lesson1-1.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Lesson 1 - Keras implementation of fast ai part 1 v2\n",
    "- This is a Keras implementation of Jeremy Howard's fastai part 1 v2 lesson 1 (original version is in PyTorch)\n",
    "- This has only been tested with the TensorFlow backend to Keras\n",
    "- https://github.com/fastai/fastai/blob/master/courses/dl1/lesson1.ipynb\n",
    "- [Lesson 1 & 2 notes](http://forums.fast.ai/t/deeplearning-lecnotes2/7515)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## [Jupyter Hotkeys](https://www.cheatography.com/weidadeyue/cheat-sheets/jupyter-notebook/)\n",
    "(1) `Shift` + `Enter`: Execute the selected cell.<br/>\n",
    "(2) `Tab`: pop up code completion or indent.<br/>\n",
    "(3) `Shift` + `Tab`: pop up tip of the function. (Simple)<br/>\n",
    "(4) `Shift` + `Tab`(2 times): pop up description of the function. (Detail)<br/>\n",
    "(5) `Shift` + `Tab`(3 times): show description of the function in an extra window.<br/>\n",
    "(6) Equal to (5), Type `?` before the function name. For example, `?np.exp` and then execute the cell.<br/>\n",
    "(7) `??function_name`. For example, `??np.exp` to show the source code of the function.<br/>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[Reload submudule in IPython](https://stackoverflow.com/questions/5364050/reloading-submodules-in-ipython)<br/>\n",
    "[autoreload](http://ipython.readthedocs.io/en/stable/config/extensions/autoreload.html)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.5/dist-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "from fastai.utils import *\n",
    "from fastai.learner import *\n",
    "from fastai.dataset import *"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Download dataset and do data augmentation\n",
    "The dataset is available at http://files.fast.ai/data/dogscats.zip. You can download it directly on your server by running the following line in your terminal. `wget http://files.fast.ai/data/dogscats.zip` You should put the data in a subdirectory of this notebook's directory, called data/. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "PATH = \"../data/dogscats/\" #update with your path to cats and dogs\n",
    "sz = 224"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Loading in Keras' off the shelf/pretrained ResNet50 model.\n",
    "\n",
    "***<span style=\"color: red\">Note: the PyTorch course uses ResNet34. After many hours spent writing code to convert PyTorch models and weights into Keras. I opted to use Keras' ResNet50.</span>***"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 23000 images belonging to 2 classes.\n",
      "Found 2000 images belonging to 2 classes.\n"
     ]
    }
   ],
   "source": [
    "arch = ResNet50\n",
    "data = ImageClassifierData.from_paths(path=PATH, tfms = tfms_from_model(arch, sz))\n",
    "learn = ConvLearner.pretrained(arch, data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "----\n",
    "## Choosing a learning rate\n",
    "The learning rate determines how quickly or how slowly you want to update the weights (or parameters). Learning rate is one of the most difficult parameters to set, because it significantly affect model performance.\n",
    "\n",
    "The method learn.lr_find() helps you find an optimal learning rate. It uses the technique developed in the 2015 paper [Cyclical Learning Rates for Training Neural Networks](https://arxiv.org/abs/1506.01186), where we simply <font color=\"blue\">keep increasing the learning rate from a very small value, until the loss starts decreasing. The purpose of learning rate finder is not to find the minimum, but figure out what learning rate is allowing us to decrease rapidly.</font> Instead we go back to the point where the lower learning rates quickly giving us a quick decrease in the loss. We can plot the learning rate across batches to see what this looks like.\n",
    "\n",
    "\n",
    "<!-- ![learning_rate](./ipython_images/lr.jpeg) -->\n",
    "<!-- img src=\"./ipython_images/lr.jpeg\" alt=\"Drawing\" style=\"width: 60%;\"/ >\n",
    "(Source: Andrew Ng’s Machine Learning course on Coursera)-->\n",
    "\n",
    "   Learning Rate           |  Learning Rate Finder        \n",
    ":-------------------------:|:-------------------------:\n",
    "<img src=\"images/lr_1.png\" alt=\"\" style=\"width: 80%;\"/>|<img src=\"images/lr_1-2.png\" alt=\"\" style=\"width: 80%;\"/>\n",
    "\n",
    "\n",
    "It's a orthogonal way to optimizers such as **Adam**, **Momentum**, and so on.<br/>\n",
    "In fact, the fast.ai library is trying to figure out the right gradient descent version. Behind the scenes, it's something actually called **Adam**. And this technique is telling us this is the best learning rate to use, given the tweaks we're using, in this case, the Adam optimizer.<br/><br/>\n",
    "**For more information**:\n",
    "* [Paper - No More Pesky Learning Rates](https://arxiv.org/pdf/1206.1106.pdf) (Less effective)\n",
    "* [Paper - Leslie N. Smith, \"Cyclical Learning Rates for Training Neural Networks\"](https://arxiv.org/abs/1506.01186)\n",
    "* [Dynamic Learning Rate](http://wiki.fast.ai/index.php/Dynamic_learning_rates)\n",
    "* [Improving the way we work with learning rate.](https://techburst.io/improving-the-way-we-work-with-learning-rate-5e99554f163b)\n",
    "* [Train your deep model faster and sharper — two novel techniques](https://hackernoon.com/training-your-deep-model-faster-and-sharper-e85076c3b047)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[Magic Function](http://ipython.org/ipython-doc/dev/interactive/tutorial.html#magic-functions)<br/>\n",
    "%%time - measure running time of a cell"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "??LR_Find()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "??LR_Updater()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'loss': 1000000000.0, 'lr': 1e-05, 'iteration': 0}\n",
      "Epoch 1/1\n",
      "204/360 [================>.............] - ETA: 1:21 - loss: 0.4194\n",
      "best loss: 0.032150\n",
      "best lr: 0.007356\n",
      "best iteration: 172\n",
      "205/360 [================>.............] - ETA: 1:20 - loss: 0.4176"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "learn.lr_find(jump=10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAgEAAAGDCAYAAACydsMvAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMS4yLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvNQv5yAAAIABJREFUeJzt3XmYXFWZ+PHvmxWIIWCICAmQIIuACkjYFBBFkEUJOigwiMyIIDr8HHdBHhnGGUdRGZdxGUBQBGWREQgaQIIQhMiShCSQQCQEMAGEEPZAAkne3x/3thRNL5Wkq2911/fzPPepqlO3br2nq7vvW+ece05kJpIkqfUMqDoASZJUDZMASZJalEmAJEktyiRAkqQWZRIgSVKLMgmQJKlFmQRILSQiro6IY6uOozetTZ0j4vSIuLCnY5KahUmA1Asi4sGIeG/VcWTmQZl5fk8fNyL2jYhVEfF8RDwXEfMi4p9X4/VrdbItX/9y+f5t25ehcXWW+oNBVQcgqWdExKDMXFFhCI9k5piICOAgYGJETM3Meb30/pdk5kd76b26VP4MIjNXVR2L1BVbAqSKRcT7I2JmRDwdEVMj4m01z50cEfeX367nRsQHa577p4i4JSK+FxFLgNPLspsj4rsR8VREPBARB9W85saI+ETN67vad1xE3FS+9+SI+HE939azMAl4Eqityw8iYmFEPBsR0yNi77L8QOCrwBHlN/hZZfmIiDg3Ih6NiIcj4j8jYuAa/HxXt85TyjpfB2zU7lh7lJ/R0xExKyL2bfc+34iIW4AXgC1XN1apt5kESBWKiJ2B84BPAiOBsyi+QQ8td7kf2BsYAfw7cGFEbFJziN2BBcDGwDdqyuZRnMC+DZxbfjPtSFf7/hq4vYzrdOCYOus0ICIOLY85v+apO4CdgNeXx/5NRKyTmdcA/0XxTf51mbljuf8vgBXAVsDOwAHAJ+qJoRvd1Xl6+dx/AH8fSxARo4HfA/9Z1uGLwP9FxKiaYx8DnAAMBx7qgVilhjIJkKp1AnBWZt6WmSvLvuvlwB4AmfmbzHwkM1dl5iXAfcBuNa9/JDP/JzNXZOaLZdlDmXlOZq4Ezgc2oUgSOtLhvhGxObArcFpmvpSZNwMTu6nLphHxNPAicDnw+cy8s+3JzLwwM5eUsZ4JDAW27ehAEbExcDDw2cxcmpmPA98Djuzi/T9SfkNv2zZdwzp/LTOXZ+ZNwFU1r/soMCkzJ5Wfx3XAtDLONr/IzDllHV/uIlapKZgESNXaAvhC7ckL2AzYFCAiPlbTVfA08BZe3US9sINj/q3tTma+UN59XSfv39m+mwJP1pR19l61HsnMDYD1gR8C76l9MiK+GBH3RMQzZV1GtKtLrS2AwcCjNXU/C3hDF+9/aWZuULM90sl+XdX5qcxcWrNv7bf5LYAPt/us9qJIItp09zOSmooDA6VqLQS+kZnfaP9ERGwBnAPsB/w5M1dGxEygtmm/UcuAPgq8PiLWqzlRblbPCzNzeUR8BZgXEYdl5hVl//+XKeoyJzNXRcRTvFKX9vVYSNEislEvDnZ8FNgwIobVJAKb18S2ELggM4/v4hguy6o+xZYAqfcMjoh1arZBFCf5EyNi9ygMi4hDImI4MIzipLIYIIpL7t7SG4Fm5kMUTd2nR8SQiNgT+MBqvP4l4EzgtLJoOEX//mJgUEScRtFi0OYxYGxEDChf/yjwB+DMiFi/HGfwpoh419rWrYuY2+r872Wd9+LVdb4Q+EBEvC8iBpaf4b4RMaZRMUmNZhIg9Z5JFP3lbdvpmTkNOB74EfAUxUC6fwLIzLkUJ9I/U5wk3wrc0ovxHg3sCSyhGAx3CcW383qdB2weER8ArgWuAf5C0cS+jFc3nf+mvF0SETPK+x8DhgBzKX42l/HqpvdG+EeKgYNPAv8G/LLticxcCEyguJJhMUX8X8L/o+rDItPWK0ndi4hLgHsz89+qjkVSzzCDldShiNi1bIIfUF7LPwG4ouq4JPUcBwZK6swbgd9SzBOwCPhU7SV/kvo+uwMkSWpRdgdIktSiTAIkSWpRLTEmYKONNsqxY8dWHYYkSb1i+vTpT2TmqO72a4kkYOzYsUybNq3qMCRJ6hURUdcCVnYHSJLUokwCJElqUSYBkiS1KJMASZJalEmAJEktyiRAkqQWZRIgSVKLMgmQJKlFmQRIktSiTAIkSWpRJgGSJLUokwBJkip0zTVw//3VvLdJgCRJFVm1Cg4/HP7nf6p5f5MASZIqsmABLF0Kb3tbNe9vEiBJUkVmzSpud9yxmvc3CZAkqSKzZ8OAAbD99tW8v0mAJEkVmTULtt0W1l23mvc3CZAkqSKzZ1c3HgBMAiRJqsSzz8IDD1Q3HgBMAiRJqsRddxW3tgRIktRiqr4yAEwCJEmqxOzZsOGGMHp0dTE0NAmIiAMjYl5EzI+Ikzt4fmhEXFI+f1tEjC3L94+I6RFxV3n7nprX3Fgec2a5vaGRdZAkqRFmzSpaASKqi6FhSUBEDAR+DBwEbA8cFRHtr4Q8DngqM7cCvgecUZY/AXwgM98KHAtc0O51R2fmTuX2eKPqIElSI6xaVYwJqHI8ADS2JWA3YH5mLsjMl4CLgQnt9pkAnF/evwzYLyIiM+/MzEfK8jnAuhExtIGxSpLUa9qmC65yPAA0NgkYDSysebyoLOtwn8xcATwDjGy3zz8AMzJzeU3Zz8uugK9FdNyQEhEnRMS0iJi2ePHitamHJEk9avbs4rY/twSstYjYgaKL4JM1xUeX3QR7l9sxHb02M8/OzPGZOX7UqFGND1aSpDrNmlVMF7zDDtXG0cgk4GFgs5rHY8qyDveJiEHACGBJ+XgMcDnwscz8+0rLmflwefsc8GuKbgdJkvqM2bNhm22qmy64TSOTgDuArSNiXEQMAY4EJrbbZyLFwD+Aw4E/ZmZGxAbA74GTM/OWtp0jYlBEbFTeHwy8H7i7gXWQJKnHzZpVfVcANDAJKPv4TwKuBe4BLs3MORHx9Yg4tNztXGBkRMwHPg+0XUZ4ErAVcFq7SwGHAtdGxGxgJkVLwjmNqoMkST2tGaYLbjOokQfPzEnApHZlp9XcXwZ8uIPX/Sfwn50cdpeejFGSpN7UDNMFt2nqgYGSJPU3bVcGNENLgEmAJEm9aNYs2GADGDOm6khMAiRJ6lWzZ1c/XXAbkwBJknrJypVFEtAM4wHAJECSpF5z333FdMG7NMkQd5MASZJ6yYwZxe3b315tHG1MAiRJ6iUzZsA668B221UdScEkQJKkXjJjRjEeYFBDZ+mpn0mAJEm9ILNIApqlKwBMAiRJ6hULFsAzz5gESJLUctoGBTbLlQFgEiBJUq+YMQMGD4Yddqg6kleYBEiS1AtmzIC3vAWGDq06kleYBEiS1GDNOCgQTAIkSWq4hQvhiSdMAiRJajnNOCgQTAIkSWq4GTNg4MDmWTiojUmAJEkNNmNGMVXwuutWHcmrmQRIktRgzTgoEEwCJElqqEcfLTaTAEmSWkyzLR9cyyRAkqQGmjEDImCnnaqO5LVMAiRJaqAZM2CbbWD48KojeS2TAEmSGmj6dNh556qj6JhJgCRJDfK3vxWzBe62W9WRdMwkQJKkBrnjjuLWJECSpBZz223FTIF2B0iS1GJuvx3e+lZYb72qI+mYSYAkSQ2walXRHbD77lVH0jmTAEmSGmD+fHj66eYdDwAmAZIkNcTttxe3JgGSJLWY226DYcOK1QOblUmAJEkNcPvtMH58cXVAszIJkCSphy1fDjNnNvegQDAJkCSpx82eDS+91NzjAcAkQJKkHtcXBgWCSYAkST3uttvgjW+EMWOqjqRrJgGSJPWw228vWgEiqo6kayYBkiT1oKefhnnzmr8rAEwCJEnqUdOmFbfNfmUAmARIktSj2gYFjh9fbRz1MAmQJKkH3XYbbLstbLBB1ZF0zyRAkqQeklkkAX1hPACYBEiS1GMeeAAeewze8Y6qI6mPSYAkST3klluKW5MASZJazNSpsP76sMMOVUdSH5MASZJ6yC23wB57NPfKgbUamgRExIERMS8i5kfEyR08PzQiLimfvy0ixpbl+0fE9Ii4q7x9T81rdinL50fEDyOafT4mSVIreOYZuPtueOc7q46kfg1LAiJiIPBj4CBge+CoiNi+3W7HAU9l5lbA94AzyvIngA9k5luBY4ELal7zU+B4YOtyO7BRdZAkqV633lpcHdBXxgNAY1sCdgPmZ+aCzHwJuBiY0G6fCcD55f3LgP0iIjLzzsx8pCyfA6xbthpsAqyfmbdmZgK/BA5rYB0kSarL1KkwYEDfmCmwTSOTgNHAwprHi8qyDvfJzBXAM8DIdvv8AzAjM5eX+y/q5piSJPW6W26Bt70Nhg+vOpL6NfXAwIjYgaKL4JNr8NoTImJaRExbvHhxzwcnSVJpxYpikqC+NB4AGpsEPAxsVvN4TFnW4T4RMQgYASwpH48BLgc+lpn31+xfuzpzR8cEIDPPzszxmTl+1KhRa1kVSZI6d9dd8PzzfWs8ADQ2CbgD2DoixkXEEOBIYGK7fSZSDPwDOBz4Y2ZmRGwA/B44OTNvads5Mx8Fno2IPcqrAj4GXNnAOkiS1K22SYJsCSiVffwnAdcC9wCXZuaciPh6RBxa7nYuMDIi5gOfB9ouIzwJ2Ao4LSJmltsbyuc+DfwMmA/cD1zdqDpIklSPqVNh001h882rjmT1RDHIvn8bP358Tmtb4FmSpB62xRbFVQGXXlp1JIWImJ6Z3S5m3NQDAyVJanaLFsFf/9r3xgOASYAkSWtl6tTitq+NBwCTAEmS1srUqbDuurDTTlVHsvpMAiRJWgu33AK77QaDB1cdyeozCZAkaQ099xzceSfstVfVkawZkwBJktbQLbfAypXwrndVHcmaMQmQJGkNTZkCgwb1zSsDwCRAkqQ1NmUK7LorDBtWdSRrxiRAkqQ1sHQp3HFH3+0KAJMASZLWyNSpxeqBJgGSJLWYKVNg4MC+OUlQG5MASZLWwJQpsMsuMHx41ZGsOZMASZJW04svwu239+2uADAJkCRptd16K7z0kkmAJEktZ8oUGDCg784U2MYkQJKk1TRlCuy8M4wYUXUka8ckQJKk1bBsGfz5z32/KwBMAiRJWi233w7Ll5sESJLUcqZMgQjYe++qI1l7JgGSJK2GKVNgxx1hww2rjmTtmQRIklSnZcuK5YP33bfqSHqGSYAkSXW65ZYiEdh//6oj6RkmAZIk1WnyZBg0CPbZp+pIeoZJgCRJdZo8GfbcE173uqoj6RkmAZIk1eHJJ2H6dHjve6uOpOeYBEiSVIcbboDM/jMeAEwCJEmqy3XXFcsG77pr1ZH0HJMASZLqMHkyvPvdxcDA/sIkQJKkbjzwANx/f/8aDwAmAZIkdev664tbkwBJklrMddfB6NHw5jdXHUnPMgmQJKkLq1YVLQHvfW+xcFB/YhIgSVIXZs2CJUv6X1cAmARIktSlyZOL2/32qzaORjAJkCSpC9ddBzvsAJtsUnUkPc8kQJKkTixdCjfdBAccUHUkjWESIElSJ264AZYvh4MPrjqSxjAJkCSpE1dfDcOGwd57Vx1JY5gESJLUgUyYNKkYEDh0aNXRNIZJgCRJHbj3Xnjwwf7bFQB1JgERsVdE/HN5f1REjGtsWJIkVevqq4vbgw6qNo5G6jYJiIh/A74CnFIWDQYubGRQkiRVbdKk4tLAzTevOpLGqacl4IPAocBSgMx8BBjeyKAkSarS888Xlwb2564AqC8JeCkzE0iAiBjW2JAkSarW9dfDyy+bBABcGhFnARtExPHAZOBnjQ1LkqTqTJoEw4fDO99ZdSSNNai7HTLzuxGxP/AssC1wWmZe1/DIJEmqQGYxKHD//WHw4Kqjaax6BgaekZnXZeaXMvOLmXldRJzRG8FJktTb5syBhQv7f1cA1NcdsH8HZf34gglJUiubNKm47c+XBrbpNAmIiE9FxF3AthExu2Z7AJhdz8Ej4sCImBcR8yPi5A6eHxoRl5TP3xYRY8vykRFxQ0Q8HxE/aveaG8tjziy3N6xOhSVJ6sqkSbDjjrDpplVH0nhdjQn4NXA18E2g9gT+XGY+2d2BI2Ig8GOKloRFwB0RMTEz59bsdhzwVGZuFRFHAmcARwDLgK8Bbym39o7OzGndxSBJ0upYsgT+9Cf46lerjqR3dNoSkJnPZOaDmXlUZj4EvEhxmeDrIqKeqRN2A+Zn5oLMfAm4GJjQbp8JwPnl/cuA/SIiMnNpZt5MkQxIktQrfv97WLUKJrQ/W/VT9QwM/EBE3Ac8AEwBHqRoIejOaGBhzeNFZVmH+2TmCuAZYGQdx/552RXwtYiIOvaXJKlbV1wBo0fDLrtUHUnvqGdg4H8CewB/ycxxwH7ArQ2NqmtHZ+Zbgb3L7ZiOdoqIEyJiWkRMW7x4ca8GKEnqe158Ea69tmgFaJWvl/UkAS9n5hJgQEQMyMwbgPF1vO5hYLOax2PKsg73iYhBwAhgSVcHzcyHy9vnKMYt7NbJfmdn5vjMHD9q1Kg6wpUktbLJk+GFF+Cww6qOpPfUkwQ8HRGvA24CfhURP6BcR6AbdwBbR8S4iBgCHAlMbLfPRODY8v7hwB/LKYo7FBGDImKj8v5g4P3A3XXEIklSl664AtZfH971rqoj6T3dzhhIMXjvReBzwNEU39a/3t2LMnNFRJwEXAsMBM7LzDkR8XVgWmZOBM4FLoiI+cCTFIkCABHxILA+MCQiDgMOAB4Cri0TgIEUUxifU2ddJUnq0MqVcNVVcMghMGRI1dH0nnqmDW771r8KOD8iBgBHAb+q47WTgEntyk6rub8M+HAnrx3byWFbZLiGJKm3/PnPsHhx61wV0KaryYLWj4hTIuJHEXFAFE4CFgAf6b0QJUlqrCuvLNYJaIVZAmt11RJwAfAU8GfgE8BXgQAOy8yZvRCbJEkNlwmXXw7veU8xJqCVdJUEbFleikdE/Ax4FNi8bMKXJKlfmDsX7r8fvvjFqiPpfV1dHfBy253MXAksMgGQJPU3V15Z3B56aLVxVKGrloAdI+LZ8n4A65aPA8jMbLFGE0lSf3T55bDrrq2xYFB7nSYBmTmwNwORJKm3LVgA06bBt79ddSTVqGeyIEmS+qXLLituP9zhxer9n0mAJKllXXop7LYbjB1bdSTVMAmQJLWk+++H6dNbtxUATAIkSS3qN78pbls5Ceh22uCIeA5ov6jPM8A04AuZuaARgUmS1EiXXgq77w5bbFF1JNWpZwGh7wOLKJbtDYpFft4EzADOA/ZtVHCSJDXCfffBnXfCmWdWHUm16ukOODQzz8rM5zLz2cw8G3hfZl4CbNjg+CRJ6nFtXQGHH15tHFWrJwl4ISI+EhEDyu0jQNvMge27CSRJanq/+Q3suSdsvnnVkVSrniTgaOAY4HHgsfL+RyNiXeCkBsYmSVKP+8tfYOZM+Ijr4XY/JqAc+PeBTp6+uWfDkSSpsewKeEU9VweMAo4Hxtbun5kfb1xYkiT1vEy46CJ4xztgzJiqo6lePVcHXAn8CZgMrGxsOJIkNc6sWTBnDvzkJ1VH0hzqSQLWy8yvNDwSSZIa7IILYPBgxwO0qWdg4O8i4uCGRyJJUgOtWAG//jUcfDCMHFl1NM2hniTgXykSgRcj4tmIeC4inm10YJIk9aTrr4e//Q2OOabqSJpHPVcHDO+NQCRJaqQLL4QRI+CQQ6qOpHl0mgRExJsz896IeHtHz2fmjMaFJUlSz3n+efjtb+Hoo2GddaqOpnl01RLweeAEoKOZlRN4T0MikiSph11+Obzwgl0B7XWaBGTmCeXtu3svHEmSet4FF8DYsfDOd1YdSXOp5xJBIuIdvHayoF82KCZJknrMI48UgwJPOQUG1DMcvoXUM2PgBRRLB8/klcmCEjAJkCQ1vYsuglWr7AroSD0tAeOB7TPTFQMlSX1KJpx/Puy6K2y7bdXRNJ96GkbuBt7Y6EAkSeppd9wBd90Fxx1XdSTNqZ6WgI2AuRFxO7C8rTAzD21YVJIk9YCf/QzWWw+OOqrqSJpTPUnA6Y0OQpKknvb888V4gI98BNZfv+pomlOXSUBEDARO9zJBSVJfc8klRSJw/PFVR9K8uhwTkJkrgVURMaKX4pEkqUf87Gew3Xaw555VR9K86ukOeB64KyKuA5a2FWbmZxoWlSRJa+Huu+HWW+HMMyGi6miaVz1JwG/LTZKkPuHcc2HwYOcG6E49qwie3xuBSJLUE5Yvh1/+Ej74QRg1qupomls9MwZuDXwT2B74+9pLmbllA+OSJGmNXHEFPPkkfOITVUfS/OqZLOjnwE+BFcC7KaYLvrCRQUmStKbOOgu22AL226/qSJpfPUnAupl5PRCZ+VBmng4c0tiwJElafXPnwg03wIknulhQPeoZGLg8IgYA90XEScDDwOsaG5YkSavvxz+GoUOdJrhe9eRJ/wqsB3wG2AX4KHBsI4OSJGl1PfNMsVjQkUc6ILBe9VwdcAdARKzKzH9ufEiSJK2+X/4Sli6Fk06qOpK+o9uWgIjYMyLmAveWj3eMiJ80PDJJkuq0ahX86Eew++4wfnzV0fQd9XQHfB94H7AEIDNnAfs0MihJklbH9dfDX/5iK8DqqmvsZGYubFe0sgGxSJK0Rn70I3jDG+DDH646kr6lniRgYUS8A8iIGBwRXwTuaXBckiTV5YEH4Kqr4IQTiisDVL96koATgX8BRlNcHrgT8OlGBiVJUr1++tNiToBPfrLqSPqeeq4OeAI4urYsIj5LMVZAkqTKPPccnH02fOhDMGZM1dH0PWs6n9Ln69kpIg6MiHkRMT8iTu7g+aERcUn5/G0RMbYsHxkRN0TE8xHxo3av2SUi7ipf88MIF4mUpFZ1zjnF/ABf+lLVkfRNa5oEdHvijYiBwI+BgygWHzoqIrZvt9txwFOZuRXwPeCMsnwZ8DXgix0c+qfA8cDW5XbgmlRAktS3vfwyfP/7sM8+sOuuVUfTN61pEpB17LMbMD8zF2TmS8DFwIR2+0wA2pYqvgzYLyIiM5dm5s0UycDfRcQmwPqZeWtmJsViRoetYR0kSX3YpZfCwoW2AqyNTscERMRzdHyyD2DdOo49Gqi9tHARsHtn+2Tmioh4BhgJPNHFMRe1O+bojnaMiBOAEwA233zzOsKVJPUVmfCd78B228HBB1cdTd/VaRKQmcN7M5CelplnA2cDjB8/vp6WC0lSHzF5MsyaBeee62qBa6ORP7qHgc1qHo8pyzrcJyIGASMoZybs4pi14z87OqYkqZ/77nfhjW+Eo4/ufl91rpFJwB3A1hExLiKGAEcCE9vtM5FXViQ8HPhj2dffocx8FHg2IvYorwr4GHBlz4cuSWpWs2bBH/4An/mMkwOtrW7nCVhTZR//ScC1wEDgvMycExFfB6Zl5kTgXOCCiJgPPEmRKAAQEQ8C6wNDIuIw4IDMnEsxUdEvKMYlXF1ukqQW8Z3vwLBhcOKJVUfS9zUsCQDIzEnApHZlp9XcXwZ0ONNzZo7tpHwa8Jaei1KS1Ffcdx9cdBF87nOw4YZVR9P3OZxCktRnfOMbMGSIlwX2FJMASVKfcP/9cOGFRTfAxhtXHU3/YBIgSeoTvvlNGDTIVoCeZBIgSWp6Dz4I558Pxx8Pm25adTT9h0mAJKnpfetbxaRAX/lK1ZH0LyYBkqSmtnAhnHceHHecywX3NJMASVJT++Y3i9uTX7MgvdaWSYAkqWktWADnnFO0ArgWXM8zCZAkNa2vfQ0GDy5u1fNMAiRJTWnmTPj1r+Gzn/WKgEYxCZAkNaVTTy2mBv7yl6uOpP9q6NoBkiStiZtugkmT4IwzYIMNqo6m/7IlQJLUVDLhlFOKLoCTTqo6mv7NlgBJUlP53e9g6lQ46yxYb72qo+nfbAmQJDWNl18uxgBssw18/ONVR9P/2RIgSWoaP/kJ3HsvXHVVsViQGsuWAElSU3jiCTj9dNh/fzjkkKqjaQ0mAZKkpvBv/wbPPQff+x5EVB1NazAJkCRV7u674X//F048EXbYoepoWodJgCSpUpnwuc/BiBHw7/9edTStxWEXkqRKXXUVTJ4MP/gBjBxZdTStxZYASVJlXngB/vVfYbvt4FOfqjqa1mNLgCSpMv/xH/Dgg3DjjcVqgepdtgRIkioxZw5897vwT/8E73pX1dG0JpMASVKvW7WquBJg/fXhO9+pOprWZXeAJKnX/fzncPPNcO65sNFGVUfTumwJkCT1qsWLi/UB9t676ApQdUwCJEm96nOfg2efhZ/+FAZ4FqqUP35JUq+58kr41a/gq191ZsBmYBIgSeoVS5bAJz8JO+4Ip55adTQCBwZKknrJ//t/RSJw7bUwZEjV0QhsCZAk9YLLL4eLLoKvfa1oCVBzMAmQJDXUE08UcwLsvDOcckrV0aiW3QGSpIbJLMYBPPVUsUiQUwM3F5MASVLDnHMO/Pa38O1vw1vfWnU0as/uAElSQ8ydC5/9LOy/P3zhC1VHo46YBEiSetyyZXDUUTBsGJx/vpMCNSu7AyRJPe4rX4HZs+F3v4NNNqk6GnXG3EyS1KN+9zv44Q/hM5+BQw6pOhp1xSRAktRjFiyAY46BnXaCM86oOhp1xyRAktQjXnwR/uEfivv/93+wzjrVxqPuOSZAkrTWMuHTn4aZM4vugC23rDoi1cOWAEnSWjvnHPjFL+C00xwH0JeYBEiS1srttxeLAx14YJEEqO8wCZAkrbGFC2HCBNh0U7jwQhg4sOqItDocEyBJWiPPPw+HHgpLlxbrAowcWXVEWl0mAZKk1bZqFXz0o8WEQL//PeywQ9URaU2YBEiSVtspp8CVVxaTAh14YNXRaE01dExARBwYEfMiYn5EnNzB80Mj4pLy+dsiYmzNc6eU5fMi4n015Q9GxF0RMTMipjUyfknSa51zTrEq4Kc/DSedVHU0WhsNawmIiIHAj4H9gUXAHRExMTPn1ux2HPBUZm4VEUcCZwBHRMT2wJHADsCmwOSI2CYzV5ave3dmPtGo2CVJHbviCjjxxOLb/w9+ABFVR6S10ciWgN2A+Zm5IDNfAi4GJrTbZwJwfnn/MmC/iIiy/OLMXJ6ZDwDzy+NJkipy001w5JF98qD5AAARhklEQVSw665w2WUwyA7lPq+RScBoYGHN40VlWYf7ZOYK4BlgZDevTeAPETE9Ik7o7M0j4oSImBYR0xYvXrxWFZGkVjd7dnElwLhxxUDAYcOqjkg9oS/OE7BXZr4dOAj4l4jYp6OdMvPszByfmeNHjRrVuxFKUj/ywAPwvvfB8OFw7bVeCtifNDIJeBjYrObxmLKsw30iYhAwAljS1Wszs+32ceBy7CaQpIZ56CF4z3tg+XK45hrYfPOqI1JPamQScAewdUSMi4ghFAP9JrbbZyJwbHn/cOCPmZll+ZHl1QPjgK2B2yNiWEQMB4iIYcABwN0NrIMktayFC4sE4Kmn4LrrnAugP2rYsI7MXBERJwHXAgOB8zJzTkR8HZiWmROBc4ELImI+8CRFokC536XAXGAF8C+ZuTIiNgYuL8YOMgj4dWZe06g6SFKrevjhIgF44okiAdhll6ojUiNE8cW7fxs/fnxOm+aUApJUj0cfhX33hUcegT/8Afbcs+qItLoiYnpmju9uv744MFCS1CAPPQT77FO0BFx9tQlAf+dVnpIkAObNg/e+t1gYaPJk2GOPqiNSo5kESJKYORMOOKCYAfDGG2HHHauOSL3B7gBJanE331yMAVhnHfjTn0wAWolJgCS1sEsvLboANt64SAC22abqiNSbTAIkqQVlwne+A0ccAePHw9SpsMUWVUel3mYSIEktZsWKYgngL38ZPvzhYhCgUwG3JpMASWohTz0FhxwCP/kJfOlLcPHFxVgAtSavDpCkFjFnDkyYAH/9K5x9Nhx/fNURqWomAZLUAq64Ao45plgC+MYb4R3vqDoiNQO7AySpH1uxAk49FT74QdhuO5g+3QRAr7AlQJL6qYcfhqOOKi79+8Qn4H/+x/5/vZpJgCT1Q9dcUzT/v/giXHghHH101RGpGdkdIEn9yLJl8IUvwEEHwSabFM3/JgDqjEmAJPUTd95ZTPzz3/8Nn/oU3HYbbLtt1VGpmZkESFIft3IlfPObsPvu8OSTxRLAP/kJrLtu1ZGp2TkmQJL6sFmzikF/06YVs//99KfO/qf62RIgSX3QsmXFpX/jx8NDDxUz/11yiQmAVo8tAZLUx1x/PXz60/CXv8Cxx8KZZ3ry15qxJUCS+oi//hU+8pFi6d+XX4Zrr4Vf/MIEQGvOJECSmtyyZfBf/1XM+HfVVfD1r8PcuXDAAVVHpr7O7gBJalKrVsFFF8FXv1q0Anzwg8Xlf2PHVh2Z+gtbAiSpCd1wA+y6K3z0o0Vz/+TJ8NvfmgCoZ5kESFITuflm2H9/eM974Ikn4IILisv/9tuv6sjUH5kESFITaDv57703zJ5djPifN69oCRjgf2o1iL9aklShW2557cn/gQfg8593xT81ngMDJamXrVwJV15ZnPCnToU3vKG4f+KJsN56VUenVmISIEm95Pnn4bzz4PvfL77tb7kl/OAHxbS/nvxVBZMASWqwBQvgrLOK7Zln4J3vhO9+FyZMgIEDq45OrcwkQJIa4KWXiib/c86B664rBvcdfnjR17/77lVHJxVMAiSpB913H/zsZ/Dzn8PixbD55sUMfx//OIweXXV00quZBEjSWnr8cbj0UvjVr+DWW4sm/kMPhRNOKEb+2+SvZmUSIElr4Pnn4YorihP/ddcVI/7f+lb41rfgYx+DTTapOkKpeyYBklSnJ5+E3/0OLr+8WMHvxReL5v4vfQn+8R+LJEDqS0wCJKkLixYV3/gvvxymTCm+8Y8eXfTxH3FEMdLfGf3UV5kESFKNl14qZvG7+mq45hq4666i/M1vhi9/GQ47DMaP98Sv/sEkQFJLy4R77y1W7bv2Wrj+eli6FAYPhr32gjPOKAb5vfnNVUcq9TyTAEktJbNYmOeGG+DGG4vt8ceL58aOLQb1HXggvPvdMHx4hYFKvcAkQFK/9vzzxVK8t95abH/+8ysn/dGji0v49t232N70JoioMlqpd5kESOo3Vq4sJuu57bZXTvqzZ8OqVcXz22wD73sf7LOPJ30JTAIk9VEvvAB33w0zZ8Kddxa3s2cX5VA05e++O5x6KuyxR3F/5MhqY5aajUmApKb20kvFt/t77im2uXNh1qyiX7/tG/6IEbDTTnD88bDjjrDbbsVAPmfqk7pmEiCpcpnw2GNw//3FNm9ecbK/5x6YP79o5m8zdiy87W3w4Q8XJ/6ddirKbNaXVp9JgKRe8dJL8NBDxbK6bSf7++8vHi9YUFyW12bQINhqK9h++2Llve22K+5vuy2st151dZD6G5MASWttxQp45BFYuLDYFi167f3HHiu+8bdZZx3YcsticN5++xW3bY/HjYMhQ6qrj9QqTAIkdWjVKliypDh5P/YY/O1vnd9//PFX+ufbDB8Om20GY8YUzfdjxhTN9m96U7G98Y3OuidVzSRAagEvvlic0J98suutbZ/Fi4sTe21ffJshQ4oT+MYbFyf2XXaBTTctTvhtJ/3NNisG60lqbiYBUpNatao4eb/wQjHhzbPPvrI999yrH7ff2p5/+unipL5sWefvM2RIcenc619fbOPGwa67vnKib9vaHo8Y4SA8qb9oaBIQEQcCPwAGAj/LzG+1e34o8EtgF2AJcERmPlg+dwpwHLAS+ExmXlvPMaVGWLECli/velu2rOOyF1549bZ0adeP28q6OnHXiiia3tdf/9XbmDHFCbv2BN9+GzkS1l3Xk7rUqhqWBETEQODHwP7AIuCOiJiYmXNrdjsOeCozt4qII4EzgCMiYnvgSGAHYFNgckRsU76mu2NqDWS+dlu1qthWriy2tvvtb3vruZdfLk7GtbeNKOvoxN6+v3tNDBoEw4YVo9vbtrbHG2742rL2+40Y8doT/fDhxXP2rUtaE41sCdgNmJ+ZCwAi4mJgAlB7wp4AnF7evwz4UUREWX5xZi4HHoiI+eXxqOOYDXXiicXMZO1Plh2dROvdqn597YjtvmrQoGLVt/a33ZUNHdpxWUfbOut0/lxX+7ad1AcPrvqnJEmv1sgkYDSwsObxImD3zvbJzBUR8Qwwsiy/td1rR5f3uzsmABFxAnACwOabb75mNehA2zeyiGIbMOCV+2uyre3rGxnDgAHFjGttt7X3q3iusxP5wIE2Z0vSmui3AwMz82zgbIDx48f32HfdM8/sqSNJklStRvYkPgxsVvN4TFnW4T4RMQgYQTFAsLPX1nNMSZJUh0YmAXcAW0fEuIgYQjHQb2K7fSYCx5b3Dwf+mJlZlh8ZEUMjYhywNXB7nceUJEl1aFh3QNnHfxJwLcXlfOdl5pyI+DowLTMnAucCF5QD/56kOKlT7ncpxYC/FcC/ZOZKgI6O2ag6SJLUn0X2h6Hh3Rg/fnxOmzat6jAkSeoVETE9M8d3t59XF0uS1KJMAiRJalEmAZIktSiTAEmSWpRJgCRJLcokQJKkFmUSIElSizIJkCSpRZkESJLUolpixsCIWAw81AOH2gh4ogeOU7X+UI/+UAewHs2mP9SjP9QBrMfa2iIzR3W3U0skAT0lIqbVMw1js+sP9egPdQDr0Wz6Qz36Qx3AevQWuwMkSWpRJgGSJLUok4DVc3bVAfSQ/lCP/lAHsB7Npj/Uoz/UAaxHr3BMgCRJLcqWAEmSWpRJQJ0i4sCImBcR8yPi5KrjqUdEbBYRN0TE3IiYExH/WpafHhEPR8TMcju46li7ExEPRsRdZbzTyrLXR8R1EXFfebth1XF2JSK2rfmZz4yIZyPis33h84iI8yLi8Yi4u6asw59/FH5Y/q3Mjoi3Vxf5Kzqpw3ci4t4yzssjYoOyfGxEvFjzmfxvdZG/Wif16PR3KCJOKT+LeRHxvmqifq1O6nFJTR0ejIiZZXlTfh5d/I/tO38bmenWzQYMBO4HtgSGALOA7auOq464NwHeXt4fDvwF2B44Hfhi1fGtZl0eBDZqV/Zt4OTy/snAGVXHuRr1GQj8DdiiL3wewD7A24G7u/v5AwcDVwMB7AHcVnX8XdThAGBQef+MmjqMrd2vmbZO6tHh71D59z4LGAqMK/+PDay6Dp3Vo93zZwKnNfPn0cX/2D7zt2FLQH12A+Zn5oLMfAm4GJhQcUzdysxHM3NGef854B5gdLVR9agJwPnl/fOBwyqMZXXtB9yfmT0xiVXDZeZNwJPtijv7+U8AfpmFW4ENImKT3om0cx3VITP/kJkryoe3AmN6PbDV1Mln0ZkJwMWZuTwzHwDmU/w/q1xX9YiIAD4CXNSrQa2mLv7H9pm/DZOA+owGFtY8XkQfO5lGxFhgZ+C2suiksjnqvGZvRi8l8IeImB4RJ5RlG2fmo+X9vwEbVxPaGjmSV/+D62ufB3T+8++rfy8fp/iW1mZcRNwZEVMiYu+qgloNHf0O9dXPYm/gscy8r6asqT+Pdv9j+8zfhklAC4iI1wH/B3w2M58Ffgq8CdgJeJSi2a3Z7ZWZbwcOAv4lIvapfTKLtrY+calLRAwBDgV+Uxb1xc/jVfrSz78jEXEqsAL4VVn0KLB5Zu4MfB74dUSsX1V8dejzv0PtHMWrk+Sm/jw6+B/7d83+t2ESUJ+Hgc1qHo8py5peRAym+OX8VWb+FiAzH8vMlZm5CjiHJmke7EpmPlzePg5cThHzY21NaeXt49VFuFoOAmZk5mPQNz+PUmc//z719xIR/wS8Hzi6/IdN2Xy+pLw/naIvfZvKguxGF79DfeqzAIiIQcCHgEvaypr58+jofyx96G/DJKA+dwBbR8S48lvckcDEimPqVtmvdi5wT2b+d015bR/UB4G727+2mUTEsIgY3nafYjDX3RSfwbHlbscCV1YT4Wp71becvvZ51Ojs5z8R+Fg5EnoP4JmaptGmEhEHAl8GDs3MF2rKR0XEwPL+lsDWwIJqouxeF79DE4EjI2JoRIyjqMftvR3fanovcG9mLmoraNbPo7P/sfSlv42qRyb2lY1iVOdfKDLQU6uOp86Y96JohpoNzCy3g4ELgLvK8onAJlXH2k09tqQY4TwLmNP28wdGAtcD9wGTgddXHWsddRkGLAFG1JQ1/edBkbQ8CrxM0Y95XGc/f4qRzz8u/1buAsZXHX8XdZhP0Ufb9vfxv+W+/1D+rs0EZgAfqDr+burR6e8QcGr5WcwDDqo6/q7qUZb/Ajix3b5N+Xl08T+2z/xtOGOgJEktyu4ASZJalEmAJEktyiRAkqQWZRIgSVKLMgmQJKlFmQRILS4ini9vx0bEP/bwsb/a7vHUnjy+pLVjEiCpzVhgtZKAcna3rrwqCcjMd6xmTJIayCRAUptvAXuX67V/LiIGRsR3IuKOcmGaTwJExL4R8aeImAjMLcuuKBd3mtO2wFNEfAtYtzzer8qytlaHKI99d0TcFRFH1Bz7xoi4LCLujYhflbOyERHfimLd9tkR8d1e/+lI/VB3Wbyk1nEyxZr07wcoT+bPZOauETEUuCUi/lDu+3bgLVksTwvw8cx8MiLWBe6IiP/LzJMj4qTM3KmD9/oQxWI3OwIbla+5qXxuZ2AH4BHgFuCdEXEPxXS4b87MjIgNerz2UguyJUBSZw6gmOd8JsXyqCMp5mwHuL0mAQD4TETMAm6lWCBla7q2F3BRFovePAZMAXatOfaiLBbDmUnRTfEMsAw4NyI+BLzQwTElrSaTAEmdCeD/ZeZO5TYuM9taApb+faeIfSkWfdkzM3cE7gTWWYv3XV5zfyUwKDNXUKyMdxnFin/XrMXxJZVMAiS1eQ4YXvP4WuBT5VKpRMQ25SqO7Y0AnsrMFyLizcAeNc+93Pb6dv4EHFGOOxgF7EMXq9uV67WPyMxJwOcouhEkrSXHBEhqMxtYWTbr/wL4AUVT/IxycN5i4LAOXncNcGLZbz+PokugzdnA7IiYkZlH15RfDuxJsTJkAl/OzL+VSURHhgNXRsQ6FC0Un1+zKkqq5SqCkiS1KLsDJElqUSYBkiS1KJMASZJalEmAJEktyiRAkqQWZRIgSVKLMgmQJKlFmQRIktSi/j/Xpdpdty0e3QAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f35243a9390>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.sched.plot_lr(size=(8,6))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[Figure Annotation](https://matplotlib.org/users/annotations_intro.html)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'loss': 0.032150164, 'lr': 0.0073564225, 'iteration': 172.0}\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAfUAAAGDCAYAAAAyM4nNAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMS4yLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvNQv5yAAAIABJREFUeJzs3XmYFNXVBvD3zIZADzuoLALKALKJOO7iriCoRCUoiUqMiCYuiYqK+dT4aTRqjCRxX1CMG64gLgjGJSAxygziwiIiDLJ9gsM2M8Cs5/vjdNE1M90zvU5317y/5+mnu6urq27V1PSpc++tW6KqICIiovSXkewCEBERUXwwqBMREXkEgzoREZFHMKgTERF5BIM6ERGRRzCoExEReQSDOlECicgcEZmQ7HI0pVi2WURuF5Hn410mouaCQZ08SUSKROTUZJdDVc9Q1WfjvVwROVFEakSkVERKRORbEbkkgu/HFDz936/0r9953AgkbpujEctxkCrHEFEkspJdAKJ0JSJZqlqVxCJsVNXuIiIAzgAwW0T+o6rfNtH6X1bVC5toXQ3y7wNR1Zpkl4UomZipU7MjImeKyBIR2S4i/xGRIa7PpojI9/7sd5mInOP67FcislBEpopIMYDb/dM+EZH7RWSbiKwRkTNc3/lYRCa6vt/QvL1FZL5/3f8SkYfDyabVvAtgKwD3tvxdRNaJyE4RKRSR4f7pIwH8AcD5/gz7S//0tiIyTUQ2icgGEfmTiGRGsX8j3eZ/+7f5fQCd6izrKP/faLuIfCkiJ9ZZz10ishDALgAHNlKuBssS4TZeJiKrRGSriMwWka7+6eI/Pjb79/vXIjLI/9ko/zFV4t+/k6NZN1FDGNSpWRGRQwE8DeByAB0BPA7LcFv4Z/kewHAAbQH8L4DnRWR/1yKOBLAawL4A7nJN+xYWkO4DMM2fOQbT0LwvAvjcX67bAVwU5jZliMjZ/mWucn20CMBQAB38y35VRPZR1fcA3A3LtH2qeoh//ukAqgD0AXAogNMBTAynDI1obJsL/Z/dCWBvW7yIdAPwDoA/+bdhMoDXRaSza9kXAZgEIBfA2hjLEhYRORnAnwGMA7C/f70z/B+fDuB4AH1hx9A4AMX+z6YBuFxVcwEMAvBhJOslCgeDOjU3kwA8rqqfqWq1v+23HMBRAKCqr6rqRlWtUdWXAXwH4AjX9zeq6oOqWqWqu/3T1qrqk6paDeBZ2A/9viHWH3ReETkAwOEAblPVClX9BMDsRralq4hsB7AbwEwA16nqF86Hqvq8qhb7y/pXAC0A9Au2IBHZF8AoAL9X1TJV3QxgKoALGlj/OH8G7Ty6RrnNt6pquarOB/CW63sXAnhXVd/1/z3eB1DgL6djuqou9W9jZQNlbbAsYXzP7ZcAnlbVxapaDuBmAEeLSC8AlbATjP6w5oDlqrrJ/71KAANEpI2qblPVxRGul6hRDOrU3PQEcL07GAHoAcCpPr3YVTW/HZZRuauE1wVZ5v85L1R1l/+lL8T6Q83bFcBW17RQ63LbqKrtALQB8A8AJ7s/FJHJIrJcRHb4t6VtnW1x6wkgG8Am17Y/DqBLA+t/RVXbuR4bQ8zX0DZvU9Uy17zubLsngJ/X+VsdBwvEjsb2UbhliURXdzlVtRSWjXdT1Q8BPATgYQCbReQJEWnjn/U82AnJWn+Tw9ERrpeoUQzq1NysA3BXnWDUSlVfEpGeAJ4EcBWAjv6A+Q0Ad/Vsom5ruAlABxFp5ZrWI5wv+rPFmwAMFpGfAYC//fxGWPVve/+27EBgW+puxzpYjUUn135po6oDo96ixm0C0F5EWrumHVCnTM/V+Vu1VtV7XPMk4zaTG2EnHAAAf/k7AtgAAKr6D1U9DMAAWDX8Df7pi1R1DOxEaRaAV5q43NQMMKiTl2WLyD6uRxYsaF8hIkf6OzW1FpHRIpILoDUsSGwBALFLxAY1RUFVdS2savl2EcnxZ3FnRfD9CgB/BXCbf1IurH18C4AsEbkNltE7fgTQS0Qy/N/fBGAegL+KSBt/O/1BInJCrNvWQJmdbf5f/zYfh9rb/DyAs0RkhIhk+v+GJ4pI90SVKYhgx9BLAC4RkaH+vhh3A/hMVYtE5HD/sZUNoAzAHgA1/u37pYi09TcT7ATAnvoUdwzq5GXvwtqbncftqloA4DJYFek2WMeyXwGAqi6DBcZPYUFvMICFTVjeXwI4GlaV+ycAL8Oy53A9DeAAETkLwFwA7wFYCasq3oPaVdWv+p+LRcRp270YQA6AZbB98xpqV3Unwi9gnde2AvgjgH86H6jqOgBjYD31t8DKfwOa9ncr2DH0LwC3AngdVttwEAJ9D9rAThy3wfZ7MYC/+D+7CECRiOwEcAXs700UV6KajNorImqMiLwMYIWq/jHZZSGi9MBMnShF+KtuD/JXfY+EZamzkl0uIkofHFGOKHXsB+ANWKer9QB+475EjYioMax+JyIi8ghWvxMREXkEgzoREZFHpF2beqdOnbRXr17JLgYREVGTKCws/ElVOzc+ZxoG9V69eqGgoCDZxSAiImoSIhLOzYoAsPqdiIjIMxjUiYiIPIJBnYiIyCMY1ImIiDyCQZ2IiMgjGNSJiIg8gkGdiIjIIxjUiYiIPIJBnYiIyCMY1ImIiDyCQZ2IiMgjmn1QX7482SUgIiKKj2Yd1F99FRgwAJg5M9klISIiil2zDupffWXP33yT3HIQERHFQ7MO6kRERF7CoE5EROQRDOpEREQewaBORETkEQkL6iLytIhsFpGg3dDE/ENEVonIVyIyLFFlISIiag4SmalPBzCygc/PAJDnf0wC8GgCy0JEROR5CQvqqjofwNYGZhkD4J9q/gugnYjsn6jyNEQ1GWslIiKKr2S2qXcDsM71fr1/Wj0iMklECkSkYMuWLXErgEjcFkVERJR0adFRTlWfUNV8Vc3v3LlzsotDRESUkpIZ1DcA6OF6390/jYiIiKKQzKA+G8DF/l7wRwHYoaqbklgeIiKitJaVqAWLyEsATgTQSUTWA/gjgGwAUNXHALwLYBSAVQB2AbgkUWUhIiJqDhIW1FV1fCOfK4ArE7V+IiKi5iYtOsolyrZt9sxL2oiIyAuadVB/6CF7fu+95JaDiIgoHpp1UHdUVCS7BERERLFjUAer34mIyBsY1ImIiDyCQZ2IiMgjGNTB6nciIvIGBnUiIiKPYFAHM3UiIvIGBnUwqBMRkTcwqBMREXkEgzoREZFHMKgTERF5BIM6ERGRRzCogx3liIjIGxjUAfz4Y7JLQEREFDsGdQCbNye7BERERLFjUCciIvIIBnUiIiKPYFAnIiLyCAZ1IiIij2BQJyIi8ggGdSIiIo9gUCciIvIIBnUiIiKPYFAnIiLyCAZ1IiIij2BQJyIi8ggGdSIiIo9gUCciIvIIBnUiIiKPYFAnIiLyCAZ1IiIij2BQJyIi8ggGdb/165NdAiIiotgwqPsxqBMRUbpjUPerrk52CYiIiGLDoO7HoE5EROmOQd2vpibZJSAiIooNgzoREZFHMKgTERF5BIM6ERGRRzCo+6kmuwRERESxYVAnIiLyCAZ1IiIij2BQ92P1OxERpTsGdb+KimSXgIiIKDYM6n4iyS4BERFRbBjUiYiIPIJB3W/XrmSXgIiIKDYJDeoiMlJEvhWRVSIyJcjnB4jIRyLyhYh8JSKjElmeuvLyAq+vuaYp10xERBR/CQvqIpIJ4GEAZwAYAGC8iAyoM9stAF5R1UMBXADgkUSVJ5iWLQOveT91IiJKd4nM1I8AsEpVV6tqBYAZAMbUmUcBtPG/bgtgYwLLQ0RE5GlZCVx2NwDrXO/XAziyzjy3A5gnIlcDaA3g1ASWpx5em05ERF6S7I5y4wFMV9XuAEYBeE5E6pVJRCaJSIGIFGzZsqXJC0lERJQOEhnUNwDo4Xrf3T/N7VIArwCAqn4KYB8AneouSFWfUNV8Vc3v3LlzgopLRESU3hIZ1BcByBOR3iKSA+sIN7vOPD8AOAUARORgWFBvslSc1e9EROQlCQvqqloF4CoAcwEsh/VyXyoid4jI2f7ZrgdwmYh8CeAlAL9SZaglIiKKRiI7ykFV3wXwbp1pt7leLwNwbCLLQERE1Fwku6McERERxUmzDuqs6CciIi9p1kGdiIjIS5p1UOftVomIyEuadVAnIiLykmYd1Nu1S3YJiIiI4qdZB/WMZr31RETkNQxrREREHsGgTkRE5BEM6kRERB7BoE5EROQRDOpEREQewaBORETkEQzqREREHsGgTkRE5BEM6n4ciIaIiNJdsw5lkycnuwRERETx06yD+llnAbfcYq95b3UiIkp3zTqoA4HbrzKoExFRumv2QZ1t6URE5BXNPqQ5mToREVG6a/ZBnYiIyCuafVBnpk5ERF7BoM6gTkREHsGgzqBOREQewaDOoE5ERB7BoM6gTkREHtHsgzoREZFXNPugzsFniIjIK5p9SHNXv3/wQfLKQUREFCsGdVdQf+yx5JWDiIgoVgzq7ChHREQe0eyDuhvv1EZEROms2Qd1ZupEROQVzT6os/c7ERF5BUMaERGRRzT7oF5enuwSEBERxUezD+qTJgVes6McERGls2Yf1Dt2THYJiIiI4qPZB3U39oQnIqJ0xqDuwup3IiJKZwzqREREHsGgTkRE5BEM6i6LFiW7BERERNFjUHcpLk52CYiIiKLHoO6SmZnsEhAREUWPQd0lOzvZJSAiIooeg7rL1q3JLgEREVH0GNSJiIg8gkGdiIjIIxjUiYiIPCKhQV1ERorItyKySkSmhJhnnIgsE5GlIvJiIstDRETkZVmJWrCIZAJ4GMBpANYDWCQis1V1mWuePAA3AzhWVbeJSJdElYeIiMjrEpmpHwFglaquVtUKADMAjKkzz2UAHlbVbQCgqpsTWB4iIiJPS2RQ7wZgnev9ev80t74A+orIQhH5r4iMDLYgEZkkIgUiUrBly5YEFZeIiCi9JbujXBaAPAAnAhgP4EkRaVd3JlV9QlXzVTW/c+fOTVxEIiKi9JDIoL4BQA/X++7+aW7rAcxW1UpVXQNgJSzIExERUYQSGdQXAcgTkd4ikgPgAgCz68wzC5alQ0Q6warjVyewTERERJ6VsKCuqlUArgIwF8ByAK+o6lIRuUNEzvbPNhdAsYgsA/ARgBtUtcnvlbbPPk29RiIiovgTVU12GSKSn5+vBQUFcV3mxInAtGn2Os12BxEReZyIFKpqfjjzJrujXEoQSXYJiIiIYsegTkRE5BEM6mCmTkRE3sCgDgZ1IiLyBgZ1ABncC0RE5AEMZ2CmTkRE3hBWUBeRg0Skhf/1iSJyTbDhXImIiCh5ws3UXwdQLSJ9ADwBG/7VM/c+Z/U7ERF5QbjhrMY/Qtw5AB5U1RsA7J+4YhEREVGkwg3qlSIyHsAEAG/7p2UnpkhNj23qRETkBeEG9UsAHA3gLlVdIyK9ATyXuGI1rXPPTXYJiIiIYpcVzkyqugzANQAgIu0B5KrqvYksWFM6/PDA682bgS5dklcWIiKiaIXb+/1jEWkjIh0ALAbwpIg8kNiiNR139fvppyevHERERLEIt/q9raruBHAugH+q6pEATk1csZLnu++SXQIiIqLohBvUs0RkfwDjEOgo5xnuTJ2XtxERUboKN4TdAWAugO9VdZGIHAjAMzmtO6hnZiavHERERLEIt6PcqwBedb1fDeC8RBWqqTGoExGRF4TbUa67iMwUkc3+x+si0j3RhUuGrLBOc4iIiFJPuNXvzwCYDaCr//GWf5onMFOPn2OOOQYAUFRUhBdfjO9Iwq+++ioGDhyIjIwMFBQU7J3+wgsvYOjQoXsfGRkZWLJkCXbt2oXRo0ejf//+GDhwIKZMmRLX8hARpZpwg3pnVX1GVav8j+kAOiewXE2KHeXi5z//+Q+A6IJ6VVVVg58PGjQIb7zxBo4//vha03/5y19iyZIlWLJkCZ577jn07t0bQ4cOBQBMnjwZK1aswBdffIGFCxdizpw5EZWJiCidhBvCikXkQhHJ9D8uBFCcyII1JWbq8ePz+QAAU6ZMwYIFCzB06FBMnToV1dXVuOGGG3D44YdjyJAhePzxxwEAH3/8MYYPH46zzz4bAwYMaHDZBx98MPr169fgPC+99BIuuOACAECrVq1w0kknAQBycnIwbNgwrF+/PtZNJCJKWeEG9V/DLmf7PwCbAIwF8KsElSmp2KYeH/fccw+GDx+OJUuW4Nprr8W0adPQtm1bLFq0CIsWLcKTTz6JNWvWAAAWL16Mv//971i5ciUAYNSoUdi4cWNU63355Zcxfvz4etO3b9+Ot956C6ecckr0G0VElOLC7f2+FsDZ7mki8nsAf0tEoZoaq98Tb968efjqq6/w2muvAQB27NiB7777Djk5OTjiiCPQu3fvvfO+++67Ua3js88+Q6tWrTBo0KBa06uqqjB+/Hhcc801OPDAA6PfCCKiFBdLXnodPBjUece2xFBVPPjggxgxYkSt6R9//DFat24dl3XMmDEjaJY+adIk5OXl4fe//31c1kNElKpiyUs9E/4YyOMvNzcXJSUle9+PGDECjz76KCorKwEAK1euRFlZWdzWV1NTg1deeWVve7rjlltuwY4dO/C3v3ni/JOIqEGxBHWNWylSCAN8fAwZMgSZmZk45JBDMHXqVEycOBEDBgzAsGHDMGjQIFx++eUhe7uHalOfOXMmunfvjk8//RSjR4+ulfXPnz8fPXr0qFW9vn79etx1111YtmwZhg0bhqFDh+Kpp56K/8YSEaUIUQ0dm0WkBMGDtwBoqapN3q0sPz9f3dcox4NqoC29Xz9gxYq4Lp6IiChqIlKoqvnhzNtgUFbV3PgUKbUxOyciIi9gX+86GOCJiChdMajXsWIFsGtXskvhDb169cJPP/2U8PU8++yzyMvLQ15eHp599tmg82zduhWnnXYa8vLycNppp2Hbtm0AgDfffBNDhgzB0KFDkZ+fj08++QQAsGTJEhx99NEYOHAghgwZgpdffjnh20FEFKsG29RTUSLa1IHaGfrddwM33xz3VTQ7vXr1QkFBATp16pSwdWzduhX5+fkoKCiAiOCwww5DYWEh2rdvX2u+G2+8ER06dMCUKVNwzz33YNu2bbj33ntRWlqK1q1bQ0Tw1VdfYdy4cVixYgVWrlwJEUFeXh42btyIww47DMuXL0e7du0Sti1ERMFE0qbOTD2I6upkl8B7HnjgAQwaNAiDBg3ae3lZWVkZRo8ejUMOOQSDBg3amw1PmTIFAwYMwJAhQzB58uQGlzt37lycdtpp6NChA9q3b4/TTjsN7733Xr353nzzTUyYMAEAMGHCBMyaNQuADWsr/jO6srKyva/79u2LvLw8AEDXrl3RpUsXbNmyJQ57gogocTgoahDZ2ckugbcUFhbimWeewWeffQZVxZFHHokTTjgBq1evRteuXfHOO+8AsFHmiouLMXPmTKxYsQIigu3btwMAZs+ejYKCAtxxxx21lr1hwwb06NFj7/vu3btjw4YN9crw448/Yv/99wcA7Lfffvjxxx/3fjZz5kzcfPPN2Lx5896yuH3++eeoqKjAQQcdFPvOICJKIGbqQTCox9cnn3yCc845B61bt4bP58O5556LBQsWYPDgwXj//fdx0003YcGCBWjbti3atm2LffbZB5deeineeOMNtGrVCgBw9tln1wvo0RKRvRk5AJxzzjlYsWIFZs2ahVtvvbXWvJs2bcJFF12EZ555BhkcQ5iIUhx/pYLgTV2aRt++fbF48WIMHjwYt9xyC+644w5kZWXh888/x9ixY/H2229j5MiRDS6jW7duWLdu3d7369evR7du3erNt++++2LTpk0ALFB36dKl3jzHH388Vq9evbdz386dOzF69GjcddddOOqoo2LZVCKiJsGgHgSDenwNHz4cs2bNwq5du1BWVoaZM2di+PDh2LhxI1q1aoULL7wQN9xwAxYvXozS0lLs2LEDo0aNwtSpU/Hll182uOwRI0Zg3rx52LZtG7Zt24Z58+bVG18esEzf6Rn/7LPPYsyYMQCAVatWweksunjxYpSXl6Njx46oqKjAOeecg4svvhhjx46N8x4hIkoMhq8gGNTja9iwYfjVr36FI444AgAwceJEHHrooZg7dy5uuOEGZGRkIDs7G48++ihKSkowZswY7NmzB6qKBx54AEDoNvUOHTrg1ltvxeGHHw4AuO2229ChQ4e967niiiuQn5+PKVOmYNy4cZg2bRp69uyJV155BQDw+uuv45///Ceys7PRsmVLvPzyyxARvPLKK5g/fz6Ki4sxffp0AMD06dMxdOjQpthlRERR4SVtfu5L2p5+GrjkkrivgoiIKGK8pC0K7huGMVMnIqJ0xKDu5+9kDYBBnYiI0hODehC8pI2IiNIRg3oQzNTjI5qx38vLy3H++eejT58+OPLII1FUVBR0vvfeew/9+vVDnz59cM899+ydvmbNGhx55JHo06cPzj//fFRUVAAArr32WgwdOhRDhw5F37596w33unPnTnTv3h1XXXVVZBtJRJRCGNSpSakqampqQn4+bdo0tG/fHqtWrcK1116Lm266qd481dXVuPLKKzFnzhwsW7YML730EpYtWwYAuOmmm3Dttddi1apVaN++PaZNmwYAmDp1KpYsWYIlS5bg6quvxrnnnltrmbfeeiuOP/74OG4pEVHTY1APooGYQ1EoKipCv379cPHFF2PQoEG1Boupyz1G+9ixY/HBBx+g7hUan3/+Ofr06YMDDzwQOTk5uOCCC/Dmm29CVfHhhx/uva7cPca720svvYTx48fvfV9YWIgff/wRp59+ejw2l4goaRjUg7j99mSXwHu+++47/Pa3v8XSpUvRs2dPTJw4EcEuTXSP5Z6VlYW2bduiuLg45DxAYLz34uJitGvXDln+9pNg48CvXbsWa9aswcknnwwAqKmpwfXXX4/7778/rttLRJQMDOou//iHPX/9dXLL4UU9e/asNdTqU089hfz8sC67jKsZM2Zg7NixyMzMBAA88sgjGDVqFLp3797kZSEiijd2CXPxJ2+UAK1btw5rPmcs9+7du6Oqqgo7duxAx44dg87jcMZ779ixI7Zv346qqipkZWUFHQd+xowZePjhh/e+//TTT7FgwQI88sgjKC0tRUVFBXw+X63Od0RE6YKZuot7VLk0G2jPM9xjtL/22ms4+eSTa91RDQAOP/xwfPfdd1izZg0qKiowY8YMnH322RARnHTSSXjttdcA1B7jHQBWrFiBbdu24eijj9477YUXXsAPP/yAoqIi3H///bj44osZ0IkobTGou7hjx1tvJa8czUGoNvVLL70UxcXF6NOnDx544IG9AXbjxo0YNWoUAGtrf+ihhzBixAgcfPDBGDduHAYOHAgAuPfee/HAAw+gT58+KC4uxqWXXrp32TNmzMAFF1xQ7ySBiMgrOPa7y4oVwMEH2+sXXwRcHaSJiIiSImXGfheRkSLyrYisEpEpDcx3noioiDR9zymXjIzgr4mIiNJBwkKXiGQCeBjAGQAGABgvIgOCzJcL4HcAPktUWcLlrpVlUCcionSTyNB1BIBVqrpaVSsAzAAwJsh8dwK4F8CeBJYlLO6gPm4cMGlS8spCREQUqUQG9W4A3EOHrfdP20tEhgHooarvJLAcYavbf+rJJ5NTDiIiomgkrZJZRDIAPADg+jDmnSQiBSJSsGXLloSViVXuRESUzhIZxjYA6OF6390/zZELYBCAj0WkCMBRAGYH6yynqk+oar6q5nfu3DlhBeaVTkRElM4SGdQXAcgTkd4ikgPgAgCznQ9VdYeqdlLVXqraC8B/AZytqom5Xi0MDOpERJTOEjZMrKpWichVAOYCyATwtKouFZE7ABSo6uyGl9D0gl2yn5cHnHFGYFx4Si2zZs3CF18sRknJVhx33Ak499yfJ7tIRERJw8FnXNasAQ48MPhnababmo1hww5Cfv5qVFUBW7YMx1tvzU92kYiI4iplBp8hSjSfrzV++UvgwguBkpKdyS4OEVFSMai7MBtPP7m5bVBSAuTmAqWlZckuDhFRUjGou9TUJLsEFCmfLxelpYDPB5SUMKgTUfPGoO7CTD395Oa2Q2mpk6nvTnZxiIiSikGd0prP1w4lJU6mzqBORM0bg7oLM/X0k5vbYW/1e1lZBdLtag4ionhiUHfp1CnZJaBI+XxtUFKSiawsICcnE7t27Up2kYiIkoZB3aVDB+Cjj5JdCoqEz+dDaWk2ACA3NwulpaVJLhERNWeffAI8/zxQXZ2c9TOo19GhQ/DppaXAtdcCW7c2bXmoYbm5uSgpyQQA+HxZKCkpSXKJiKg5e+YZixXJukFYwoaJTVeh/hBTpwJ/+xtQVgY88UTTlolCs0zdgnpubgYzdSJKqsJC4LDDkncvEWbqdYQK6mvX2nNlZdOVhRpnmbr99/h8wkydiJJmzx5g6VJg2LDklYFBPUybNtkz7+SWWixTt9d2rTozdSJKjq++AqqqLFNPFgb1OkJ1bigubtpyUHgsU7ehAH2+GgZ1IkqawkJ7ZlBPIaGGinWCOjP11GKZuv3RcnOrWf1ORElTWGidrXv2TF4ZGNTrYKaeXixTrwIA+HzVzNSJKGkWL05uJzmAQb2eUJn6tm32zEw9tVimbr0Xc3MrePtVIkqK8nLgm2+S20kOYFCvJ1kDBlB0WrRoAUBQXg74fIrS0h3JLhIRNUNff21XRyWzPR1gUK9n332TXQKKlM/XYu+d2kpKODoQETW9VOgkBzCo19OrV/1pyRoZiMKTm9ty753amKkTUTIUFgLt2wO9eye3HAxXYfD5kl0CaojP19KVqTOoE1HTW7zY2tOT3e+KQT0MubmB18n+g1F9ubmtXZk6O8oRUdOqqLA29WR3kgMY1MPiDuqUepxR5SxT5yVtRNS0vvnGAnuy29MBBvWgNm+u/d5d/c5MPfXk5rZxZeoM6kTUtFKlkxzAoB5U587AeecF3jNTT20+XxuUllpQLynZleziEFEzU1gItG0LHHRQskvCoB6SexAaZuqpLTe3PUpKnBu67El2cYiomUmVTnIAg3pI7qCenZ28clDjfL52rkydQZ2Imk4DVdMsAAAgAElEQVRlpd2dLRU6yQEM6iE5QX3CBGD37sD0VDgTo9qcoN6yJVBZWY1K3vSeiJrI0qU2RGwqtKcDDOohOUF97FhANblloYZZR7kciAA+XzY7yxFRk0mlTnIAg3pITlDPyKg9Hjwz9dRjl7RlAQByc7MY1ImoyRQWWn+ePn2SXRLDoB5CqKDOrD312O1XMwEAPl8m76lORE3G6SSXKsOJp0gxUo87qFdVBabzLm6pxzJ1O5RzczOYqRNRk6iqAr78MnU6yQEM6iE5wbtupu4O8JQaLFO31xyAhoiayrJlwJ49qdOeDjCoh8RMPX1Ypm7tIrm5yup3ImoSqdZJDmBQD8kJ6pmZzNRTnWXq9gfz+WqYqRNRkygstNrBvn2TXZIABvUQQmXqDOqpxzJ1O/PKza1mpk5ETWLxYuDQQ1OnkxzAoB5SqN7vrH5PPZap29mWz1fJTJ2IEq6qCliyJLU6yQEM6iGFCurM1FNP69atsWtXJWpqgNzcKpSU8J7qRJRYK1bYaKOp1J4OMKiHlKiOci++CHz+eWzLoNoyMjLQqlUOysqc3u/bkl0kIvK4VOwkBzCoh+QO6m3a2OtOnSzAV1dHPwjNDTcADz0UnzJSgM+3D0pLbWSnkhIGdSJKrMJCoHVroF+/ZJekNgb1ENzXqb/xBvCXvwB5eTZ9v/2AQYPCW8411wAffRR4v2cPwCbf+MvNbYmSEidT35Hs4hCRxy1eDAwdaldIpRIG9RDcmXqPHsDkyUBWlmXqP/1kgw5s2QLMmhV6GaqWlb/7bmBaeTmDeiL4fK1cmTqDOhElTnU18MUXqddJDgCykl2AVOUO6o6sLMu0HaNGAQUFwM6dFkzqqqqywL5rV2BaeTlQVpaYMjdnubk+V6bOS9qIKHG+/dZ+11OtPR1gph7SffdZNXteXmBaTg5QVBR4//XX9hwqSJeX27NzP/aaGgv0zNTjz+ezoG6ZOncwESVOqnaSAxjUQxo5Eti0CWjVKjCtRQub5nCCdqig7mT1TlB35mdQj7/c3LYoLXUy9V2Nf4GIKEqFhUDLlkD//skuSX0M6hHIyQm83mefwOtQQbpupt7YSQBFz+dr62pTZ1AnosRxOsllpWADNoN6BFq0CLx2t603FtSdNnVm6omTm9ve1aa+p/EvEBFFoaYmdTvJAQzqEXEHdbdoMnWnI55Xbd9u/RGaaqAdn6+9q/q9HBrtQAJERA1YudJ+81OxPR1gUI9IqKDuVKdXVABPP2094oH6Qd2d3TvTvGrVKnt8/HHTrM/na4OSkkxkZQE5OZnY7fUdTERJkcqd5ABe0haRxjL1jz8GLr3UXldWhq5+d77TunVCipkSiovt+fvvm2Z9ubm5+OGHbADVyM3NQklJCVq5ezkSEcVBYaH1qRowINklCY6ZegQaC+o7XfcR2bUrdO9393e8autWe26qoG63X83yv87indqIKCEWLwYOOSQ1O8kBDOoRaSyouweZ2bUrdJs64P0e8E0d1O32qxn+1xlh31Odd90jonDV1FhQT9VOckCCg7qIjBSRb0VklYhMCfL5dSKyTES+EpEPRKRnIssTq+zs2u/PO8+et24F1q2rHagbC+peTySd6vcffrC+Bolmmbr4X0tYmfrKldYEsnhxoktHRF6wahVQUpK67elAAoO6iGQCeBjAGQAGABgvInVbIb4AkK+qQwC8BuC+RJUnnsRiB7KyLCjcey9wwAH1q98ba1P3MidTr6kB1q5N/PosU3deI6xM/dNP7YTj008TXDgi8oRU7yQHJDZTPwLAKlVdraoVAGYAGOOeQVU/UlWn0vq/ALonsDwxc+7c1rKlPWdm2iVUDvdoc+6gXlVlj+ZU/e5k6kDTVMFbpq7+1zVhZerLl9vzt98msmRE5BWFhdYMO3BgsksSWiKDejcA61zv1/unhXIpgDnBPhCRSSJSICIFW7ZsiWMRI+NcW+6MJpeZWbsH+8aNgdfujnKAVcE3t0x9//3tdSKCet3L0C1Tr/a/jiyor1gR79IRkRctXgwMGVK/KTaVpERHORG5EEA+gL8E+1xVn1DVfFXN79y5c9MWzqVupp6RUfvubHWDujuI133v9Ux961Y7m23ZMv5B/ZlnrLljw4bANMvUq/yvq8KqfmemTkThUk39TnJAYoP6BgA9XO+7+6fVIiKnAvgfAGeranndz1OJE9SdMeAzMoA2bQKfu4NM3SDe3DL14mKgY0fgwAPjH9T/8x9g/XpgwoRA7Yll6pX+15WNZurl5Vauli2tM98uDhdPRA34/ntgx47Ubk8HEhvUFwHIE5HeIpID4AIAs90ziMihAB6HBfTNCSxLXDhB3ZGZCbRtG3i/YUPg/uvNPahv3WpB/aCD4h/Ui4qsCeSDD4CpU21aTk4OVK3jm89Xg5KS7Q0uY+VKOyEYOTLwnogolHToJAckMKirahWAqwDMBbAcwCuqulRE7hCRs/2z/QWAD8CrIrJERGaHWFxKcLJCp/d73Uy9shLo1MleNxbUk1H9vmZN7TIkSk0NsG0b0KGDBfXVq+u3gcdizRpgzBjgnHOAm28GliwBRAS5ufvsvad6aem2BpfhVL3/7Gf2zCp4ImpIYaHV0g4alOySNCyhbeqq+q6q9lXVg1T1Lv+021R1tv/1qaq6r6oO9T/ObniJyTVlCnD++cDEifa+bqYOhA7q7o5zbds2faa+Y4e1cT/+eNOsq6YmENR37659ZUAsqquturx3b+DJJ4HOnYHx423/+nz77L2pS0lJ40FdBDjzTHtmUCeihixeDAweXPsW3KkoJTrKpYtOnYAZMwKXsdXN1AELMkDDvd/bt2/6oL54sZWhKa4Zd65Rd6rfgfhVwW/caDUivXrZ8p991nqvT54M5Oa2cmXqOxtczvLldmLQoYN1umMPeCIKJV06yQEM6lFx2tYzMupn6m3b2qA0oarfc3Is6DR19bvTHuS+fjxRnHU4mToQv6BeVGTPvXvb86mnWkB/9FEgK6uVK1NvPKgffLC97t+fmToRhbZmjTUppnp7OsCgHhWnbT0zs36m3qqVPUJd0rbPPhZ0djYcc+LOuR3sTz81PF9lJXD11da7PFpOpt6hA9Czp538xCuor1ljz716BaZdc43zKteVqYe+pK262oK4E9T79bP3vAU7EQWTLp3kAAb1qIwda4H5sssCA9E46gb1zEyb7mTqLVpYQPruu6Ytc7iZ+rJlwEMPAW+8Ef263NXvOTlWvR3PoC5iJwuO/fazaSK5rkw9dFWI02HQnamXldW+JJFS0yefWN+QJI5BRc1QYaENODN4cLJL0jgG9SgccIAN6j9gQOASNkfdoN6unU13B/VDDrHOXk7wS7Tt2+1GBEDjmfpm/4WFsQRhd/U7EN/L2oqKgK5da98xLzsb2HdfoKqqrStTD33hudPz3Z2pA6yCTwf33GMnnrNT+joZ8prCQuv1HupOnamEQT1GdYN6aakF9RdeAF591TrFAcCVV9oPUYsWwNChNu2rr+x59+7EVv06dyHLy2s8U3cyIOckIBrOyYqz7fEM6mvW1K56d3TrBpSXt3Vl6rtDLoNBPT0VFQHvvmuv33knqUWhZiSdOskBDOox69rVnocPt+eVK2tf8uBk6oAFVCdTB4Avv7QA2KlTYn+knKr3ESOss0dD9xCPV6budBgELKgXF9ulbrEqKgp0knPr1g0oLe3gytT31J/Jb/lyG5fe+dt07WonAuwBn9oefzxwGeK8eU0z5gLR2rX2O50O7ekAg3rMTjvNAvIrr9j7o46ywVYcrVpZINp3X3vfooW1AXfpYkG9qMiq6pcsia0c27cDzz0H/O1vVkPgVlho2a2TkTZU7e9k6qtX1x9BL1xbtwaq3oH49YCvrLT71ofK1HfsaIfSUuvnUFlZjaoQZy/unu+ABQqnsxylpvJyYNo04OyzgSuusD4Q8+cnu1SUykpL4/M/nU6d5AAG9bgYNcoC9dq1wJ//bAHWUV1tnbqcwOa0yRxyiAV1JzOOtZPWX/8KXHwxcO21wLhxwI8/Bj4rKLADsmNHe99QFbxTnsrK6HvAO0PEOuIV1NevtysPgmXq3bsDu3e3wc6dORABfL7soOO/q9YP6gCDeqp7/XU74fzNb4CTTrITt7ffTnapKFWpWofmQYNsOOlYFBZareOQIfEpW6IxqMfRAQdY1btT7QwA//d/9uzO1AHrZLdyZfyC+qpVlsF++KG9dw7kbdssmObnB0a7czrLVVQATzxROyN39yqOtl29uDgxmbpzOVuo6nfAh5IS2/k+X1bQO7Vt2mSXE9YN6v3720kZb+ySmh59FOjTx8YlaNUKOOUUC+q8DJGCmTMHmDvXTv7OOw9YujT6ZRUW2hUXda90SlUM6gmwbJllzkAgqO+3nz07Qb1rV6secgLdhg0WUJzryffsqZ1tN2btWrsj2vHHW0CdN8+mL1xoz4cdFgjqTqY+Zw5w+eW1qzE3bwb69rXX0QbhutXvubk20l68gnqo6ne7Tj3Tv87MoJn6smX2HCxTB5r+UsPmZOPG6EZS/Ppru5TtiisCHVPPPNOaiFi7QnVVVgLXX28dg7/4wu7EOGpU4Lc4EunWSQ5gUE+IvDyrAgcCI8c5mbpz3brTwe7LL+15wwbgkUeAI4+0auZbbrHlhDusa1GRVfNnZlo2M2+eHZBPPWUB9fjjA1XiTqbuBEl3NfuWLcChh9rJR7SZet3qd8CyrPnzY7uMr6jItq9Hj/qfOZn69u12tx2fLyNopu70fB8woPZ0J6izs1xiVFUBhx9ubeKRZtePPmpZ0q9+FZg2apQ9swqe6nr8cfs/vv9++915+237zTvzzMhH8ly3zr6bLu3pAIN6wuy/f+33Tqbu9AB3grrTQW7zZuC//7U2448/ts53JSWWnTT2I1hebtXKTgZ7+un2/r33gLfeshvQtGhRP1N3hlx1V/1v3mwnINHeB726OnCHNrc//MHWd9xxdo1+NNassbZzd/OGI1D9bkHdesDXTwuXL7de785JliMvjzd2SaT58y1T/+gju9wzXCUl1gH0/PNrnygecIC1cTKok9u2bcAf/wicfDJw1lk27bDDgJdftqx9/PjIOgCnWyc5gEE9YTIzgQsvBJ5/3t47QcTpROcEdScTVw20gz//vJ1pHnKIBeZLLrEe7VdfDbz2WmB+J8tft86enVHWTjvNni+7zOabNMnet2plGY+TqTvrdoJ6ebm1N3fpYu3g0WTqO3bYOusG9TPPtDaujRuBY46xKtVIhbqcDbDhelu2zMXOnXYG5PNpyEz94IMDt891tGplgYJBPTFefdX2cX4+cN119uMbjueftyr73/62/mdnnmnV8u6OqdS83XmnHVsPPFD7f/zMM4EHH7Qk5/e/D7+2qLDQfsudy5DTAYN6Aj33HPDLX9prJ1N3foDcmXx2du3P5s615+nTgauusrPMceNs+Fanrf5f/7JBbBYtCmTcTlA/4ADr+LVhAzB6dO026E6dApl63aDudJLr3Nmqrb7/PvKqUmfZdavfAeDEE4EFC2yZxx0Xeft1qIFnHPvt50NJiRU4N7cmZKZetz3d0b9/cqrfS0sD9xPwoupqG3Z49Gi7XW5xsdXcNEbVmqSGDbOq+7pGj7ZlO/8v1LytXGmB+9JLgwfh3/7W2tofegj4+9/DW2ZhoTXVtWwZ37ImEoN6E6mbqbdpY5kLYJddOI4+2p67dLED88EHLeguWGCZvxPAnd6c//1vIDi7A56Trf/mN7XL0bFj6Ezd6YnvZOplZZF11gNq38wlmMGDrYf+zp2R/RiXl1uWHypTB4Bu3XJRVmbXpvt81fUy9a1bbXtCBfV+/eyHoSl7VFdW2gnUtdc23Tqb2vz5dmz9/Od2Ivq731m752efNfy9hQuBb76xY7huzQpg/U86dmQVPJkbbrCayDvvDD3PffdZb/jrrmv8/haqFtTTqZMcwKDeZJyg7lwyJRKogncfNM4dx049NfBD1qGDZbb9+lkPzt27AwPcfPGFBeeMDKddObCc22+3UeTcnEy9pCQQgENl6kDkVfDum7mE0revndREkhU7JyANBfXu3X3YtcuCem5uZb1Mve7wsHX162dZ88aN4ZcrVosW2YnGgw8GrnzwmldfDfRABoD//V879q+4ouHRDR95xEYmHD8++OeZmbbMOXOiHyiJvOHDD20Y7j/8IVArGkxGhtWgHnmk1aI2dGK5YYP9JqZTezrAoN5kWrWyDlr33x+Y5gT1gw+2jmwtWgDnngv8+tfB2xCdTPyHHwKd2L74wrL3bt0C1fiABeU//jHQ297hZOpOkOzTx04UqqvrZ+pA5J3l6t7MJRiRyKu6nRqKhqrfe/ZsjYqKStTUAD5fVb3br4bq+e7o39+em7IK/sMPbX907mwZqdeCk7vqvXVrm5aba9WfS5ZYVWgwP/5o/UcmTAh8L5jRo+2YayzrJ++qrraarp49w6vxatkSePNN+/0966zaI4C6pWMnOYBBvUlt22ZtOg6nXX3ffe0AGzDABq+ZNg049tj633cCWlFR4EBcutTaphsKdm5Opu4E9WOOsX+KH3+snak7l8dFm6k3FNSByIN6QwPPOHr0yERmZjZ27bLAUVJS+/q55cvtH9p921a3ZNzY5cMPrZll6lTL1J98sunW3RQWLLBj6+c/rz393HOBM84Abr01+MBLTz9tTRN1m4/qGjHCjtNo7p1QVcWOkV7wzDN2c6z77gt/gJguXezmQFVVVtsT7FLbwkLL7J0bcKULBvUkcjL1Ll2sQ9zVVzc8vxO4v//eglzPnvbD9/nnoQNVXR072gHsnBQcc4w9b9hgmXp2tlV5Rnsf9Lp3aAvF6cgXpIN6UGvWWNnqXiroZrUVLVFSYjdoKS2t3S16+XIL3HXvrOdwbuzSVD/0e/YA//mPXX4zfrwNf/qHPwRqTLzAqXofPbr2dBHL0quqrDeyW3U18Nhjtj+c2pNQ2rWzmylF065+9dVWS+bVZo/mYOdOG9Pj2GPrnzg2pl8/y9jXrAHOOaf+DYIKC+34cPo+pQsG9SRyB/XrrrNL1xqbPzvbAkF5uXX4AKzndLhBvVMn6wCyZIkFbqdqyWk/6tw50Jbfp0/kmXpxsf3Q1q32r8v5sQ43gLoH1wmlWzcgM3MflJY6mXr9oB6qPR0I3NilqarfP/3U/o4nn2zrfvhhO8m56aamWX+iVVfbmO3uqne3Aw+0H+TXXrN2ccecOdbEFKwJKpjRoy1Ti2T8g88/t856qtb3hNLTn/9sNUFTpwbvTNmY4cPtKqP5863Z091JdvHi9Kt6BxjUk+q446zK3Wm/bkxGhgU2Z3z300+3zBIIv/rd6cBWWGiZePfu9t7J1Lt0CcwbzX3Q6w4RG0qk7ddr1jRc9Q7YtmRktHZl6jv3flZWZk0ODQV1oGlv7PLhh3aS4ty29+CDrXlm+vTA8L7pLFTVu9vkyXYsXHWVdQAFrIPc/vsDY8aEt54zz7Rn517rjamutk56++1nJ1DvvGMdFuNh6VI76V61KjBmAyVGUZEF8wsvDH7JY7jGjwfuvht48UXgttts2saN1tco3Xq+AwzqSXXUUfYjkJsb/nd69bLR4gDLpJ3rMSPJ1AFbb69eFsSzsoIH9T59LEhHMrRrcXHDPd8dBx1kAS3coF5U1PiJS5cuFtQDmXogqC9ebD+wjf2T9utnGV9T3Njlo49sMJY2bQLTbr3VhsH9zW8a7hmeDkJVvbu1aGFBfPVq+2FdvdoGXLrsstodPxvSr58dT+FWwT/yiHUwnTrVmjs6dIhPtr58uQ2xfOyxNkJhu3bWxtujhx13I0fanRQnTwbuvddG1kvEPeHLy21sC+ckyatuuskSnT//OfZlTZliI2/+6U/WnyNdO8kBQJABNymVOYEtM9My7UMPtawu0ky9qspOBDIyLCtyqt/z8gLzOq9XrrQTkHCEm6nn5Fj1azhBvazMTjgay9QzM4Hs7FyUlFiALy0NDPTsZGKNndH372/B/7vvEjuKVGmp9diePLn29NatrWf4uefaZW7pev26U/U+alTDvdcBazu/6CILdMuX2zF52WXhr0vEThyeeMJOxhpqA920yar8TzvNBnQSsb/BH/5gf48jjwx/vW6qlv37fNZxa8cO+3/avNkezusVK+zZCbhvvgnMmBG6n0ekampsjPwZM+xmTY89Fp/lxotqdNXkdS1cCLzyil3h49Q2xkLETvZ++MFG4DzpJJuWbp3kAACqmlaPww47TJuzP/1JFVDt3dvez5+vOmKEakVFeN9fs8a+D6jecYdNO+oo1VNOUW3dWvXaawPzrlhh802fHn75DjpI9Re/CG/es85SHTSo8fmWLrVyvPhi4/P26DFCZ8yALl8O7dev697p55+vesABjX9/5Upb1+9+1/i8sZgzx9Yzb179z2pqVM84QzU3V3XDhsiWu3Wr/Q1/+ik+5YzWxx/b9s2YEd78P/6o2q6dfeeccyJf37x59t233254vl/8QjUnR/XbbwPTdu5U7dhRdeTIyNfrmD7d1v/EE+HNX1qq+uc/23duvTX69db1hz/YMg891J5nzozfsmP1xBP2N/7zn8P/vQqmulr18MNVu3a1/RhPO3aoDh5s++7gg+O77FgAKNAwY2TSg3Skj+Ye1F94wf5qp5wS3fd37gwE9X/+06add55qz5427e67A/NWVKhmZanefHP4y2/XTvWqq8Kb94Yb7Ae2qqrh+d5+28r26acNz1dZWal9+pyrTz4JXbcO2rlzrn7wwQf63nvvae/eVTp2bHjluvJKVRHVTz4Jb/5o3Hijana2allZ8M9XrVJt0UJ1/PjIlvs//2P76sorYy9jLK68UnWffVRLSsL/zuOPW9k/+CDy9e3Zo+rzqV5xReh5/vUvW/5tt9X/7J577LP//CfydRcXq3bqpHr00RZwwlVTo/rrX9t6X3gh8vXW9eSTtqxJk1TLy1UPO8xOViI9MUyEd95RzciwQAyoHnKI6qJF0S3ruedsGc8+G98yOtatU+3eXfXyyxOz/GgwqHvYwoX2V7vssui+X1NjgRRQ/fe/bdo11wQC/VNP1Z6/Xz/Vc88Nb9lVVaF/NIOZNs3mX7Wq4fkefNDm+7//Cz3Pz342QjMyRFu3ztbXXoPu3g098cRWesIJbRWAAhv13nvDK1dJiWqvXqp5eaq7doX3nUjl56sOH97wPE7W9c034S1z5047qWrRQjUz02pakqGqSnW//cI/bhw1Narffx/9es85x36Ma2rqf7Znjx3LBx4Y/G9aUmKB+fTTI1/vZZfZ/v7yy8i/W16uesIJ9jeL5oTCMXeulWHkSNXKSpu2YoVqq1aqp54a2clGvBUWWi3gsGG2n994Q3X//S3IX3ttZCd+paWq3brZ/08it6msLLAfUwGDuodt2mR/tfvui34Z++9vyygqsvf33hsI6m++WXves89WHTgwvOVu2WLL+Pvfw5vfOUEJVWW6Y4fqZ5/Zj3XLlsF/rB233jpFr7suW+seMsuXQ7t27aSA6kcfhVcuVcsWAdXrr294nt/8xrLS3/3O5r3pJqtObehEZds2+0H74x8bLsOWLfajfOGF4ZX5/vsD+zM3V3XMmPC+F2///rdGVPUeL85JYrDg6jRbvftu6O/fd5/Ns3Bh+Ot0juHJkyMvr+Onn1T79FHt0iXwPxmJL7+0v/chh9iJndsTT1j57r8/+vLFoqjITvAOOEB148bA9O3brVYFsFrCOXPCW97tt9t35s9PSHFTFoO6x73zjv1TRGvwYDurd85En38+ENTrVnHfcINlEY1VkataOyVg1WPh+Omn4D84zz0XqKZzHiec0PCyli5dqt26tdLq6tqHzBNPQPPzf6Yi9X/wGnP55RZ86+6Tykqr5haxKt8OHVTbtrVspEUL3Vu9GCqTePNNrVVT0pDrrrO/VWMZ7J49ts9OPtne3323rePjjxtfR7xddVXkVe/x4Jzw3nVX7emrV1t5zjuv4e+XllpgPe208NZXUWH/Sz16xL6ty5fbMTRoUGTH6fr1VjvRrZtVG9dVU2MnxdnZqosXx1bGSG3bpjpggG1XqNqmBQtU+/e3v9svfmF9K0JZv95Ocn/+88SUN5UxqFODTjzRzo4dH30UCJ51M8ynnrLp4VSLfvqpzfvOO+GXpXNn1YkTA+8rK60mYeBAC0wzZ1o1YjhVYYMH99J//7v2IXPRRa108ODHdMCA8Mvk2LHDMoz+/VV377ZpGzaoHn+8beevfx28Tdzp9/DSS8GX+7vfWZDZs6fxMmzYYM0lkyY1PJ+Tpc6da+937bIf+0RXU9ZVXW1/v0ir3uMlP1/1mGMC72tqVEePthOuYEGvLqe2I5z+FH/5i8a1M9r779sJ3OjR4Z1E79ypOnSoZelLloSe76ef7ISvf//QfTjirbxc9aST7GTiww8bnnfPHqu1ys62E+Tp04PXyl18sf0vrF6dkCKnNAZ1atC0aVqrfdnp8Q3UzxIWLNBGqy0dToe2//43/LIMH6563HGB97Nm2TJmzQp/GY677rpDr7iihboPmV69WmuHDst1woTIl6dqQRKwavW5c+0kpFWrQCfDYKqrVYcMsSrVYL18Bw+2ds5wXXGF/ZitXx/886oqay8+9NDaP4b//KeGfdVAvMyf3/AJTaLdfrvVoGzZYu9nztSIqp+dbL2xjqhr19pxcNZZDTcLRerRR6287qtQgqmstPbzzEzV995rfLn/+pftl4Y6EsZLTY3qRRdpRLV2qnaVyzHH2PdOPbV2grFoUeD/sDliUKeIlJXZkbDPPvV/oDZvts+mTm18OU4QWbky/HVPnGgdlByjR1umF00nle+//147d95HKyvtcFm3Dtqpk0+BGn344ciX5y5jRob9KA4apLpsWePfeest2xePP157urM/61YRN2T1avvx/v3vg3/++soSQqsAABFRSURBVOu2zJdfrj29utoCfc+egZqGREtW1bujoED3XtlRUmJV44MHR3YJ1V//qo22244ZY0E9mjbwxjgdV0NdHldTYzU3kVxCp2pNadGeMEfitttsPXfeGfl3q6tVH3nEah9atrTko6LCTvy7dLHas+aIQZ0i1q6d/QDWVVOj2r59w2f4W7aoPvCAZaaAXSsdLqe686efrHo0I8N6fUfriCP663vv2eHy4ovQY489RQHVzz+Pfpnbt1vP3csuC7/6sqbGso6uXWv3tn7lFQ3r8ry6Lr7YfuQ2b66/niOOsPEBglXZOh3+YulYGS6n6j2a68zjWYb99rNxCW68UcOuSncrK1Pdd1+rPg7G6RMR7tUUkXKy8Kys4Jf3OR1bI7nUVNWqxIcNS+xlbk8/bWW79NLYajDWrbMTJ6cjXbAT5OaEQZ0iNnCgXdcazFFHhf6Be+CBwCVyRx0V+fW2TpX9woU2GE647fehTJ36V73ggpaqCr3iihZ60kl/1Zyc8Nqv483pBf6XvwSmXXGFZSGR1kQsX241BXVPeD780Nbx2GOhvzt6tHVWcqqkE8Wpem/K6v5gLr3U2tCzsqzfQzSmTtWgHQ1LS62fxcCBsQ2g0pjt220d7drVHijn5ZetXBdcEF1fiRUr7OQwEZe5zZtn+/z00+Ozb2pqVF97zU7Shg0Lr5+BVzGoU8TuvtsG4AhmwgTLOOtauNAy61GjVL/+Orr1rlplR+GTT9oZebSD6jjWr1+vPl8L3bMHOnBgrh52WIEefnhsy4zFyJHW+ce5WqFfP9tf0Rg7VrVNG+tV7Dj9dMsqG6peX7rU/k7XXBPdeh1Lltg1x1u2BM/Crr7aev9HepVBvDnt6B06RH8is2uX1TqceGLt6U72v2BB7OVszOrV1jSVl2cD3Hzyie3f446LrTnFGeTnr3+NX1mdy+qGDIl/FfmePYkbLyJdMKhTXDmXR7l/rHfssAFaeveO7Z+4qsoy/aFDNW7XNvfqdag+9RS0TZt91OerTOroaosX23bdeqt1dIvlmmFnWX/6U+33oU7G3C6/3LKoSPo7uDk9vZ1Hy5aqfftaxvfrX1vv5X33TW7Vu6OkxIb4jHWUtr//XWuNb/DVV7YPL7005iKG7ZNP7P/jmGOs2rxv39iHAHYuc8vJUf3ii9jLuH69XVIX6rI6ih2DOsWV0xGroCAw7aKLrPNWLKNgOQYNsuV37BifavJRox7Vjh0z9LjjjlQgsrHrE2HcOKsOdjpgFRZGv6xRo2w/lZZau3GbNuGNWbBpk11TH82lZk5V9LhxdixMnWrXz48dq3rkkZbRitg8b7wR+fJT1e7dVkN1/PFWVX3ssbbvm3pcfacDaqdOjY++GK54Xea2Y4eNydDYZXUUGwZ1iqtvvrEjxcl8nLGXb789PssfO9aWd9118Vne/fdv0YyMDD311P9VILze6on07bd2ApSTY50OY2kbdEYwu/JKq1KP5BKfO++0706bFv53nCF6zzuv4X4A5eU2jG88L+9KBc72T5hgz08/nZxyvPGGNaPE0/vv2zZFe5lbRYXdTCozMzA+AiUGgzrF1Z49FkBuu80u1crOtuvL4zU2snMJTLyC71tvqWZm/kYPPniJ5uYmd9xrx8SJto3xqJ4+8URbVosWtYfebExZWeC7F17YeLPJY4/ZvGPGJLZTWCrbvduqlQE75lPhWIqnyZM1qsvcamoCx3QkJ4kUnUiCutj86SM/P18LCgqSXYxm56CD7B7oq1cDQ4YA778PtGsXn2UXFwOFhcDpp8dneV98AQwbZq9POgn48MP4LDcW69cDgwbZvdInTIhtWf/6l90PfNIk4PHHI/tudTVw993A7bfb/elfein4PeanTQMmTgTOPNPui56TE1uZ09n06XZf+08+AQYOTHZp4quiAjjqKLuP+O9+B5SV2WPXrsDrug/3Z7fcAtx5Z7K3wvtEpFBV88Oal0GdwjFqFDBnDnDoocAHHwDt2ye7RKFt2QJ06WKvb7wRuPfe5JbHUVERn+CoCrz6KnDqqUCHDtEtY+FC4Be/ADZutCB//fVARoZ99uyzwCWXACNGALNmAS1axF7mdFdZCWRnJ7sUibFiBXD88fZ/k5MDtGoFtG4d/OH+rF8/4LLLAJFkb4H3RRLUsxJdGPKGs84Cdu8GXnsttQM6AHTqZD9OFRXAEUckuzQB8cp2RYBx42JbxrHHAkuW2I/yjTda9v/ss3bCdsklwCmnAG+8wYDu8GpAB4D+/e3kTtXb29lcMFMnT+rdGygqAtauBQ44INmlSV2qwJNPWtVr69bAtm3ACScAb79tWRkRJV8kmXpGogtDlAzdugH77gv06JHskqQ2EWubLyiwfXXSScBbbzGgE6UrVr+TJ119NbBjB9v7wjVwILB4sb3mPiNKXwzq5Ennn5/sEqQfBnOi9MfqdyIiIo9gUCciIvIIBnUiIiKPYFAnIiLyiIQGdREZKSLfisgqEZkS5PMWIvKy//PPRKRXIstDRETkZQkL6iKSCeBhAGcAGABgvIgMqDPbpQC2qWofAFMBpMiAnkREROknkZn6EQBWqepqVa0AMAPAmDrzjAHwrP/1awBOEeGFNURERNFIZFDvBmCd6/16/7Sg86hqFYAdADomsExERESelRYd5URkkogUiEjBli1bkl0cIiKilJTIoL4BgHvk7e7+aUHnEZEsAG0BFNddkKo+oar5qprfuXPnBBWXiIgovSUyqC8CkCcivUUkB8AFAGbXmWc2gAn+12MBfKjpdts4IiKiFJGwsd9VtUpErgIwF0AmgKdVdamI3AGgQFVnA5gG4DkRWQVgKyzwExERURQSekMXVX0XwLt1pt3mer0HwM8TWQYiIqLmQtKttltEtgBYG8dFdgLwUxyX1xxxH8aO+zB23Iex4f6LXaL2YU9VDatDWdoF9XgTkQJVzU92OdIZ92HsuA9jx30YG+6/2KXCPkyLS9qIiIiocQzqREREHsGgDjyR7AJ4APdh7LgPY8d9GBvuv9glfR82+zZ1IiIir2CmTkRE5BGeC+qx3MNdRG72T/9WREaEu0wvSdD+KxKRr0VkiYgUNM2WJE+0+1BEOorIRyJSKiIP1fnOYf59uEpE/uH1uxkmaB9+7F/mEv+jS9NsTXLEsA9PE5FC//FWKCInu77D47D259Hsw8Qeh6rqmQds5LrvARwIIAfAlwAG1JnntwAe87++AMDL/tcD/PO3ANDbv5zMcJbplUci9p//syIAnZK9fWmwD1sDOA7AFQAeqvOdzwEcBUAAzAFwRrK3NQ334ccA8pO9fWmwDw8F0NX/ehCADTwO47oPE3ocei1Tj+Ue7mMAzFDVclVdA2CVf3nhLNMrErH/mpuo96GqlqnqJwD2uGcWkf0BtFHV/6r9KvwTwM8SuhXJFfd92AzFsg+/UNWN/ulLAbT0Z6Q8DmPch01RaK8F9Vju4R7qu+Es0ysSsf8AQAHM81dDTUpAuVNJLPuwoWWub2SZXpKIfeh4xl/leavHq47jtQ/PA7BYVcvB4zAe+9CRsOMwoWO/E/kdp6ob/G1H74vIClWdn+xCUbPzS/9xmAvgdQAXwbJNCkJEBgK4F8DpyS5LugqxDxN6HHotU4/lHu6hvhvOMr0iEfsPquo8bwYwE96ulo9lHza0zO6NLNNLErEP3cdhCYAXweMw5D4Uke6w/9WLVfV71/w8DkPME+Y+TPhx6LWgHss93GcDuMDfdtQbQB6sU0g4y/SKuO8/EWntPyOFiLSGnbF+0wTbkiyx7MOgVHUTgJ0icpS/qu5iAG/Gv+gpI+77UESyRKST/3U2gDPB4zDoPhSRdgDeATBFVRc6M/M4jH0fNslxmKyehYl6ABgFYCWs1+L/+KfdAeBs/+t9ALwK68j1OYADXd/9H//3voWrV2ewZXr1Ee/9B+s5+qX/sdTr+y8O+7AIwFYApbA2vAH+6fmwf/7vATwE/8BRXn3Eex/CesUXAvjKfxz+Hf6rM7z6iHYfArgFQBmAJa5HFx6Hse/DpjgOOaIcERGRR3it+p2IiKjZYlAnov9v795CrKriOI5/f6NZ6Theyiazi6BZSIWghmGIwihdkAglk15EeogiKSt7KILIsrKHEEm60c0bqCBlgg5eymzEwmlGG+2tQsqM0kybyPTfw/ofZjudmaPDcc64+X9gM3vvs/Zaa5+Z4b/XXufsfwghJyKohxBCCDkRQT2EEELIiQjqIYQQQk5EUA+hCyQd7+b23pE0ukx1nfJHVO6T9Il/p7az8gMlPdyFdiRpq6Qa3+7W96wznilrXIkyqyVd3119CqEcIqiH0AP406g6ZGYPmllLmZprNbMxZnYT6fvcj5QoP5CUjepc3QU0mdmxLhzbEywDFlS6EyGciwjqIZSJpCGS1kn6ypeJvv9WSQ2SGiV9KekG3z9H0seStgJbJE32EeRaSQckrSgke8iOLJVyhb8oqUnSLkm1vn+Eb++VtPAsR8YNeJIKSdWStkja43UUMlK9DIzw0f1iL/uUn2OzpOc7qPsBijxxzEfwi/1OwV5Js3x/laQ3/NzrJW2UNLPI8fMktXjbqzN9f8/ra5Y0w/cvk/S1pG876qekaf772SNpjaRqf2kHUFfqgiuEHqXST+yJJZYLcQGOF9m3kpS8BuBaYL+v1wC9fb0OWOfrc0hPPBvs25NJWZ6uJl1wN2Tq247nYCZlvZvu668Cz/r6BmC2rz9UrI/ZvpPyRa8B7vDt3qTUmgCXk56SJWA4sC9z/DTgLX+tytudVKSdH4D+RdqdAdR7+7XAj8BQ0mM2N3qdVwJHgJlF6v0JuNjXB/rPV4DXM2UG+c/BmXPdDtySfT/9PD8H+vn+p4HnMvXUA2Mr/fcWSyxnu8QVaAjlUweMVlsmxRof9Q0APvD5WQMuyhxTb2a/Z7Z3m9lBAEnfkALqF+3a+YcUSCE9cnKqr99GW37rlcBrHfTzUq97GLCfFLggBemXJE0CTvvrtUWOn+ZLo29Xk5713z7z3mBLSSvaux1YZWangF8kfQaM9/1rzOw0cEjStg763wyskLQeWO/76kjP5gbAzI746n1K6X57ky4cRvvxBRN8307/vfUhXUwVHAauIr3PIfR4EdRDKJ8qYIKZ/Z3dKWkpsM3M7pU0nDRKLDjRro5szuVTFP8fPWlmVqJMZ1rNbIykvsAm0pz6EtLt8iGkkelJSd+Tnm3dnoBFZvZmiXb+lVTlQbqc7gYmAdOBZyTdXKyQUmKhJ4HxZnZE0vv8/3xEurCa3UFblwCtZel1CN0g5tRDKJ/NwKOFDUljfHUAbSkb55zH9neRbm1DZtTaETP7C5gHPKG2tJGHPaBPAa7zon8C/TOHbgLmFuaeJQ2TdEWRJr4jJfRpbwcwS1IvSUNIAXo3sBOY4XPrtaTpiDNIqgKuMbNtpFvlA0h3CurJfOBP0iDStMcJ4A+v784ifdkFTJQ00o/rJ2lU5vVR5DubW8iZCOohdE1fSQczy3xSgBznH9RqIc1rQ5r3XiSpkfN7d+wxYL6kZmAkaX6+U2bWSLodPRtYQer/XlJazQNe5jfS7el9khab2WbS7f0GL7uWM4N+wacUCcykHNPNpMx9W4EFZnYIWEf6jEELsBzYU+QcegHLvd1GYImZHQUWAoO8j03AFDNr8jIHvL8729WFmf1KutBa5e9bA3AjgF8ItHrfQrggRJa2EHLCb6e3mplJup/0obl7Sh13HvszFPjQzKaWLNx2TLWZHZd0GWn0PrFSQVXS48AxM3u3Eu2H0BUxpx5CfowFlvrX4I4CcyvZGTP7WdLbkmrs7L+rvkHpYTh9gBcqPEo+CnxUwfZDOGcxUg8hhBByIubUQwghhJyIoB5CCCHkRAT1EEIIISciqIcQQgg5EUE9hBBCyIkI6iGEEEJO/AdKlW0pXoGHPAAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f35243a9358>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.sched.plot(log_scale=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "As you can see in the plot above it appears the optimal learning rate to set is <span style=\"color:blue\">0.01</span>.\n",
    "### What the Fastai library does:\n",
    "* uses the `Adam` optimizer. In Keras here, we only implemented in `SGD` (list other optimizers in TODO). \n",
    "* tries to find the fastest way to converge to a solution."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<hr style = 'background-color:green; border-width:0; color:green; height:1px; lineheight:0; display: inline-block; text-align: left; width:100%;' />\n",
    "## Data Augmentation on the training data\n",
    "Note that the pretrained Keras ResNet50 model is not quite as accurate as the pretrained PyTorch ResNet50 model.\n",
    "\n",
    "If you try training for more epochs, you’ll notice that we start to <font color=\"red\">**overfit** (the training loss is much lower than the validation loss), which means that our model is learning to recognize the specific images in the training set, rather than generalizing such that we also get good results on the validation set.</font> <font color=\"blue\">One way to fix this is to effectively create more data, through data augmentation.</font> This refers to randomly changing the images in ways that shouldn’t impact their interpretation, such as horizontal flipping, zooming, and rotating.\n",
    "\n",
    "The pictures we saw where learn each different layer has these activations that basically look for anthing from the pictures. Literally what happens is that the last layers of CNN have these things called **activations**. <font color=\"red\">Activation literally is a number that says the features like eyeball of a dog is in a location of the image with certain level of confidence and its probability.</font>\n",
    "<img src=\"images/activations_1.png\" alt=\"\" style=\"width: 70%;\"/>\n",
    "\n",
    "A pretrained network is one where it's already learned to recognize certain things. In the case of the given image above, it's learned to recognize __1.5M images in ImageNet dataset__. \n",
    "\n",
    "### Note: this augmentation won’t doing anything because of precompute\n",
    "In fast.ai, we take __second last layer__, so one which is like got all of the information necessary to figure out what kind of thing it is, and we can <font color=\"blue\">save those activations, so we save the activations for every image, and call it **pre-computed activations**.</font><br/> In other words, we are using a pretrained network. We can take the 2nd last layer and save those activations. There is this level of “dog space” “eyeballs” etc. For efficiency, we use **Precomputed Activations** where we cache the activation values of the last of the Later Layers for each training image instead of recalculating each image each time it is in a minibatch, for about a 10x speedup.\n",
    "\n",
    "<font color=\"green\">When we'd like to create a new classifier which can basically take advantage of these pre-computed activations, we can quickly train, and remember all the details shortly, or train a linear model based on those. </font><br/>\n",
    "\n",
    "<font color=\"red\">When we create a learner, in order to use data augmentation, we have to set `learn.precompute = False` which sets all but the last layer to *frozen*. That means that It's still only updating the weights in the last layer.</font><img src=\"images/aug_1.jpg\" alt=\"\" style=\"width: 60%;\"/>\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "tfms = tfms_from_model(arch, sz, crop_type=CropType.RANDOM, \n",
    "                       rotation_range=10, \n",
    "                       zoom_range=0.1, \n",
    "                       horizontal_flip=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<font color=\"red\">Note that **activation precomputation** is not implemented yet in this Keras library.</font>\n",
    "We can quickly train a simple linear model based on these saved precomputed numbers. So the first time you run a model, it will take some time to calculate and compile. Then afterwards, it will train much faster."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 23000 images belonging to 2 classes.\n",
      "Found 2000 images belonging to 2 classes.\n"
     ]
    }
   ],
   "source": [
    "#need to add random lighten\n",
    "data = ImageClassifierData.from_paths(path=PATH, tfms = tfms)\n",
    "learn = ConvLearner.pretrained(arch, data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<hr style = 'background-color:green; border-width:0; color:green; height:1px; lineheight:0; display: inline-block; text-align: left; width:100%;' />\n",
    "## Learning Rate Annealing\n",
    "Learning Rate Annealing is decreasing the learning rate as you train, hoping to find the nadir of the cost vs iteration function.\n",
    "### Why do we use the learning rate that isn’t the lowest point?\n",
    "Each time we iterate, we will double the learning rate. The purpose of this to find what learning rate is helping use to decrease quickly. The learning rate is going too high.\n",
    "<img src=\"images/lr_2.jpg\" alt=\"\" style=\"width: 40%;\"/>\n",
    "Three simple annealing methods start the learning rate at a constant (as found above) and gradually decrease the learning rate each mini-batch: \n",
    "* <font color=\"blue\">Step-wise Annealing</font> uses discrete decreases, sometimes in a manual or hacky way.\n",
    "* <font color=\"blue\">Linear Annealing</font> decreases learning rate a set amount per iteration.\n",
    "* <font color=\"blue\">Cosine Annealing</font> uses the cosine curve from 0 to 90 degrees to reduce slowly then quickly then slowly again. <br/><br/>\n",
    "A more complicated method would good results is <font color=\"blue\">Stochastic Gradient Decent With Restarts (SGDR)</font> which periodically resets a simple (usually cosine) linear annealing curve every one or more epochs. This hopes to settle in **broad valleys** corresponding to good solutions that aren’t too specific to the specific training images. \n",
    "\n",
    "These advanced SGDR routines work better than the approaches of a **Grid Search** (systematic search for good hyperparameter values) for a learning rate or **Ensemble** (run multiple models and take best result) of different starting points.\n",
    "\n",
    "Compare the different learning rates over time with different methods.\n",
    "\n",
    "|  Simple Annealing Choices |  Annealing With Restarts Choices|  \n",
    "|:-------------------------:|:-------------------------:|\n",
    "|<img src=\"images/lr_4.png\" style=\"width: 100%;\"/>|<img src=\"images/lr_5.png\" style=\"width: 100%;\"/>|\n",
    "\n",
    "#### What is that `cycle_len` parameter? \n",
    "What we've done here is used a technique called stochastic gradient descent with restarts (SGDR)</font>, a variant of learning rate annealing, <font color=\"blue\">which gradually decreases the learning rate as training progresses.</font> This is helpful because as we get closer to the optimal weights, we want to take smaller steps. The number of epochs between resetting the learning rate is set by `cycle_len`, and the number of times this happens is referred to as `the number of cycles`.\n",
    "\n",
    "|  High dimensional Space   |  Cosine Annealing         |  SGDR Annealing           |\n",
    "|:-------------------------:|:-------------------------:|:-------------------------:|\n",
    "|<img src=\"images/lr_6.png\" alt=\"\" style=\"width: 100%;\"/>|<img src=\"images/lr_7.png\" alt=\"\" style=\"width: 100%;\"/>|<img src=\"images/lr_8.png\" alt=\"\" style=\"width: 100%;\"/>|\n",
    "\n",
    "<!--img src=\"images/lr_ann_3.png\" alt=\"\" style=\"width: 30%; float: right;\"/-->\n",
    "\n",
    "#### What do you do when you have more than one minima?\n",
    "Sometimes one minima will be better than others (based on how well it generalizes). So sharply changing the learning rate has the idea that if we suddenly jump up the learning rate, we will get out of “narrow” minimum and find the most “generalized” minimum.\n",
    "\n",
    "<img src=\"images/lr_3.png\" alt=\"\" style=\"width: 40%;\"/>\n",
    "\n",
    "However, we may find ourselves in a part of the weight space that isn't very resilient - that is, <font color=\"blue\">small changes to the weights may result in big changes to the loss.</font> We want to encourage our model to find parts of the weight space that are both accurate and stable. Therefore, from time to time we increase the learning rate (this is the 'restarts' in 'SGDR'), which will <font color=\"blue\">force the model to jump to a different part of the weight space if the current area is </font>\"<font color=\"red\">spikey</font>\". Here's a picture of how that might look if we reset the learning rates 3 times (in this paper they call it a `cyclic LR schedule`):\n",
    "<img src=\"images/sgdr.png\" alt=\"\" style=\"width: 80%;\"/><br/>\n",
    "<center>(From the paper [Snapshot Ensembles](https://arxiv.org/abs/1704.00109)).</center>\n",
    "As the figure shown here, where it is in high dimensional space, we got lots of different flat points that are actual local minima."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "360/360 [==============================] - 264s 734ms/step - loss: 0.0840 - acc: 0.9658 - val_loss: 0.0664 - val_acc: 0.9800\n"
     ]
    }
   ],
   "source": [
    "learn.fit(.005, 1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Evaluate function may run a little slow.  I find that Keras generator's are not the fastest compared to the ones in the fastai PyTorch library.  Generally I try to load the entire dataset into memory, but since this repository is meant to be shared I opted for generators so that it works for a wider audience."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Fit for 3 epochs with a Cyclical Learning Rate"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epochs: 7\n",
      "Epoch 1/7\n",
      "360/360 [==============================] - 265s 736ms/step - loss: 0.0601 - acc: 0.9766 - val_loss: 0.0791 - val_acc: 0.9740\n",
      "Epoch 2/7\n",
      "360/360 [==============================] - 259s 719ms/step - loss: 0.0581 - acc: 0.9789 - val_loss: 0.0767 - val_acc: 0.9780\n",
      "Epoch 3/7\n",
      "213/360 [================>.............] - ETA: 1:37 - loss: 0.0480 - acc: 0.9824"
     ]
    }
   ],
   "source": [
    "learn.fit(.01, 3, cycle_len=1, cycle_mult=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYwAAAEKCAYAAAAB0GKPAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\ngg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7feb38159be0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.sched.plot_lr()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "learn.save('keras_resnet50.hdf5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Fine-tuning and differential learning rate annealing\n",
    "Now that we have a good final layer trained, we can try fine-tuning the other layers. To tell the learner that we want to unfreeze the remaining layers, just call (surprise surprise!) `unfreeze()`.\n",
    "#### In general you can only freeze layer from ‘n’ and on\n",
    "Note that the other layers have already been trained to recognize ImageNet photos (whereas our final layers where randomly initialized), so we want to be careful of not destroying the carefully tuned weights that are already there.\n",
    "\n",
    "Generally speaking, the earlier layers (as we’ve seen) have more **general-purpose features**. <font color=\"blue\">Therefore we would expect them to need less fine-tuning for new datasets.</font> For this reason we will use different learning rates for different layers: the first few layers will be at `1e-4`, the middle layers at `1e-3`, and our `Fully Connected (FC)` layers we’ll leave at `1e-2` as before. We refer to this as differential learning rates, although there’s no standard name for this technique in the literature that we’re aware of.\n",
    "#### Specifying learning rates\n",
    "We are going to specify ‘differential learning rates’ for different layers. We are grouping the blocks (ResNet blocks) in different areas and assigning different learning rates.\n",
    "\n",
    "<font color=\"red\">Reminder: we unfroze the layers and now we are retraining the whole set. The learning rate is smaller for early layers and making them larger for the ones farther away.</font>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "learn.unfreeze()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/7\n",
      "360/360 [==============================] - 223s - loss: 0.1533 - acc: 0.9847 - val_loss: 0.1801 - val_acc: 0.9855\n",
      "Epoch 2/7\n",
      "360/360 [==============================] - 219s - loss: 0.1379 - acc: 0.9854 - val_loss: 0.2030 - val_acc: 0.9800\n",
      "Epoch 3/7\n",
      "360/360 [==============================] - 220s - loss: 0.0975 - acc: 0.9893 - val_loss: 0.1973 - val_acc: 0.9795\n",
      "Epoch 4/7\n",
      "360/360 [==============================] - 219s - loss: 0.0953 - acc: 0.9892 - val_loss: 0.2123 - val_acc: 0.9790\n",
      "Epoch 5/7\n",
      "360/360 [==============================] - 220s - loss: 0.0722 - acc: 0.9912 - val_loss: 0.1732 - val_acc: 0.9790\n",
      "Epoch 6/7\n",
      "360/360 [==============================] - 219s - loss: 0.0722 - acc: 0.9916 - val_loss: 0.1342 - val_acc: 0.9845\n",
      "Epoch 7/7\n",
      "360/360 [==============================] - 220s - loss: 0.0614 - acc: 0.9921 - val_loss: 0.1560 - val_acc: 0.9850\n"
     ]
    }
   ],
   "source": [
    "learn.fit([0.0001, .001, .01], 3, cycle_len=1, cycle_mult=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false,
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAZIAAAEKCAYAAAA4t9PUAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\niriQmJlZERcSMzMr4kJiZmZFfgJIzZ7EZ7oq7QAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7fea98619b00>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.sched.plot_lr()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Test Time Augmentation (TTA)\n",
    "\n",
    "In Natural Language Processing (NLP), we can try to replace words with synonyms. But the appropriate data augmentation for non-image domain is still under-researched/under-developed. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 19 s, sys: 284 ms, total: 19.3 s\n",
      "Wall time: 6.77 s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[0.17039716272840053, 0.98450000000000004]"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "learn.model.evaluate_generator(data[1], steps=math.ceil(data[1].samples/data[1].batch_size), workers=4 )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "learn.save('keras_resnet50_2.hdf5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### ResNet to ResNext \n",
    "See video at [1:51:07](https://youtu.be/JNxcznsrRb8?t=1h51m7s).<br/>\n",
    "Since we changed our model from `ResNet` to `ResNext`, we have larger architecture which has more parameters, it wil be overfitting quickly. Thus, here we don't use `cycle_mult` to double the cycle length."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Train with Larger Images\n",
    "Most ImageNet models are trained on either 224x224 or 299x299 images.\n",
    "If we pre-train a few epochs on our model with small size images like 224x224 at first, which is sort of overfit, then pass in images with larger size like 299x299 to train a few more epochs, it will make our model not overfit by chaning the size of images. To the model, the images are totally different although conceptually the model picks the same kinds of pictures or things. This trick training on small images for a few epochs and then switching to bigger images and then continuing training is an amazingly effectie way to avoid overfitting.\n",
    "\n",
    "For more information:\n",
    "* [Progressive Growing of GANs for Improved Quality, Stability, and Variation](https://arxiv.org/abs/1710.10196)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [tf-py36]",
   "language": "python",
   "name": "Python [tf-py36]"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}