{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Lesson 3 Notes"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Minimum code to run image recognition\n",
    "This assumes we have already set up our directories and found the learning rate using the learning rate finder `lr_find()`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from fastai.conv_learner import *\n",
    "PATH = \"data/dogscats/\"\n",
    "size = 224\n",
    "batch_size = 64"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Downloading: \"https://download.pytorch.org/models/resnet50-19c8e357.pth\" to /home/paperspace/.torch/models/resnet50-19c8e357.pth\n",
      "100%|██████████| 102502400/102502400 [00:02<00:00, 48204227.51it/s]\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "2b36284262b94ac6aedaafa6a62c78d9",
       "version_major": 2,
       "version_minor": 0
      },
      "text/html": [
       "<p>Failed to display Jupyter Widget of type <code>HBox</code>.</p>\n",
       "<p>\n",
       "  If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean\n",
       "  that the widgets JavaScript is still loading. If this message persists, it\n",
       "  likely means that the widgets JavaScript library is either not installed or\n",
       "  not enabled. See the <a href=\"https://ipywidgets.readthedocs.io/en/stable/user_install.html\">Jupyter\n",
       "  Widgets Documentation</a> for setup instructions.\n",
       "</p>\n",
       "<p>\n",
       "  If you're reading this message in another frontend (for example, a static\n",
       "  rendering on GitHub or <a href=\"https://nbviewer.jupyter.org/\">NBViewer</a>),\n",
       "  it may mean that your frontend doesn't currently support widgets.\n",
       "</p>\n"
      ],
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Epoch', max=3), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[0.      0.04423 0.02491 0.98877]                             \n",
      "[1.      0.0434  0.02632 0.98975]                             \n",
      "[2.      0.03641 0.02626 0.98877]                             \n",
      "\n",
      "CPU times: user 15min 19s, sys: 2min 43s, total: 18min 2s\n",
      "Wall time: 8min 17s\n"
     ]
    }
   ],
   "source": [
    "# Set up our transforms, data, and learner\n",
    "# Do our initial learning for our final layer\n",
    "# Note we do not call precompute=true. This is just an optimization that lets us some weights that have been computed before. We can always leave it out though.\n",
    "\n",
    "transforms = tfms_from_model(resnet50, size, aug_tfms=transforms_side_on, max_zoom=1.1)\n",
    "data = ImageClassifierData.from_paths(PATH, tfms=transforms, bs=batch_size)\n",
    "learn = ConvLearner.pretrained(resnet50, data)\n",
    "%time learn.fit(0.01, 3, cycle_len=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "78000de63ce64bef96778dee9241cbe3",
       "version_major": 2,
       "version_minor": 0
      },
      "text/html": [
       "<p>Failed to display Jupyter Widget of type <code>HBox</code>.</p>\n",
       "<p>\n",
       "  If you're reading this message in the Jupyter Notebook or JupyterLab Notebook, it may mean\n",
       "  that the widgets JavaScript is still loading. If this message persists, it\n",
       "  likely means that the widgets JavaScript library is either not installed or\n",
       "  not enabled. See the <a href=\"https://ipywidgets.readthedocs.io/en/stable/user_install.html\">Jupyter\n",
       "  Widgets Documentation</a> for setup instructions.\n",
       "</p>\n",
       "<p>\n",
       "  If you're reading this message in another frontend (for example, a static\n",
       "  rendering on GitHub or <a href=\"https://nbviewer.jupyter.org/\">NBViewer</a>),\n",
       "  it may mean that your frontend doesn't currently support widgets.\n",
       "</p>\n"
      ],
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Epoch', max=1), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[0.      0.02143 0.02226 0.98975]                             \n",
      "\n",
      "CPU times: user 9min 35s, sys: 1min 41s, total: 11min 16s\n",
      "Wall time: 7min 55s\n"
     ]
    }
   ],
   "source": [
    "# Unfreeze the other layers and learn again using differential learning rates\n",
    "# (Lower rates for more general layers)\n",
    "learn.unfreeze()\n",
    "\n",
    "# This will be explained later. General guideline: If using a model > 34, and similar data to image-net, use this.\n",
    "# It causes batch normalization moving averages to not be updated\n",
    "learn.bn_freeze(True)\n",
    "\n",
    "# Do our learning again with differential learning rates\n",
    "% time learn.fit([1e-5,1e-4,1e-2], 1, cycle_len=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 1min 55s, sys: 20.8 s, total: 2min 16s\n",
      "Wall time: 1min\n"
     ]
    }
   ],
   "source": [
    "# Use test time augmentation to get our predictions\n",
    "# Recall test time augmentation means transforming our test images to increase\n",
    "# the chance that they are identified correctly\n",
    "%time log_predictions,y = learn.TTA()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([0, 0, 0, ..., 1, 1, 1])"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# y is an array of len 2000\n",
    "y"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(5, 2000, 2)"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 2000 images, can be in one of two classes, and we take a prediction for 1 image + 4 augmentations\n",
    "log_predictions.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[[0.99999, 0.00001],\n",
       "        [0.9999 , 0.0001 ],\n",
       "        [1.     , 0.     ],\n",
       "        ...,\n",
       "        [0.00016, 0.99984],\n",
       "        [0.0002 , 0.9998 ],\n",
       "        [0.00008, 0.99992]],\n",
       "\n",
       "       [[0.99994, 0.00006],\n",
       "        [0.99993, 0.00007],\n",
       "        [0.99999, 0.00001],\n",
       "        ...,\n",
       "        [0.00003, 0.99997],\n",
       "        [0.0001 , 0.9999 ],\n",
       "        [0.00012, 0.99988]],\n",
       "\n",
       "       [[0.99999, 0.00001],\n",
       "        [0.99994, 0.00006],\n",
       "        [0.99998, 0.00002],\n",
       "        ...,\n",
       "        [0.00025, 0.99975],\n",
       "        [0.00003, 0.99997],\n",
       "        [0.00005, 0.99995]],\n",
       "\n",
       "       [[0.99999, 0.00001],\n",
       "        [0.99962, 0.00038],\n",
       "        [0.99999, 0.00001],\n",
       "        ...,\n",
       "        [0.00015, 0.99985],\n",
       "        [0.00017, 0.99983],\n",
       "        [0.00003, 0.99997]],\n",
       "\n",
       "       [[0.99995, 0.00005],\n",
       "        [0.99834, 0.00166],\n",
       "        [1.     , 0.     ],\n",
       "        ...,\n",
       "        [0.00004, 0.99996],\n",
       "        [0.0002 , 0.9998 ],\n",
       "        [0.00007, 0.99993]]], dtype=float32)"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.exp(log_predictions)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "??learn.TTA()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.016303548492258166"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Finally get our log loss and accuracy\n",
    "# We convert the predictions from a log scale via np.exp\n",
    "probabilities = np.mean(np.exp(log_predictions), 0)\n",
    "metrics.log_loss(y, probabilities)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.994"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "accuracy(np.mean((log_predictions),0), y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}