfe5ee",
   "metadata": {},
   "source": [
    "Let's load some example data from a single session and compute gradients of the output phone probabilities with respect to the input channels:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "cf3e25e0",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "36 / 207\n",
      "72 / 207\n",
      "108 / 207\n",
      "144 / 207\n",
      "180 / 207\n",
      "216 / 207\n",
      "1 / 40\n",
      "36 / 314\n",
      "72 / 314\n",
      "108 / 314\n",
      "144 / 314\n",
      "180 / 314\n",
      "216 / 314\n",
      "252 / 314\n",
      "288 / 314\n",
      "324 / 314\n",
      "2 / 40\n",
      "36 / 181\n",
      "72 / 181\n",
      "108 / 181\n",
      "144 / 181\n",
      "180 / 181\n",
      "216 / 181\n",
      "3 / 40\n",
      "36 / 234\n",
      "72 / 234\n",
      "108 / 234\n",
      "144 / 234\n",
      "180 / 234\n",
      "216 / 234\n",
      "252 / 234\n",
      "4 / 40\n",
      "36 / 345\n",
      "72 / 345\n",
      "108 / 345\n",
      "144 / 345\n",
      "180 / 345\n",
      "216 / 345\n",
      "252 / 345\n",
      "288 / 345\n",
      "324 / 345\n",
      "360 / 345\n",
      "5 / 40\n",
      "36 / 254\n",
      "72 / 254\n",
      "108 / 254\n",
      "144 / 254\n",
      "180 / 254\n",
      "216 / 254\n",
      "252 / 254\n",
      "288 / 254\n",
      "6 / 40\n",
      "36 / 184\n",
      "72 / 184\n",
      "108 / 184\n",
      "144 / 184\n",
      "180 / 184\n",
      "216 / 184\n",
      "7 / 40\n",
      "36 / 170\n",
      "72 / 170\n",
      "108 / 170\n",
      "144 / 170\n",
      "180 / 170\n",
      "8 / 40\n",
      "36 / 286\n",
      "72 / 286\n",
      "108 / 286\n",
      "144 / 286\n",
      "180 / 286\n",
      "216 / 286\n",
      "252 / 286\n",
      "288 / 286\n",
      "9 / 40\n",
      "36 / 123\n",
      "72 / 123\n",
      "108 / 123\n",
      "144 / 123\n",
      "10 / 40\n",
      "36 / 229\n",
      "72 / 229\n",
      "108 / 229\n",
      "144 / 229\n",
      "180 / 229\n",
      "216 / 229\n",
      "252 / 229\n",
      "11 / 40\n",
      "36 / 150\n",
      "72 / 150\n",
      "108 / 150\n",
      "144 / 150\n",
      "180 / 150\n",
      "12 / 40\n",
      "36 / 223\n",
      "72 / 223\n",
      "108 / 223\n",
      "144 / 223\n",
      "180 / 223\n",
      "216 / 223\n",
      "252 / 223\n",
      "13 / 40\n",
      "36 / 288\n",
      "72 / 288\n",
      "108 / 288\n",
      "144 / 288\n",
      "180 / 288\n",
      "216 / 288\n",
      "252 / 288\n",
      "288 / 288\n",
      "324 / 288\n",
      "14 / 40\n",
      "36 / 248\n",
      "72 / 248\n",
      "108 / 248\n",
      "144 / 248\n",
      "180 / 248\n",
      "216 / 248\n",
      "252 / 248\n",
      "15 / 40\n",
      "36 / 298\n",
      "72 / 298\n",
      "108 / 298\n",
      "144 / 298\n",
      "180 / 298\n",
      "216 / 298\n",
      "252 / 298\n",
      "288 / 298\n",
      "324 / 298\n",
      "16 / 40\n",
      "36 / 213\n",
      "72 / 213\n",
      "108 / 213\n",
      "144 / 213\n",
      "180 / 213\n",
      "216 / 213\n",
      "17 / 40\n",
      "36 / 297\n",
      "72 / 297\n",
      "108 / 297\n",
      "144 / 297\n",
      "180 / 297\n",
      "216 / 297\n",
      "252 / 297\n",
      "288 / 297\n",
      "324 / 297\n",
      "18 / 40\n",
      "36 / 327\n",
      "72 / 327\n",
      "108 / 327\n",
      "144 / 327\n",
      "180 / 327\n",
      "216 / 327\n",
      "252 / 327\n",
      "288 / 327\n",
      "324 / 327\n",
      "360 / 327\n",
      "19 / 40\n",
      "36 / 181\n",
      "72 / 181\n",
      "108 / 181\n",
      "144 / 181\n",
      "180 / 181\n",
      "216 / 181\n",
      "20 / 40\n",
      "36 / 294\n",
      "72 / 294\n",
      "108 / 294\n",
      "144 / 294\n",
      "180 / 294\n",
      "216 / 294\n",
      "252 / 294\n",
      "288 / 294\n",
      "324 / 294\n",
      "21 / 40\n",
      "36 / 221\n",
      "72 / 221\n",
      "108 / 221\n",
      "144 / 221\n",
      "180 / 221\n",
      "216 / 221\n",
      "252 / 221\n",
      "22 / 40\n",
      "36 / 160\n",
      "72 / 160\n",
      "108 / 160\n",
      "144 / 160\n",
      "180 / 160\n",
      "23 / 40\n",
      "36 / 149\n",
      "72 / 149\n",
      "108 / 149\n",
      "144 / 149\n",
      "180 / 149\n",
      "24 / 40\n",
      "36 / 231\n",
      "72 / 231\n",
      "108 / 231\n",
      "144 / 231\n",
      "180 / 231\n",
      "216 / 231\n",
      "252 / 231\n",
      "25 / 40\n",
      "36 / 346\n",
      "72 / 346\n",
      "108 / 346\n",
      "144 / 346\n",
      "180 / 346\n",
      "216 / 346\n",
      "252 / 346\n",
      "288 / 346\n",
      "324 / 346\n",
      "360 / 346\n",
      "26 / 40\n",
      "36 / 165\n",
      "72 / 165\n",
      "108 / 165\n",
      "144 / 165\n",
      "180 / 165\n",
      "27 / 40\n",
      "36 / 297\n",
      "72 / 297\n",
      "108 / 297\n",
      "144 / 297\n",
      "180 / 297\n",
      "216 / 297\n",
      "252 / 297\n",
      "288 / 297\n",
      "324 / 297\n",
      "28 / 40\n",
      "36 / 177\n",
      "72 / 177\n",
      "108 / 177\n",
      "144 / 177\n",
      "180 / 177\n",
      "29 / 40\n",
      "36 / 258\n",
      "72 / 258\n",
      "108 / 258\n",
      "144 / 258\n",
      "180 / 258\n",
      "216 / 258\n",
      "252 / 258\n",
      "288 / 258\n",
      "30 / 40\n",
      "36 / 254\n",
      "72 / 254\n",
      "108 / 254\n",
      "144 / 254\n",
      "180 / 254\n",
      "216 / 254\n",
      "252 / 254\n",
      "288 / 254\n",
      "31 / 40\n",
      "36 / 147\n",
      "72 / 147\n",
      "108 / 147\n",
      "144 / 147\n",
      "180 / 147\n",
      "32 / 40\n",
      "36 / 337\n",
      "72 / 337\n",
      "108 / 337\n",
      "144 / 337\n",
      "180 / 337\n",
      "216 / 337\n",
      "252 / 337\n",
      "288 / 337\n",
      "324 / 337\n",
      "360 / 337\n",
      "33 / 40\n",
      "36 / 241\n",
      "72 / 241\n",
      "108 / 241\n",
      "144 / 241\n",
      "180 / 241\n",
      "216 / 241\n",
      "252 / 241\n",
      "34 / 40\n",
      "36 / 180\n",
      "72 / 180\n",
      "108 / 180\n",
      "144 / 180\n",
      "180 / 180\n",
      "216 / 180\n",
      "35 / 40\n",
      "36 / 246\n",
      "72 / 246\n",
      "108 / 246\n",
      "144 / 246\n",
      "180 / 246\n",
      "216 / 246\n",
      "252 / 246\n",
      "36 / 40\n",
      "36 / 169\n",
      "72 / 169\n",
      "108 / 169\n",
      "144 / 169\n",
      "180 / 169\n",
      "37 / 40\n",
      "36 / 322\n",
      "72 / 322\n",
      "108 / 322\n",
      "144 / 322\n",
      "180 / 322\n",
      "216 / 322\n",
      "252 / 322\n",
      "288 / 322\n",
      "324 / 322\n",
      "38 / 40\n",
      "36 / 289\n",
      "72 / 289\n",
      "108 / 289\n",
      "144 / 289\n",
      "180 / 289\n",
      "216 / 289\n",
      "252 / 289\n",
      "288 / 289\n",
      "324 / 289\n",
      "39 / 40\n",
      "36 / 206\n",
      "72 / 206\n",
      "108 / 206\n",
      "144 / 206\n",
      "180 / 206\n",
      "216 / 206\n",
      "40 / 40\n"
     ]
    }
   ],
   "source": [
    "datasetIdx = -2\n",
    "\n",
    "dataset   = list(nsd.tfValDatasets[datasetIdx])[0]\n",
    "layerIdx  = nsd.args['dataset']['datasetToLayerMap'][datasetIdx]\n",
    "gradients = list()\n",
    "for idx in range(dataset['inputFeatures'].shape[0]):\n",
    "    data      = dataset['inputFeatures'][idx, :dataset['nTimeSteps'][idx], :]\n",
    "    getter    = GradientGetter(nsd, data)\n",
    "    gradients.extend(getter())\n",
    "    \n",
    "    print(idx+1, '/', dataset['inputFeatures'].shape[0])\n",
    "    # break # uncomment this line for a much faster demo"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "597e1307",
   "metadata": {},
   "source": [
    "We've now computed our Jacobians! These matrices track the gradient of each (logit) output with respect to the input neural features (hence why we use \"gradient\" interchangeably - the Jacobian is just a matrix tracking all of these 256 inputs x 40 outputs gradients). \n",
    "\n",
    "After averaging across timepoints, each output logit (representing a phoneme probability) will now have an associated set of weights on the input channels. These reflect how to budge each channel's activity to maximize the probability of a given class."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "01d04e47",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7fb43c0c7c40>"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAABTwAAAGJCAYAAACn2/xCAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8pXeV/AAAACXBIWXMAAA9hAAAPYQGoP6dpAABeHElEQVR4nO3deVxUZf//8fegAm6AKKsbmOaKYnhLqKUpCWV9M8nUNJe4zUxtoU375ZbdaZZlLt3mHW7dbnln3WlmmamZopZGWilfNb0tFZcIUEsBOb8/+jq3AwMOOMzG6/l4zEPnnOuc8zmHuc6c+ZzrXJfJMAxDAAAAAAAAAOABvJwdAAAAAAAAAADYCwlPAAAAAAAAAB6DhCcAAAAAAAAAj0HCEwAAAAAAAIDHIOEJAAAAAAAAwGOQ8AQAAAAAAADgMUh4AgAAAAAAAPAYJDwBAAAAAAAAeAwSngAAAAAAAAA8BglPAAAAXLfNmzfLZDJp8+bNzg6lGGuxDR06VBEREU6LyR0cPXpUJpNJixYtKveyr732mv0DAwAAuAYSngAAABVs0aJFMplMJb527Njh7BABp1m3bp0mTZrk7DAAAIAHqersAAAAACqLF198UZGRkcWmN23a1AnRVG7/+Mc/VFhY6OwwXFrjxo31xx9/qFq1ahW6nXXr1mnu3LkkPQEAgN2Q8AQAAHCQO+64Qx06dHB2GJAqPInnCUwmk3x9fZ0dBgAAQJnxSDsAAICLmDhxory8vLRx40aL6Q8//LC8vb313XffSZLy8vI0YcIExcTEyN/fXzVr1tQtt9yiTZs2WSx3dT+Kc+fOVZMmTVSjRg317NlTP//8swzD0JQpU9SgQQNVr15d99xzj7KysizWERERobvuukufffaZoqOj5evrq1atWmn16tU27dPOnTuVmJgof39/1ahRQ127dtW2bdtsWnb27Nlq3bq1atSooTp16qhDhw5atmyZef5//vMfPfroo2revLmqV6+uunXrqm/fvjp69Og1122tD8/CwkLNnDlTrVu3lq+vr0JCQjRixAj99ttvFuWuHJOvvvpKHTt2lK+vr5o0aaIlS5YU2052draefPJJRUREyMfHRw0aNNDgwYN19uxZnT9/XjVr1tTjjz9ebLlffvlFVapU0dSpU0vch5tuukl9+vSxmBYVFSWTyaS9e/eap61cuVImk0n79+83Tzt+/LgeeughhYSEyMfHR61bt9aCBQss1lVSH56rVq1Sq1at5OvrqzZt2uiDDz4otU/U+fPn64YbbpCPj4/+8pe/6OuvvzbPGzp0qObOnStJFt08XLFixQrFxMSodu3a8vPzU1RUlN58880SjwkAAIBEC08AAACHycnJ0dmzZy2mmUwm1a1bV5L0wgsvaM2aNUpOTta+fftUu3Ztffrpp/rHP/6hKVOmqF27dpKk3NxcvfPOOxowYICGDx+uc+fOKTU1VQkJCdq1a5eio6MttrF06VLl5eVpzJgxysrK0vTp03X//fere/fu2rx5s5577jkdOnRIs2fP1tNPP10s8XXw4EH169dPjzzyiIYMGaKFCxeqb9++Wr9+vW6//fYS9/eLL77QHXfcoZiYGHMyd+HCherevbu2bt2qjh07lrjsP/7xDz322GO677779Pjjj+vixYvau3evdu7cqQceeECS9PXXX2v79u3q37+/GjRooKNHj+rvf/+7unXrph9//FE1atSw+W8jSSNGjNCiRYs0bNgwPfbYYzpy5IjmzJmjb7/9Vtu2bbNoFXro0CHdd999Sk5O1pAhQ7RgwQINHTpUMTExat26tSTp/PnzuuWWW7R//3499NBDuummm3T27Fl99NFH+uWXXxQdHa17771XK1eu1Ouvv64qVaqY1798+XIZhqGBAweWGO8tt9yi5cuXm99nZWXphx9+kJeXl7Zu3aq2bdtKkrZu3aqgoCC1bNlSknTq1CndfPPNMplMGj16tIKCgvTJJ58oOTlZubm5euKJJ0rc5scff6x+/fopKipKU6dO1W+//abk5GTVr1/favlly5bp3LlzGjFihEwmk6ZPn64+ffrop59+UrVq1TRixAidOHFCGzZs0Lvvvmux7IYNGzRgwAD16NFDr7zyiiRp//792rZtm9UkMQAAgJkBAACACrVw4UJDktWXj4+PRdl9+/YZ3t7exl//+lfjt99+M+rXr2906NDByM/PN5cpKCgwLl26ZLHcb7/9ZoSEhBgPPfSQedqRI0cMSUZQUJCRnZ1tnj5u3DhDktGuXTuL9Q4YMMDw9vY2Ll68aJ7WuHFjQ5Lx/vvvm6fl5OQYYWFhRvv27c3TNm3aZEgyNm3aZBiGYRQWFhrNmjUzEhISjMLCQnO533//3YiMjDRuv/32Uo/ZPffcY7Ru3brUMr///nuxaWlpaYYkY8mSJSXGZhiGMWTIEKNx48bm91u3bjUkGUuXLrVY3/r164tNv3JMvvzyS/O006dPGz4+PsZTTz1lnjZhwgRDkrF69epicV45Jp9++qkhyfjkk08s5rdt29bo2rVrqfu/atUqQ5Lx448/GoZhGB999JHh4+Nj/M///I/Rr18/i3Xde++95vfJyclGWFiYcfbsWYv19e/f3/D39zcf1yufn4ULF5rLREVFGQ0aNDDOnTtnnrZ582ZDksXxvLJs3bp1jaysLPP0f//734YkY82aNeZpo0aNMqz9LHn88ccNPz8/o6CgoNTjAAAAUBSPtAMAADjI3LlztWHDBovXJ598YlGmTZs2mjx5st555x0lJCTo7NmzWrx4sapW/e+DOVWqVJG3t7ekPx/DzsrKUkFBgTp06KA9e/YU227fvn3l7+9vfh8bGytJGjRokMV6Y2NjlZeXp+PHj1ssHx4ernvvvdf83s/PT4MHD9a3336rzMxMq/uanp6ugwcP6oEHHtCvv/6qs2fP6uzZs7pw4YJ69OihL7/8stRBgwICAvTLL79YPP5cVPXq1c3/z8/P16+//qqmTZsqICDA6nEozapVq+Tv76/bb7/dHOvZs2cVExOjWrVqFesuoFWrVrrlllvM74OCgtS8eXP99NNP5mnvv/++2rVrZ3Hsrrjy2HZ8fLzCw8O1dOlS87zvv/9ee/fu1aBBg0qN+cr2v/zyS0l/tuT8y1/+ottvv11bt26V9Ocj9d9//725rGEYev/993X33XfLMAyLfU1ISFBOTk6Jx+7EiRPat2+fBg8erFq1apmnd+3aVVFRUVaX6devn+rUqVMs5quPU0kCAgJ04cIFbdiw4ZplAQAArsYj7QAAAA7SsWNHmwYteuaZZ7RixQrt2rVLL7/8slq1alWszOLFizVjxgwdOHBA+fn55unWRoFv1KiRxfsryc+GDRtanV60z8qmTZta9KsoSTfeeKOkP/t5DA0NLbbNgwcPSpKGDBlifSf15yP+VyfDrvbcc8/p888/V8eOHdW0aVP17NlTDzzwgDp37mwu88cff2jq1KlauHChjh8/LsMwLNZdFgcPHlROTo6Cg4Otzj99+rTF+6LHVJLq1KljcewOHz6spKSkUrfr5eWlgQMH6u9//7t+//131ahRQ0uXLpWvr6/69u1b6rIhISFq1qyZtm7dqhEjRmjr1q267bbbdOutt2rMmDH66aeftH//fhUWFpoTjWfOnFF2drbmz5+v+fPn27SvV/znP/+R9OfnoaimTZtaTZQWPU5X/t5FP2PWPProo3rvvfd0xx13qH79+urZs6fuv/9+JSYmXnNZAABQuZHwBAAAcDE//fSTOWG4b9++YvP/+c9/aujQoerdu7eeeeYZBQcHmwe4OXz4cLHyV/cNacv0qxOH5XWl9earr75arE/RK65uJVhUy5YtlZGRobVr12r9+vV6//339dZbb2nChAmaPHmyJGnMmDFauHChnnjiCcXFxcnf318mk0n9+/cvtfVoSfEGBwdbtLS8WlBQkMV7ex67wYMH69VXX9WHH36oAQMGaNmyZbrrrrssWuWWpEuXLtq4caP++OMP7d69WxMmTFCbNm0UEBCgrVu3av/+/apVq5bat28v6b9/l0GDBpWYjL7S96c9XM9xCg4OVnp6uj799FN98skn+uSTT7Rw4UINHjxYixcvtluMAADA85DwBAAAcCGFhYUaOnSo/Pz89MQTT+jll1/WfffdZzEa97/+9S81adJEq1evtmh5OXHixAqJ6dChQzIMw2Jb//u//ytJJY7MfcMNN0j68/H3+Pj4cm23Zs2a6tevn/r166e8vDz16dNHf/vb3zRu3Dj5+vrqX//6l4YMGaIZM2aYl7l48aKys7PLvK0bbrhBn3/+uTp37mzxqPz1uOGGG/T9999fs1ybNm3Uvn17LV26VA0aNNCxY8c0e/Zsm7Zxyy23aOHChVqxYoUuX76sTp06ycvLS126dDEnPDt16mROPAYFBal27dq6fPlymf8ujRs3lvTn56Eoa9NsVbT18NW8vb1199136+6771ZhYaEeffRRvf322xo/frzVlqYAAACSRB+eAAAALuT111/X9u3bNX/+fE2ZMkWdOnXSyJEjLUZ3v5K8urqV3M6dO5WWllYhMZ04cUIffPCB+X1ubq6WLFmi6Ohoq4+zS1JMTIxuuOEGvfbaazp//nyx+WfOnCl1m7/++qvFe29vb7Vq1UqGYZgf4a9SpUqxloKzZ8/W5cuXbdqvq91///26fPmypkyZUmxeQUFBuZKoSUlJ+u677yyO3RVF437wwQf12WefaebMmapbt67uuOMOm7Zx5VH1V155RW3btjW3Cr3lllu0ceNGffPNNxZ9jVapUkVJSUl6//33rSZjS/u7hIeHq02bNlqyZInF33TLli1WWyLbqmbNmpJU7BgX/Qx4eXmZW59eunSp3NsDAACejxaeAAAADvLJJ5/owIEDxaZ36tRJTZo00f79+zV+/HgNHTpUd999tyRp0aJFio6ONvdnKEl33XWXVq9erXvvvVe9evXSkSNHNG/ePLVq1cpqcvF63XjjjUpOTtbXX3+tkJAQLViwQKdOndLChQtLXMbLy0vvvPOO7rjjDrVu3VrDhg1T/fr1dfz4cW3atEl+fn5as2ZNicv37NlToaGh6ty5s0JCQrR//37NmTNHvXr1Uu3atc3H4d1335W/v79atWqltLQ0ff7556pbt26Z97Fr164aMWKEpk6dqvT0dPXs2VPVqlXTwYMHtWrVKr355pu67777yrTOZ555Rv/617/Ut29fPfTQQ4qJiVFWVpY++ugjzZs3T+3atTOXfeCBB/Tss8/qgw8+0MiRI1WtWjWbttG0aVOFhoYqIyNDY8aMMU+/9dZb9dxzz0mSRcJTkqZNm6ZNmzYpNjZWw4cPV6tWrZSVlaU9e/bo888/V1ZWVonbe/nll3XPPfeoc+fOGjZsmH777TfNmTNHbdq0KfdnLyYmRpL02GOPKSEhQVWqVFH//v3117/+VVlZWerevbsaNGig//znP5o9e7aio6PVsmXLcm0LAABUDiQ8AQAAHGTChAlWpy9cuFCNGzfWkCFDVK9ePc2cOdM8r1mzZpo6daoef/xxvffee7r//vs1dOhQZWZm6u2339ann36qVq1a6Z///KdWrVqlzZs32z3uZs2aafbs2XrmmWeUkZGhyMhIrVy5UgkJCaUu161bN6WlpWnKlCmaM2eOzp8/r9DQUMXGxmrEiBGlLjtixAgtXbpUr7/+us6fP68GDRroscce0wsvvGAu8+abb6pKlSpaunSpLl68qM6dO+vzzz+/ZlwlmTdvnmJiYvT222/r+eefV9WqVRUREaFBgwZZDJZkq1q1amnr1q2aOHGiPvjgAy1evFjBwcHq0aOHGjRoYFE2JCREPXv21Lp16/Tggw+WaTu33HKLVq1apS5dupinxcTEqEaNGiooKFBsbGyxbe3atUsvvviiVq9erbfeekt169ZV69at9corr5S6rbvvvlvLly/XpEmTNHbsWDVr1kyLFi3S4sWL9cMPP5Qp7iv69OmjMWPGaMWKFfrnP/8pwzDUv39/DRo0SPPnz9dbb72l7OxshYaGql+/fpo0aZK8vHhQDQAAlMxk2KNXegAAAHikiIgItWnTRmvXrnV2KB7v3nvv1b59+66rP0xniY6OVlBQkDZs2ODsUAAAAOjDEwAAAHC2kydP6uOPPy5z605Hy8/PV0FBgcW0zZs367vvvlO3bt2cExQAAEARPNIOAAAAOMmRI0e0bds2vfPOO6pWrdo1H/V3tuPHjys+Pl6DBg1SeHi4Dhw4oHnz5ik0NFSPPPKIs8MDAACQRMITAAAAcJotW7Zo2LBhatSokRYvXlziqPeuok6dOoqJidE777yjM2fOqGbNmurVq5emTZtWrsGiAAAAKgJ9eAIAAAAAAADwGPThCQAAAAAAAMBjkPAEAAAAAAAA4DHow9MBCgsLdeLECdWuXVsmk8nZ4QAAAAAAAABuxTAMnTt3TuHh4fLyKr0NJwlPBzhx4oQaNmzo7DAAAAAAAAAAt/bzzz+rQYMGpZYh4ekAtWvXlvTnH8TPz8/J0QAAAAAAAADuJTc3Vw0bNjTn2UpDwtMBrjzG7ufnR8ITAAAAAAAAKCdbuotk0CIAAAAAAAAAHoOEJwAAAAAAAACPQcITAAAAAAAAgMegD08AAAAAAADASS5fvqz8/Hxnh+ESqlWrpipVqlz3ekh4AgAAAAAAAE5w/vx5/fLLLzIMw9mhuASTyaQGDRqoVq1a17UeEp4AAAAAAACAg12+fFm//PKLatSooaCgIJtGH/dkhmHozJkz+uWXX9SsWbPraunpdn14zp07VxEREfL19VVsbKx27dpVavlVq1apRYsW8vX1VVRUlNatW2cxf/Xq1erZs6fq1q0rk8mk9PT0Yuvo1q2bTCaTxeuRRx6x524BAAAAAACgEsnPz5dhGAoKClL16tXl6+tbqV/Vq1dXUFCQDMO47kf83SrhuXLlSqWkpGjixInas2eP2rVrp4SEBJ0+fdpq+e3bt2vAgAFKTk7Wt99+q969e6t37976/vvvzWUuXLigLl266JVXXil128OHD9fJkyfNr+nTp9t13wAAAAAAAFD5VPaWnVez17EwGW7USUBsbKz+8pe/aM6cOZKkwsJCNWzYUGPGjNHYsWOLle/Xr58uXLigtWvXmqfdfPPNio6O1rx58yzKHj16VJGRkfr2228VHR1tMa9bt26Kjo7WzJkzyxV3bm6u/P39lZOTIz8/v3KtAwAAAAAAAJ7j4sWLOnLkiCIjI+Xr6+vscFxCacekLPk1t2nhmZeXp927dys+Pt48zcvLS/Hx8UpLS7O6TFpamkV5SUpISCixfGmWLl2qevXqqU2bNho3bpx+//33EsteunRJubm5Fi8AAAAAAAAAFc9tBi06e/asLl++rJCQEIvpISEhOnDggNVlMjMzrZbPzMws07YfeOABNW7cWOHh4dq7d6+ee+45ZWRkaPXq1VbLT506VZMnTy7TNgAAAAAAAABcP7dJeDrTww8/bP5/VFSUwsLC1KNHDx0+fFg33HBDsfLjxo1TSkqK+X1ubq4aNmzokFgBAAAAjzfJ/6r/5zgvDgAAKkDE2I8dur2j03qVa7m0tDR16dJFiYmJ+vhj6zEvX75cgwYN0iOPPKK5c+deT5hl4jaPtNerV09VqlTRqVOnLKafOnVKoaGhVpcJDQ0tU3lbxcbGSpIOHTpkdb6Pj4/8/PwsXgAAAIC9RYz92PwCAABwpNTUVI0ZM0ZffvmlTpw4UWKZZ599VsuXL9fFixcdFpvbJDy9vb0VExOjjRs3mqcVFhZq48aNiouLs7pMXFycRXlJ2rBhQ4nlbZWeni5JCgsLu671AAAAAAAAAO7m/PnzWrlypUaOHKlevXpp0aJFxcocOXJE27dv19ixY3XjjTeW2DVkRXCbhKckpaSk6B//+IcWL16s/fv3a+TIkbpw4YKGDRsmSRo8eLDGjRtnLv/4449r/fr1mjFjhg4cOKBJkybpm2++0ejRo81lsrKylJ6erh9//FGSlJGRofT0dHM/n4cPH9aUKVO0e/duHT16VB999JEGDx6sW2+9VW3btnXg3gMAAAAAAADO995776lFixZq3ry5Bg0apAULFsgwDIsyCxcuVK9eveTv769BgwYpNTXVYfG5VcKzX79+eu211zRhwgRFR0crPT1d69evNw9MdOzYMZ08edJcvlOnTlq2bJnmz5+vdu3a6V//+pc+/PBDtWnTxlzmo48+Uvv27dWr15/9FfTv31/t27fXvHnzJP3ZsvTzzz9Xz5491aJFCz311FNKSkrSmjVrHLjnAAAAAAAAgGtITU3VoEGDJEmJiYnKycnRli1bzPMLCwu1aNEic5n+/fvrq6++0pEjRxwSn8komn6F3eXm5srf3185OTn05wkAAAC7ubrvzvIOOOCWGLQIAOABLl68qCNHjigyMlK+vr7m6a4+aFFGRobatGmj48ePKzg4WJI0evRo5eTk6N1335Ukffrpp3rggQeUmZmpatWqSZLuuusutW/fXlOmTClx3SUdE6ls+TVGaQcAAAAAAABgk9TUVBUUFCg8PNw8zTAM+fj4aM6cOfL391dqaqqysrJUvXp1c5nCwkLt3btXkydPlpdXxT50TsITAAAAAAAAwDUVFBRoyZIlmjFjhnr27Gkxr3fv3lq+fLn69u2rf//731qxYoVat25tnn/58mV16dJFn332mRITEys0ThKeAAAAAAAAAK5p7dq1+u2335ScnCx/f3+LeUlJSUpNTdXFixdVt25d3X///TKZTBZl7rzzTqWmppLwBAAAAAAAACoLV+6XOzU1VfHx8cWSndKfCc/p06dr9+7dGjlyZLFk55UyDz74oM6ePat69epVWJwkPAEAAAAAAABc05o1a0qc17FjR11rbPT7779f999/v73DKqZiewgFAAAAAAAAAAci4QkAAAAAAADAY5DwBAAAAAAAAOAxSHgCAAAAAAAA8BgkPAEAAAAAAAB4DBKeAAAAAAAAADwGCU8AAAAAAAAAHoOEJwAAAAAAAACPQcITAAAAAAAAgMeo6uwAAAAAAAAAAPyfSf4O3l6OY7fnALTwBAAAAAAAAFAmaWlpqlKlinr16mUx/ejRozKZTFZfO3bscEhstPAEAAAAAAAAUCapqakaM2aMUlNTdeLECYWHh1vM//zzz9W6dWuLaXXr1nVIbCQ8AQAAAAAAANjs/PnzWrlypb755htlZmZq0aJFev755y3K1K1bV6GhoU6Jj0fagcpmkv9/XwAAAAAAAGX03nvvqUWLFmrevLkGDRqkBQsWyDAMZ4dlRsITAAAAAAAAgM1SU1M1aNAgSVJiYqJycnK0ZcsWizKdOnVSrVq1LF6OwiPtAAAAAAAAAGySkZGhXbt26YMPPpAkVa1aVf369VNqaqq6detmLrdy5Uq1bNnSKTGS8AQAAAAAAABgk9TUVBUUFFgMUmQYhnx8fDRnzhzztIYNG6pp06bOCJFH2gEAAAAAAABcW0FBgZYsWaIZM2YoPT3d/Pruu+8UHh6u5cuXOztESbTwBAAAAAAAQNGBbSflOCcOuLS1a9fqt99+U3Jysvz9LT8zSUlJSk1NVWJioiTp119/VWZmpkWZgIAA+fr6VnicJDwBAAAAAAAAV+HCyebU1FTFx8cXS3ZKfyY8p0+frtzcXElSfHx8sTLLly9X//79KzxOEp4AAAAAAAAArmnNmjUlzuvYsaMMw5Ak87/OQh+eAAAAAAAAADwGCU8AAAAAAAAAHoOEJwAAAAAAAACPQcITAAAAAAAAgMcg4QkAAAAAAADAY5DwBAAAAAAAAOAxqjo7AMBZIsZ+bP7/0Wm9nBiJk03yv+r/Oc6LAwAAAAAAwA5IeAIAAAAAAMCl0EgJ14NH2gEAAAAAAAB4DFp4AgAAAAAAAC4ianGUQ7e3b8g+h27PEdyuhefcuXMVEREhX19fxcbGateuXaWWX7VqlVq0aCFfX19FRUVp3bp1FvNXr16tnj17qm7dujKZTEpPTy+2josXL2rUqFGqW7euatWqpaSkJJ06dcqeuwUAAAAAAAC4jbS0NFWpUkW9ell2OXD06NESc2zdunXTE088UeGxuVXCc+XKlUpJSdHEiRO1Z88etWvXTgkJCTp9+rTV8tu3b9eAAQOUnJysb7/9Vr1791bv3r31/fffm8tcuHBBXbp00SuvvFLidp988kmtWbNGq1at0pYtW3TixAn16dPH7vsHAAAAAAAAuIPU1FSNGTNGX375pU6cOOHscCy4VcLz9ddf1/DhwzVs2DC1atVK8+bNU40aNbRgwQKr5d98800lJibqmWeeUcuWLTVlyhTddNNNmjNnjrnMgw8+qAkTJig+Pt7qOnJycpSamqrXX39d3bt3V0xMjBYuXKjt27drx44dFbKfAAAAAAAAgKs6f/68Vq5cqZEjR6pXr15atGiRs0Oy4DYJz7y8PO3evdsiMenl5aX4+HilpaVZXSYtLa1YIjMhIaHE8tbs3r1b+fn5Futp0aKFGjVqVOJ6Ll26pNzcXIsXAAAAAAAA4Anee+89tWjRQs2bN9egQYO0YMECGYbh7LDM3CbhefbsWV2+fFkhISEW00NCQpSZmWl1mczMzDKVL2kd3t7eCggIsHk9U6dOlb+/v/nVsGFDm7cHAAAAAAAAuLLU1FQNGjRIkpSYmKicnBxt2bLFokynTp1Uq1Yti9fWrVsdEp/bJDzdybhx45STk2N+/fzzz84OCQAAAAAAALhuGRkZ2rVrlwYMGCBJqlq1qvr166fU1FSLcitXrlR6errFq0OHDg6JsapDtmIH9erVU5UqVYqNjn7q1CmFhoZaXSY0NLRM5UtaR15enrKzsy1aeZa2Hh8fH/n4+Ni8DQAAAAAAAMAdpKamqqCgQOHh4eZphmHIx8fHYtychg0bqmnTphbLVq9e3SExuk0LT29vb8XExGjjxo3maYWFhdq4caPi4uKsLhMXF2dRXpI2bNhQYnlrYmJiVK1aNYv1ZGRk6NixY2VaDwAAAAAAAODOCgoKtGTJEs2YMcOi5eZ3332n8PBwLV++3NkhSnKjFp6SlJKSoiFDhqhDhw7q2LGjZs6cqQsXLmjYsGGSpMGDB6t+/fqaOnWqJOnxxx9X165dNWPGDPXq1UsrVqzQN998o/nz55vXmZWVpWPHjunEiROS/kxmSn+27AwNDZW/v7+Sk5OVkpKiwMBA+fn5acyYMYqLi9PNN9/s4CMAAAAAAAAAOMfatWv122+/KTk5Wf7+/hbzkpKSlJqaqsTERCdF919ulfDs16+fzpw5owkTJigzM1PR0dFav369eWCiY8eOycvrv41WO3XqpGXLlumFF17Q888/r2bNmunDDz9UmzZtzGU++ugjc8JUkvr37y9JmjhxoiZNmiRJeuONN+Tl5aWkpCRdunRJCQkJeuuttxywxwAAAAAAAKhM9g3Z5+wQSpSamqr4+PhiyU7pz4Tn9OnTlZub64TILLlVwlOSRo8erdGjR1udt3nz5mLT+vbtq759+5a4vqFDh2ro0KGlbtPX11dz587V3LlzyxIqAAAAAAAA4DHWrFlT4ryOHTvKMAxJMv9blLXcXUVwmz48AQAAAAAAAOBaSHgCAAAAAAAA8BgkPAEAAAAAAAB4DLfrwxMAAAAAAACe7ajvA1e9y3FaHHBPtPAEAAAAAAAAnKSkAX4qI3sdC1p4AgAAAHCIiLEfW7w/Oq2XkyIBAMD5qlSpIknKy8tT9erVnRyNa8jLy5P032NTXiQ8AQAAAAAAAAerWrWqatSooTNnzqhatWry8qrcD2IXFhbqzJkzqlGjhqpWvb6UJQlPAAAAAAAAwMFMJpPCwsJ05MgR/ec//3F2OC7By8tLjRo1kslkuq71kPAEAAAAAAAAnMDb21vNmjUzP8pd2Xl7e9ulpSsJTwAAAAAAAFcwyb/Ie0Ynrwy8vLzk6+vr7DA8SuXuHAAAAAAAAACAR6GFJwAAAADAvmilBgBwIlp4AgAAAAAAAPAYtPAEAAC4XrRkAgAA7ubq6xeuXeBhaOEJAAAAAAAAwGPQwhMAAKCycfEWqRFjP7Z4f3RaLydFAgAAAHdEC08AAAAAAAAAHoMWngAAAAAAAHApUZGNzP/f58Q44J5o4QkAAAAAAADAY9DCEwAqCxfvsw8AKruoxVHm/+8bQlsWAACA8iLhCQAAAMAp7DZAFTf1AADAVUh4AvZ29QU3F9sAAAAAAAAORcITAAAAgFthIAsAAFAaEp4AAACAm6rdcuxV78r5ODgAAICHIeEJj0An/wAAAAAAAJBIeAJuy26d/AMAgP+iL24AAAC35+XsAAAAAAAAAADAXkh4AgAAAAAAAPAYJDwBAAAAAAAAeAz68ATcxNV9dtJfJwBPwzkOcDL6LgUAAB6EhCcAAAAAwLOR1IcDcAMXcB0kPAEAAAAAANzF1Ql8yW5J/KjIRhbv99llrYBz0IcnAAAAAAAAAI9BC08AAAAPwyN1AAAAqMxIeAIAgMrNlfp1c6VYAACVEjfNAHgCEp4AgNKRgAEAAAAAuBESngAAAICnqKCBLAAAANyJ2w1aNHfuXEVERMjX11exsbHatWtXqeVXrVqlFi1ayNfXV1FRUVq3bp3FfMMwNGHCBIWFhal69eqKj4/XwYMHLcpERETIZDJZvKZNm2b3fQMAAAAAAID7ixj7scULjuVWCc+VK1cqJSVFEydO1J49e9SuXTslJCTo9OnTVstv375dAwYMUHJysr799lv17t1bvXv31vfff28uM336dM2aNUvz5s3Tzp07VbNmTSUkJOjixYsW63rxxRd18uRJ82vMmDEVuq8AAAAAAADuLmpxlMULcAS3Sni+/vrrGj58uIYNG6ZWrVpp3rx5qlGjhhYsWGC1/JtvvqnExEQ988wzatmypaZMmaKbbrpJc+bMkfRn686ZM2fqhRde0D333KO2bdtqyZIlOnHihD788EOLddWuXVuhoaHmV82aNSt6dwEAAAC3xg9cAADgDG6T8MzLy9Pu3bsVHx9vnubl5aX4+HilpaVZXSYtLc2ivCQlJCSYyx85ckSZmZkWZfz9/RUbG1tsndOmTVPdunXVvn17vfrqqyooKCgx1kuXLik3N9fiBQAAAAAAACsm+Vu+gOvkNoMWnT17VpcvX1ZISIjF9JCQEB04cMDqMpmZmVbLZ2ZmmudfmVZSGUl67LHHdNNNNykwMFDbt2/XuHHjdPLkSb3++utWtzt16lRNnjy5bDsIh6vdcuxV73o5LQ4AgJ0wWAsAAIBbuLrl/74h+8q1jqL9Yh6dxu96/JfbJDydKSUlxfz/tm3bytvbWyNGjNDUqVPl4+NTrPy4ceMslsnNzVXDhg0dEisAACibqMhG5v+X73Lbch3Xsx53c/UPjev6kXF1sppENWBf1C+gcqiEdf2o7wNFpjhuv21Jtlo2spJoaOVYbpPwrFevnqpUqaJTp05ZTD916pRCQ0OtLhMaGlpq+Sv/njp1SmFhYRZloqOjS4wlNjZWBQUFOnr0qJo3b15svo+Pj9VEKHA9aJGKClEJL4wAd2ePFhEAAACAJ3ObPjy9vb0VExOjjRs3mqcVFhZq48aNiouLs7pMXFycRXlJ2rBhg7l8ZGSkQkNDLcrk5uZq586dJa5TktLT0+Xl5aXg4ODr2SUAAAAAAAAAduY2LTylPx8tHzJkiDp06KCOHTtq5syZunDhgoYNGyZJGjx4sOrXr6+pU6dKkh5//HF17dpVM2bMUK9evbRixQp98803mj9/viTJZDLpiSee0EsvvaRmzZopMjJS48ePV3h4uHr37i3pz4GPdu7cqdtuu021a9dWWlqannzySQ0aNEh16tRxynFwVbQ4AQAAcH1FR0znug0AAHgat0p49uvXT2fOnNGECROUmZmp6OhorV+/3jzo0LFjx+Tl9d9Gq506ddKyZcv0wgsv6Pnnn1ezZs304Ycfqk2bNuYyzz77rC5cuKCHH35Y2dnZ6tKli9avXy9fX19Jfz6evmLFCk2aNEmXLl1SZGSknnzySYs+OgEAAAAAAAC4BrdKeErS6NGjNXr0aKvzNm/eXGxa37591bdv3xLXZzKZ9OKLL+rFF1+0Ov+mm27Sjh07yhUrAAAAAAAAKgHGR3ApbpfwBAAAKC9bRtQEAAAA4N5IeAIAAAAA7CoqspHFe3qKBQA4EglPAAAAAPBQVyceSToCACoLEp4AAACAK7q6LzCJ/sAAXD/6GARQSZDwBAAAAP4PreEAAADcHwlPAABsQYsIAABQVq7eUtvV44PNuGFXfi4/qCW/Q8qFhCcAAAAAACg/EjKSpKO+D1z1rvIeB8AVkPAEnCBqcZT5//uGlO/+W+2WY4tMcbG7UAAAAEBpbGldSAtEAEA5kPAEPNjViVWp/MlVAKhotIgAAAAAYC8kPAHAU/FoEQCUihuDAFAcNyEBeAISnkAlQ2fWboBHtwAA9sQNMAAAUMmQ8AQAd0RSFAAAAAAAq0h4Ai4oYuzHFu+PTmNAIocjoYgyot6irKw9Tm2vVviWA9vxWUTZXX1O85TzmSfuE1zf1ed1yT2esOKJMKCC8BvToUh4AgAAwO2QvHIBPCoPONXVN87og7gUJJngwqjHFYeEJwAAqDQsB2KQGIwBAAAX5so3Vkik2pU7toaGayPhCVRybv/IChcanoO/JVyEK43c7UqxAKhAfAcCAGBXJDwBF2TZ95pE/2uACyr641TLnBIGnMSVW5x4AL4HARdDQhZwO9w0RWVX5oTnsWPH1LBhQ5lMJovphmHo559/VqNGjUpYEnBD/KAFro0fQYBzUQeByoG6bjN3S/S4W7x8FlEZ0Xe4+ylzwjMyMlInT55UcHCwxfSsrCxFRkbq8uXLdgsOAADgCjp1BxzHspWt9R92bt8tjgdyu8RZeZBsA1xfJWg4RL/wrq/MCU/DMIq17pSk8+fPy9fX1y5BAa6qUlxEwmPRETgqwtV3uyU73vHmBy2uwZaEHOynwuo6AI/ADRDATrgGthubE54pKSmSJJPJpPHjx6tGjRrmeZcvX9bOnTsVHR1t9wABh+HEArgf6i0AAAAAoAibE57ffvutpD9beO7bt0/e3t7med7e3mrXrp2efvpp+0cIAIAd0VIbAAAAwPXg6TnXZ3PCc9OmTZKkYcOG6c0335Sfn1+FBQUAAI9P2hmtYQEAqHhFv28jGdTXYxT925aTPfokL2+yjf7QXR+JVPspcx+eCxcurIg4UFlUgs6L4YLKkehx5hcNX3IA3AX9aMKdWA4wUXHXoTxJAACA85U54XnhwgVNmzZNGzdu1OnTp1VYWGgx/6effrJbcADKhjt2bsADW9lVVILWkSMfFqs7Hvh3ciXObL3LDYVScFMSlQ3n+uty9bn86LRelfccUln324UU+ywCgMqR8PzrX/+qLVu26MEHH1RYWJjVEdsBALgmD/uBUJEJWi7kATibI29CAUBJKqoFNV0pAZ6nzAnPTz75RB9//LE6d+5cEfEA9sEde7fEI2AAAADu6eoW9FzBAQCcrcwJzzp16igwMLAiYgEA9+dCrRb54QHgenhEaxcXOifjT5b9vkrn9k9zUiRwV3RNAsDein430S+5ZyhzwnPKlCmaMGGCFi9erBo1alRETABwfRz0A9cjkgHlQF+xuF48GgsA8FRFE7LlRXc2lRdJfcA+bEp4tm/f3qKvzkOHDikkJEQRERGqVq2aRdk9e/bYN0IATufMBNe1LvY84oKAFkiVG39/XCePOA/C5fE5g8PwveixSOICcCSbEp69e/eu4DAAeLrK2hoSsED/whWKhEzJaJltZyRk7IbHCAH7KXqup398AJWZTQnPiRMnVnQcAABYIHnleCTFAKCCkCQHUIkwlgBcQZn78AQAAMW5eoLW1eMDHKHo0wa1WzopEJTKlW++8MQKALg4bjDh/5RrlPar+/O8wmQyydfXV02bNtXQoUM1bNgwuwQIAK7ClX+ASfSL5K64Aw4AcCRXv54BrlexR/mdFIc9cZ0PlF2ZE54TJkzQ3/72N91xxx3q2LGjJGnXrl1av369Ro0apSNHjmjkyJEqKCjQ8OHD7R4wgApGH4MAAAC4Ck8JAFbqwZFjTooEgC3KnPD86quv9NJLL+mRRx6xmP7222/rs88+0/vvv6+2bdtq1qxZJDwBwN2Q8AYAuBgGXvEczmyldq0nKvicAeVXGVqOWw6yRytbd1DmhOenn36qV155pdj0Hj166KmnnpIk3XnnnRo7tuiIi6h0SJz8ieMAB3H3L+Gid80BoCzoWxHXQhcigOOQQIar4Pqg8vIq6wKBgYFas2ZNselr1qxRYGCgJOnChQuqXbv29Udnxdy5cxURESFfX1/FxsZq165dpZZftWqVWrRoIV9fX0VFRWndunUW8w3D0IQJExQWFqbq1asrPj5eBw8etCiTlZWlgQMHys/PTwEBAUpOTtb58+ftvm9wQ5P8LV+oVI76PmDxckdRkY3ML6A8arcca/FC5RYx9mOLF4DrwDUmXAWfRQBuqMwtPMePH6+RI0dq06ZN5j48v/76a61bt07z5s2TJG3YsEFdu3a1b6SSVq5cqZSUFM2bN0+xsbGaOXOmEhISlJGRoeDg4GLlt2/frgEDBmjq1Km66667tGzZMvXu3Vt79uxRmzZtJEnTp0/XrFmztHjxYkVGRmr8+PFKSEjQjz/+KF9fX0nSwIEDdfLkSW3YsEH5+fkaNmyYHn74YS1btszu+wj3VxlaD3DHFnAsy4Q6rcSB8iIpD5dB4kgSfYO6Av4GFYyn/QCnKXPCc/jw4WrVqpXmzJmj1atXS5KaN2+uLVu2qFOnTpJkfrTd3l5//XUNHz7cPAL8vHnz9PHHH2vBggVWH6F/8803lZiYqGeeeUaSNGXKFG3YsEFz5szRvHnzZBiGZs6cqRdeeEH33HOPJGnJkiUKCQnRhx9+qP79+2v//v1av369vv76a3Xo0EGSNHv2bN1555167bXXFB4eXiH7CqDycKUkubtd9FZUvO52HFCyytCnFIDy4VzveNzAA+AKuD6sHMqc8JSkzp07q3PnzvaOpVR5eXnavXu3xo0bZ57m5eWl+Ph4paWlWV0mLS1NKSkpFtMSEhL04YcfSpKOHDmizMxMxcfHm+f7+/srNjZWaWlp6t+/v9LS0hQQEGBOdkpSfHy8vLy8tHPnTt17773Ftnvp0iVdunTJ/D43N7dc+wwArqA8CVlnDkrgiVwpKQ4AgMu5uhUdLehgA65V/8Q1JjyZTQnP3Nxc+fn5mf9fmivl7O3s2bO6fPmyQkJCLKaHhITowIEDVpfJzMy0Wj4zM9M8/8q00soUfVy+atWqCgwMNJcpaurUqZo8ebKNe+Y5it4ZsXbXvOgJteij0dbWWfTui7XHqfcdOVbqtlWOZazFa8syxe4QWbnoKnasbHhEvOi2y12myDRb/m7F1lvkvVT8LlnR98UfI+xVrMy5/dOKrbfoQDy2HKtinzNrLTiKXBgXXa+1WMrzmSn2GbeyTNFtWd2nosfXlr/TNf7W1qZZ2+9i8dnwOSs2gJKVx3mKHqui27F28WnLZ+Za5wyb9rE8n6vy/t2KnTOKH6vSzhklnRfLc6yKfcZtWG9ZzweSrNYDW+rttc6dttRbm86LthyHot8Ptvz9bTnX26neXnOfbPgb2HLeLrpea8tc63xry/G0+t1fpIzVeK/xPV7e72hnXc8U3ZbVOmhln655HKyVKcdn0Za/k01/gyJlbPmM23TtZ8M+XevvVLtl0SVsu54penwjLlp2jVX082utTlr929pwTr7W97hN53obvuvL811qt+9bGz7jReMrtt/l+D6zdj4oz/WMLcfXlnOGTdfs14j3yrpLW8bauelanzPp2vtkLQ5b/rZF12vLeaboMrZ8fxX7DNlQb+11PVNsH8t7fVCO3wu21INr1X9bfrva9FksxzWltb/tteqKTfWtHOciWz4PNp0XbThvw8aEZ506dXTy5EkFBwcrICBAJpOpWBnDMGQymXT58mW7B+luxo0bZ9GyNDc3Vw0bNnRiRACc5ZrJeABwI9Z+aKI4jpPjXfNHMeCJ3Kw1K+dGlKaiWtpa+9xdfdPpaIVs9f8UraN8NzmUTQnPL774wjwC+6ZNmyo0oJLUq1dPVapU0alTpyymnzp1SqGhoVaXCQ0NLbX8lX9PnTqlsLAwizLR0dHmMqdPn7ZYR0FBgbKyskrcro+Pj3x8fGzfOQBOV/SLMMJKGYd9MQJwea5059xePxBcaZ8qLUclL9wsSeJI7paQcbt4Oc8AABzEpoTn1SOuV8To67bw9vZWTEyMNm7cqN69e0uSCgsLtXHjRo0ePdrqMnFxcdq4caOeeOIJ87QNGzYoLi5OkhQZGanQ0FBt3LjRnODMzc3Vzp07NXLkSPM6srOztXv3bsXExEj6MwFcWFio2NjYitlZAG7L2qMY7rwdALA3zl8AysKZ5wx3SygDAP6rXIMWbd26VW+//bZ++uknrVq1SvXr19e7776ryMhIdenSxd4xmqWkpGjIkCHq0KGDOnbsqJkzZ+rChQvmUdsHDx6s+vXra+rUqZKkxx9/XF27dtWMGTPUq1cvrVixQt98843mz58vSTKZTHriiSf00ksvqVmzZoqMjNT48eMVHh5uTqq2bNlSiYmJGj58uObNm6f8/HyNHj1a/fv3Z4R2F8ZjTYBz8QMBZeWJrX6oBwDgGNxIcSxXP95F+8g9aqWMLf2dwjq7/f152gAVrMwJz/fff18PPvigBg4cqD179phHI8/JydHLL7+sdevW2T3IK/r166czZ85owoQJyszMVHR0tNavX28edOjYsWPy8vIyl+/UqZOWLVumF154Qc8//7yaNWumDz/8UG3atDGXefbZZ3XhwgU9/PDDys7OVpcuXbR+/Xr5+vqayyxdulSjR49Wjx495OXlpaSkJM2aNavC9hMAinK7izAuYACX5okJZgBAEVyPAajEypzwfOmllzRv3jwNHjxYK1asME/v3LmzXnrpJbsGZ83o0aNLfIR98+bNxab17dtXffv2LXF9JpNJL774ol588cUSywQGBmrZsmUlzgcAW1lLMtitg24uagFUMHfry9jVWyG5EpLgqGyKnR98HyhWxt3OeSgDN7tudrvGD4ALKHPCMyMjQ7feemux6f7+/srOzrZHTKhEXGkE6/Jc6PNDCtdiyyM1NilyUVZRoxgCAADXwHUmPI6bJRkBV0UC3DZlTniGhobq0KFDioiIsJj+1VdfqUmTJvaKC4CTeOLFNa1Wrg/JVQAl4scrKhl3/JFZGb7Hi+2ji5+b3PFz5CpsaUzgSp/5yvI7xJWOeWXFeaW4Mic8hw8frscff1wLFiyQyWTSiRMnlJaWpqefflrjx4+viBgBAJWAq18oFb0Z4Ik3B4oq70V6ZTg2wNX4kQG4Ple/zgBgH1yH4gqbE55HjhxRZGSkxo4dq8LCQvXo0UO///67br31Vvn4+Ojpp5/WmDFjKjJWAJUQPyJRGkf+eHHlH0rUEwCAp6gsLeIAlJ0rdYlnDecv12JzwvOGG25Q48aNddttt+m2227T/v37de7cOZ0/f16tWrVSrVq1KjJOAK7EymNC9ki4uHJCyZoKHYAIdlP0sxnhnDAqNVe/OAXKih80jsV3KwBXUPRcFLV4rJMicS7OyXAXNic8v/jiC23evFmbN2/W8uXLlZeXpyZNmqh79+7q3r27unXrppCQkIqMFQBQwSrrBUxl3e9iHNTnGcki9+AJLYftUbc94TjA9XFeRFF8JgDg+tic8OzWrZu6desmSbp48aK2b99uToAuXrxY+fn5atGihX744YeKihUA+OEJVBL80APgyTzxRpsn7lN5uPpx4PsVQGVR5kGLJMnX11fdu3dXly5ddNttt+mTTz7R22+/rQMHDtg7PgBuisQkroULbsDzlPfcz/kAAFBe/O4AYE2ZEp55eXnasWOHNm3apM2bN2vnzp1q2LChbr31Vs2ZM0ddu3atqDgBAADgSA7q4gAA3FllvGHj6q1YAUAqQ8Kze/fu2rlzpyIjI9W1a1eNGDFCy5YtU1hYWEXGBwBlVhkvPAEAAGC7Ykm7SU4JA8B14HcfSmNzwnPr1q0KCwszD1DUtWtX1a1btyJjAwDArXERBoiWoteBcwjgQFbOVUWTojw6XblEXFxm/v9RG5fhvA24DpsTntnZ2dq6das2b96sV155RQMGDNCNN96orl27mhOgQUFBFRkr3AwneyfgRyU83Ln905wdAgAApeJxX1yTC12zu9LntbImlIvud4RzwsC1uFC9hW1sTnjWrFlTiYmJSkxMlCSdO3dOX331lTZt2qTp06dr4MCBatasmb7//vsKCxYoEScfAHBZV7eQkGxvJQGgkuK6DgBQTq6UxC+KRmGOVa5R2qU/E6CBgYEKDAxUnTp1VLVqVe3fv9+esQG4mgde/NNaD6g4FdVKojJcqBVN0NbWWCdFAsAdVNZWaY7iyskLAIDrsjnhWVhYqG+++UabN2/Wpk2btG3bNl24cEH169fXbbfdprlz5+q2226ryFgBuDF+DMAe7PGjhx9OAAB3wPcV4Npc/Sawvc4h/I6Du7I54RkQEKALFy4oNDRUt912m9544w1169ZNN9xwQ0XGBxTDCRdAaThH2I5j5Viu/sPIFtx0AADAfZVnICbAXdmc8Hz11Vd122236cYbb6zIeADAo3jkD3sP7F4BAAAAcFlcfwNlZnPCc8SIERUZByDJM1q/lAetrADYmzPv4Htkoh8oRbGBuXwfcFIksDeu0QAAzlZZ8yTXq9yDFgFAaWxJtpAUATyADS0OKqquF01ERFTIVgDAw9hw3ubHNQDA3ZHwBK4DCTug8qL+Vx608ALgNnjsFQCuGzd9PAMJTwAAAOA6FLsBMskpYcBJKusgICQEAACujIQnAAAOcm7/NGeHAMARaGVXqdDiH45g09MGnHvsirpdiVB3PBIJTwCujS8fAAAAAJUEidaKxzGuHEh4AgAAAAAAq+i+AIA7IuEJ90OLPwAAKhV+bANA5UR3QADKi4QnUAq+YCsvHnOwM25UVCyOr10xAA3g/mzq7xAAAHgsEp4AYCOSoG7KA5OBdmvt5oHHBigVn3k4Ca2UAfvgehyArbycHQAAAAAAAAAA2AstPOHyuCMOANePFhEAKpOIi8ss3h91ThgA3AVPAAAeh4QnAAskRYCKQ/2qWBxfVIgK+hFMH5MAADgAyexKi4QnAPvgiwQA4MTvAhKIAAAAuIKEJ+ACeOwKgL3R2hFwLSRk3QQ3cAH3Q70FYAUJTzgU/XECAAAAAACgIpHwBFAhaF0GAC6E1i8AAACei2u9Ykh4Aqic+EIAAAAAXBqNKACUl5ezAwAAAAAAAAAAe6GFJwDAs9B6FwAsnNs/zdkhwIUxoBbsgZaYAFyN27TwzMrK0sCBA+Xn56eAgAAlJyfr/PnzpS5z8eJFjRo1SnXr1lWtWrWUlJSkU6dOWZQ5duyYevXqpRo1aig4OFjPPPOMCgoKzPM3b94sk8lU7JWZmVkh+4nK6ei0XhYvAAAAAACAYiblWL5gldu08Bw4cKBOnjypDRs2KD8/X8OGDdPDDz+sZcuWlbjMk08+qY8//lirVq2Sv7+/Ro8erT59+mjbtm2SpMuXL6tXr14KDQ3V9u3bdfLkSQ0ePFjVqlXTyy+/bLGujIwM+fn5md8HBwdXzI4CAACg7Fz9gt/V44N74nMFAIBVbpHw3L9/v9avX6+vv/5aHTp0kCTNnj1bd955p1577TWFh4cXWyYnJ0epqalatmyZunfvLklauHChWrZsqR07dujmm2/WZ599ph9//FGff/65QkJCFB0drSlTpui5557TpEmT5O3tbV5fcHCwAgICbIr30qVLunTpkvl9bm7udew9AAAAAKcomlCc5O+cOIDKgAQ+ADtyi0fa09LSFBAQYE52SlJ8fLy8vLy0c+dOq8vs3r1b+fn5io+PN09r0aKFGjVqpLS0NPN6o6KiFBISYi6TkJCg3Nxc/fDDDxbri46OVlhYmG6//XZzC9GSTJ06Vf7+/uZXw4YNy7zPAADgGnicBwAAAIAVbtHCMzMzs9gj5FWrVlVgYGCJfWlmZmbK29u7WKvMkJAQ8zKZmZkWyc4r86/Mk6SwsDDNmzdPHTp00KVLl/TOO++oW7du2rlzp2666Sar2x43bpxSUlLM73Nzc0l6AgAAANycAAAADuDUhOfYsWP1yiuvlFpm//79DorGuubNm6t58+bm9506ddLhw4f1xhtv6N1337W6jI+Pj3x8fBwVIgAAAAAAAID/49SE51NPPaWhQ4eWWqZJkyYKDQ3V6dOnLaYXFBQoKytLoaGhVpcLDQ1VXl6esrOzLVp5njp1yrxMaGiodu3aZbHclVHcS1qvJHXs2FFfffVVqXEDqCC0DAHgYBEXLQdIPOqcMAAAAADYyKkJz6CgIAUFBV2zXFxcnLKzs7V7927FxMRIkr744gsVFhYqNjbW6jIxMTGqVq2aNm7cqKSkJEl/jrR+7NgxxcXFmdf7t7/9TadPnzY/Mr9hwwb5+fmpVatWJcaTnp6usLCwMu0r4NZIMgIAAKCy45oYANyGW/Th2bJlSyUmJmr48OGaN2+e8vPzNXr0aPXv3988Qvvx48fVo0cPLVmyRB07dpS/v7+Sk5OVkpKiwMBA+fn5acyYMYqLi9PNN98sSerZs6datWqlBx98UNOnT1dmZqZeeOEFjRo1yvxI+syZMxUZGanWrVvr4sWLeuedd/TFF1/os88+c9rx8CjOvGjgggUAAAAAAMDjuEXCU5KWLl2q0aNHq0ePHvLy8lJSUpJmzZplnp+fn6+MjAz9/vvv5mlvvPGGueylS5eUkJCgt956yzy/SpUqWrt2rUaOHKm4uDjVrFlTQ4YM0Ysvvmguk5eXp6eeekrHjx9XjRo11LZtW33++ee67bbbHLPjAAAA8Dh0lQAAAFBx3CbhGRgYqGXLlpU4PyIiQoZhWEzz9fXV3LlzNXfu3BKXa9y4sdatW1fi/GeffVbPPvts2QMGAAAAAAAA4HBuk/AEAAAAAFscndbL2SEAAAAnIuEJAAAAOJgnJuT2HTnm7BAAAAAkkfBEJcEFOK6JQawAAMC1cL0AAIBbIOEJAIAr4cc0AAAoiusDACgTEp5AWXChAQDuhfM2AAAAUOmQ8AQAAAAAAA5Ft2MAKpKXswMAAAAAAAAAAHsh4QkAAAAAAADAY/BIOwAAAABUVvR1DADwQLTwBAAAAAAAAOAxSHgCAAAAAAAA8Bg80g4AAAAAAADXRhccKAMSngAAAACuHz9EAQCAi+CRdgAAAAAAAAAegxaeAByDVh8AAFy3fUeOOTsEAICbOjqtl7NDAByGhCcAAAAAALAfGjsAcDIeaQcAAAAAAADgMUh4AgAAAAAAAPAYJDwBAAAAAAAAeAwSngAAAAAAAAA8BoMWAQAAAAAAoDgGoIKbIuEJAAAAAAAci0QagArEI+0AAAAAAAAAPAYJTwAAAAAAAAAeg0faAQAASnF0Wq/yLcijegAAAIBT0MITAAAAAAAAgMcg4QkAAAAAAADAY5DwBAAAAAAAAOAxSHgCAAAAAAAA8BgkPAEAAAAAAAB4DBKeAAAAAAAAADxGVWcHALiyo9N6OTsEAAAAAAAAlAEtPAEAAAAAAAB4DBKeAAAAAAAAADwGCU8AAAAAAAAAHoOEJwAAAAAAAACP4TYJz6ysLA0cOFB+fn4KCAhQcnKyzp8/X+oyFy9e1KhRo1S3bl3VqlVLSUlJOnXqlEWZxx57TDExMfLx8VF0dLTV9ezdu1e33HKLfH191bBhQ02fPt1euwUAAAAAAADAjtwm4Tlw4ED98MMP2rBhg9auXasvv/xSDz/8cKnLPPnkk1qzZo1WrVqlLVu26MSJE+rTp0+xcg899JD69etndR25ubnq2bOnGjdurN27d+vVV1/VpEmTNH/+fLvsFwAAAAAAAAD7qersAGyxf/9+rV+/Xl9//bU6dOggSZo9e7buvPNOvfbaawoPDy+2TE5OjlJTU7Vs2TJ1795dkrRw4UK1bNlSO3bs0M033yxJmjVrliTpzJkz2rt3b7H1LF26VHl5eVqwYIG8vb3VunVrpaen6/XXX79mwhUAAAAAAACAY7lFC8+0tDQFBASYk52SFB8fLy8vL+3cudPqMrt371Z+fr7i4+PN01q0aKFGjRopLS2tTNu+9dZb5e3tbZ6WkJCgjIwM/fbbb1aXuXTpknJzcy1eAAAAAAAAACqeW7TwzMzMVHBwsMW0qlWrKjAwUJmZmSUu4+3trYCAAIvpISEhJS5T0noiIyOLrePKvDp16hRbZurUqZo8ebLN24ADTMpxdgQAAAAAAABwAKe28Bw7dqxMJlOprwMHDjgzxHIZN26ccnJyzK+ff/7Z2SEBAAAAAAAAlYJTW3g+9dRTGjp0aKllmjRpotDQUJ0+fdpiekFBgbKyshQaGmp1udDQUOXl5Sk7O9uileepU6dKXKak9RQd2f3K+5LW4+PjIx8fH5u3AQAAAAAAAMA+nJrwDAoKUlBQ0DXLxcXFKTs7W7t371ZMTIwk6YsvvlBhYaFiY2OtLhMTE6Nq1app48aNSkpKkiRlZGTo2LFjiouLsznGuLg4/b//9/+Un5+vatWqSZI2bNig5s2bW32cHQAAAAAAAIDzuMWgRS1btlRiYqKGDx+uXbt2adu2bRo9erT69+9vHqH9+PHjatGihXbt2iVJ8vf3V3JyslJSUrRp0ybt3r1bw4YNU1xcnHmEdkk6dOiQ0tPTlZmZqT/++EPp6elKT09XXl6eJOmBBx6Qt7e3kpOT9cMPP2jlypV68803lZKS4vgDAQAAAAAAAKBUbjFokSQtXbpUo0ePVo8ePeTl5aWkpCTNmjXLPD8/P18ZGRn6/fffzdPeeOMNc9lLly4pISFBb731lsV6//rXv2rLli3m9+3bt5ckHTlyRBEREfL399dnn32mUaNGKSYmRvXq1dOECRP08MMPV/AeAwAAANfAwIwAAADFmAzDMJwdhKfLzc2Vv7+/cnJy5Ofn5+xwAAAA4K4m+V/1f9uSnRFjPzb//+i0XvaOCABwPa4+r0vcyAJKUZb8mtu08AQAAAAAAPAoJDiBCuEWfXgCAAAAAAAAgC1IeAIAAAAAAADwGDzSDgAAAHgw+u0EAACVDS08AQAAAAAAAHgMEp4AAAAAAAAAPAYJTwAAAAAAAAAeg4QnAAAAAAAAAI9BwhMAAAAAAACAxyDhCQAAAAAAAMBjkPAEAAAAAAAA4DFIeAIAAAAAAADwGCQ8AQAAAAAAAHgMEp4AAAAAAAAAPAYJTwAAAAAAAAAeg4QnAAAAAAAAAI9BwhMAAAAAAACAxyDhCQAAAAAAAMBjkPAEAAAAAAAA4DFIeAIAAAAAAADwGCQ8AQAAAAAAAHgMEp4AAAAAAAAAPAYJTwAAAAAAAAAeg4QnAAAAAAAAAI9BwhMAAAAAAACAxyDhCQAAAAAAAMBjkPAEAAAAAAAA4DFIeAIAAAAAAADwGCQ8AQAAAAAAAHgMEp4AAAAAAAAAPAYJTwAAAAAAAAAeg4QnAAAAAAAAAI9BwhMAAAAAAACAxyDhCQAAAAAAAMBjkPAEAAAAAAAA4DFIeAIAAAAAAADwGFWdHQAAAAAAG03KcXYEAAAALs9tWnhmZWVp4MCB8vPzU0BAgJKTk3X+/PlSl7l48aJGjRqlunXrqlatWkpKStKpU6csyjz22GOKiYmRj4+PoqOji63j6NGjMplMxV47duyw5+4BAAAAAAAAsAO3SXgOHDhQP/zwgzZs2KC1a9fqyy+/1MMPP1zqMk8++aTWrFmjVatWacuWLTpx4oT69OlTrNxDDz2kfv36lbquzz//XCdPnjS/YmJirmt/AAAAAAAAANifWzzSvn//fq1fv15ff/21OnToIEmaPXu27rzzTr322msKDw8vtkxOTo5SU1O1bNkyde/eXZK0cOFCtWzZUjt27NDNN98sSZo1a5Yk6cyZM9q7d2+JMdStW1ehoaH23jUAAAAAAAAAduQWLTzT0tIUEBBgTnZKUnx8vLy8vLRz506ry+zevVv5+fmKj483T2vRooUaNWqktLS0MsfwP//zPwoODlaXLl300UcflVr20qVLys3NtXgBAAAAAAAAqHhukfDMzMxUcHCwxbSqVasqMDBQmZmZJS7j7e2tgIAAi+khISElLmNNrVq1NGPGDK1atUoff/yxunTpot69e5ea9Jw6dar8/f3Nr4YNG9q8PQAAAAAAAADl59SE59ixY60OCHT168CBA84MUfXq1VNKSopiY2P1l7/8RdOmTdOgQYP06quvlrjMuHHjlJOTY379/PPPDowYAAAAAAAAqLyc2ofnU089paFDh5ZapkmTJgoNDdXp06ctphcUFCgrK6vEfjVDQ0OVl5en7Oxsi1aep06duu6+OGNjY7Vhw4YS5/v4+MjHx+e6tgEAAAAAAACg7Jya8AwKClJQUNA1y8XFxSk7O1u7d+82j47+xRdfqLCwULGxsVaXiYmJUbVq1bRx40YlJSVJkjIyMnTs2DHFxcVdV9zp6ekKCwu7rnUAAAAAAAAAsD+3GKW9ZcuWSkxM1PDhwzVv3jzl5+dr9OjR6t+/v3mE9uPHj6tHjx5asmSJOnbsKH9/fyUnJyslJUWBgYHy8/PTmDFjFBcXZx6hXZIOHTqk8+fPKzMzU3/88YfS09MlSa1atZK3t7cWL14sb29vtW/fXpK0evVqLViwQO+8847DjwMAAAAAAACA0rlFwlOSli5dqtGjR6tHjx7y8vJSUlKSZs2aZZ6fn5+vjIwM/f777+Zpb7zxhrnspUuXlJCQoLfeestivX/961+1ZcsW8/sric0jR44oIiJCkjRlyhT95z//UdWqVdWiRQutXLlS9913XwXuLQAAAAAAAIDyMBmGYTg7CE+Xm5srf39/5eTkyM/Pz9nhAAAAAAAAAG6lLPk1p47SDgAAAAAAAAD25DaPtLuzK41oc3NznRwJAAAAAAAA4H6u5NVseVidhKcDnDt3TpLUsGFDJ0cCAAAAAAAAuK9z587J39+/1DL04ekAhYWFOnHihGrXri2TyeTscCpUbm6uGjZsqJ9//pn+SoEKRn0DHIf6BjgWdQ5wHOob4DjUt+tjGIbOnTun8PBweXmV3ksnLTwdwMvLSw0aNHB2GA7l5+dH5QUchPoGOA71DXAs6hzgONQ3wHGob+V3rZadVzBoEQAAAAAAAACPQcITAAAAAAAAgMcg4Qm78vHx0cSJE+Xj4+PsUACPR30DHIf6BjgWdQ5wHOob4DjUN8dh0CIAAAAAAAAAHoMWngAAAAAAAAA8BglPAAAAAAAAAB6DhCcAAAAAAAAAj0HCEwAAAAAAAIDHIOEJu5k7d64iIiLk6+ur2NhY7dq1y9khAW5v0qRJMplMFq8WLVqY51+8eFGjRo1S3bp1VatWLSUlJenUqVNOjBhwL19++aXuvvtuhYeHy2Qy6cMPP7SYbxiGJkyYoLCwMFWvXl3x8fE6ePCgRZmsrCwNHDhQfn5+CggIUHJyss6fP+/AvQDcw7Xq29ChQ4t95yUmJlqUob4Btpk6dar+8pe/qHbt2goODlbv3r2VkZFhUcaW68hjx46pV69eqlGjhoKDg/XMM8+ooKDAkbsCuDxb6lu3bt2Kfcc98sgjFmWob/ZFwhN2sXLlSqWkpGjixInas2eP2rVrp4SEBJ0+fdrZoQFur3Xr1jp58qT59dVXX5nnPfnkk1qzZo1WrVqlLVu26MSJE+rTp48TowXcy4ULF9SuXTvNnTvX6vzp06dr1qxZmjdvnnbu3KmaNWsqISFBFy9eNJcZOHCgfvjhB23YsEFr167Vl19+qYcffthRuwC4jWvVN0lKTEy0+M5bvny5xXzqG2CbLVu2aNSoUdqxY4c2bNig/Px89ezZUxcuXDCXudZ15OXLl9WrVy/l5eVp+/btWrx4sRYtWqQJEyY4Y5cAl2VLfZOk4cOHW3zHTZ8+3TyP+lYBDMAOOnbsaIwaNcr8/vLly0Z4eLgxdepUJ0YFuL+JEyca7dq1szovOzvbqFatmrFq1SrztP379xuSjLS0NAdFCHgOScYHH3xgfl9YWGiEhoYar776qnladna24ePjYyxfvtwwDMP48ccfDUnG119/bS7zySefGCaTyTh+/LjDYgfcTdH6ZhiGMWTIEOOee+4pcRnqG1B+p0+fNiQZW7ZsMQzDtuvIdevWGV5eXkZmZqa5zN///nfDz8/PuHTpkmN3AHAjReubYRhG165djccff7zEZahv9kcLT1y3vLw87d69W/Hx8eZpXl5eio+PV1pamhMjAzzDwYMHFR4eriZNmmjgwIE6duyYJGn37t3Kz8+3qHstWrRQo0aNqHuAHRw5ckSZmZkWdczf31+xsbHmOpaWlqaAgAB16NDBXCY+Pl5eXl7auXOnw2MG3N3mzZsVHBys5s2ba+TIkfr111/N86hvQPnl5ORIkgIDAyXZdh2ZlpamqKgohYSEmMskJCQoNzdXP/zwgwOjB9xL0fp2xdKlS1WvXj21adNG48aN0++//26eR32zv6rODgDu7+zZs7p8+bJFxZSkkJAQHThwwElRAZ4hNjZWixYtUvPmzXXy5ElNnjxZt9xyi77//ntlZmbK29tbAQEBFsuEhIQoMzPTOQEDHuRKPbL2/XZlXmZmpoKDgy3mV61aVYGBgdRDoIwSExPVp08fRUZG6vDhw3r++ed1xx13KC0tTVWqVKG+AeVUWFioJ554Qp07d1abNm0kyabryMzMTKvfgVfmASjOWn2TpAceeECNGzdWeHi49u7dq+eee04ZGRlavXq1JOpbRSDhCQAu7I477jD/v23btoqNjVXjxo313nvvqXr16k6MDAAA++rfv7/5/1FRUWrbtq1uuOEGbd68WT169HBiZIB7GzVqlL7//nuLfuABVIyS6tvV/U1HRUUpLCxMPXr00OHDh3XDDTc4OsxKgUfacd3q1aunKlWqFBvR79SpUwoNDXVSVIBnCggI0I033qhDhw4pNDRUeXl5ys7OtihD3QPs40o9Ku37LTQ0tNgAfQUFBcrKyqIeAtepSZMmqlevng4dOiSJ+gaUx+jRo7V27Vpt2rRJDRo0ME+35ToyNDTU6nfglXkALJVU36yJjY2VJIvvOOqbfZHwxHXz9vZWTEyMNm7caJ5WWFiojRs3Ki4uzomRAZ7n/PnzOnz4sMLCwhQTE6Nq1apZ1L2MjAwdO3aMugfYQWRkpEJDQy3qWG5urnbu3GmuY3FxccrOztbu3bvNZb744gsVFhaaL2QBlM8vv/yiX3/9VWFhYZKob0BZGIah0aNH64MPPtAXX3yhyMhIi/m2XEfGxcVp3759FjcaNmzYID8/P7Vq1coxOwK4gWvVN2vS09MlyeI7jvpmXzzSDrtISUnRkCFD1KFDB3Xs2FEzZ87UhQsXNGzYMGeHBri1p59+WnfffbcaN26sEydOaOLEiapSpYoGDBggf39/JScnKyUlRYGBgfLz89OYMWMUFxenm2++2dmhA27h/Pnz5jvr0p8DFaWnpyswMFCNGjXSE088oZdeeknNmjVTZGSkxo8fr/DwcPXu3VuS1LJlSyUmJmr48OGaN2+e8vPzNXr0aPXv31/h4eFO2ivANZVW3wIDAzV58mQlJSUpNDRUhw8f1rPPPqumTZsqISFBEvUNKItRo0Zp2bJl+ve//63atWub+wD09/dX9erVbbqO7Nmzp1q1aqUHH3xQ06dPV2Zmpl544QWNGjVKPj4+ztw9wKVcq74dPnxYy5Yt05133qm6detq7969evLJJ3Xrrbeqbdu2kqhvFcLZw8TDc8yePdto1KiR4e3tbXTs2NHYsWOHs0MC3F6/fv2MsLAww9vb26hfv77Rr18/49ChQ+b5f/zxh/Hoo48aderUMWrUqGHce++9xsmTJ50YMeBeNm3aZEgq9hoyZIhhGIZRWFhojB8/3ggJCTF8fHyMHj16GBkZGRbr+PXXX40BAwYYtWrVMvz8/Ixhw4YZ586dc8LeAK6ttPr2+++/Gz179jSCgoKMatWqGY0bNzaGDx9uZGZmWqyD+gbYxlpdk2QsXLjQXMaW68ijR48ad9xxh1G9enWjXr16xlNPPWXk5+c7eG8A13at+nbs2DHj1ltvNQIDAw0fHx+jadOmxjPPPGPk5ORYrIf6Zl8mwzAMRyZYAQAAAAAAAKCi0IcnAAAAAAAAAI9BwhMAAAAAAACAxyDhCQAAAAAAAMBjkPAEAAAAAAAA4DFIeAIAAAAAAADwGCQ8AQAAAAAAAHgMEp4AAAAAAAAAPAYJTwAAAAAAAAAeg4QnAAAAXJbJZNKHH37o7DBsMnToUPXu3dvZYQAAAFR6JDwBAADgFJmZmRozZoyaNGkiHx8fNWzYUHfffbc2btzo7NAqxObNm2UymZSdnW3x3mQyycvLS/7+/mrfvr2effZZnTx50rnBAgAAuLGqzg4AAAAAlc/Ro0fVuXNnBQQE6NVXX1VUVJTy8/P16aefatSoUTpw4ICzQ3SYjIwM+fn5KTc3V3v27NH06dOVmpqqzZs3KyoqytnhAQAAuB1aeAIAAMDhHn30UZlMJu3atUtJSUm68cYb1bp1a6WkpGjHjh0WZc+ePat7771XNWrUULNmzfTRRx+Z512+fFnJycmKjIxU9erV1bx5c7355psWy1951Py1115TWFiY6tatq1GjRik/P99cJiIiQi+//LIeeugh1a5dW40aNdL8+fMt1vPzzz/r/vvvV0BAgAIDA3XPPffo6NGj130sgoODFRoaqhtvvFH9+/fXtm3bFBQUpJEjR173ugEAACojEp4AAABwqKysLK1fv16jRo1SzZo1i80PCAiweD958mTdf//92rt3r+68804NHDhQWVlZkqTCwkI1aNBAq1at0o8//qgJEybo+eef13vvvWexjk2bNunw4cPatGmTFi9erEWLFmnRokUWZWbMmKEOHTro22+/1aOPPqqRI0cqIyNDkpSfn6+EhATVrl1bW7du1bZt21SrVi0lJiYqLy/PfgdHUvXq1fXII49o27ZtOn36tF3XDQAAUBmQ8AQAAIBDHTp0SIZhqEWLFjaVHzp0qAYMGKCmTZvq5Zdf1vnz57Vr1y5JUrVq1TR58mR16NBBkZGRGjhwoIYNG1Ys4VmnTh3NmTNHLVq00F133aVevXoV6yv0zjvv1KOPPqqmTZvqueeeU7169bRp0yZJ0sqVK1VYWKh33nlHUVFRatmypRYuXKhjx45p8+bN139QirhybOzRghQAAKCyoQ9PAAAAOJRhGGUq37ZtW/P/a9asKT8/P4uWj3PnztWCBQt07Ngx/fHHH8rLy1N0dLTFOlq3bq0qVaqY34eFhWnfvn0lbsdkMik0NNS8ne+++06HDh1S7dq1LZa5ePGiDh8+XKb9scWVY2Qymey+bgAAAE9HwhMAAAAO1axZM5lMJpsHJqpWrZrFe5PJpMLCQknSihUr9PTTT2vGjBmKi4tT7dq19eqrr2rnzp02r8OWMufPn1dMTIyWLl1aLL6goCCb9qMs9u/fL+nPvkUBAABQNiQ8AQAA4FCBgYFKSEjQ3Llz9dhjjxXrxzM7O7tYP54l2bZtmzp16qRHH33UPK0iWlzedNNNWrlypYKDg+Xn52f39V/tjz/+0Pz583XrrbdWSDIVAADA09GHJwAAABxu7ty5unz5sjp27Kj3339fBw8e1P79+zVr1izFxcXZvJ5mzZrpm2++0aeffqr//d//1fjx4/X111/bPd6BAweqXr16uueee7R161YdOXJEmzdv1mOPPaZffvnlutZ9+vRpZWZm6uDBg1qxYoU6d+6ss2fP6u9//7udogcAAKhcaOEJAAAAh2vSpIn27Nmjv/3tb3rqqad08uRJBQUFKSYmpkyJvhEjRujbb79Vv379ZDKZNGDAAD366KP65JNP7BpvjRo19OWXX+q5555Tnz59dO7cOdWvX189evS47hafzZs3l8lkUq1atdSkSRP17NlTKSkpCg0NtVP0AAAAlYvJKGuv8QAAAAAAAADgonikHQAAAAAAAIDHIOEJAAAAAAAAwGOQ8AQAAAAAAADgMUh4AgAAAAAAAPAYJDwBAAAAAAAAeAwSngAAAAAAAAA8BglPAAAAAAAAAB6DhCcAAAAAAAAAj0HCEwAAAAAAAIDHIOEJAAAAAAAAwGOQ8AQAAAAAAADgMf4/ATP2mVzNkD4AAAAASUVORK5CYII=",
      "text/plain": [
       "<Figure size 1600x400 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Import phone index ordering so we know which entry corresponds to which phone\n",
    "from neuralDecoder.datasets.speechDataset import PHONE_DEF\n",
    "\n",
    "# collapse gradients across timepoints and remove CTC blank component (index 40):\n",
    "jacobians = np.dstack(gradients).squeeze().mean(axis=1)[:40, :]\n",
    "\n",
    "# some plotting stuff\n",
    "import matplotlib.pyplot as plt\n",
    "import matplotlib.cm as cm\n",
    "from matplotlib.patches import Rectangle\n",
    "\n",
    "plt.figure(figsize=(16, 4))\n",
    "\n",
    "_ = plt.bar(np.arange(256), jacobians[0, :], label = PHONE_DEF[0])\n",
    "_ = plt.bar(np.arange(256), jacobians[1, :], label = PHONE_DEF[1])\n",
    "_ = plt.bar(np.arange(256), jacobians[2, :], label = PHONE_DEF[2])\n",
    "plt.xlabel('Channel ID')\n",
    "plt.ylabel('Weight')\n",
    "plt.title('Example saliency weights')\n",
    "plt.legend()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1e582577",
   "metadata": {},
   "source": [
    "Next we'll average across time and compute a similarity score between sets of gradients."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "8e2d3923",
   "metadata": {},
   "outputs": [],
   "source": [
    "# First define consonant and vowel groupings: \n",
    "\n",
    "CONSONANTS_GRPS = [['P', 'B', 'M', 'F', 'V'], # labial\n",
    "                   ['TH', 'DH'],              # dental\n",
    "                   ['T', 'D', 'N', 'S', 'Z'], # alveolar\n",
    "                   ['CH', 'JH', 'SH', 'ZH'],  # post-alveolar\n",
    "                   ['NG', 'K', 'G'],          # velar/palatal\n",
    "                   ['Y', 'L', 'R', 'W'],      # approximant\n",
    "                   ['HH']]                    # glottal fricative\n",
    "\n",
    "VOWELS_GRPS = [['OY', 'OW', 'AO', 'AA', 'AW', 'AY', 'AE', 'EH', 'EY', 'IY', 'IH', 'AH', 'UW', 'ER', 'UH']]\n",
    "\n",
    "\n",
    "CONSONANTS = np.concatenate(CONSONANTS_GRPS)\n",
    "VOWELS     = np.concatenate(VOWELS_GRPS)\n",
    "\n",
    "VOWELS_IDX     = [PHONE_DEF.index(ph) for ph in VOWELS]\n",
    "CONSONANTS_IDX = [PHONE_DEF.index(ph) for ph in CONSONANTS]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "4c7b795e",
   "metadata": {},
   "outputs": [],
   "source": [
    "metric       = 'similarity' # distance metric between gradients\n",
    "centerGroups = True         # toggle mean subtraction within consonants and vowels separately\n",
    "\n",
    "\n",
    "# optional - center within groups to highlight within-group structure\n",
    "if centerGroups:\n",
    "    for grp_idx in [CONSONANTS_IDX, VOWELS_IDX]:\n",
    "        jacobians[grp_idx, :] -= np.mean(jacobians[grp_idx, :], axis = 0, keepdims=True)\n",
    "\n",
    "# compute similarity between pairs of phoneme gradients:\n",
    "dists = getSimilarity(jacobians, subtractMean=False, metric = metric)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ad8b48fe",
   "metadata": {},
   "source": [
    "Finally let's plot the resulting distance/similarity matrices:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "c65cd8d8",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAA0IAAAHbCAYAAADmqCrWAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8pXeV/AAAACXBIWXMAAA9hAAAPYQGoP6dpAACrPUlEQVR4nOzdeVxU1fsH8M+wDZuACgYogoiCC+aKuAJu4JaauZuiZqammVmKLWqLZGVZmmaJon7V1DIXLLcEtUSNDE1FXFFwlx2EYTu/P4z5MbLI3JkLAp/363VfOmfO85zDMMzMM/fecxVCCAEiIiIiIqIaxKCyJ0BERERERFTRWAgREREREVGNw0KIiIiIiIhqHBZCRERERERU47AQIiIiIiKiGoeFEBERERER1TgshIiIiIiIqMZhIURERERERDUOCyEiIiIiIqpxWAgRUZUXGBgIFxcXjTaFQoGFCxdWyny0FRoaCoVCgbi4OL3mLelxqSj6evzleGwWLlwIhUKh0ebi4oLAwEC9jQEAERERUCgUiIiI0GveqoaPAxE9q1gIEZFk169fx+uvv46mTZvC3Nwc5ubmaN68OaZPn46zZ89W9vRkt3nzZixbtqyyp0FVSFV/zlT1+RMRFaUQQojKngQRVT1hYWEYMWIEjIyMMGbMGDz//PMwMDDAxYsXsWPHDty4cQPXr1+Hs7Oz7HMJDAxERESExl6D7OxsGBkZwcjISLZxBwwYgHPnzum8tyI0NBQTJkzA9evX9boHJzc3FwUFBVAqlXrLWV4KhQILFizQea9Qfn4+cnNzoVQqi+3FkSovLw95eXkwNTVVt7m4uMDX1xehoaF6GQMACgoKkJOTAxMTExgYPP7eUV/PmcoiZf4lPQ5ERM8C+T4hEFG1dfXqVYwcORLOzs74/fff4eDgoHH/kiVLsHLlyqd+6MnMzISFhYUscyz6IbemKXxcjY2NK3sqOjM0NIShoaFec8pdIGdnZ6s/9Nfk5yEfByJ61vGrGSLS2meffYbMzEysW7euWBEEPP6gOXPmTDg5OanbAgMDYWlpiatXr6Jfv36oVasWxowZAwA4duwYhg0bhoYNG0KpVMLJyQlvvvkmsrKyiuXeuXMnWrZsCVNTU7Rs2RK//PJLiXMs6RyVW7duYeLEiXjuueegVCrRokULrF27VqNP4fkM27ZtwyeffIIGDRrA1NQUPXv2xJUrV9T9fH19sXfvXty4cQMKhQIKhaJce3POnz+PHj16wMzMDA0aNMDHH3+MgoKCEvv+9ttv6NatGywsLFCrVi30798f58+f1+hT1uNa9Byh3Nxc1KlTBxMmTCg2TlpaGkxNTTFnzhx1m0qlwoIFC+Dm5qb+nbzzzjtQqVQasSqVCm+++Sbs7OxQq1YtvPDCC0hISHjq41Bo+fLlaNGiBczNzVG7dm20b98emzdvVt9f0jlCLi4uGDBgACIiItC+fXuYmZnB09NTfQ7Kjh074OnpCVNTU7Rr1w7//POPxpglnSP0pKSkJMyZMweenp6wtLSElZUV+vbtizNnzmj0K3y+/Pjjj3jvvfdQv359mJubIy0trdi5MaU9ZzIyMmBhYYE33nij2DwSEhJgaGiI4ODgUucaFxcHhUKBL774At9++y1cXV1hbm6OPn36ID4+HkIIfPTRR2jQoAHMzMwwaNAgJCUlaeTYtWsX+vfvD0dHRyiVSjRu3BgfffQR8vPz1X3Kes5r8zgUOnnyJPr164fatWvDwsICrVq1wtdff63R5+LFi3jppZdQp04dmJqaon379ti9e7dGn9zcXCxatAhNmjSBqakp6tati65du+LgwYOlPmZERAD3CBGRBGFhYXBzc0PHjh21isvLy4O/vz+6du2KL774Aubm5gCA7du349GjR5g6dSrq1q2LU6dOYfny5UhISMD27dvV8QcOHMDQoUPRvHlzBAcHIzExERMmTECDBg2eOva9e/fg7e0NhUKB119/HXZ2dvjtt98wadIkpKWlYdasWRr9P/30UxgYGGDOnDlITU3FZ599hjFjxuDkyZMAgHfffRepqalISEjAV199BQCwtLQscw53796Fn58f8vLyMG/ePFhYWOD777+HmZlZsb4bN27E+PHj4e/vjyVLluDRo0dYtWoVunbtin/++Uej6CrtcS3K2NgYQ4YMwY4dO7B69WqYmJio79u5cydUKhVGjhwJ4PGhTC+88AL++OMPvPrqq2jWrBn+/fdffPXVV7h06RJ27typjn3llVfwv//9D6NHj0bnzp1x+PBh9O/fv8zHodAPP/yAmTNn4qWXXsIbb7yB7OxsnD17FidPnsTo0aPLjL1y5QpGjx6NKVOmYOzYsfjiiy8wcOBAfPfdd5g/fz6mTZsGAAgODsbw4cMRGxur1WFZ165dw86dOzFs2DA0atQI9+7dw+rVq+Hj44MLFy7A0dFRo/9HH30EExMTzJkzByqVSuPxLVTac8bS0hJDhgzB1q1b8eWXX2rsAduyZQuEEOritiybNm1CTk4OZsyYgaSkJHz22WcYPnw4evTogYiICMydOxdXrlzB8uXLMWfOHI0vAUJDQ2FpaYnZs2fD0tIShw8fxgcffIC0tDR8/vnnZc5f28cBAA4ePIgBAwbAwcEBb7zxBuzt7RETE4OwsDB1QXj+/Hl06dIF9evXV/+9bNu2DYMHD8bPP/+MIUOGAHhc2AYHB+OVV16Bl5cX0tLSEBUVhdOnT6N3795PfdyIqAYTRERaSE1NFQDE4MGDi92XnJwsHjx4oN4ePXqkvm/8+PECgJg3b16xuKL9CgUHBwuFQiFu3LihbmvdurVwcHAQKSkp6rYDBw4IAMLZ2VkjHoBYsGCB+vakSZOEg4ODePjwoUa/kSNHCmtra/UcwsPDBQDRrFkzoVKp1P2+/vprAUD8+++/6rb+/fsXG7css2bNEgDEyZMn1W33798X1tbWAoC4fv26EEKI9PR0YWNjIyZPnqwRf/fuXWFtba3RXtbjOn78eI357d+/XwAQe/bs0ejXr18/4erqqr69ceNGYWBgII4dO6bR77vvvhMAxJ9//imEECI6OloAENOmTdPoN3r06GKPf0kGDRokWrRoUWafdevWaTw2Qgjh7OwsAIjjx48X+9nMzMw0njOrV68WAER4eLi6bcGCBeLJtz9nZ2cxfvx49e3s7GyRn5+v0ef69etCqVSKDz/8UN1W+HxxdXUt9jwuvK/o2KU9Zwrn/9tvv2m0t2rVSvj4+BTr/+S8AAg7OzuNv42goCABQDz//PMiNzdX3T5q1ChhYmIisrOz1W0l/Q1OmTJFmJuba/Qrbf7aPA55eXmiUaNGwtnZWSQnJ2v0LSgoUP+/Z8+ewtPTU2P8goIC0blzZ9GkSRN12/PPPy/69+9fyqNDRFQ6HhpHRFpJS0sDUPLeD19fX9jZ2am3b7/9tlifqVOnFmsrukckMzMTDx8+ROfOnSGEUB/WdOfOHURHR2P8+PGwtrZW9+/duzeaN29e5pyFEPj5558xcOBACCHw8OFD9ebv74/U1FScPn1aI2bChAka32Z369YNwOM9BVL9+uuv8Pb2hpeXl7rNzs6u2Lf9Bw8eREpKCkaNGqUxV0NDQ3Ts2BHh4eHFcpf0uD6pR48esLW1xdatW9VtycnJOHjwIEaMGKFu2759O5o1awYPDw+N8Xv06AEA6vF//fVXAMDMmTM1xnly71ppbGxskJCQgL/++qtc/Ytq3rw5OnXqpL5duHeyR48eaNiwYbF2bX9vSqVSvQcpPz8fiYmJsLS0hLu7e7HnCgCMHz++xD175dWrVy84Ojpi06ZN6rZz587h7NmzGDt2bLlyDBs2TONvo/BnHzt2rMY5UR07dkROTg5u3bqlbis69/T0dDx8+BDdunXDo0ePcPHixXL/HOV5HP755x9cv34ds2bNgo2NjcZ9hYcsJiUl4fDhwxg+fLh6Pg8fPkRiYiL8/f1x+fJl9fxtbGxw/vx5XL58udzzJCICeGgcEWmpVq1aAICMjIxi961evRrp6em4d+9eiR/ejIyMSjyM7ebNm/jggw+we/duJCcna9yXmpoKALhx4wYAoEmTJsXiS/twWujBgwdISUnB999/j++//77EPvfv39e4XfTDNADUrl0bAIrNryR3797VuG1tbQ0zMzPcuHGjxMMJ3d3dNW4XfqArLDyeZGVlpXG7tMf1SUZGRhg6dCg2b94MlUoFpVKJHTt2IDc3V6MQunz5MmJiYmBnZ1dinsLH6saNGzAwMEDjxo3L/HlKM3fuXBw6dAheXl5wc3NDnz59MHr0aHTp0uWpsU/+fgoLgKLnpRVtL8/vraiCggJ8/fXXWLlyJa5fv65xrkzdunWL9W/UqJFW+Z9kYGCAMWPGYNWqVXj06BHMzc2xadMmmJqaYtiwYeXKoctjcv78ebz33ns4fPiw+suOQoV/g+VRnsfh6tWrAICWLVuW2ufKlSsQQuD999/H+++/X2Kf+/fvo379+vjwww8xaNAgNG3aFC1btkRAQABefvlltGrVqtzzJqKaiYUQEWnF2toaDg4OOHfuXLH7Cj/kl7a0btFv2Qvl5+ejd+/eSEpKwty5c+Hh4QELCwvcunULgYGBpS4koI3CHGPHjsX48eNL7PPkh6bSVioT5bjiwJMLSKxbt06ri3UWznfjxo2wt7cvdv+TK56V9LiWZuTIkVi9ejV+++03DB48GNu2bYOHhweef/55jfE9PT3x5ZdflpjjyQ/WUjVr1gyxsbEICwvDvn378PPPP2PlypX44IMPsGjRojJjS/v96PJ7K2rx4sV4//33MXHiRHz00UeoU6cODAwMMGvWrBKfk7rsDSo0btw4fP7559i5cydGjRqFzZs3Y8CAARp7ecoi9TFJSUmBj48PrKys8OGHH6Jx48YwNTXF6dOnMXfuXK3+BvXxOAD//zcwZ84c+Pv7l9jHzc0NANC9e3dcvXoVu3btwoEDB7BmzRp89dVX+O677/DKK6/oZT5EVD2xECIirfXv3x9r1qzBqVOnNA7zkuLff//FpUuXsH79eowbN07d/uSKT4XXIyrp8JfY2Ngyxyhc0Sw/Px+9evXSab5Flbby2JNzb9GiBYDHP0N55l+4h6VevXp6nS/w+EOjg4MDtm7diq5du+Lw4cN49913i41/5swZ9OzZs8zV1ZydnVFQUICrV69q7AV62u+jKAsLC4wYMQIjRoxATk4OXnzxRXzyyScICgqq1CWXf/rpJ/j5+SEkJESjPSUlBba2tpLzlvV4tmzZEm3atMGmTZvQoEED3Lx5E8uXL5c8VnlFREQgMTERO3bsQPfu3dXt169fL9ZXH9dyKnx+nzt3rtTnt6urK4DHi3yU52+gcEXECRMmICMjA927d8fChQtZCBFRmXiOEBFp7Z133oG5uTkmTpyIe/fuFbtfm2/fC7+tLhojhCi2jK6DgwNat26N9evXaxyqc/DgQVy4cOGpYwwdOhQ///xziXuyHjx4UO75FmVhYVHiYUO9evXS2Ar3EPXr1w8nTpzAqVOnNMYuel4IAPj7+8PKygqLFy9Gbm6u3uYLPD4E66WXXsKePXuwceNG5OXlaRwWBwDDhw/HrVu38MMPPxSLz8rKQmZmJgCgb9++AIBvvvlGo8+yZcvKNZfExESN2yYmJmjevDmEECX+3BXJ0NCw2PN4+/btGufVSFHac6bQyy+/jAMHDmDZsmWoW7eu+jGWU0l/gzk5OVi5cmWxvk+bf3m0bdsWjRo1wrJly5CSkqJxX+Ec6tWrB19fX6xevRp37twplqPo38CTzyNLS0u4ubkVW+qdiOhJ3CNERFpr0qQJNm/ejFGjRsHd3R1jxozB888/DyEErl+/js2bN8PAwKBc5614eHigcePGmDNnDm7dugUrKyv8/PPPJZ7TERwcjP79+6Nr166YOHEikpKS1NehKemcpaI+/fRThIeHo2PHjpg8eTKaN2+OpKQknD59GocOHSp2XZXyaNeuHbZu3YrZs2ejQ4cOsLS0xMCBA0vt/84772Djxo0ICAjAG2+8oV4+29nZGWfPnlX3s7KywqpVq/Dyyy+jbdu2GDlyJOzs7HDz5k3s3bsXXbp0wYoVK7Seb6ERI0Zg+fLlWLBgATw9PdGsWTON+19++WVs27YNr732GsLDw9GlSxfk5+fj4sWL2LZtG/bv34/27dujdevWGDVqFFauXInU1FR07twZv//+u8b1lsrSp08f2Nvbo0uXLnjuuecQExODFStWoH///upz0SrLgAED8OGHH2LChAno3Lkz/v33X2zatEm9p0Kqpz1nRo8ejXfeeQe//PILpk6dWiEXxe3cuTNq166N8ePHY+bMmVAoFNi4cWOJX2ho+5wviYGBAVatWoWBAweidevWmDBhAhwcHHDx4kWcP38e+/fvBwB8++236Nq1Kzw9PTF58mS4urri3r17iIyMREJCgvqaTs2bN4evry/atWuHOnXqICoqCj/99BNef/113R8cIqreKn6hOiKqLq5cuSKmTp0q3NzchKmpqTAzMxMeHh7itddeE9HR0Rp9x48fLywsLErMc+HCBdGrVy9haWkpbG1txeTJk8WZM2cEALFu3TqNvj///LNo1qyZUCqVonnz5mLHjh3FlokWovjy2UIIce/ePTF9+nTh5OQkjI2Nhb29vejZs6f4/vvv1X0Kl/rdvn27RmzhEsVF55ORkSFGjx4tbGxsSlzCuyRnz54VPj4+wtTUVNSvX1989NFHIiQkpNgS0YVz8ff3F9bW1sLU1FQ0btxYBAYGiqioKHWfsh7Xkh4XIR4vQezk5CQAiI8//rjE2JycHLFkyRLRokULoVQqRe3atUW7du3EokWLRGpqqrpfVlaWmDlzpqhbt66wsLAQAwcOFPHx8eVaPnv16tWie/fuom7dukKpVIrGjRuLt99+WyN/actnl7RcMgAxffp0jbbC39vnn3+ubivv8tlvvfWWcHBwEGZmZqJLly4iMjJS+Pj4aCxnXdrzpeh9RZfPLs9zpl+/fsWWBy9LST9jWXMrfEz/+usvdduff/4pvL29hZmZmXB0dBTvvPOOeknv8sxf28dBCCH++OMP0bt3b1GrVi1hYWEhWrVqJZYvX67R5+rVq2LcuHHC3t5eGBsbi/r164sBAwaIn376Sd3n448/Fl5eXsLGxkb9GvTJJ5+InJyccj1+RFRzKYTQ8gxSIiIiks2QIUPw77//lnvPGhERScNzhIiIiJ4Rd+7cwd69e/Hyyy9X9lSIiKo9niNERERUya5fv44///wTa9asgbGxMaZMmVLZUyIiqva4R4iIiKiSHTlyBC+//DKuX7+O9evXl3j9KCIi0i+eI0RERERERDUO9wgREREREVGNw0KIiIiIiIhqHBZCRERERERU47AQIiIiIiKiGoeFEBERERER1TgshIiIiIiIqMZhIURERERERDUOCyEiIiIiIqpxWAgREREREVGNw0KIiIiIiIhqHBZCRERERERU47AQIiIiIiKiGoeFEBERERER1TgshIiIiIiIqMZhIURERERERDUOCyEiIiIiIqpxWAgREREREVGNw0KIiIiIiIhqHBZCRERERERU47AQIiIiIiKiGoeFEBERERER1TgshIiIiIiIqMZhIURERERERDUOCyEiIiIiIqpxWAgREREREVGNw0KIqAqJiIiAQqFAREREZU+FiIiIqEpjIUQ12gsvvABzc3Okp6eX2mfMmDEwMTFBYmJiBc6MiIiqoqtXr2LKlClwdXWFqakprKys0KVLF3z99dfIysqq7Ok90y5cuICFCxciLi6usqdCNQQLIarRxowZg6ysLPzyyy8l3v/o0SPs2rULAQEBqFu3bgXPjoiIqpK9e/fC09MT27Ztw8CBA7F8+XIEBwejYcOGePvtt/HGG29U9hSfaRcuXMCiRYtYCFGFMarsCRBVphdeeAG1atXC5s2bMW7cuGL379q1C5mZmRgzZkwlzI6IiKqK69evY+TIkXB2dsbhw4fh4OCgvm/69Om4cuUK9u7dW4kzJKIncY8Q1WhmZmZ48cUX8fvvv+P+/fvF7t+8eTNq1aqFF154AdeuXcOwYcNQp04dmJubw9vbW+NNTQgBW1tbzJ49W91WUFAAGxsbGBoaIiUlRd2+ZMkSGBkZISMjQ9128eJFvPTSS6hTpw5MTU3Rvn177N69+6k/w+XLlzF06FDY29vD1NQUDRo0wMiRI5GamirxUSEiIm199tlnyMjIQEhIiEYRVMjNzU29RygvLw8fffQRGjduDKVSCRcXF8yfPx8qlUojxsXFBQMGDMAff/wBLy8vmJqawtXVFRs2bNDol5ubi0WLFqFJkyYwNTVF3bp10bVrVxw8eFCj3+HDh9GtWzdYWFjAxsYGgwYNQkxMjEafhQsXQqFQ4MqVKwgMDISNjQ2sra0xYcIEPHr0SKPvunXr0KNHD9SrVw9KpRLNmzfHqlWriv3s5fk5QkNDMWzYMACAn58fFAqFxjmxUVFR8Pf3h62tLczMzNCoUSNMnDixrF8J0VOxEKIab8yYMcjLy8O2bds02pOSkrB//34MGTIEaWlp6Ny5M/bv349p06bhk08+QXZ2Nl544QX1YXUKhQJdunTB0aNH1TnOnj2rLkj+/PNPdfuxY8fQpk0bWFpaAgDOnz8Pb29vxMTEYN68eVi6dCksLCwwePDgUg/bA4CcnBz4+/vjxIkTmDFjBr799lu8+uqruHbtmkbhRURE8tqzZw9cXV3RuXPnp/Z95ZVX8MEHH6Bt27b46quv4OPjg+DgYIwcObJY3ytXruCll15C7969sXTpUtSuXRuBgYE4f/68us/ChQuxaNEi+Pn5YcWKFXj33XfRsGFDnD59Wt3n0KFD8Pf3x/3797Fw4ULMnj0bx48fR5cuXUo8FG348OFIT09HcHAwhg8fjtDQUCxatEijz6pVq+Ds7Iz58+dj6dKlcHJywrRp0/Dtt99q/XN0794dM2fOBADMnz8fGzduxMaNG9GsWTPcv38fffr0QVxcHObNm4fly5djzJgxOHHixFMfa6IyCaIaLi8vTzg4OIhOnTpptH/33XcCgNi/f7+YNWuWACCOHTumvj89PV00atRIuLi4iPz8fCGEEJ9//rkwNDQUaWlpQgghvvnmG+Hs7Cy8vLzE3LlzhRBC5OfnCxsbG/Hmm2+qc/Xs2VN4enqK7OxsdVtBQYHo3LmzaNKkibotPDxcABDh4eFCCCH++ecfAUBs375dvw8KERGVW2pqqgAgBg0a9NS+0dHRAoB45ZVXNNrnzJkjAIjDhw+r25ydnQUAcfToUXXb/fv3hVKpFG+99Za67fnnnxf9+/cvc9zWrVuLevXqicTERHXbmTNnhIGBgRg3bpy6bcGCBQKAmDhxokb8kCFDRN26dTXaHj16VGwcf39/4erqqtFW3p9j+/btGu9xhX755RcBQPz1119l/oxE2uIeIarxDA0NMXLkSERGRmp8K7Z582Y899xz6NmzJ3799Vd4eXmha9eu6vstLS3x6quvIi4uDhcuXAAAdOvWDfn5+Th+/DiAx3t+unXrhm7duuHYsWMAgHPnziElJQXdunUD8HjP0+HDh9Xfvj18+BAPHz5EYmIi/P39cfnyZdy6davEuVtbWwMA9u/fX+yQBSIiqhhpaWkAgFq1aj2176+//goAGodRA8Bbb70FAMXOI2revLn6/QIA7Ozs4O7ujmvXrqnbbGxscP78eVy+fLnEMe/cuYPo6GgEBgaiTp066vZWrVqhd+/e6jkV9dprr2nc7tatGxITE9U/K/D48PJCqampePjwIXx8fHDt2rVih2eX5+cojY2NDQAgLCwMubm5T+1PVF4shIgA9WIImzdvBgAkJCTg2LFjGDlyJAwNDXHjxg24u7sXi2vWrBkA4MaNGwCAtm3bwtzcXF30FBZC3bt3R1RUFLKzs9X3FRZVV65cgRAC77//Puzs7DS2BQsWAECJ5y8BQKNGjTB79mysWbMGtra28Pf3x7fffsvzg4iIKpCVlRUAlHkphkI3btyAgYEB3NzcNNrt7e1hY2Ojfj8p1LBhw2I5ateujeTkZPXtDz/8ECkpKWjatCk8PT3x9ttv4+zZsxpjAij1fezhw4fIzMwsc9zatWsDgMa4f/75J3r16qU+58jOzg7z588HgGLvQ+X5OUrj4+ODoUOHYtGiRbC1tcWgQYOwbt26YudUEWmLhRARgHbt2sHDwwNbtmwBAGzZsgVCCK1XizM2NkbHjh1x9OhRXLlyBXfv3kW3bt3QtWtX5Obm4uTJkzh27Bg8PDxgZ2cH4PGCCgAwZ84cHDx4sMTtyTfMopYuXYqzZ89i/vz5yMrKwsyZM9GiRQskJCRIfDSIiEgbVlZWcHR0xLlz58odo1AoytXP0NCwxHYhhPr/3bt3x9WrV7F27Vq0bNkSa9asQdu2bbFmzZpyz0fbca9evYqePXvi4cOH+PLLL7F3714cPHgQb775JoD/f2/T5ucojUKhwE8//YTIyEi8/vrruHXrFiZOnIh27dppLDpEpC0WQkT/GTNmDM6dO4ezZ89i8+bNaNKkCTp06AAAcHZ2RmxsbLGYixcvqu8v1K1bN5w6dQqHDh2Cra0tPDw8UKdOHbRo0QLHjh3DsWPH0L17d3V/V1dXAI+LqF69epW4Pe1wC09PT7z33ns4evQojh07hlu3buG7777T+TEhIqLyGTBgAK5evYrIyMgy+zk7O6OgoKDYYWz37t1DSkqKxvuJNurUqYMJEyZgy5YtiI+PR6tWrbBw4UL1mABKfR+ztbWFhYWFVuPt2bMHKpUKu3fvxpQpU9CvXz/06tVL43A5bT2tOPT29sYnn3yCqKgobNq0CefPn8ePP/4oeTwiFkJE/ync+/PBBx8gOjpaY29Qv379cOrUKY03uMzMTHz//fdwcXFB8+bN1e3dunWDSqXCsmXL0LVrV/ULe7du3bBx40bcvn1b4zjpevXqwdfXF6tXr8adO3eKzevBgwelzjktLQ15eXkabZ6enjAwMOAhA0REFeidd96BhYUFXnnlFdy7d6/Y/VevXsXXX3+Nfv36AQCWLVumcf+XX34JAOjfv7/WYycmJmrctrS0hJubm/p9wMHBAa1bt8b69es1VhQ9d+4cDhw4oJ6TNgr38BTdo5Oamop169ZpnatQYTH25KqnycnJxfYctW7dGgD4Xkc64QVVif7TqFEjdO7cGbt27QIAjUJo3rx52LJlC/r27YuZM2eiTp06WL9+Pa5fv46ff/4ZBgb//51Cp06dYGRkhNjYWLz66qvq9u7du6uvr1C0EAKAb7/9Fl27doWnpycmT54MV1dX3Lt3D5GRkUhISMCZM2dKnPPhw4fx+uuvY9iwYWjatCny8vKwceNGGBoaYujQoXp7bIiIqGyNGzfG5s2bMWLECDRr1gzjxo1Dy5YtkZOTg+PHj2P79u0IDAzEG2+8gfHjx+P7779HSkoKfHx8cOrUKaxfvx6DBw+Gn5+f1mM3b94cvr6+aNeuHerUqYOoqCj89NNPeP3119V9Pv/8c/Tt2xedOnXCpEmTkJWVheXLl8Pa2lq950gbffr0gYmJCQYOHIgpU6YgIyMDP/zwA+rVq1fil3rl0bp1axgaGmLJkiVITU2FUqlEjx49sHnzZqxcuRJDhgxB48aNkZ6ejh9++AFWVlaSijgitUpcsY7omfPtt98KAMLLy6vYfVevXhUvvfSSsLGxEaampsLLy0uEhYWVmKdDhw4CgDh58qS6LSEhQQAQTk5OJcZcvXpVjBs3Ttjb2wtjY2NRv359MWDAAPHTTz+p+zy5fPa1a9fExIkTRePGjYWpqamoU6eO8PPzE4cOHdLhUSAiIqkuXbokJk+eLFxcXISJiYmoVauW6NKli1i+fLn6Egm5ubli0aJFolGjRsLY2Fg4OTmJoKAgjUsoCPF42emSlsX28fERPj4+6tsff/yx8PLyEjY2NsLMzEx4eHiITz75ROTk5GjEHTp0SHTp0kWYmZkJKysrMXDgQHHhwgWNPoXLZz948ECjfd26dQKAuH79urpt9+7dolWrVsLU1FS4uLiIJUuWiLVr1xbrV96fQwghfvjhB+Hq6ioMDQ3V73enT58Wo0aNEg0bNhRKpVLUq1dPDBgwQERFRRXLSaQNhRDlOEuNiIiIiIioGuE5QkREREREVOOwECIiIiIiohqHhRAREREREdU4LISIiIiIiKjGYSFEREREREQ1DgshIiIiIiKqcarUBVULCgpw+/Zt1KpVCwqForKnQ0RUowghkJ6eDkdHR42LCNd0fG8iIqocur4vValC6Pbt23BycqrsaRAR1Wjx8fFo0KBBZU/jmcH3JiKiyiX1falKFUK1atX673/xAKz0nj/16kO959RQu7as6bNyDGXLbZZwWbbcAHD0ThPZcnc/v0q23ACAmzfly/3xx/LlBnAhVr7nTPPja2TLDQB48EC21LfHB8mWGwAcDe7KlvvMPXvZcmdmpqFvX6cir8UE/P97098ALCt3KlVD6/Yw2P67ViH1Mq5JHi4211VyrPvFXdICddhjGtt0oORY951LJMVdfWmu5DEbXzsoORbdu0uP/fJLaXGTJ0seMtPMVnKsxbF90gKzsiSPiebNpcfa2UmPNTaWFBaXqP17S0ZGGrp0kf6+VKUKof8/5MAKchRCVrVUes+pOYD+51yUsZyFkKW8b+8WFvI9NlamprLlBgAolfLllvk5Y2kp33PGysxMttwAABl/r+m15H3crQwyZcttmSnv3AHw8K8nFD4elgAMYIEYeGoVX99R2riGOvz5WlhIj7WxzCt337yYGIjMJ57vhkYw0PJvzEohvfi2zJH+N2Flbi4tUIdCyNJSh/lKfF3UaUypjxGg23uc1PcAHb7IMTSXPl8LqY+TLq+3unx20+ULLxMTaUPmSB9T6vtSlSqEiIiI9C0wMBDr168HABgbG6Nhw4YYN24c5s+fDyMj7d4mY+CJFxCpVUzQeK26q+nyOaV1a+mxfb2Ty933YZ8+yI2Kkj4YEZGMWAgREVGNFxAQgHXr1kGlUuHXX3/F9OnTYWxsjKAgeQ+TJCKiysNlf4iIqMZTKpWwt7eHs7Mzpk6dil69emH37t2VPS0iIpIR9wgRERE9wczMDImJiSXep1KpoFL9/zmlaWlpFTUtIiLSI+4RIiIi+o8QAocOHcL+/fvRo0ePEvsEBwfD2tpavXHpbCKiqkmvhVB8fDwmTpwIR0dHmJiYwNnZGW+88QYSExMxadIkeHp6IicnRyPm119/hYmJCU6fPq3PqRAREZVbWFgYLC0tYWpqir59+2LEiBFYuHBhiX2DgoKQmpqq3uLj4yt2skREpBd6K4SuXbuG9u3b4/Lly9iyZQuuXLmC7777Dr///js6deqEBQsWID09HQsWLFDHpKSkYPLkyXj//ffRtm1bfU2FiIhIK35+foiOjsbly5eRlZWF9evXw6KUNaaVSiWsrKw0NiIiqnr0do7Q9OnTYWJiggMHDsDsv+uHNGzYEG3atEHjxo0RHByMdevWwd/fH4MHD0bHjh0xa9Ys1K9fv9RVeXgcNhERVQQLCwu4ublV9jSIiKgC6WWPUFJSEvbv349p06api6BC9vb2GDNmDLZu3QpfX19MmzYN48ePx/bt27Ft2zZs2LCh1Os08DhsIiIiIiKSg14KocuXL0MIgWbNmpV4f7NmzZCcnIwHDx4gODgYADBy5EgsXrwYHh4epeblcdhERERERCQHvS6fLYR4ah8zMzPMmTMHb775Jt54440y+yqVSiiVSn1Nj4iIqJjQ0NDKngIREVUCvRRCbm5uUCgUiImJwZAhQ4rdHxMTg9q1a8POzu7xoEZGMDQ0hEKh0MfwREREz4T6jkDQeO1i/jtQQmuLF0uLAwBdjjSPvFi73H3ts4xg+kSbcWYKbKP3azdo69ba9S+iGe5Jjv3HdaikuDZG/0oes9nFXyTHomVLSWFN0v6WPmafPtJj166VHtuunbS4pCTJQ1oUFEiO/dthgKS4drf3SB4Trq6SQzPzpO+IeGKB6HJrhOtax6QhXdpg/9HLoXF169ZF7969sXLlSmRlZWncd/fuXWzatAkjRoxg4UNERERERM8EvS2fvWLFCqhUKvj7++Po0aOIj4/Hvn370Lt3b9SvXx+ffPKJvoYiIiIiIiLSid4KoSZNmiAqKgqurq4YPnw4GjdujFdffRV+fn6IjIxEnTp19DUUERERERGRTvRWCAGAs7MzQkNDcffuXeTk5ODmzZv45ptvULduXY1+gYGBSElJ0efQREREWgsMDIRCoVBvdevWRUBAAM6ePVvZUyMiIpnptRAiIiKqagICAnDnzh3cuXMHv//+O4yMjDBggLQTm4mIqOpgIURERDWaUqmEvb097O3t0bp1a8ybNw/x8fF48OBBZU+NiIhkpNfrCBEREVVlGRkZ+N///gc3N7dih3UXUqlUUKlU6ttpaWkVNT0iItKjKlkIpV59CKtaqqd31JKinp3ecxa1cqWs6TFunHy5Vxx0ly85gNdt/idf8qq8UMeOHbKmb2lmJl9yXS5UUh5jx8qW+sfvZUsNAHjxRQfZcrexS5Atd5qpbtdreFaFhYXB0tISAJCZmQkHBweEhYXBwKDkgyaCg4OxaNGiipwiERHJgIfGERFRjebn54fo6GhER0fj1KlT8Pf3R9++fXHjxo0S+wcFBSE1NVW9xcfHV/CMiYhIH6rkHiEiIiJ9sbCwgJubm/r2mjVrYG1tjR9++AEff/xxsf5KpRJKpfSrrhMR0bOBe4SIiIiKUCgUMDAwQFZWVmVPhYiIZMQ9QkREVKOpVCrcvXsXAJCcnIwVK1YgIyMDAwcOrOSZERGRnFgIERFRjbZv3z44ODxewKJWrVrw8PDA9u3b4evrW7kTIyIiWbEQIiKiGis0NBShoaF6y2doCNSqpV3M4sXSxpo/X1ocAAQGSo8tZVXxkpmW0JaTAyRoubphr17a9S8qI0NyaKtW0uK2bPOUPOaogAaSY3HzprS4BjqMee2a9Ngi5+ZpzdhYUtgNU+mr4Dr/FSY5tp1WfzhFtG8veUxER0sOtTAxkR5rbi4pLrOB9r+bTFPdLl/Ac4SIiIiIiKjGYSFEREREREQ1jqRCKD4+HhMnToSjoyNMTEzg7OyMN954A4mJiQCAefPmwcPDQyPm4sWLUCgUCHxif3xoaCiUSiVX5yEiIiIiogqjdSF07do1tG/fHpcvX8aWLVtw5coVfPfdd/j999/RqVMnJCUlwc/PD7GxsepVeAAgPDwcTk5OiIiI0MgXHh4Ob29vmMl5hXsiIiIiIqIitC6Epk+fDhMTExw4cAA+Pj5o2LAh+vbti0OHDuHWrVt499130bVrVxgbG2sUPREREZg+fTqSkpIQFxen0e7n56ePn4WIiKhcAgMDoVAo8NprrxW7b/r06SUewUBERNWLVoVQUlIS9u/fj2nTphXbg2Nvb48xY8Zg69atMDc3R4cOHRAeHq6+PyIiAj179kSXLl3U7deuXcPNmzdLLYRUKhXS0tI0NiIiIn1wcnLCjz/+qHFodnZ2NjZv3oyGDRtW4syIiKgiaFUIXb58GUIINGvWrMT7mzVrhuTkZDx48AB+fn7qPUIXLlxAdnY22rRpg+7du6vbIyIiYGpqCm9v7xLzBQcHw9raWr05OTlpM10iIqJStW3bFk5OTtixY4e6bceOHWjYsCHatGlTahy/pCMiqh4kLZYghHhqH19fX1y6dAl37txBREQEunbtCkNDQ/j4+GgUQp07d4ZSqSwxR1BQEFJTU9VbfHy8lOkSERGVaOLEiVi3bp369tq1azFhwoQyY/glHRFR9aBVIeTm5gaFQoGYmJgS74+JiUHt2rVhZ2eHLl26wMTEBOHh4QgPD4ePjw8AoEOHDnj48CGuXbuGiIgI9OjRo9TxlEolrKysNDYiIiJ9GTt2LP744w/cuHEDN27cwJ9//omxY8eWGcMv6YiIqgetCqG6deuid+/eWLlyZbHlru/evYtNmzZhxIgRUCgUMDMzQ8eOHREREYEjR47A19cXAGBsbAxvb2+EhIQgPj6eCyUQEVGlsbOzQ//+/REaGop169ahf//+sLW1LTOGX9IREVUPWh8at2LFCqhUKvj7++Po0aOIj4/Hvn370Lt3b9SvXx+ffPKJuq+fnx9+/PFHZGdno23btup2Hx8fLF++HBYWFujQoYN+fhIiIiIJJk6ciNDQUKxfvx4TJ06s7OkQEVEF0boQatKkCaKiouDq6orhw4ejcePGePXVV+Hn54fIyEjUqVNH3dfPzw/p6eno0qULjIyM1O0+Pj5IT09XL7NNRERUWQICApCTk4Pc3Fz4+/tX9nSIiKiCGD29S3HOzs4IDQ19aj9fX98SF1bw8fEp14ILREREcjM0NFSf+2poaFjJsyEioooiqRAiIiKqTnieDxFRzcNCiIiIapynHdWwc+dOSXktLIDWrbWLkbr6dmCgtDgAcHSUHjtqVPn7zogDXJ5oy6tTD4mDJ2k1pmm2Vt01mJhbS4413vI/SXGj3N0lj/nbCennTvdtLfGBOnRI8phYv1567NKl0mPPnpUU5py4S/qYTZtKj83JkRaXkSF9TFNTyaH/FDwvOdbGRFrcAQl/bk+s3aY1SdcRIiIiIiIiqspYCBERERERUY1TNQ+Nq10bkOF47pUr9Z5Sw7Rp8uaX8+Lmr0/JlS85AOyQcfXAy5flyw0APXvKlzsqSr7cANCnj3y5ZX7cH2TXki338OGypQYA3L8vX+7zdxrIlvvRozTZchMREVU07hEiIqIaLzAwEAqFoth25cqVyp4aERHJpGruESIiItKzgIAArFu3TqPNzs6ukmZDRERyYyFEREQEQKlUwt7evrKnQUREFYSFEBERkRZUKhVUKpX6dloaz50iIqqKeI4QERERgLCwMFhaWqq3YcOGldgvODgY1tbW6s1JzpVqiIhINtwjREREBMDPzw+rVq1S37awsCixX1BQEGbPnq2+nZaWxmKIiKgKYiFERESEx4WPm5vbU/splUoolcoKmBEREcmJh8YREREREVGNo7dCKDIyEoaGhujfv3+x+7KysrBgwQI0bdoUSqUStra2GDZsGM6fP6+v4YmIiIiIiMpNb4VQSEgIZsyYgaNHj+L27dvqdpVKhV69emHt2rX4+OOPcenSJfz666/Iy8tDx44dceLECX1NgYiIiIiIqFz0co5QRkYGtm7diqioKNy9exehoaGYP38+AGDZsmWIjIzEP//8g+effx4A4OzsjJ9//hkdO3bEpEmTcO7cOSgUimJ5uUQpERFVhNDQ0MqeAhERVTC9FELbtm2Dh4cH3N3dMXbsWMyaNQtBQUFQKBTYvHkzevfurS6CChkYGODNN9/EmDFjcObMGbRu3bpY3uDgYCxatEgfUyQiIpKdjWUe+nonaxUTebG2pLHq1pUUBgAYNUp67JYtWowDwOWJNqOCHNTNStBqzGMXGmjVv6hu9WIlx17uOFZSXL16kodEXzftnj9F3cl2kBTncO2a5DExZYr02L//lh47aJC0uCtXpI+5b5/02GnTpMXdvy99zIwMyaFJSdKHtbKSFjflVaF1TFqawJtvShsP0NOhcSEhIRg79vGLRUBAAFJTU3HkyBEAwKVLl9CsWbMS4wrbL126VOL9QUFBSE1NVW/x8fH6mC4REREREdVwOhdCsbGxOHXqFEb99/WSkZERRowYgZCQEHUfIbSv8IDHS5RaWVlpbERERERERLrS+dC4kJAQ5OXlwdHRUd0mhIBSqcSKFSvQtGlTxMTElBhb2N60aVNdp0FERERERFRuOu0RysvLw4YNG7B06VJER0ertzNnzsDR0RFbtmzByJEjcejQIZw5c0YjtqCgAF999RWaN29e7PwhIiIiuQ0cOBABAQEl3nfs2DEoFAqcPXu2gmdFREQVRac9QmFhYUhOTsakSZNgbW2tcd/QoUMREhKCY8eOYdeuXRg4cCCWLl2Kjh074t69e1i8eDFiYmJw6NChEleMIyIiktOkSZMwdOhQJCQkoEEDzZPx161bh/bt26NVq1aVNDsiIpKbTnuEQkJC0KtXr2JFEPC4EIqKisKlS5dw+PBhjBs3DvPnz4ebmxsCAgJgaGiIEydOwNvbW5cpEBERSTJgwADY2dkVWzo7IyMD27dvx6RJkypnYkREVCF02iO0Z8+eUu/z8vLSWCTh448/xscff6zLcERERHpjZGSEcePGITQ0FO+++6766ITt27cjPz9fvQjQk3iNOyKi6kEvy2cTERFVRRMnTsTVq1fVl3wAHh8WN3To0BKPdgAeX+PO2tpavTk5OVXUdImISI9YCBERUY3l4eGBzp07Y+3atQCAK1eu4NixY2UeFsdr3BERVQ8shIiIqEabNGkSfv75Z6Snp2PdunVo3LgxfHx8Su3Pa9wREVUPLISIiKhGGz58OAwMDLB582Zs2LABEydO5GqmREQ1gM4XVK0MWTmGMM4x1HveceP0nlKD3IeRDxwoX+4LF4zlSw7AdfAI2XLfvy9bagCAU0bJFwzWi6go+XIDwJ078uX+6CP5cgOIOiBf7r4mv8uXHIBTjx6y5RaQ7wN8dV0TwNLSEiNGjEBQUBDS0tIQGBgoOVdeTAwe9umjVYx9lsS3YlNpYQAwI056bMlLSJTs+beigEaabXkWsXgYP0irMRvkm2jVv6iHD7MAAEZGTWFjI+/rEhFVLVWyECIiItKnSZMmISQkBP369YOjo6PkPCIzE7lafoGhQz0jmUtFxTYC0EKzSSATuarTWo2py2OUm6tDMBFVayyEiIioxuvUqZPGJR+IiKj6YyFERERE1Z4wNUOuq7tWMU1O/iFtsAwLaXEAjqS1kRzbtKnEwMGDJY+Jc+ekx47Q4bD4K1ckhf0D6Y+vaYD02GZn/5IWGBsrecyYdmMlx/Zs9UByLB4+lBZ3RUJZkpEhbaz/cLEEIiIiIiKqcbhHiIiISB9atwcMtX9bNc5MkTZeTo60OAB5depJjjUqKP+4eRaxEMjUaMvNMsKZa+21GtPNTavusDT+F0YGmU/vSEQ1GgshIiIiPTDY/jsMaml/TSHb6P3SBkxIkBYHIHFw6ReMfZq6WeUf92H8oGILI5y51h4vvB6p1ZgbNmjVHe3rdYKN8oR2QURU4/DQOCIiqlYUCkWZ28KFCxEXFweFQoHo6Ohi8b6+vpg1a1aFz5uIiCqW3gqhyMhIGBoaon///qX22bJlCwwNDTF9+nR9DUtERKThzp076m3ZsmWwsrLSaJszZ05lT5GIiJ4BeiuEQkJCMGPGDBw9ehS3b98utc8777yDLVu2IDs7W19DExERqdnb26s3a2trKBQKjTZLS0ud8qtUKqSlpWlsRERU9eilEMrIyMDWrVsxdepU9O/fH6GhocX6XL9+HcePH8e8efPQtGlT7NixQx9DExERVajg4GBYW1urNycnp8qeEhERSaCXQmjbtm3w8PCAu7s7xo4di7Vr1xa7MN26devQv39/WFtbY+zYsQgJCXlqXn7rRkREcurcuTMsLS01tmPHjpUZExQUhNTUVPUWHx9fQbMlIiJ90kshFBISgrFjH1+0KSAgAKmpqThy5Ij6/oKCAoSGhqr7jBw5En/88QeuX79eZl5+60ZERHLaunUroqOjNbb27cte2lmpVMLKykpjIyKiqkfnQig2NhanTp3CqFGjAABGRkYYMWKExh6fgwcPIjMzE/369QMA2Nraonfv3li7dm2ZufmtGxERycnJyQlubm4am5mZWWVPi4iIKoDO1xEKCQlBXl4eHB0d1W1CCCiVSqxYsQLW1tYICQlBUlKSxptLQUEBzp49i0WLFsHAoOR6TKlUQqlU6jpFIiIiIiIiDToVQnl5ediwYQOWLl2KPn36aNw3ePBgbNmyBcOGDcOuXbvw448/okWLFur78/Pz0bVrVxw4cAABAQG6TIOIiIiIiEgrOhVCYWFhSE5OxqRJk2Btba1x39ChQxESEoLs7GzUrVsXw4cPh0Kh0OjTr18/hISEsBAiIiIiIqIKpVMhFBISgl69ehUrgoDHhdBnn32Gv//+G1OnTi1WBBX2efnll/Hw4UPY2trqMhUiIqJiAgMDERgYWKzdxcWl2OqmhSIiIiSNVS/jGqwUtbQPbN1a0njo1UtaHABTHS7ld+xCg3L3bZBvAtMn2tzcgA0btBtz3Djt+u9eAbRvodmmyM+H8aNUrfJcfq6rdgPrQVs36bHHj0uLc/hjs/RBmzaVHnvzpvRYidejbJOwR/qYupD6Obd3b+lD6rISQF6e9FgbG0lh+fUctI/RcUVpnQqhPXtKfzJ5eXmV+iZTaPjw4Rg+fLguUyAiIiIiItKaXpbPJiIiIiIiqkpYCBERERERUY3DQoiIiKqVwMBAKBQKKBQKGBsb47nnnlNfu66goEDdz8XFBcuWLSsWv3DhQrSWet4OERFVGSyEiIio2gkICMCdO3cQFxeH3377DX5+fnjjjTcwYMAA5OlyEjAREVUbOl9QtTKYJVyGmaWl3vOuOOiu95xFvT4lV9b8Fy4Yy5a7eXPZUgMAxM142XJfuOAkW24AcOzVTLbchiWsdqVXZSx4orPISPlyA+jbo4N8yQs6y5cb0G01nqdQbNZh9aen5c7Kki23vimVStjb2wMA6tevj7Zt28Lb2xs9e/ZEaGgoXnnllUqeIRERVTbuESIiohqhR48eeP7557Fjxw6d8qhUKqSlpWlsRERU9bAQIiKiGsPDwwNxcXHq23PnzoWlpaXGtnjx4jJzBAcHw9raWr05Ocm715mIiOTBQoiIiGoMIYTGBb7ffvttREdHa2yvvfZamTmCgoKQmpqq3uLj5Tu0l4iI5FMlzxEiIiKSIiYmBo0aNVLftrW1hZubm0afOnXqlJlDqVRCqVTKMj8iIqo43CNEREQ1wuHDh/Hvv/9i6NChlT0VIiJ6BnCPEBERVTsqlQp3795Ffn4+7t27h3379iE4OBgDBgzAuHHjKnt6NZql8b9oX6+TVjG7V2g3xj8+UdhnrtmmMLgAo+zeWuXJNjPUbmA9MH4kPTajsbS43+o9PrzTJckOsyIGSJ8AURXDQoiIiKqdffv2wcHBAUZGRqhduzaef/55fPPNNxg/fjwMDHgwRGUyMsiEjfKEVjHtW2g3xj5z4KrZk62ZQMFf2iWq+DoIKHh6l1JZSAu7IzGOqKpjIURERNVKaGgoQkNDn9qv6OpxRS1cuBALFy7U65yIiOjZo1MhFBkZia5duyIgIAB79+4FAIwcORIpKSnYt2+fut++ffvQt29fLFiwQOPNZeHChVi7di1u3rypyzSIiIgqXWyuKyxzrLSOa4Z70gbMyJAWB8DE3FpybLd6seXu+/BhFnLlvZY46ZOjIzB3rnYx0dHSxyvly4hySUiQFBbrPV7ykO6O6ZJjceiQtDgdLsBd0PQ5ybHIzpYcmm7b6OmdSmAgYUgdpvl4TF2CQ0JCMGPGDBw9ehS3b98GAPj5+eHPP/9EXpFfXHh4OJycnBAREaERHx4eDj8/P12mQEREREREpDXJe4QyMjKwdetWREVF4e7duwgNDcX8+fPh5+eHjIwMREVFwdvbGwAQERGBefPm4a233kJ2djZMTU2RnZ2NkydPYsKECaWOoVKpoFKp1Ld59W4iIqKqw8ioqcZtYVrsxJ1yUeTna9ff4AKATI02wxwjJJ5tr1Uej2ZadVcTQlocABjqcF6SlUn5vx6PM7iEbIUOKzMQVQOSC6Ft27bBw8MD7u7uGDt2LGbNmoWgoCA0bdoUjo6OCA8Ph7e3N9LT03H69GmEhYVh+fLliIyMhJ+fH44fPw6VSlXmHqHg4GAsWrRI6hSJiIioEtnYfKRxO9fVXVIe40epWvU3yu5dbGGExLPtcfKFSK3yLPhVq+5qOTnS4gDAxkZ6rE+Dq+XuO0v5Ei4aRksfjKgakHxoXEhICMaOHQsACAgIQGpqKo4cOQLg8eFxhYfBHTt2DE2bNoWdnR26d++ubo+IiECjRo3g7Oxc6hi8ejcREclJoVCUuXHRBCKi6kvSHqHY2FicOnUKv/zyy+MkRkYYMWIEQkJC4OvrC19fX8yaNQu5ubmIiIiAr68vAMDHxwerV68G8LgQetr5Qbx6NxERyenOnTvq/2/duhUffPABYmP/fzEAS0vLypgWERFVAEmFUEhICPLy8uDo6KhuE0JAqVRixYoV8PPzQ2ZmJv766y+Eh4fj7bffBvC4EJo4cSKSkpJw8uRJTJkyRT8/BRERkQT29vbq/1tbW0OhUGi0ERFR9aV1IZSXl4cNGzZg6dKl6NOnj8Z9gwcPxpYtW/Daa6/ByckJu3fvRnR0NHx8fAAA9evXR/369bF06VLk5ORwxTgiIqpyuJAPEVH1oHUhFBYWhuTkZEyaNAnW1prXIRg6dChCQkLw2muvwc/PDytXroSbmxuee+7/1zH38fHB8uXL1YsqEBERVSVcyIeIqHrQerGEkJAQ9OrVq1gRBDwuhKKionD27Fn4+fkhPT1dfX5QIR8fH6Snp3NvEBERVUlcyIeIqHrQeo/Qnj17Sr3Py8sL4r/F81u1aoXAwMBifcaPH4/x46Vf1ZeIiKgycSEfIqLqQfLy2URERERERFUVCyEiIiIiIqpxWAgREREREVGNw0KIiIgIQGBgIFJSUip7GkREVEEkXVC1sh290wQWFlZ6z/u6zf/0nlPDDmNZ07sOHiFbbnFT3lWRFA2dZMt94YJsqQEAhrfle2z2ZMi7uqJ3kHz57fBAttwAgEeP5Mu9b598uQHg5k35ctvaypc7L0++3NWA+8VdsDI31zruH9ehksZr1UpSGADAeIv097vLHcdKjm1y8g9pYz7XVav+2WaGgKFmm0czYMGv2o3br592/Qt9+KG0OAB4v95qybGJLct/ofq8XFNAPNGYkwOk3NJqzO13tPvdFDUsLURybGzXSZLidLpe8sqV0mPnzJEUlphi+PROpdDpQ76dneRQU4lryRj/uFHrmPysLGmD/Yd7hIiIiIiIqMZhIURERERERDUOCyEiIiIiIqpxWAgREVGNFBgYCIVCAYVCAWNjYzz33HPo3bs31q5di4KCgsqeHhERyYyFEBER1VgBAQG4c+cO4uLi8Ntvv8HPzw9vvPEGBgwYgDwuDkFEVK1VyVXjiIiI9EGpVML+v2Wk6tevj7Zt28Lb2xs9e/ZEaGgoXnnllWIxKpUKKpVKfTstLa3C5ktERPrDPUJERERF9OjRA88//zx27NhR4v3BwcGwtrZWb05O8i3/T0RE8pGlEIqMjIShoSH69+8PANi4cSMsLCxw5coVjX63b99G7dq1sWLFCjmmQUREJImHhwfi4uJKvC8oKAipqanqLT5e3uusERGRPGQphEJCQjBjxgwcPXoUt2/fxssvvwx/f38EBgZqnIA6efJktGvXDtOnT5djGkRERJIIIaBQKEq8T6lUwsrKSmMjIqKqR++FUEZGBrZu3YqpU6eif//+CA0NBQCsXr0aly5dwpdffgkACA0NxZ9//ol169aV+mZDRERUGWJiYtCoUaPKngYREclI74XQtm3b4OHhAXd3d4wdOxZr166FEAJ2dnb4/vvv8f777+PgwYN488038fXXX5d5bLVKpUJaWprGRkREJKfDhw/j33//xdChQyt7KkREJCO9F0IhISEYO3YsgMfLkqampuLIkSMAgMGDB2P48OEICAiAj48Pxo8fX2YunpBKRERyUqlUuHv3Lm7duoXTp09j8eLFGDRoEAYMGIBx48ZV9vSIiEhGei2EYmNjcerUKYwaNQoAYGRkhBEjRiAkJETd5/3330dBQQHee++9p+bjCalERCSnffv2wcHBAS4uLggICEB4eDi++eYb7Nq1C4aGhpU9PSIikpFeryMUEhKCvLw8ODo6qtuEEFAqlVixYgWsra1hZPR4yMJ/y6JUKqFUKvU5RSIiIgCPz1UtPI9VLwwMHm9aamP0r6ThtmzzlBQHAKPc3SXH1qsnORTIsNAhWDdCADk52sV8+KG0sT74QFocALx/+wXJsZculb9vtj0A0yca8/KABw+0GrNfP+nPJWzQ8hdShJubxCE3SB4SEyx0eP4eOiQpzKqHv+QhjTOSJcficpz0ccvxGb9Eo0drH5OWBkyZIm086HGPUF5eHjZs2IClS5ciOjpavZ05cwaOjo7YsmWLvoYiIiIiIiLSid72CIWFhSE5ORmTJk2CtbW1xn1Dhw5FSEgIXnvtNX0NR0REREREJJne9giFhISgV69exYog4HEhFBUVhbNnz+prOCIiIiIiIsn0Vgjt2bMHe/fuLfE+Ly8vCCHQqlUruLi4QAiB1q1b62toIiIirQUGBkKhUODTTz/VaN+5cyevb0dEVAPofflsIiKiqsLU1BRLlixBcrIOJxUTEVGVxEKIiIhqrF69esHe3h7BwcGVPRUiIqpgLISIiKjGMjQ0xOLFi7F8+XIkJCSUK0alUiEtLU1jIyKiqoeFEBER1WhDhgxB69atsWDBgnL1Dw4OhrW1tXpzcnKSeYZERCQHvV5QtaJ0P78KVqZPXgVMD+rU0X/Ooi5fljX9/fvy5b5wQd43+gsX5MvdvLl8uQHg00/le2xeekm21AAAu7B1suVemjRBttwA0LatfLn95P6G38pKvtwFBVUzdyVbsmQJevTogTlz5jy1b1BQEGbPnq2+nZaWxmKIiKgK4h4hIiKq8bp37w5/f38EBQU9ta9SqYSVlZXGRkREVU+V3CNERESkb59++ilat24Nd3f3yp4KERFVAO4RIiIiAuDp6YkxY8bgm2++qeypEBFRBWAhRERE9J8PP/wQBdX4XCgiIvp/PDSOiIhqpNDQ0GJtLi4uUKlUFT8ZIiKqcCyEiIiI9CC26UBYWmq/cEKzi79IGm9UQANJcQDw24kOkmP7uiVLjj2S1kZSXFs37fobPwLwxI49Q0PAxka7PO/XW61dQGHc7RckxQGAwtFBcuyOHeXvm1MXwBML8Oaa1sKdxl21GtPhp/Va9S/q53pTJccOPRUpKW5CS+kffc+ZvS45tmXmSUlxxof3Sx4THh7SY8t5XbUS3bghKexWHU+tY9LTDSWNVYiHxhERERERUY3DQoiIiIiIiGocvRZCkZGRMDQ0RP/+/TXa4+LioFAoStxOnDihzykQERERERE9lV4LoZCQEMyYMQNHjx7F7du3i91/6NAh3LlzR2Nr166dPqdARESklQcPHmDq1Klo2LAhlEol7O3t4e/vjz///LOyp0ZERDLS22IJGRkZ2Lp1K6KionD37l2EhoZi/vz5Gn3q1q0Le3v7cudUqVQaq/ekpaXpa7pEREQAgKFDhyInJwfr16+Hq6sr7t27h99//x2JiYmVPTUiIpKR3gqhbdu2wcPDA+7u7hg7dixmzZqFoKAgKBQKyTmDg4OxaNEifU2RiIhIQ0pKCo4dO4aIiAj4+PgAAJydneHl5VXJMyMiIrnp7dC4kJAQjB07FgAQEBCA1NRUHDlyRKNP586dYWlpqbGVJSgoCKmpqeotPj5eX9MlIiJSvxft3Lmz3NcPUqlUSEtL09iIiKjq0UshFBsbi1OnTmHUqFEAACMjI4wYMQIhISEa/bZu3Yro6GiNrSxKpRJWVlYaGxERkb4YGRkhNDQU69evh42NDbp06YL58+fj7NmzpcYEBwfD2tpavTk5OVXgjImISF/0UgiFhIQgLy8Pjo6OMDIygpGREVatWoWff/4Zqamp6n5OTk5wc3PT2IiIiCrT0KFDcfv2bezevRsBAQGIiIhA27ZtERoaWmJ/Hq1ARFQ96FwI5eXlYcOGDVi6dKnGnp4zZ87A0dERW7Zs0cc8iYiIZGNqaorevXvj/fffx/HjxxEYGIgFCxaU2JdHKxARVQ86L5YQFhaG5ORkTJo0CdbW1hr3DR06FCEhIQgICAAAJCYm4u7duxp9bGxsYGpqqus0iIiI9KZ58+bYuXNnZU+DiIhkpPMeoZCQEPTq1atYEQQ8LoSioqLUJ5L26tULDg4OGhvfaIiIqLIkJiaiR48e+N///oezZ8/i+vXr2L59Oz777DMMGjSosqdHREQy0nmP0J49e0q9z8vLC0IIAFD/S0RE9KywtLREx44d8dVXX+Hq1avIzc2Fk5MTJk+eXOxaeEREVL3o7TpCREREVY1SqURwcDCCg4MreypERFTBWAgRERHpgfvOJbCScs5ry5bSBrx5U1ocgL6tsyXH3sl2kBzbtKm0uOPHteuf0RiAhWablUk2fBpc1SpPYssp2g38n0uXJIUBAHbskB774ovl79txN2DTXrPNOCMZDrFh2g363zUkpeiXIzkUW3Z2khQ3cqT0MVvGXZce7CBxpeQc6Q+Sqo70v1VlSorkWNSrJymsvnWG1jFpCu1jitLbBVWJiIiIiIiqChZCRERERERU41TNQ+Nu3gSUysqehfZ69pQ1vVNGjGy5HXs1ky03ABjelu+ChJ9+Ku9V3+fNky93r17y5QaAq90nyJZb+ZtsqeWnyyEB5TF9uny5o6Ply52ZKV9uIiKiCsY9QkREVCNFRERAoVCUuvn5+VX2FImISEZVc48QERGRjjp37ow7d+4Ua9+9ezdee+01TJs2rRJmRUREFYWFEBER1UgmJiawt7fXaIuJicGcOXMwf/58DBs2rMQ4lUoFlUqlvl140XAiIqpaeGgcERERgJSUFAwaNAi+vr746KOPSu0XHBwMa2tr9ebkJO95iEREJA8WQkREVOMVFBRg9OjRMDIywqZNm6BQKErtGxQUhNTUVPUWHy/fYi9ERCQfHhpHREQ13vz58xEZGYlTp06hVq1aZfZVKpVQVsWVS4mISAMLISIiqtF+/PFHfPHFF9i7dy+aNGlS2dMhIqIKopdD4wIDA0tcejQgIAAA4OLigmXLlhWLW7hwIVq3bq2PKRAREWktOjoakyZNwqeffgp/f//Kng4REVUgve0RCggIwLp16zTaeOgAERE9qx4+fIjBgwfD19cXY8eOxd27dzXuNzQ0hJ2dXSXNjvQtzuASZilf0iomL9dU0ljZ9k/vU5qcutJjO+4uf9/45+1wES9rtJ0xM8NJ97PaDZp0U7v+RRQIyaFIaiMt7pvEx/+2MLLHD9YvSp8AVQt6K4SUSmWxZUiJiIieVXv37sWNGzdw48YNODg4FLvf2dkZcXFxFT8xkkW24hEuGkZrFyT1g7q0+knnWJv25e97ES8jFfU12lINgbuW6doNmqdlf32xlhZ2JVe/06Cq7Zk+R4jXaiAiIrmMHz8e48eP11u+qy/NhaWlldZxTdL+ljZggwbS4gDg0CHJoQ7Xrkkfd/BgaWP+sVmr/r/Vi8cdC0lDUU2Rlw9FSrJWIXdMG0keziErQVLcH3HS/867mmr382kwNJQemyux2rx3T/uYdN0Kcb0tnx0WFgZLS0uNbfHixer7586dW+b9JeG1GoiIiIiISA562yPk5+eHVatWabTVqVNH/f+3334bgYGBGvd/8803OHr0aKk5g4KCMHv2bPXttLQ0FkNERERERKQzvRVCFhYWcHNzK/V+W1vbYvcXLZRKwms1EBERkbZckooscuHoKD1RTo60uLw8yUPmmpZ9HauyGGeU/1CoM2ZmSH3i6CeDXIGCMw21GtNBh4c3K0t6bF0tFpVIsLwHlZHE3yVVa8/0OUJERERS3b17F5988gn27t2LW7duoV69emjdujVmzZqFnj17wsXFBbNmzcKsWbM04hYuXIidO3ciOjq6UuZNupsVMeD/b8ydKz1Ryi1pcQ8eSB7yTuOukmMdYsPK3fek+9liCyMUnGkIvDBNqzFHvqlVdw0XL0qPHTGi/H0/8lqJqzbx0gejaktvhZBKpSq29KiRkRFsbW31NQQREVG5xMXFoUuXLrCxscHnn38OT09P5ObmYv/+/Zg+fTou6vIJjIiIqgW9FUL79u0rtvyou7s732yIiKjCTZs2DQqFAqdOnYKFxf8vH9aiRQtMnDixEmdGRETPCr2sGhcaGgohRLGtsAiKi4srdugB8PjwAx56QERE+pSUlIR9+/Zh+vTpGkVQIRsbG53yq1QqpKWlaWxERFT16G35bCIiomfBlStXIISAh4fHU/vy0g5ERDUXCyEiIqpWhBDl7vv2228jOjpaY3vttdfKjAkKCkJqaqp6i4/nSdhERFURV40jIqJqpUmTJlAoFOU6R5WXdiAiqrm4R4iIiKqVOnXqwN/fH99++y0yMzOL3Z+SklLxkyIiomdO1dwj9PHHgJWV/vPu2KH/nEVFRVXZ/IaBgbLlBoA9GX6y5X7pJdlSAwB69ZIvd/v28uUGgBLWMNGbr2bdkC85AGRny5db7gc+MlK+3Pb28uU2NHx6n2fEt99+iy5dusDLywsffvghWrVqhby8PBw8eBCrVq1CTExMZU+RiIgqWdUshIiIiMrg6uqK06dP45NPPsFbb72FO3fuwM7ODu3atcOqVasqe3pERPQMYCFERETVkoODA1asWIEVK1aUeH9cXFyJ7QsXLsTChQvlmxgRET0TWAgRERHpQeNrB2Flbq59YJ8+0ga8dk1aHACsXy89dsoU6bHnzkmLa9pU+pg6XK9w+52ukuL69XOXPKbDTzr8bsaOLX/fpJtAXrrm2I7AyDe1G/Krr7TrX9S2bdJjf/21/H3TPADYaLblFBgiPqO2VmOamGjVXcPWow0kxeny1NfpNJLKuD7ahg3ax6hUOg3JxRKIiIiIiKjGYSFEREREREQ1DgshIiIiIiKqcVgIERFRjRAYGIjBgwcX+39RERERUCgUvNYQEVENoNdCKDAwEAqFotjWq1cv2NvbY/HixcVihg8fDm9vb+Tn5+tzKkRERERERKXS+6pxAQEBWLdunUabUqnEsWPHMGzYMAwcOBCenp4AgO3btyMsLAz//PMPDKvQhfqIiIiIiKhq03shpFQqYV/Clc1feOEFjB49GuPHj8fJkyeRkpKC6dOn49NPP4W7u/RlJomIiCqSSqWCqsiSrWmVscwsERHprEKvI/T111/D09MTH330EWJiYtCyZUvMmDGj1P58syEiIrmEhYXB0tJSo608h2kHBwdj0aJFck2LiIgqiN4XSyh8Yym6FZ4bZGVlhXXr1mHx4sU4cOAA1q1bB4VCUWqu4OBgWFtbqzcnJyd9T5eIiGooPz8/REdHa2xr1qx5alxQUBBSU1PVW3x8fAXMloiI9E3ve4T8/PywatUqjbY6deqo/9+jRw94e3ujdevWcHZ2LjNXUFAQZs+erb6dlpbGYoiIiPTCwsICbm5uGm0JCQlPjVMqlVAqlXJNi4iIKojeC6GS3liKDWpkBCOjpw/NNxsiIiIiIpIDryNEREREREQ1jt73CKlUKty9e1dzECMj2Nra6nsoIiKicisoKCjX0QhERFQz6P0dYd++fXBwcNBoc3d3x8WLF/U9FBERUbndv39ffeh2aGhoiX18fX0hhKjAWRERUWXRayEUGhpa6ptLUREREfocloiIqFTJycn4888/ERERgddee02+gbp3B6ystI9bu1baeE85H7dMS5dKj/37b+mxI0ZIi7t5U/qYcXGSQ4elhUgL3JAjecyf602VHNtPi2ELSqj3s7IAbb+33rZNu/5FDR8uPfbTT8vf97BZ8TaT1AdwOrVeqzHje4zXqn9RI1rFSIo7mNBM8pg6efRIemxiorS4efO0j0lLA5YskTYeKvg6QkRERBVt4sSJ+Ouvv/DWW29h0KBBlT0dIiJ6RrAQIiKiau2XX36p7CkQEdEziKvGERERERFRjcNCiIiIqqUHDx5g6tSpaNiwIZRKJezt7eHv748///wTAODi4oJly5YVi1u4cCFat25dsZMlIqIKx0PjiIioWho6dChycnKwfv16uLq64t69e/j999+RKPVEXiIiqlaqZCF0IdYQlpaGes/b0qyEZUX0qU8fefPfuSNf7j175MsNwDvIT7bcdmHrZMsNAFe7T5At96xZsqUGAJTwZbjezJ7tLF9yAE6P/pEvedOm8uUGgNq15cst59LPaWny5dazlJQUHDt2DBEREfDx8QEAODs7w8vLq5JnRkREzwoeGkdERNWOpaUlLC0tsXPnTqhUKr3mVqlUSEtL09iIiKjqYSFERETVjpGREUJDQ7F+/XrY2NigS5cumD9/Ps6ePavRb+7cueqiqXBbvHhxmbmDg4NhbW2t3pycnOT8UYiISCYshIiIqFoaOnQobt++jd27dyMgIAARERFo27atxoW/3377bURHR2tsT7voalBQEFJTU9VbfHy8zD8JERHJoUqeI0RERFQepqam6N27N3r37o33338fr7zyChYsWIDAwEAAgK2tLdzc3DRi6tSpU2ZOpVIJpVIp15SJiKiCcI8QERHVGM2bN0dmZmZlT4OIiJ4B3CNERETVTmJiIoYNG4aJEyeiVatWqFWrFqKiovDZZ59h0KBBlT09IiJ6BrAQIiKiasfS0hIdO3bEV199hatXryI3NxdOTk6YPHky5s+fX9nTIyKiZ4DshVBgYCCSk5ORmZkJQ0ND7N+/X+P+lStXYv78+Th37hwaNGgg93SIiKgGUCqVCA4ORnBwcKl94uLiSmxfuHAhFi5cKM/EiIjomVEhe4QUCgXWrVsHT09PrF69GlOmTAEAXL9+He+88w5WrVrFIoiIiKq2L78ETE21j2vXTtp4xsbS4gDgiWXEtaLLoYVXrkiLy86WPmZCguTQ2K6TJMU9sf6GVoaeipQcu2Vnp3L3TWoDwFqzrW5dYMQI7cb89Vft+hf16afSY+fN06JzcwB2T7TZ2gKjR2s1ptOvu7Tqr6F7d0lhSTr8qeLRI8mhN8ybSY5t0FVarOGh/U/v9CQdz/mssMUSnJyc8PXXX2POnDm4fv06hBCYNGkS+vTpg5dffrmipkFERERERFSx5wiNHz8ev/zyCyZOnIgXX3wR586dw/nz50vtr1KpNK4Izqt3ExERERGRPlT48tnff/89zp07h1mzZuH777+Hnd2T+yr/H6/eTURE2oqIiIBCoSh18/PzQ1xcHBQKBaKjo4vF+/r6YtasWRU+byIiqlgVXgjVq1cPU6ZMQbNmzTB48OAy+/Lq3UREpK3OnTvjzp07xbbVq1dDoVBg2rRplT1FIiJ6BlTK8tlGRkYwMnr60Lx6NxERacvExAT29vYabTExMZgzZw7mz5+PYcOGlbpiHBER1Ry8jhAREVVrKSkpGDRoEHx9ffHRRx/pnI/nrxIRVQ8VfmgcERFRRSkoKMDo0aNhZGSETZs2QaFQaNzfuXNnWFpaamzHjh0rMyfPXyUiqh64R4iIiKqt+fPnIzIyEqdOnUKtWrWK3b9161Y0a6Z5zYsxY8aUmTMoKAizZ89W305LS2MxRERUBcleCIWGhhZr41W7iYhIbj/++CO++OIL7N27F02aNCmxj5OTE9yeuPqlmZlZmXl5/ioRUfXAQ+OIiKjaiY6OxqRJk/Dpp5/C39+/sqdDRETPIB4aR0RE1crDhw8xePBg+Pr6YuzYsbh7967G/YaGhpU0MyIiepawECIiompl7969uHHjBm7cuAEHB4di9zs7OyMiIqLiJ0b0jEuwvIePvFZqFZPmIX28w2UfhVq25lr0zfEADnprNJ1TmKCTYYp2Y6ZoM+gT9qqe3qcEiSkPAQBOxkZ41cZG+vhUIhZCRERUrYwfPx7jx49/aj8hRIntLJKoplIZ5eCqjZYXr7eRZSpPZ6dF34PeQKLmlyIZAE4gT8tBLbXsX8SDkl9vni5X+pj0VFWyEGp+fA2snnIyqyRyr/pz+bK8+fVwfYxSRUbKlxuAHR7Ilntp0gTZcgOA8jf5cn8164Z8yQHMnu0sW+6GDWVLDQBYvryNbLlfb/2HbLkBADdk/L2mp8uX+9Ej+XJXB5MnAyWsTPdUSUmShrth6i4pDgCcE3dJjsWVK5JD/4G0v9s2CXskjxnr/fSiuDRPXJe33DZskDwkJrSU/tFs5Mjy9/0mEbjCz9hVRl2zTIxoqmWRat5K8njOl/6WHJvfoJ20wPr1tY/JyJA21n+4WAIREREREdU4LISIiIiIiKjGqZKHxhERERGRdC2Mihz3l5cvOU9OgfRVGE1SdTgs3ta23F3PKUxQ7AAqkQ/cMNVqSE9Prbpr0OaMjvNJecjMk3pOEWmDhRAREVUrgYGBWL9+PYKDgzFv3jx1+86dOzFkyBD1IglCCKxZswZr167F+fPnUVBQAGdnZ/Tq1QszZswodqFVourkB+sX1f9XpCRLzhOfUVtyrNOp9ZJjMXp0ubt2MkwpvjDCDVPgm/IXUwCwLkqr7ho8tFhdr8fOhzh1nydwVQQeGkdERNWOqakplixZguTkkj/gCSEwevRozJw5E/369cOBAwdw4cIFhISEwNTUFB9//HEFz5iIiCoa9wgREVG106tXL1y5cgXBwcH47LPPit2/detW/Pjjj9i1axdeeOEFdXvDhg3h7e1d6tLaRERUfXCPEBERVTuGhoZYvHgxli9fjoSEhGL3b9myBe7u7hpFUFEKhaLU3CqVCmlpaRobERFVPbIVQoGBgRg8eHCx/xcVEREBhUKBlJQUuaZBREQ11JAhQ9C6dWssWLCg2H2XLl2Cu7vmdXhmzZoFS0tLWFpaokGDBqXmDQ4OhrW1tXpzkvsadEREJAvuESIiompryZIlWL9+PWJiYp7a991330V0dDQ++OADZJRxkb6goCCkpqaqt/h4LS9ySEREzwQWQkREVG11794d/v7+CAoK0mhv0qQJYmNjNdrs7Ozg5uaGevXqlZlTqVTCyspKYyMioqrnmS6EeBw2ERHp6tNPP8WePXsQGRmpbhs1ahRiY2Oxa9euSpwZERFVpgpbNS4sLAyWlpYabfn5ZV/AKzg4GIsWLZJzWkREVM15enpizJgx+Oabb9RtI0eOxI4dOzBy5EgEBQXB398fzz33HG7cuIGtW7fC0FD6RSKJiKhqqLA9Qn5+foiOjtbY1qxZU2YMj8MmIiJ9+PDDD1FQUKC+rVAosHXrVixbtgy//vorevbsCXd3d0ycOBFOTk74448/KnG2RERUESpsj5CFhUWxq3SXtKRpUUqlEkqlUs5pERFRNRMaGlqszcXFBSqVSqPNwMAAU6ZMwZQpUypoZkRE9CzhBVWJiIj0INPMFobm2i+cYFFkT5U2nP8KkxQHAGjaVHrsvn2SQ00D2kgfVyJ3x3TpwStXSgqbYGEhechzZq9Ljm0Zd11S3B3TRpLHNDGRHIr4HuMlxzr9qsX5fSnNAWienuHpCayL0m7M9u2161/UyJHl73u1+HSB/AKgjNUsS/LXaemH+HaAtNclADC8LfEIrvr1tY/Rcf2AZ3qxBCIiIiIiIjmwECIiIiIiohpHtkPjih6jXdLx2gDg6+sLIYRcUyAiIiIiIioR9wgREVGNFhgYiMGDB2u0/fTTTzA1NcXSpUsrZ1JERCQ7LpZARERUxJo1azB9+nR89913mDBhQmVPh4iIZMI9QkRERP/57LPPMGPGDPz4448sgoiIqjnuESIiIgIwd+5crFy5EmFhYejZs2ep/VQqlcY1idJ0XL6ViIgqB/cIERFRjffbb7/hs88+w65du8osggAgODgY1tbW6s3JyamCZklERPpUNfcIPXgAmJrqP+/YsfrPWcSD7Fqy5o86IF/uvj06yJccAB49ki1127aypZZfdras6Z0e/SNb7uXL5b1w4owZ8uV+PfSqfMkB4IkT8/XqxAn5cisU8uWuZK1atcLDhw+xYMECeHl5wdLyyasZ/r+goCDMnj1bfTstLY3FEBFRFcQ9QkREVOPVr18fERERuHXrFgICApCenl5qX6VSCSsrK42NiIiqHhZCREREAJydnXHkyBHcvXv3qcUQERFVfSyEiIiI/uPk5ISIiAjcv38f/v7+XAiBiKgaYyFERERURIMGDRAREYGHDx+yGCIiqsaq5mIJREREehIaGlqsrX79+rh06VLFT4aIiCoMCyEiIiI9sDi2Dxbm5lrH/e0wQNJ47erWlRQHAMjJkR47bZrk0GZn/5IWaGsreUwcOiQ9ds6cCh+zZeZJybFwcJMWlpUgecitRxtIjh3RKkZyLLp3L3/fvSrggdBoMjMDPDy0G3LkSO36F/Xjj1p0ngngyYUrLS2A9u21GrOhDjuz822lrxZcUCAtzvj4Ee2DMjOlDfYfHhpHREREREQ1jt4KocjISBgaGqJ///4a7XFxcVAoFIiOji4W4+vri1mzZulrCkREREREROWit0IoJCQEM2bMwNGjR3H79m19pSUiIpLN3bt38cYbb8DNzQ2mpqZ47rnn0KVLF6xatQqPZLzQMxERVT69nCOUkZGBrVu3IioqCnfv3kVoaCjmz5+vj9RERESyuHbtGrp06QIbGxssXrwYnp6eUCqV+Pfff/H999+jfv36eOGFFyp7mkREJBO9FELbtm2Dh4cH3N3dMXbsWMyaNQtBQUFQKBQ65VWpVFCpVOrbXMKUiIj0Zdq0aTAyMkJUVBQsLCzU7a6urhg0aBCEEGVEExFRVaeXQ+NCQkIwduxYAEBAQABSU1Nx5Ijmyg+dO3eGpaWlxnbs2LEy8wYHB8Pa2lq9OTk56WO6RERUwyUmJuLAgQOYPn26RhFUVGlf5qlUKqSlpWlsRERU9ehcCMXGxuLUqVMYNWoUAMDIyAgjRoxASEiIRr+tW7ciOjpaY2v/lGUAg4KCkJqaqt7i4+N1nS4RERGuXLkCIQTc3d012m1tbdVf1s2dO7fEWH5JR0RUPeh8aFxISAjy8vLg6OiobhNCQKlUYsWKFeo2JycnuLlprm9vZmZWZm6lUgmlUqnrFImIiMrl1KlTKCgowJgxYzQOzS4qKCgIs2fPVt9OS0tjMUREVAXpVAjl5eVhw4YNWLp0Kfr06aNx3+DBg7FlyxYEBAToNEEiIiJ9c3Nzg0KhQGxsrEa7q6srgLK/qOOXdERE1YNOh8aFhYUhOTkZkyZNQsuWLTW2oUOHFjs8joiI6FlQt25d9O7dGytWrECmjlcmJyKiqkmnQigkJAS9evWCtbV1sfuGDh2KqKgonkRKRETPpJUrVyIvLw/t27fH1q1bERMTg9jYWPzvf//DxYsXYWhoWNlTJCIiGel0aNyePXtKvc/Ly0u99GhpS5BGREToMjwREZFkjRs3xj///IPFixcjKCgICQkJUCqVaN68OebMmYNp06ZV9hSJSCbnk/LQY+dDrWKuNtdhwJnl7+oUmQuzXZptOcb5GP91ulZD5uZL39ttrEOFIPXCA4r0x+daujXKwftv35M+AS3o5TpCREREVZGDgwOWL1+O5cuXV/ZUiKgCZeYJnLqfq12QpQ4DahFrtguwuKHZJgD8i3wtB9W2f2UrexE1ObAQIiIi0oesLEDChcTb3S796IoyPeUSFGXKyJAee/++9NgnFqcot969pY+Zlyc5NDFF2uGRVj38JY9pfHi/5Fjk5EgK+yOugeQhmzaVHIqDCc0kxyadLX/fxJSHALQseqjSZBhY4S9zn/L1LdDtFBy9XFCViIiIiIioKmEhRERERERENU6VPDTu9vggpNey0nveH7/Xe0oNw4fLm7+vye/yJS/oLF9uANi3T7bUfnKvXJiSIl9uXQ59KQ9djml4itdb/yFbbgB4PfSqbLkVgeNlyw0AJ07Il/v4BemH5DxNdjZXASWiqsepyJn/dc10WC4/v0B6rKVFubvmGOcXW3Ag3wjIzjPWakjv1tla9S8qOctUcqw2l1lLuJwHVZbU5RV0VyULISIiIn0QQqB3794wNDTE/v2a52asXLkS8+fPx7lz59CggfRzKIiocr1qY6P+/4im8dIT6XJunRZfbI7/Or3YwgjZeca4BFuthozddlmr/kWFxUp/zXNyKn/fj8c9xNWzlXf+Fg+NIyKiGkuhUGDdunU4efIkVq9erW6/fv063nnnHSxfvpxFEBFRNcVCiIiIajQnJyd8/fXXmDNnDq5fvw4hBCZNmoQ+ffrg5ZdfruzpERGRTHhoHBER1Xjjx4/HL7/8gokTJ+LFF1/EuXPncP78+RL7qlQqqFQq9e00uc9DJCIiWbAQIiIiAvD999+jRYsWOHr0KH7++WfY2dmV2C84OBiLFi2q4NkREZG+8dA4IiIiAPXq1cOUKVPQrFkzDB48uNR+QUFBSE1NVW/x8TqcfE1ERJWGe4SIiIj+Y2RkBCOjst8alUollNqsD0tERM8krfcI+fr6YtasWcXaQ0NDYWNjg3379kGhUODu3bsa9zs4OMDFxUWjLS4uDgqFAr//LuP1b4iIiIiIiJ6g90PjunbtCiMjI0RERKjbYmJikJWVheTkZMTFxanbw8PDoVQq0aVLF31Pg4iIiIiIqFR6L4QsLS3RoUMHjUIoIiICXbt2RZcuXYq1e3t7w9RU+tVriYiIiIiItCXLYgl+fn4IDw9X3w4PD4evry98fHw02iMiIuDn51dqHpVKhbS0NI2NiIhILgsXLkR0dHRlT4OIiCqALIsl+Pn5YfHixbhz5w4cHBxw5MgRvP3228jLy8OqVasAANeuXcPNmzfLLIS4RCkREVUZzZsDlpbax7m6ShtPl4JNlyMxMjIkh8a0GyspzlaHr20Lmj4nOVbqhyTjjGTJY8LDQ3Koqo6DpLiupjrM18pKeqwuHj2SFmfeSvKQf502lBzbUIvv8nPzMwHka7R5t85G7LbLWo2paNpEq/5FiWN/SI6Fefn/5swNzABoPq6WeSnokPR3ueLTMjO1mVkxsuwR6ty5M0xMTBAREYELFy4gKysLbdu2Rfv27fHgwQNcv34dERERMDMzg7e3d6l5uEQpERERERHJQesvO6ysrJCamlqsPSUlBdbW1gAAc3NzeHl5ITw8HElJSejatSsMDQ1haGiIzp07Izw8HOHh4ejSpQtMTExKHYtLlBIRERERkRy03iPk7u6O06dPF2s/ffo0mjZtqr7t5+eHiIgIREREwNfXV93evXt3RERE4MiRI2UeFkdERERERCQXrQuhqVOn4tKlS5g5cybOnj2L2NhYfPnll9iyZQveeustdT8/Pz9cvnwZ+/fvh4+Pj7rdx8cHO3fuRHx8PAshIiJ6JgQGBmLw4MGVPQ0iIqpAWh8a5+rqiqNHj+Ldd99Fr169kJOTAw8PD2zfvh0BAQHqfp06dYJSqYQQAu3atVO3d+zYEbm5uepltomIiIiIiCqapAVROnTogAMHDpTZx9TUFNnZ2cXalUolsrKypAxLRERERESkF7Isn01ERFRdqVQqqFQq9W1e446IqGqSZflsIiKi6io4OBjW1tbqzcnJqbKnREREErAQIiIi0gKvcUdEVD3w0DgiIiIt8Bp3RETVA/cIERERERFRjVMl9wg5GtyFlUGm3vO++KKD3nMWdf++rOnh1KOHfMnz8uTLDQA3b8qX28pKvtwAMH26fLkjI+XLDQC1a8uX+8YN+XIDgIzXfDlxQrbUAABvb/lyr1kjX25jY/lyPwtSU1MRHR2t0Va3bl2eA0REVE1VyUKIiIhI3yIiItCmTRuNtkmTJmGNnNUlERFVGhZCRERU44WGhiI0NLSyp0FERBWIhRAREZE+2NkBtWppHZaZJ23hBQsTE0lxAPBPwfOSY5OSJIeiZ6sH0gJ1OTy7hIu7l5udnbS4y3HSx0xIkByqTEmRFmhoKHlM6HIdrUePJIfeMG8mKc750t+Sx+yAAsmx+bYdyt3XuIRP58lZpgiLbaDVmOLYH1r1L0rRravk2NdeK3/f63ceAsjVaEtV2OB3I/9yxWca6XYdNy6WQERERERENQ4LISIiIiIiqnFYCBERUY0VGBgIhUIBhUIBY2NjNGrUCO+88w6ydTmcioiIqgSeI0RERDVaQEAA1q1bh9zcXPz9998YP348FAoFlixZUtlTIyIiGeltj1DRb9WKbgEBAQAAFxcXdZu5uTk8PT25JCkREVU6pVIJe3t7ODk5YfDgwejVqxcOHjxY2dMiIiKZ6XWPUOG3akUplf+/Gs6HH36IyZMn49GjR9i+fTsmT56M+vXro2/fvvqcBhERkSTnzp3D8ePH4ezsXGoflUoFlUqlvp2my6pZRERUafR6jlDht2pFt9pFrlxfq1Yt2Nvbw9XVFXPnzkWdOnX4rRsREVWqsLAwWFpawtTUFJ6enrh//z7efvvtUvsHBwfD2tpavTk5OVXgbImISF8q5RyhgoIC/PLLL0hOToZJGddB4LduREQkNz8/P6xatQqZmZn46quvYGRkhKFDh5baPygoCLNnz1bfTktLYzFERFQF6XWPUOG3akW3xYsXq++fO3cuLC0toVQq8dJLL6F27dp45ZVXSs3Hb92IiEhuFhYWcHNzw/PPP4+1a9fi5MmTCAkJKbW/UqmElZWVxkZERFWPXgshPz8/REdHa2yvFbm87Ntvv43o6GgcPnwYHTt2xFdffQU3N7dS8wUFBSE1NVW9xcfH63O6REREGgwMDDB//ny89957yMrKquzpEBGRjPRaCBV+q1Z0q1Onjvp+W1tbuLm5oVu3bti+fTtmzpyJCxculJqP37oREVFFGzZsGAwNDfHtt99W9lSIiEhGlXZBVScnJ4wYMQJBQUGVNQUiIqJijIyM8Prrr+Ozzz5DZmZmZU+HiIhkotfFElQqFe7evas5gJERbG1tS+z/xhtvoGXLloiKikL79u31ORUiIqKnCg0NLbF93rx5mDdvXsVOhoiIKpRe9wjt27cPDg4OGlvXrl1L7d+8eXP06dMHH3zwgT6nQUREREREVCa97REKDQ0t9Zs1AIiLiyuxfd++ffqaAhERUeUxNgbKuCREaXJypA1nYW4uLRCAjfbTVNPpdN2HD6XF2dhIHjLdtpHkWFPl0/uUxNhIh49XN25Ij61XT1pcbq70MXWRmCg5tEHXZpLi8hu0kzym4W3pi3YVFJS/ryihTakEtF482fw5LQP+X5G1zrT23Xfl79sUgMUTbQUFQHnXqsnOLv9YJam0c4SIiIiIiIgqCwshIiIiIiKqcVgIERFRjfDdd9+hVq1ayMvLU7dlZGTA2NgYvr6+Gn0jIiKgUChw9erVCp4lERFVFBZCRERUI/j5+SEjIwNRUVHqtmPHjsHe3h4nT55EdpGDzcPDw9GwYUM0bty4MqZKREQVQK/LZ1eUM/fsYZmp/4urtrFL0HvOos7faSBrfgGFbLkVmzfLlhsAUMoS63qhzRmKUkRHy5fb3l6+3AAgSjolU0/S0+XLDQAnTsiW+vgFf9lyA8CaNfLlfuUV+XJXde7u7nBwcEBERAS8vb0BPN7zM2jQIBw+fBgnTpxQ7xmKiIiAn59fJc6WiIjkxj1CRERUY/j5+SE8PFx9Ozw8HL6+vvDx8VG3Z2Vl4eTJk6UWQiqVCmlpaRobERFVPSyEiIioxvDz88Off/6JvLw8pKen459//oGPjw+6d++OiIgIAEBkZCRUKlWphVBwcDCsra3Vm5PWa9oSEdGzgIUQERHVGL6+vsjMzMRff/2FY8eOoWnTprCzs4OPj4/6PKGIiAi4urqiYcOGJeYICgpCamqqeouPl35tESIiqjxV8hwhIiIiKdzc3NCgQQOEh4cjOTkZPj4+AABHR0c4OTnh+PHjCA8PR48ePUrNoVQqoVRKvNImERE9M7hHiIiIahQ/Pz9EREQgIiJCY9ns7t2747fffsOpU6e4UAIRUQ3APUJERFSj+Pn5Yfr06cjNzVXvEQIAHx8fvP7668jJyWEhRETPlITLefh43EOtYswNzCSPd/2OdmMV1VSLvnORChfkabQpbyhQb075Vp7NyNdthVoWQkREVKP4+fkhKysLHh4eeO6559TtPj4+SE9PVy+zTUT0rFBlCVw9m6tllKEOI2o71v+z0KKvC/LQDPmajdkAYvNK7P8kgeyndyqDTofG+fr6YtasWcXaQ0NDYWNjAwBYuHAhWrduXaxPXFwcFAoFouW8BgsREdETXFxcIIRATEyMRruzszOEELh48WIlzYyIiCoS9wgRERHpQVxiLdTKqaV1XCNclzReZgN3SXEAcOB/kkMx5VUdLsR8RdrHjvx60vfQGejwhbHxjxulBY4eLXnMW3U8JcfWt86QFnjvnuQxsWGD9Nh58ySHGh7aLy2wfn3JY+oSa3z8SLn7KtKdAEg/rI3Kj4slEBERERFRjfNM7xFSqVRQqVTq27x6NxERERFVZ26NctT/zzCwkpzHMi9FcmyqwkZybEFB+fsqbyjw5Gk+j6DABZiWGdcM2bCADnun/1MhhdC///4LS0tLjTYhnj754OBgLFq0SK5pERERERE9U95/+/8PVfzL3KeMnmXrkPS35Njfjfwlx2Zllb9vvTnpxRZGuABTvIBGZcbtxnW0hxYDlaJCDo1zd3dHdHS0xvbrr78+NY5X7yYiIm0FBgZi8ODBxdojIiKgUCiQkpKi8f8nubi4YNmyZbLPk4iIKpdOe4SsrKyQmpparD0lJQXW1tbq2yYmJnBzc9Mc2OjpQ/Pq3UREREREJAed9gi5u7vj9OnTxdpPnz6Npk21uZwSERERERFRxdFpj9DUqVOxYsUKzJw5E6+88gqUSiX27t2LLVu2YM+ePfqaIxER0TODC/kQEVUPOhVCrq6uOHr0KN5991306tULOTk58PDwwPbt2xEQEKCvORIREWklLCys2CI9+fn5xfo1aNCgWNujR4/KzM2FfIiIqgedV43r0KEDDhw4UOr9CxcuxMKFC4u1F17Zm4iISN/8/PywatUqjbaTJ09i7NixGm3Hjh1DrVqaF0H19fUtM3dQUBBmz56tvp2WlgYnJyfdJkxERBXumb6OEBERkRQWFhbFFulJSEgo1q9Ro0awsbHRaHvaYj5cyIeIqHqokOWziYiIiIiIniXcI0RERERERM+EZsjGblx/ah99YCFERERERETPBAsItEdWhYxVpQqhwsUVMjPlWao0zTRdlryFHj2Sd4lVOVdwVWTJ/ITMy5Mvd0GBfLkBIDNTvtyGhvLlBuR90jxl5S2dKRSypc7Olvdv1dhY1vQyevy4POsL3YSGhpbY7uvrq5570f8/KS4uTqvxCvNkZEh73qRB2ntPpqn056kuL+lpaTr8/jMyJIXl6/Bala3DF8f5Uh8oHeabni79dT9NIe3xRboOn3+KLCWvNV3eg6S+90p8DgKolPlmFEgfM02HzyeZRhXzN5eRnw6hw96dDDz+fUp9X1KIZ/0drYiEhASuzENEVMni4+NLXHa6puJ7ExFR5ZL6vlSlCqGCggLcvn0btWrVgqIc3wYXLmkaHx8PKysrvc9Hzvyce+Xkr6q55c7PuVdO/mdt7kIIpKenw9HREQYGXGunUFnvTbr8DisjlvN9dmM5X863ssd8Fuer6/tSlTo0zsDAQFK1Z2VlJcuHiIrIz7lXTv6qmlvu/Jx75eR/luZubW0t2zyqqvK8N+nyO6yMWM732Y3lfOWN5XzljZVjTF3el/iVHhERERER1TgshIiIiIiIqMap1oWQUqnEggULZLsCuJz5OffKyV9Vc8udn3OvnPxVee70mC6PcWXEcr7PbiznK28s5ytvbGXN92mq1GIJRERERERE+lCt9wgRERERERGVhIUQERERERHVOCyEiIiIiIioxmEhRERERDVKXl5eZU+BSLLk5OTKnkK1USMKIb7gVS451uPgGh9l4+NTMj4uRPJITExEQUGBTjmSk5ORnZ2tpxmV7vz58wgODkZ6errsY+mLLo+tSqXS2zzK+xpa+Hvka67+PXz4EJ6enjh16lRlT0Vr+nwu6ku1L4QuX76MMWPGID8/X9ZxcnJyAOj2YlWWqvhicuXKFZw4cQIKhUKveS9fvowNGzYAgN5/r0lJSRXyRiyXO3fuAIDeH/OKeP7JOUbhByyFQiH7zyLXawCR3B49eiQpLiUlBe7u7ti8ebPkse/du4cRI0bgiy++0Ok1+Gl/32fOnIGnpyeMjY1Rq1YtyeNIcePGDUmvP3FxcVizZg2ioqK0jr18+TKGDBmC+Ph4rWOLunXrFlJTU6FQKJ76GhceHg5nZ2fEx8dXyGtuSSqqqAZQ4QV1eno6DAwMqtxlDxISEtC2bVskJCTonEufz6lqXwhdv34dp06dQlZWlmwfUBISEuDp6YmrV6/CwMBAL78gIQQKCgqQn58PIYTeP9gWjiEXlUqFDz74AFOnTlV/ONeXRYsWYe3atQAAQ0NDveU9ffo0GjdujHPnzuktZ0lyc3MB6P8Dc2JiIvz9/XHkyBG95r18+TLGjh0r2/Pl8uXLOHXqlGxvmEU/YGVlZcn6xnzp0iV88sknAOQpiOSaN4s3/bt16xa2bt2KLVu24PTp01rFCiF0+l3HxcVh9+7dWsX8/fffaNWqFW7evKn1eObm5ujWrRt2796NtLQ0reMBoHbt2qhbty5+++03fPvtt1p/iC3PB/ULFy6gU6dO+OCDDzBv3jxJ85RKpVJh5MiRcHV11ep3+++//8Lf3x/79u3D/fv3tRrz7Nmz8PLywr59+3Q6lCo9PR2TJ09G//79kZycDAMDgzJfMxo2bAhnZ2f06NEDCQkJkl9zr1+/jpUrV2odp6+iGnj6a+7Fixfx4osvav05JyEhAZs2bcIPP/yApKQkrWIbNWoEBwcH/PrrrwCqzuu3EALZ2dmYP3+++nOQtjIyMpCXl6fX9/FqXwjVq1cPDx48QHx8PAwM5PlxU1JSYGhoiG+++QaPHj3SuWi5dOkSZsyYge7du6N9+/YYMGAATp48qd7rpKusrCwA+t9rUJRSqUTv3r2Rnp6Ow4cPA9DfH2u3bt3U/9fXH8KZM2fg6+uLSZMmoX379nrNXdStW7cwYMAAnD179qlvJtrKz8/HrVu3tH5RfZrTp0/jn3/+AaD/x6SgoABfffUVvL29ERkZKUuRUvQD1sqVK2Utho4fP44FCxbgjz/+0MvrTVZWFlJSUtR/++X5NlZbV69exahRo/DgwQO95q3Jzp49i+7du+PDDz/EuHHjMGbMGGzbtq1csZcuXcLMmTMxdOhQLF26VOuxb9++jQ4dOmDevHnYtGlTuWLOnDkDPz8/DBw4EA0bNtR6TBMTE/Ts2ROHDx/Gw4cPAWj3ep+fnw8TExOEhISgRYsW2L59O1auXFnuD7Hl+aB+7tw5+Pj4wMXFBQsXLgSg/WHzjx49wsOHDxEREYFbt25pVfSZmJjg888/h6WlJdq1a1eu15+LFy/Cx8cHL774IlasWIF+/fqVe7zo6Gh4e3tj1KhR8PT0xLVr1wBIew23sLDAsGHDYGhoiLFjxyIpKanM96/GjRtj06ZNcHJyQpcuXXD79m2tX3Pz8vKwceNGLFy4EMuWLdNqvroU1bdv38aePXuwfPnycn2eO3HiBDIzM+Hg4FDuMc6fP48BAwZg3759uHLlCurUqVPu2MLHvGHDhurfqVyfbQvFx8cjJCQEP/zwA/744w/JeRwdHTFlyhScOXMGBw8e1Do+ISEBgwcPxvbt2/VbDIlqKi8vTxQUFIhHjx6JNm3aiIMHD8o63vz580WTJk3E8ePH1eNLcebMGVG3bl0xcuRI8e6774rJkyeLFi1aCAsLC7Fs2TKRnJys0zwvXLggfH19xcyZM0VycrLIzMwUQgiRn5+vU95CBQUFGreHDx8uXFxc9DrOsWPHhLW1tbh8+bLOuYQQIjo6WpiZmYn58+drtN+9e1cv+Yv6/fffRefOnUWXLl3EuXPnhBD6eUwKcwwZMkR89tlnQgjpz8EnnTp1Spibm4sLFy7oJd+T7t27J8aNGycsLCzEH3/8IYQo/jySqvAxyMzMFJMnTxYdO3YUS5cuFVlZWXodp1BKSooYMmSI6Nmzp4iPj9cp1/nz58WgQYNEy5YtxaBBg8Qvv/yin0k+4eLFi0KpVIoBAwaIhw8fyjJGTXLmzBlhbm4u3nnnHXHr1i0RFhYmevToIdq0aSOuXLlSZmx0dLSws7MTgwcPFiNHjhTGxsbi888/12r88PBwYWBgIDp06CAGDRokQkNDyzXfJ1//VCpVucYr+jfUpk0bMXLkSK3mW9J4bdu2FR4eHhp/q2XJz88Xa9euFd27dxf9+vUTiYmJ6nYhHj+u5ubmwtfXVzg6OoqZM2eqY8v7OhkbGyvGjRsnPDw8hKmpqbCxsRGjR48Wf/31V7l/xvz8fBEZGSk8PDxEmzZtynz9ycrKEsOGDRPTp0/XaM/JyRHx8fHi4sWLpcYWvqe9++67QgghHB0dxfLly8s9z6IKHx+VSiU2b94sfH19xQsvvCBSUlLUP9OTP2Phz3X//n1hZGQkPD09RUJCghBCu9fcq1evinnz5gl3d/dy/x087TW/LOfOnRMdO3YU48ePF++9916JeZ+0ePFi0b59+3L/XOfOnRO1a9cW7733nkhNTVW37969W5w+fbrEmKtXr4oVK1aImJgYcfPmTSGEEJs2bRK9e/cWKpXqqc/h8vzspTlz5oxwdnYWXl5eom7duqJx48Zi+/bt5Y5PSkrSuJ2SkiJatWol+vbtq/VcsrKyRIcOHYS3t7f4+eefRU5OjhBC9/fxalcIJSQkiPj4ePUHbyGE8Pf313jh04fk5GSNJ1dBQYFo1aqV6NWrl0abNhISEoSbm5uYO3euRnt8fLwYN26cUCqV4n//+58QQvqH3E8//VQ8//zzomvXrqJ79+5i8uTJ4s8//9ToI/VJdfnyZbFq1Sr1C54Q//8zBQYGSsophND4XQrxuEBp2LCh+OeffyTnLHThwgVhbGws3n//fY32Dz74QNSvX1+kpaXpPMaTfvvtN9GvXz/RsWNHERMTI4TQvRgqjJ86darw9fUVQuj+4pCfny/y8vLE/fv3hbu7uzhx4oRO+Z5U+BzOz88XDx48EFOnThXW1tbqDxb6KlJK+4D16NEjvYyTm5urcXvjxo3C1dVVrFu3Tggh7XcbHR0tbGxsxMSJE8WSJUtEgwYNRMuWLdXFs76dPXtWODo6it69e4v79+/LMkZNcPPmTWFrayuGDRum0f79998LCwsL9d97Sc6cOaPxhUx+fr54/fXXxaxZs7T+IDNx4kTRunVrMXToUNGjRw+xcePGMuc7fPhwjfavvvpKzJkzp9T3mezsbI3bhX8Dn332mWjXrp264Cvrbys2Nlb89NNPxfotW7ZM2NjYCH9/f+Hl5SW++OKLMn/+p31QP3nypDA2NhYLFy4UeXl5YvXq1cLW1larYujMmTPCwcFBvPbaayI0NFTExMSIuXPnCjc3N+Hh4aH+AudJd+7cEZGRkRptOTk54uTJk6JJkyZlFkO5ubmiW7duGgXMvn37xKxZs4SVlZVo1KiR6NmzZ7H4v//+W9SqVUujsPX29hYffvihEKL8r3dFX9eKvoaOGDFCKJVKMXDgQPWH3MOHD4vffvtNCKH5WL700kuiWbNmol27dsLZ2Vn95dDT5lD0NfPGjRvi448/Fg0bNhQrV64s19ylFNWFBcr7778vbt++rW7fsWOH+guiwnkVzfPhhx9qfO4rS2Jioujevbt4/fXXNdo//fRToVAoRI8ePYoVQzk5OWL48OGiYcOGolGjRsLKykoEBAQINzc38dxzz6k/b5X2PpOQkCCGDRsmDh8+XK45FlX4Jcm8efNEZmamOHjwoKhfv77o37+/yMzMfOp725UrV4Stra0YNGiQuHfvnvqz3MmTJ4WpqalWX/IUPq/S09NF7969Rbt27fRWDFWrQigpKUl4eXkJpVIp2rZtKwYNGiTmzp0rRo0aJSZNmiSEePxg6frB59q1a8LJyUkMGzZM4w/m2LFjonbt2mLJkiWS8v7444+iU6dO4sGDB0IIzSd2amqqGDJkiHBwcFB/2yXFvn37hLe3t0hMTBSHDx8WM2bMEDY2NmLGjBnqIquQNh/gVCqV6N69u1AoFKJ9+/YiPDxcfd/nn38uWrVqJfbt26f1fBMSEkT9+vVF7969xcsvvyxCQkLE6dOnRbt27cSyZcu0zvekr7/+WigUCrFjxw51W3BwsHjuuedEWFiYzvmLKvrGEhUVJdzd3UXnzp3Ve1q0/cB85coVsXLlSrF//371B4/vv/9e9OvXT91HynP97t27Ii4uTiPW399f/Q2Zrn8/Rb8FK/qG9cEHHwiFQiGsrKzUxbmUsbT5gKVrMRQTEyO6desmjh49qs4lxOMPC87Ozuo2bX63586dE5aWluKDDz5Qt/34449CoVCIbdu2afTVpYAujC18g3n48KEwMTERY8eOVb8GkXauX78uOnToIF544QVx7NgxdfuBAweEra2tOHPmTIlxpRVQI0aMEK1btxYeHh4iICBArF+/vszxCwuUvXv3isDAQLF//37x4osviu7duxd7fX9yvoUf5oODg4WVlZXGa3hR165dE4MHDxZr167VeM4L8fhLu9q1a4sFCxaUOc+8vDzx+eefC4VCITZt2qRu//TTT0Xt2rXVf/+TJ08WXl5e4quvvio2Vnk/qB85ckTMmDFDfX9KSopWxVDhh8GgoKBiX3ps3bpVtGnTRnh5eRU7QuHmzZuibt26QqFQCF9fXxEUFCR+//139evfqVOnRJs2bUSrVq1KfP1JTU0VHh4eYvLkyeLixYti8eLFwt3dXQwdOlR8/fXXIiQkRLi5uYnZs2erYxITE0XdunU12oQQYvDgwWLMmDFl/pxFXblyRQQFBRUr3L/44gtRu3ZtMXfuXNG1a1fRr18/ERcXJ3r27Ck6deqkLoaEEOLFF18ULVu2FHFxcSI+Pl50795duLi4lLlnKC0tTf17LvyAK4QQ77zzjjA1NRWurq7iiy++KBYXGxur3kshpah+8OCB8PLyElOmTNFoX7JkiVAoFMLDw0P9BdHNmzfFsGHDxIEDB4QQQixYsECMGDFCCKH5mlrSa/OFCxdE48aNxeHDh9X3r1q1ShgbG4tvv/1W9O7dW/Tr10/8/fffGnGFBcSlS5fErl27xPLly8WwYcNEixYtRP/+/dVHr5T0u7169aro1KmT6N+/f6kFe0lKe03q0KGDaNq0qfqLhrJcunRJ2NjYCIVCIfr06SOWLVsm/v33XyGEELNnzxbt27cXJ0+efGqewt9p0S9gnJ2dRadOnfRSDFWbQig7O1uoVCpx8OBBcejQIbFnzx4xbdo00a9fP1G7dm1hZGSk8WFXlw9z586dE3Z2dqJu3bqiUaNGYt26deoPs1OmTBE+Pj7qX7Y2goKCxP+1d+ZxUVVtHJ+LgCyCbIILgooyrAIhoCCICgzDGgihiKGmZJpKiuJO5laY5fa+uVvpm69blJVkuWRSbrymIMtACAihkIpkEALD7/2Dz73NnX2GGZc83790mDlz5t57znl+z/Oc57i5uUn9m1AoRG5uLgwMDPDNN9+o3Xega1KcOnUqMyFcv36dmbBDQkLw0UcfqZzWIxQKsXv3bvD5fLz++uvo378/li9fjp9//plJT5wwYQLzfmWv/++//46dO3di48aN4PF4CAgIQJ8+fUBRFCIiIiQ8k+qwdOlS6Onp4dtvv0V2djYsLCyYSU4UUQNeWerq6nD9+nXm//Tkl52djV69esHHxwd+fn4qp8n9+uuvTGTP1NQU9vb28PPzg4+PDyiKwpkzZ5j3qvKsV1ZWYty4cTAyMkJQUBASExOxbt06jBs3jmU0qDt+6urqEB4ejh07drBef/fdd2FhYYGjR49i6tSpMDY2ZowhVY391atXq2RgiUccVSEzMxMURYHL5WL9+vVM+lpjYyMcHByQkpKiUnsdHR3w9vZG3759WdGfpUuXgqIobNu2DWfOnMGdO3dYn1Plfty+fZtJrxA1ID/44AOYmJjA1NQUPB6PiCE1KSsrQ3h4OMLCwlBcXIxHjx6hT58+WLx4sczPyBIkRkZGWLNmDfbs2QNnZ2cMGzaMNZ8AXfdTdG0DulKSnJycsH37djQ0NCA+Ph7BwcFSxRDd35iYGMycORPW1tY4deqUzL4WFxcjKioKurq6CAoKwtKlS/HHH38wc/GGDRvg5uYmM3Xr5s2beOedd/DXX38xc++pU6ekzr2tra2YNWsWhg0bhu3btzOvV1RUKG2o01EL0ZStpqYmpcSQNGOws7OTJYh27doFU1NT7Nq1i/k7AFRVVcHT0xNcLhcjRoxAamoqDAwM4OnpiSlTpuDw4cM4cuQIuFwuxo4dK3UMnzlzBrq6urC3t4eJiQl27NjBCK62tjaEhYUhNTWVef/jx49x5coVid8zbdo0loOMZtOmTUw6vyjHjx+Hvr4+5s+fz3wfPYf+8MMPALpSs9zd3cHj8XD27FnExMQgJCQEeXl5SE5OhpubGyoqKpg2KyoqEBwcDAMDA6lR57q6Ovj7+2P37t149OgR8/qGDRtgaWmJTz/9FMuWLQOXy2WJodbWVmzbtq1bovrHH3+Eu7s769p98sknMDMzwwcffICxY8fC1dUV9fX1jLDg8/n43//+h6VLl2LKlCkSv4dGVNAdOHAAPXr0YN3rmpoa/PjjjwCAwsJCjB8/Hr6+viwbTNb8npOTg4CAAPB4POaaSlsv6THO4/FYYki03Y6ODlRVVTH/lzYnrV+/HhRFwdfXF9HR0Zg2bRq2bduG2tpaCTFCj5EtW7bgrbfewvLlyzFr1iz4+PggNzcXV65cAZfLxerVq+X2mxZKotdx69atMDMzg4eHh0YiQ/8IIVRcXIzY2FgMHz4curq68Pb2xtq1a5m/l5SUYMmSJTAxMcGhQ4eY19W5YEKhEC0tLVi1ahXeffddvPvuu0hKSkJUVBS++OIL3Lp1CxYWFli3bh3zfmXZuHEjTExMGCNFvH9tbW0wMDDA7t27Ve63aF9yc3MRGhrKRLNmzpyJIUOG4Mcff0RqaipcXFzg5OSkssioqamBv78/9u7di+LiYiZNa9OmTbh48SJ0dHTw6aefKtVWSUkJVqxYwRqYNKWlpTh16hT69OmDhIQEtQQKwF70Fi1aBIqioKuri9OnT0u8d/Xq1XjnnXdYg1ERbW1t8Pb2Bp/PZ+WRr1u3DpaWljh9+jTOnDkDPp8PPz8/FBUVAVD8zFy7dg2WlpbYt28fgK4Ft6SkBB9++CGysrLg5OSE1NRUmROeLH777TekpKTg/fffR0FBAfbs2YPY2FjExsZCX18f/fv3Z0UkVB0/bW1tuHv3LqKjozF69GjGu52dnQ1zc3NmH19DQwNSUlJgZmbGLLqqsmzZMrUMLFVpbm5GUlISxo8fj8zMTIwfPx4LFixAe3s73n//ffj7++PEiRNKtSUQCFBQUICrV6/C0dEREydORFVVFSOaY2Nj8fbbb8PS0hIBAQEICQnB9u3bUVlZqXR///rrL0RFRWH48OEsI2XDhg2wsLDA+fPnUV5ejv79+yM8PJzsGVKTsrIy8Pl8jBkzBubm5khPT2f+Jmt8iwqSGTNmSAiS6upqUBSFnTt3Mq+JRh0iIiJw+PBhCAQCAF17DgIDA9HQ0IDi4mLEx8cjJCQEe/bskfhugUCA0NBQGBoaSvW4S+PGjRtIS0uDg4MD7OzskJGRgcLCQuTn52PgwIFMRF30916/fh0URWH9+vXMa7LmXnp+bm1txfz583Hr1i3mb0eOHIGBgQHS09PlGur+/v6IiYmR+hyLiqG33npL6m+UFeED2PNfUFAQy9FHU15ejri4OMTGxuLSpUuorq7GoUOHEBAQAF9fXxgZGcHd3R0URSEuLk5qH27fvo38/HwJx4RQKERiYiJWrFghM9OFvoYbN25EUFAQq9+rVq2Cjo4OCgoKmPc3NDTg6tWruHv3Lk6dOoUBAwZg+fLlyMjIgKWlJWsOraiogIODA0xMTBATE4Nr164hMjIS/fr1w4ABA5iIgeg6W1ZWhtmzZ8uMSkVGRsLd3Z1J5aSFLT0OKisrkZmZCVdXV6xZswYCgQB8Ph91dXVYvny52nP+1q1bYWNjw0qFP3bsGBOZuXLlCoKCgmBvb49Hjx6hvLwcPB4P8fHx8Pb2xksvvYRXX30VU6dOxfTp05GSkoLU1FQkJCRg/vz5zBi4cOECevbsiePHj7PuBX0/gS5h7ePjI+HsEoV+r1AoxH//+1+MHTsWI0eOlOu8kiWGgC4BnZ6ejsTERJZjUHxO6tOnD44ePYrq6mrk5ORg7dq1sLGxga2tLaKiotDZ2ckSsQDwww8/IDw8HCdPnkRLSwu2bdvGCMzw8HCYmZnJDBykpaWBoijW2NuwYQPMzc1x8eJFdHZ2MmlyOTk5Ktlnojz3QqigoAC9e/fGnDlzsGfPHhw/fhyxsbHQ1dVFeHg4o0rLy8uRnp4OCwsLhZtHpSFubJ8/fx5cLhfff/89Hj58iH//+98wMTHB+vXrkZKSgp49e7K8C/Kg+3jq1CnY2NggIyNDYhLp6OhAWVkZPD09Vd6rQQ8aetA1Nzdj+PDhWLduHebMmYN+/foxfRUKhSgqKkJ1dbXCdqurqyUW1bNnz6JXr17Iy8tDW1sbzp07Bzs7O4SHh8PBwQFDhgxhDH5ZtLW1MZGNoUOHYuHChfjss88k3vfzzz/D3NwckyZNUipMC3QJrGXLlqGqqkrCIFm3bh0oipLYCJiVlQWKoiTC1fIQCAQoKirCxYsXGaO2pqaGmaBPnjzJvPfbb79FdHQ0uFyu3D0EwN+bfhctWiTzPadOnYKvry9eeeUVqd4+WZSXl2P06NHg8/kSz25+fj7mz58PLpfLuhfKiqGamhokJCTg+vXruHPnDpKTkxEUFIT4+HhYWVmxIlhAVyQwNjYWAwYMkPDeyUMZcSvPwFKV9vZ27N27F7Nnz8b58+fx008/YdCgQZg8eTI2bdqEwYMH44033lDYzi+//AIjIyNs3boVQNeCOXjwYLi6usLMzIyVplRbW4u8vDyEh4fDz89P6f7THsbTp09j3LhxGDNmDFpbW7F9+3aWsQF0ee3p1IPupOK+yJSVlWHcuHGwt7fH+fPnmdcV7ZsRFySdnZ1oa2tDbW0tPDw8WPNTVVUVRowYgVGjRuGll17CjBkzYG9vj507d+Lw4cOIiopi5pqioiKEhIQgOjpaqvPo119/RVhYGPh8PsvwkNff1tZWNDY2IiMjAwEBAdDT00NWVhasrKzg5eXFMoqKiopgaGgoNW2O9jSLz73iqWiiHDhwAP3798eyZcukGupCoRCHDh2Ci4sLXnnlFakCtKmpCbt37wZFUViyZInU7xE1IGVdl+DgYCQnJ0v9fGlpKXg8HkJDQ1nzamNjIxPl8PLykrlRXhqPHz/GihUr0L9/f5SVlSl8//79+2FjY8NEx1atWgUDAwPk5+cz7ykqKkJAQABCQ0MZUfbxxx/D2toaenp6zJ5H4O85dMeOHXB1dUVUVBQaGhpQWFgIHo8Hf39/1p4UadeebqOyshLbt29nBG1iYiI8PT3xyiuvsIQtTVVVFd588024urrC1NQUFEUxKacZGRlKz/mi6d/79u2Dvr6+XNskOzsbPj4+jK1RWloKPp+PXr16wdLSErNmzUJYWBh4PB4mTJiA2NhYREREsIRmTU0NrK2tERMTI9XJCwALFy5EYmKiwv3JdN87OzvxySefgM/nK7TbpImhx48f480330SPHj2k7rum5yRZe3ru3buHo0ePory8HHfu3MHAgQOxbNkyVl/WrFkDKysrJi3ywoULmD59OiIjI0FRFKKjo5kCZ0DXM/Hw4UP89ddfmDJlCoyNjVFeXo5t27ZJrFUtLS3g8/lwcHDAV199Jff3y+K5FkINDQ3w8vKSmMAaGhqwfft2GBsbszaBVldXY8aMGRg4cCD++OMPpY24uro6eHp6YtGiRWhpaWE+t3nzZtZEVFRUhKioKISGhoKiKEyaNEmmQpVlkE+YMAGGhobYsGGDRLWNFStWwNXVVa6nQFr7or+TXli+/PJL6OjosIoOKHs9Ojs78fjxY8yYMYPxVAsEAsaTkJWVhcmTJzOGV0tLCzIyMjBs2DCYm5uz9lXJIjs7Gx988AG+++47ZGVlwdzcHCkpKfjXv/7F6ufFixdBURRSU1MV9l9UYA0bNgwZGRk4fPgw6z0LFiyAnp4eE2JftWoVevbsqZIIUmTU0ka/6CJ/4sQJJCYmyvXuy6ru9N133zFeYJpvvvkGAQEB4PP5SuXg0pSVlYHH44HH40ksQAUFBUhLS4ObmxtrUVSGiooK+Pn5ISIiAjdv3kRdXR2Sk5NhamrKymcXFTL37t3Db7/9prBtdcStPANLHrdu3UJ8fDxKS0sZgVZRUYHAwEBmT09LSwveeustzJs3j0njlFdlhxa39DxGP8f5+flwcHBAYGAgazEVRVkHQHNzM0aOHAk/Pz8AXakgtIezV69eUp02N27cgLOzs1JOEYJ0ysvLZXphZSEqSOiUGQBYuXIlBg8ezGQM0JSVlSE+Ph4vv/wyPv/8c+Tk5CA4OBgvv/wyKIqCn58fkwJZWloqN+1ZntdYEb///jv279+PMWPGwMjICObm5ky6TmFhIaysrODs7My8X3xtXLx4MfT09KQ6vYCuZ/j333/H6dOnGYPqxIkTsLKyQo8ePaQa6h0dHTh69KjcefXhw4f4+OOPJeZQUWRdF6FQiJqaGvD5fMbBKm0dkjevAqrNRwcOHMC8efNgY2OjtHg6efIkrK2t0d7ejjVr1kiIoJs3b8LMzIwxYEX7c+zYMfTt2xcLFixgbJ3Ozk7mGu/fvx+jR49GUlIS7t69ixs3biAiIgJjx45lGazSrktBQQEcHR0RFxeHL774gnk9OTkZFEVh8eLFrOgHTW5uLgwNDfHGG2/A3d2dVVFTGVHd2tqKkSNHws7ODp2dnbhz5w7s7OwQFxfH7Lehn0/6e+fOnYuUlBSWY668vByRkZEIDQ2VOUeLc+zYMejr62PKlCks4dXU1IRFixbB3Nxc6aI4omJI2cJOos/yuXPnsHjxYhgaGsp9lmQ5ScTHcGNjI1avXo3evXtj3Lhx+PDDD5m/paamIjU1lVmz7t69i7NnzyIyMpJ17dra2pjqjo2NjWhtbUVSUhJ0dHRgYGDAsmfoZ7ClpQVxcXFqOzWfayF07do1uLm5obCwkFWBCuia3NauXQsjIyNW/nRNTY3KZZGLi4uRmZkJU1NTjBo1Cnv37mUWlpkzZ2LRokWM1/T+/fs4fvw44uLiZHoXpBnkopN/dHQ0jIyMEBISgpycHOzYsQNz586FiYmJUhOftPbFN1kLBAJ4eHjg7bffBqBaFTrRHOiTJ0/C0dERQ4cOxcqVK3H37l0mbPzll18yn+no6MDVq1clFnFZnDt3DqampkxKWV1dHd5++20YGRnBz88Pu3btYnLQ8/Pz5ZYSFUWawJo8eTJLYGVlZcHQ0BChoaEwNjZmLRiKUGTUBgUFsQa96OT8559/ymxXVnWnNWvWYODAgUwkSXSxycnJQUhICKuKnzLIM4Zu3ryJ5ORk+Pr6oqmpSaX0ONoYCAsLQ2FhIerr65GcnAx/f39WlFaVZ1Edcasu1dXV+P7778HlcjFo0CAsWbKEEQl5eXno1asXU6Shvb0dly9fxvz580FRlMy5QJa4zc3NRVNTEy5duoQhQ4Zg4sSJrOdQVSHX3t6OL7/8Em5ubuDxeAC6xFBgYCDc3d0Z54T4tVe2hDJBNmVlZYiKisLIkSMlKojJ+ww9Bq9du4b33nsPBgYGMud/2jsdFhYGgUCAP//8ExcvXkRUVBSTZqTsWFW1v+Lt1tfX4/Lly0zqpSqlqzMzM2FsbMyk/dKIl642MTFBcnIybt++jby8PFhbWyM9PZ0VHVFnTZOHrMhQZmYmPDw8FO6rFf28eKVWZSktLUVwcDDi4uJUOs6guroaXC6XSXMWnUvu37+P0aNHS1TWFZ1jDhw4gAEDBmDevHlM5Eb0momKoYaGBty4cQORkZEICQmRmRpcUlICc3NzLFmyRKrDa/LkyXBycsKnn37KKjojrTQ47XSkUSSqOzs7ceHCBbi5ucHHxwdAl4AyMTFBWloaaw/Tw4cPkZmZCSsrK6nXXCAQMCJX1HEhfo1oOjo6sGPHDujq6sLJyQnTp0/H66+/jqioKPTt21elyKCs71AEPcbNzc2hr6+vlKNXFSdJUVEREhISMHToUAQHB6O0tBRHjhxBamqqxFE20vpfWFiIESNGwM3NDQ8ePEBzczPmzJkDPT09iaqy6jo1RXmuhdD+/fthYGDA/F/8gt66dQu9e/dW+RwGmpaWFrS2tqK+vh7t7e24c+cOYmNj4eLigtDQUNTW1uLgwYOIjo6WqAak6OZIM8iTkpKYgbt69WqMHj0aenp6cHJyQnx8vEoFGORFVGixuHnzZlhaWirldacpKSnB0qVLWcpbKBRi9uzZ8Pb2hoODAy5evIj09HQMHTpUIl9UFTIyMjB58mSmqENSUhKz/yUoKAh6enpK57PTyBJYhoaGjMAqKyvDpk2boKurq9KkpK5Rq8yCLWsztZWVFatSD8D2nMkTV/KQN+kVFxcrFZWU1S4thkQjQ6NGjVJ6/5g4qojbvXv3qvUdzc3NCAoKgq+vL4Cukqmenp6ws7PDiRMn0NbWhi1btmDSpEkSnmXxyC6NPHFra2vLeAXz8vIwZMgQTJ48Wa0S5vTz0N7ejtzcXDg5OSEsLAzA35GhoKAgxmuuqfOnCH9TUlKChIQElaJrtLFCpyYpcsiUlZUhLCwMYWFhKkdzNNFfaVy9elXl0tVz5syBtbU1k74nq3T14MGDweVyUVFRgW+//Rb9+vVjGeraQJpA7dWrl0QBC3mfV1UUi1NfX690JJjm9u3boCgK+vr6En0tKiqCg4MDzp8/LxFRF917dPDgQdjZ2WHatGmMyFUkhkaOHMmq2Ecj74ykyspKZn1JS0uDo6MjDh48iObmZpVKg9OiWtacT5/pNGzYMAQGBgLoSqfu3bs3nJ2dsXnzZmbfjLW1tVxbQPS+KjtHX7p0CfHx8UzRoyVLlmj12RWntLQUMTExKh3JoMrze//+fXz99dfw8vLCkCFDsGTJEnh7eyMtLU3mZ+j7JxQKUVJSglGjRmHEiBFobGxES0sLJk6c2K1CSrJ4roXQhQsXYGBgwHhhpeHl5cXaqKos9OZSFxcX6Orqws3NDRs3bkRbWxuOHDmCESNGoE+fPti6dSucnJykbpSUhyyDXE9PD+PHj8fu3btx8+ZN3Lt3D62trSpXtpIXUfHx8cGuXbtw+vRpeHp6Ijs7Wymvgrj3XXzvTn5+PpKSkmBsbIwFCxbA1NRU7sZIRRw9ehSjRo2CUCjEa6+9BhsbG2bQlpaWYsuWLWqdqyJLYL366quMwDp06JBKi82TMGrFqzv16dNHanUnRXuwVP2+7ngwZbUrLoZeffVVODs7s4qZKIsy4lYgEGDdunWwsrJSq7hGe3s7vvjiC7i4uCAqKgpA1zOYkpICHR0dpKWlYdmyZZg2bRqzEZaOpsgaW8qIW3rs5OXlwczMDFOnTlW6iIlo2oKoGDp58iRLDP3www8YM2YMxo0bxyqgQNAs6kTXVDVWZEUt1EET0cDz58+zRI+ypavr6+sBKC5d7eHhAV9fX/z55584cuQI7O3t8dprr2n1OVZVoIqjKZGpCh0dHVizZo3UPaj/+c9/oKuryzJCxWlubkZtbS1ycnLg7OzMyqoRnd/27dvHiKF79+6hsrJSqlNY0RlJtra2iI+PB9CVdWNjY4O9e/eqXBo8MzMTlpaWaGpqknum06BBg5hiEseOHUN4eDhsbW3h5eWF+fPnK5Vxos59fdpOJ3WKC6jzO9PT0xEeHo4BAwaAoiiJgl+iJc1F+7Rw4UJQFAUvLy88ePAAra2tmDRpUrcKKUnjuRZCsjae0QP5wYMH8Pf3l3mYnCzECzB8/vnniI2NRY8ePZCQkMDctHnz5mHUqFHw8PAARVFM+UxlkWWQp6SkMAZ5dnY2APXCn/IiKsHBwdDX1wdFUXJzo8WR5n2fNGkSdu/ezfTxs88+w+jRo0FRFFxdXVX2XokSFBQEHR0d9O/fX2mvmyIUCawPP/xQZYGlbaOWRtZmavrar1y5Era2tmhsbNTIgaSa8GDKapcWQ0VFRaitrcXMmTNVqoAmijLi9ujRo2pt/BePqHC5XISGhjJ/37dvH0JCQpjqT46Ojko7LpQVt01NTfj111+V9hhWVVUhPT2dFakW/x0uLi7MBu/c3Fx4e3sjMjIS7e3tGjvMltB9VDVWtDVmu4ui0tXiBrOypauNjY2ZtXfHjh1wcXFROf1dVdTxpovyNFJOZRndP/30k0KH8pYtW5g5T9p6Lk0MTZkyhUnLFv9uZc5Isre3Z0orp6amoqSkROXS4FeuXMGDBw+UOtPJ3d2difgDYPquSuqVqvdV9Lo9T3Ousr9T9DedO3cOmZmZMDExYQlyWQe+vvfee7C0tMSePXvg7e3NpMm1trYiMjJS5UJK8niuhRDwd737KVOmSExKK1aswKBBg2RW55CGogIMhoaGLK//yZMnMXfuXPTt21ep6i2iyDPIBQKB2hEPZdovKirCvn37VI4eyPK+GxgYwNfXF7t27cK9e/dQV1eHjz76SCWRJQo9gL755hs4OjoymyE1NVloQ2Bpy6gVR95mavFNsJpAWx7MsrIyREREwM/PDyUlJd3yjmlD3MqLqDg7OyMwMJB5HgsKCrBz505YWFiAoiiVrpUicbt8+XL069dPpTTTgoICDB48GLNmzWL9btGT0T/55BO4uroywv3rr78mhRH+ITyNqIMqaLp09csvv8xq+0mgbqneZ43a2lqpDmXRa7xw4UIsWrRI7oH0oq/v3bsXfn5+TFVZdc9IklWJD1BcGpyiKGY/rrJnOg0bNgzBwcHPrUB5FhG/fuLjkz6XKSIiguVEtrCwYPYTFRcXw8vLC56enrh//z4eP36s8t5neTz3Qkh04xmXy8X06dOxfPlyJCcnw9zcXOWNZ8oWYBD1nrS0tKgd9dCGQa7t9uVFmuh9Tdu2bdNI2Pfu3bsYOnQoVqxY0e22AO0LLG0YtdKQtZla0yKIRlsezJKSEkyYMEEjBpsmn3VlIipubm4SZ3/U19erVblGkbgVPYdKWX755RempLKoGKLH5cOHD2FtbY3Nmzer3Dbh2edZL3ShrdLVxHBVnePHj6Nnz54Slcyam5uxdOlS2NvbK+XUFL32kZGRiI2Nlfv+7pyRRKNMaXBAM2c6EbQDPc5jY2NlOpFLSkowaNAg+Pv7a2xvEM1zL4Ro6I1nrq6uCAgIwOzZsxWeySINbRdgEG9XWwa5NttX5H3fvHlztyJZ4hw4cADGxsYqlYFWhKYFlijaMGql0d1c9WeF7hps2njWlY2oDB8+XGo5dHXQhri9du0aI4ZEDZz29nY8evQIPB6PVXqWQHiSPInS1QTFCIVCViWzadOm4Y033kBMTIzCQgHi0Pdg9uzZmDhxosrzu6pnJCkqDS6KNs50ImgGWYc6i4oegUDQrXP/ZPGPEUIAWAcyqasYtVmAQRraNMi12b62I1mi1NbWIjg4WGF5UlXRhsCieVIRm+7mqv+T0PSzrmxERXTDb3fRhrilxdC0adOYBb6trQ1ZWVkYPHjwM5s+RXgxeBKlqwnKcfnyZSQkJMDT0xOBgYHIzMxUOeUf6DpTKiAgQKVKt4B6ZyTJKw0uDU2e6UTQLLLOK9J0BEicf5QQ0kRep7YKMMhDmwa5ptvXdiRLFqJVRTSFtgQWzZOK2PxTctU1gabH0tOIqGhD3F6/fh2BgYFwcXFBXFwcEhISYGtrK/UkcQLhWaS7pasJyqGpSmaqrtnqnpEkrzS4LLRVEZXQfbpzqLO6/KOEkKbQdAEGRWjbINdG+9qOZD0ptCGwRCERmyeLNp71pxFR0Ya4ra6uxubNmzFhwgSsXbtW7UImBMLT4p+SDvws8zQLBahzRpK80uDyeFarKxKe/L2hAIBDYCEUCjl79uzhvPnmmxwHBwdOQEAAp1+/fpzKykpObm4u58yZMxwvLy+NfmdrayvHwMBAo21qu/2DBw9yZs2axTl79izH19dXo23/k2hvb+fo6ek97W68MGjjWb9x4wZn7ty5nPv373O4XC6nR48enEuXLnG++uorjqenp0a/i0AgyEYgEHAWL17MWb9+PcfV1fVpd4fwDCAUCjk9evRQ+XOlpaWclStXcjZt2sSxs7PTQs8I6vIk7w0RQnK4fPkyJzs7myMQCDhmZmYcDw8Pzty5czlOTk5Pu2vPBL/99hsnJSWFc+DAAY6tre3T7g6BoFVu377NycnJ4Vy4cIHj5eXFSUxM5Dg6Oj7tbhEILxzEuUTQFG1tbRx9ff2n3Q2CFJ7UvSFCSAFCoZCjo6PDoSiK09nZydHR0XnaXXqm0HYki0AgEAgEAoFA0AZECCkAAIeiKIl/EwgEAoFAIBAIhOcXIoQIBAKBQCAQCATCCwfJ8yIQCAQCgUAgEAgvHEQIEQgEAoFAIBAIhBcOIoQIBAKBQCAQCATCCwcRQgQCgUAgEAgEAuGFgwghAoFAIBAIBAKB8MJBhBCBQCAQCAQCgUB44SBCiEAgEAgEAoFAILxwECFEIBAIBAKBQCAQXjiIECIQCAQCgUAgEAgvHEQIEQgEAoFAIBAIhBeO/wNW6LBwQ5OHnwAAAABJRU5ErkJggg==",
      "text/plain": [
       "<Figure size 1000x500 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize = (10, 5))\n",
    "plt.suptitle('Gradient-derived similarity matrices')\n",
    "\n",
    "for i, (ph_type, PHONES_GRPS, subset) in enumerate(zip(['Vowels', 'Consonants'], [VOWELS_GRPS, CONSONANTS_GRPS], [VOWELS_IDX, CONSONANTS_IDX])):\n",
    "    PHONES_SUBSET  = np.asarray(subset)\n",
    "    PHONES         = list(np.concatenate(PHONES_GRPS))\n",
    "    dist_subset    = dists[subset, :][:, subset]\n",
    "    \n",
    "    plt.subplot(1, 2, i+1)\n",
    "    plt.imshow(dist_subset, clim = [-1, 1], cmap = 'bwr_r')\n",
    "    plt.xticks(ticks=np.arange(len(PHONES_SUBSET)), labels=PHONES, rotation=45)\n",
    "    plt.yticks(ticks=np.arange(len(PHONES_SUBSET)), labels=PHONES)\n",
    "    plt.title(ph_type)\n",
    "    \n",
    "    if len(PHONES_GRPS) > 1:\n",
    "        groups      = [[PHONES.index(ph) for ph in grp] for grp in PHONES_GRPS]\n",
    "        groupColors = cm.hsv(np.linspace(0,1,len(groups)+1)[0:len(groups)])*0.9\n",
    "\n",
    "        currentAxis = plt.gca()\n",
    "        for gIdx in range(len(groups)):\n",
    "            thisIdx = groups[gIdx]\n",
    "            currentAxis.add_patch(Rectangle((thisIdx[0] - 0.5, thisIdx[0] - 0.5), len(thisIdx), len(thisIdx),\n",
    "                                  alpha=1, facecolor='none', edgecolor=groupColors[gIdx,:], linewidth=4))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8f5c2d5a",
   "metadata": {},
   "source": [
    "Looking alright! In the paper, we actually compute similarity matrices for each session (with SmoothGrad as well, mentioned in `GradientGetter` notes) and then average the resulting matrices together. \n",
    "\n",
    "Nevertheless we can see some articulatory structure in these maps without all the fancy bells and whistles. For instance, the labial fricatives (F, V) are more similar to each other than the bilabials (P, B, and M) and vice-versa. And together these labials cluster more closely than compared to other phonemes."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c933501c-2a9f-4ab6-9ab3-1e8923723788",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}