7_multiple_gpus.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Train a ConvNet using multiple GPUs\n",
    "\n",
    "This notebook shows how to train a deep model using multiple GPUs. In this notebook, we use CIFAR-10 dataset and train a VGG-like deep model using 2 GPUs.\n",
    "\n",
    "First, **please make sure that you are running this notebook with an environment that has at least 2 GPUs**.\n",
    "\n",
    "Let's import necessary packages:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import chainer\n",
    "from chainer import cuda, Function, gradient_check, report, training, utils, Variable\n",
    "from chainer import datasets, iterators, optimizers, serializers\n",
    "from chainer import Link, Chain, ChainList\n",
    "import chainer.functions as F\n",
    "import chainer.links as L\n",
    "from chainer.training import extensions\n",
    "from chainer.datasets import get_cifar10"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Dataset preparation\n",
    "\n",
    "Chainer provides a utility to retrieve and handle the CIFAR-10/100 dataset. In this notebook, we use the dataset class provided by Chainer to use CIFAR-10."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "batchsize = 128\n",
    "\n",
    "train, test = get_cifar10()\n",
    "train_iter = iterators.MultiprocessIterator(train, batchsize)\n",
    "test_iter = iterators.MultiprocessIterator(test, batchsize, repeat=False, shuffle=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Model definition\n",
    "\n",
    "Let's define a deep model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "class VGG(chainer.ChainList):\n",
    "\n",
    "    def __init__(self, n_class):\n",
    "        super(VGG, self).__init__(\n",
    "            ConvBlock(64),\n",
    "            ConvBlock(64, True),\n",
    "            ConvBlock(128),\n",
    "            ConvBlock(128, True),\n",
    "            ConvBlock(256),\n",
    "            ConvBlock(256),\n",
    "            ConvBlock(256),\n",
    "            ConvBlock(256, True),\n",
    "            LinearBlock(),\n",
    "            LinearBlock(),\n",
    "            L.Linear(None, n_class)\n",
    "        )\n",
    "\n",
    "    def __call__(self, x):\n",
    "        for f in self.children():\n",
    "            x = f(x)\n",
    "        return x\n",
    "    \n",
    "class ConvBlock(chainer.Chain):\n",
    "\n",
    "    def __init__(self, n_ch, pool_drop=False):\n",
    "        w = chainer.initializers.HeNormal()\n",
    "        super(ConvBlock, self).__init__()\n",
    "        with self.init_scope():\n",
    "            self.conv = L.Convolution2D(None, n_ch, 3, 1, 1, nobias=True, initialW=w)\n",
    "            self.bn = L.BatchNormalization(n_ch)\n",
    "        self.pool_drop = pool_drop\n",
    "\n",
    "    def __call__(self, x):\n",
    "        h = F.relu(self.bn(self.conv(x)))\n",
    "        if self.pool_drop:\n",
    "            h = F.max_pooling_2d(h, 2, 2)\n",
    "            h = F.dropout(h, ratio=0.25)\n",
    "        return h\n",
    "\n",
    "class LinearBlock(chainer.Chain):\n",
    "\n",
    "    def __init__(self):\n",
    "        w = chainer.initializers.HeNormal()\n",
    "        super(LinearBlock, self).__init__()\n",
    "        with self.init_scope():\n",
    "            self.fc = L.Linear(None, 1024, initialW=w)\n",
    "\n",
    "    def __call__(self, x):\n",
    "        return F.dropout(F.relu(self.fc(x)), ratio=0.5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Then instantiate the model and setup a optimizer."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "model = L.Classifier(VGG(n_class=10))\n",
    "optimizer = optimizers.Adam()\n",
    "optimizer.setup(model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Create ParallelUpdater\n",
    "\n",
    "To use multiple GPUs for training, use `ParallelUpdater` instead of `StarndardUpdater` for the updater given to a trainer. So what you need to do is just replacing `StandardUpdater` with `ParallelUpdater`. That's all!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch       main/loss   val/main/loss  main/accuracy  val/main/accuracy  elapsed_time\n",
      "\u001b[J1           2.06062     1.72784        0.248561       0.345728           27.6125       \n",
      "\u001b[J2           1.51435     1.37587        0.431706       0.499407           53.2946       \n",
      "\u001b[J3           1.22403     1.08813        0.555529       0.636175           78.8932       \n",
      "\u001b[J4           1.01321     0.991406       0.639666       0.684731           106.35        \n",
      "\u001b[J5           0.901679    0.875154       0.685582       0.737144           136.299       \n",
      "\u001b[J6           0.782395    0.828507       0.732292       0.741792           165.523       \n",
      "\u001b[J7           0.710225    0.84949        0.761949       0.747033           194.874       \n",
      "\u001b[J8           0.642385    0.633132       0.790144       0.812104           224.745       \n",
      "\u001b[J9           0.574841    0.63775        0.809303       0.812698           254.329       \n",
      "\u001b[J10          0.510813    0.638685       0.831562       0.80449            283.289       \n",
      "\u001b[J11          0.472714    0.576256       0.847075       0.809929           312.512       \n",
      "\u001b[J12          0.443705    0.486639       0.857297       0.846816           341.53        \n",
      "\u001b[J13          0.392901    0.50913        0.87448        0.834454           370.354       \n",
      "\u001b[J14          0.351432    0.482847       0.8875         0.842662           398.908       \n",
      "\u001b[J15          0.33375     0.446264       0.890066       0.855815           428.561       \n",
      "\u001b[J16          0.3024      0.455186       0.901963       0.851859           457.791       \n",
      "\u001b[J17          0.266332    0.451124       0.912084       0.862441           487.5         \n",
      "\u001b[J18          0.253685    0.501251       0.915601       0.842662           516.447       \n",
      "\u001b[J19          0.2364      0.432909       0.924279       0.862638           545.575       \n",
      "\u001b[J20          0.208632    0.506819       0.933824       0.862638           575.64        \n",
      "\u001b[J21          0.194269    0.43255        0.93762        0.870451           604.736       \n",
      "\u001b[J22          0.170368    0.466          0.945393       0.866891           633.722       \n",
      "\u001b[J23          0.166821    0.529847       0.946811       0.862441           662.984       \n",
      "\u001b[J24          0.151152    0.463431       0.951082       0.874407           692.154       \n",
      "\u001b[J25          0.141251    0.459501       0.954963       0.872033           720.543       \n",
      "\u001b[J26          0.128121    0.573401       0.960038       0.858089           749.405       \n",
      "\u001b[J27          0.130896    0.456771       0.959095       0.872132           778.601       \n",
      "\u001b[J28          0.117692    0.522442       0.963035       0.870748           807.629       \n",
      "\u001b[J29          0.113074    0.490545       0.963875       0.872923           837.338       \n",
      "\u001b[J30          0.106436    0.453312       0.965705       0.88123            866.936       \n",
      "\u001b[J31          0.0936333   0.583216       0.96923        0.86966            897.011       \n",
      "\u001b[J32          0.0980956   0.519949       0.969511       0.870253           926.474       \n",
      "\u001b[J33          0.0901761   0.511225       0.971427       0.87945            955.517       \n",
      "\u001b[J34          0.087176    0.518954       0.972506       0.878263           985.729       \n",
      "\u001b[J35          0.0817941   0.573686       0.974119       0.877967           1015.35       \n",
      "\u001b[J36          0.0820182   0.511111       0.974465       0.878461           1044.64       \n",
      "\u001b[J37          0.07492     0.585388       0.977502       0.882812           1073.62       \n",
      "\u001b[J38          0.0726181   0.539151       0.977364       0.879846           1102.61       \n",
      "\u001b[J39          0.0736187   0.49998        0.977701       0.881527           1131.62       \n",
      "\u001b[J40          0.0609967   0.594223       0.98105        0.879747           1160.68       \n",
      "\u001b[J41          0.0701719   0.537505       0.978141       0.881329           1191.11       \n",
      "\u001b[J42          0.0657121   0.536152       0.980059       0.883999           1219.69       \n",
      "\u001b[J43          0.0588336   0.571397       0.98105        0.884494           1248.67       \n",
      "\u001b[J44          0.0629678   0.6295         0.981418       0.87589            1277.73       \n",
      "\u001b[J45          0.0551307   0.625942       0.982777       0.874802           1306.83       \n",
      "\u001b[J46          0.0584306   0.558424       0.98129        0.881626           1335.72       \n",
      "\u001b[J47          0.0591954   0.557666       0.982417       0.875989           1364.9        \n",
      "\u001b[J48          0.0581805   0.637992       0.982131       0.879055           1393.83       \n",
      "\u001b[J49          0.0556543   0.586167       0.984255       0.879549           1423.11       \n",
      "\u001b[J50          0.0531761   0.571888       0.983696       0.879846           1452.54       \n",
      "\u001b[J51          0.0499855   0.617992       0.984014       0.879648           1481.04       \n",
      "\u001b[J52          0.0532928   0.650794       0.983336       0.880835           1510.73       \n",
      "\u001b[J53          0.055306    0.572665       0.982976       0.882615           1539.85       \n",
      "\u001b[J54          0.0483244   0.622639       0.985857       0.88123            1570.31       \n",
      "\u001b[J55          0.0501615   0.596857       0.985294       0.881725           1600.52       \n",
      "\u001b[J56          0.0472725   0.604612       0.986258       0.882318           1630.87       \n",
      "\u001b[J57          0.0484204   0.626023       0.985214       0.879055           1661.5        \n",
      "\u001b[J58          0.0465541   0.612769       0.986213       0.884098           1690.94       \n",
      "\u001b[J59          0.0433752   0.607433       0.98758        0.877176           1720.28       \n",
      "\u001b[J60          0.0518179   0.569884       0.985534       0.880736           1750.26       \n",
      "\u001b[J61          0.0512358   0.596257       0.985334       0.884691           1779.3        \n",
      "\u001b[J62          0.0445089   0.616687       0.986338       0.879252           1807.78       \n",
      "\u001b[J63          0.0402594   0.728541       0.988171       0.875198           1836.64       \n",
      "\u001b[J64          0.0463596   0.602059       0.987139       0.883208           1865.08       \n",
      "\u001b[J65          0.0379239   0.638789       0.98929        0.879747           1894.15       \n",
      "\u001b[J66          0.0435649   0.628246       0.986973       0.879252           1923.19       \n",
      "\u001b[J67          0.0397172   0.65113        0.988662       0.879055           1951.86       \n",
      "\u001b[J68          0.0437917   0.641258       0.987132       0.880934           1981.48       \n",
      "\u001b[J69          0.0421342   0.605103       0.987652       0.884593           2010.2        \n",
      "\u001b[J70          0.0338352   0.631827       0.990385       0.888252           2038.75       \n",
      "\u001b[J71          0.0374422   0.61708        0.98949        0.881329           2066.91       \n",
      "\u001b[J72          0.0385588   0.617529       0.988141       0.885285           2095.65       \n",
      "\u001b[J73          0.0356159   0.663951       0.98957        0.885581           2126.64       \n",
      "\u001b[J74          0.0368221   0.738188       0.988771       0.879945           2155.04       \n",
      "\u001b[J75          0.0376984   0.657186       0.989784       0.876879           2183.84       \n",
      "\u001b[J76          0.0408007   0.598325       0.987932       0.879055           2212.58       \n",
      "\u001b[J77          0.0378266   0.643631       0.988331       0.880439           2241.51       \n",
      "\u001b[J78          0.0343585   0.684672       0.990024       0.881131           2270.22       \n",
      "\u001b[J79          0.0369227   0.657258       0.98929        0.886076           2299.7        \n",
      "\u001b[J80          0.0313397   0.773905       0.990946       0.876978           2329.74       \n",
      "\u001b[J81          0.0364844   0.67806        0.98941        0.887559           2359.12       \n",
      "\u001b[J82          0.0320268   0.649919       0.990529       0.882714           2388.19       \n",
      "\u001b[J83          0.0374646   0.593741       0.988902       0.886472           2417.71       \n",
      "\u001b[J84          0.0326638   0.688821       0.990569       0.875396           2447.42       \n",
      "\u001b[J85          0.0356872   0.722809       0.98965        0.87589            2477.61       \n",
      "\u001b[J86          0.0337024   0.651198       0.990064       0.882516           2507.77       \n",
      "\u001b[J87          0.0334121   0.609174       0.990249       0.885581           2537.2        \n",
      "\u001b[J88          0.0259913   0.681316       0.991947       0.879846           2565.96       \n",
      "\u001b[J89          0.0262138   0.691931       0.992048       0.883505           2595.56       \n",
      "\u001b[J90          0.0343821   0.695963       0.98969        0.887164           2624.64       \n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\u001b[J91          0.0263309   0.635631       0.992468       0.889636           2653.49       \n",
      "\u001b[J92          0.0341086   0.616935       0.990329       0.879648           2682.73       \n",
      "\u001b[J93          0.0306678   0.775979       0.991648       0.875989           2711.84       \n",
      "\u001b[J94          0.0269641   0.671159       0.991386       0.885186           2742.81       \n",
      "\u001b[J95          0.0318551   0.657127       0.990649       0.88212            2771.69       \n",
      "\u001b[J96          0.0261267   0.649264       0.991546       0.884296           2800.52       \n",
      "\u001b[J97          0.0299986   0.698202       0.991089       0.886867           2829.12       \n",
      "\u001b[J98          0.0338298   0.71409        0.990969       0.879351           2858.17       \n",
      "\u001b[J99          0.0307153   0.65424        0.991587       0.889043           2886.97       \n",
      "\u001b[J100         0.0284084   0.648113       0.992088       0.885384           2915.83       \n"
     ]
    }
   ],
   "source": [
    "devices = {\n",
    "    'main': 0,\n",
    "    'gpu1': 1,\n",
    "}\n",
    "\n",
    "updater = training.ParallelUpdater(\n",
    "    train_iter, optimizer, devices=devices)\n",
    "\n",
    "trainer = training.Trainer(\n",
    "    updater, (100, 'epoch'), out='multi_gpu_result')\n",
    "\n",
    "trainer.extend(extensions.LogReport())\n",
    "trainer.extend(extensions.PrintReport(['epoch', 'main/loss', 'val/main/loss', 'main/accuracy', 'val/main/accuracy', 'elapsed_time']))\n",
    "trainer.extend(extensions.Evaluator(test_iter, model, device=devices['main']), name='val')\n",
    "trainer.extend(extensions.PlotReport(['main/loss', 'val/main/loss'], x_key='epoch', file_name='loss.png'))\n",
    "trainer.extend(extensions.PlotReport(['main/accuracy', 'val/main/accuracy'], x_key='epoch', file_name='accuracy.png'))\n",
    "\n",
    "trainer.run()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAycAAAG0CAYAAAAoxFhtAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from IPython.display import Image\n",
    "Image('mnist_result/loss.png')\n",
    "Image('mnist_result/accuracy.png')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ChainerCV\n",
    "\n",
    "[ChainerCV](https://github.com/chainer/chainercv) is a collection of tools to train and run neural networks for computer vision tasks using Chainer. It provides many kinds of models, their pre-trained weights, data augomentation algorithms, dataset abstraction, evaluation tools, etc.\n",
    "\n",
    "In this notebook, let's use ChainerCV to apply some transformations to the training images for data augmentation."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Requirement already satisfied: chainercv in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/chainercv-0.6.0-py3.6-linux-x86_64.egg\n",
      "Requirement already satisfied: chainer>=2.0 in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainercv)\n",
      "Requirement already satisfied: Pillow in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainercv)\n",
      "Requirement already satisfied: mock in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: six>=1.9.0 in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: numpy>=1.9.0 in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: filelock in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: protobuf>=2.6.0 in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: nose in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: olefile in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from Pillow->chainercv)\n",
      "Requirement already satisfied: pbr>=0.11 in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages (from mock->chainer>=2.0->chainercv)\n",
      "Requirement already satisfied: setuptools in /home/shunta/.pyenv/versions/anaconda3-4.4.0/lib/python3.6/site-packages/setuptools-27.2.0-py3.6.egg (from protobuf>=2.6.0->chainer>=2.0->chainercv)\n"
     ]
    }
   ],
   "source": [
    "%%bash\n",
    "pip install chainercv"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Chainer's `TransformDataset` takes two arguments, the first one is a dataset object and the second one is a transformation function. Each datum in the dataset is transformed by using the transformation function. The function should take an input and perform some transformations using any kind of libraries or just array maniputation to the input ndarray, and then return the resulting transformed input.\n",
    "\n",
    "Let's train the same model with the `TransformDataset` that apply\n",
    "\n",
    "- Random cropping of (28, 28)-sized sub-region\n",
    "- Random LR-flipping\n",
    "- Random color augmentation\n",
    "\n",
    "using `chainercv.transforms.random_crop`, `chainercv.transforms.random_flip`, and `chainercv.transforms.pca_lighting`, respectively."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch       main/loss   val/main/loss  main/accuracy  val/main/accuracy  elapsed_time\n",
      "\u001b[J1           1.98431     1.57125        0.273497       0.420787           27.4356       \n",
      "\u001b[J2           1.49766     1.36127        0.446811       0.519778           54.3845       \n",
      "\u001b[J3           1.22926     1.06972        0.565385       0.645669           83.2746       \n",
      "\u001b[J4           1.04492     1.01885        0.638107       0.653778           110.789       \n",
      "\u001b[J5           0.920094    0.890986       0.689738       0.717662           138.658       \n",
      "\u001b[J6           0.835819    0.95429        0.722035       0.683248           166.987       \n",
      "\u001b[J7           0.751259    0.736902       0.74976        0.772844           194.712       \n",
      "\u001b[J8           0.683263    0.723143       0.77496        0.788964           222.172       \n",
      "\u001b[J9           0.636065    0.643785       0.79176        0.812302           250.009       \n",
      "\u001b[J10          0.576232    0.594172       0.813619       0.815467           277.722       \n",
      "\u001b[J11          0.544391    0.547082       0.823678       0.824763           305.61        \n",
      "\u001b[J12          0.513232    0.515735       0.832721       0.844838           334.097       \n",
      "\u001b[J13          0.483268    0.502613       0.846627       0.848299           361.894       \n",
      "\u001b[J14          0.441759    0.458496       0.859736       0.851661           388.927       \n",
      "\u001b[J15          0.416805    0.448545       0.866129       0.851365           415.767       \n",
      "\u001b[J16          0.391952    0.494758       0.873077       0.835146           442.869       \n",
      "\u001b[J17          0.35353     0.474451       0.884311       0.847706           469.965       \n",
      "\u001b[J18          0.328768    0.418238       0.893582       0.868473           496.852       \n",
      "\u001b[J19          0.316272    0.400021       0.899159       0.872033           524.897       \n",
      "\u001b[J20          0.303331    0.396322       0.900895       0.876879           552.724       \n",
      "\u001b[J21          0.282228    0.378844       0.909127       0.876978           580.15        \n",
      "\u001b[J22          0.25917     0.37692        0.914103       0.879945           607.911       \n",
      "\u001b[J23          0.238414    0.421849       0.921276       0.869858           634.995       \n",
      "\u001b[J24          0.234552    0.391722       0.923838       0.878066           662.711       \n",
      "\u001b[J25          0.21528     0.429548       0.930107       0.881032           691.39        \n",
      "\u001b[J26          0.20277     0.439254       0.934223       0.873022           719.256       \n",
      "\u001b[J27          0.205687    0.417644       0.933494       0.881725           747.172       \n",
      "\u001b[J28          0.181115    0.503979       0.941376       0.862144           774.588       \n",
      "\u001b[J29          0.17157     0.412516       0.943654       0.883208           802.642       \n",
      "\u001b[J30          0.165172    0.449125       0.949119       0.878659           830.812       \n",
      "\u001b[J31          0.160274    0.458656       0.94765        0.886867           859.353       \n",
      "\u001b[J32          0.149734    0.428269       0.951042       0.890724           887.395       \n",
      "\u001b[J33          0.146231    0.443984       0.953285       0.88479            915.214       \n",
      "\u001b[J34          0.133489    0.459568       0.956202       0.888252           944.388       \n",
      "\u001b[J35          0.135224    0.522137       0.957171       0.859474           972.041       \n",
      "\u001b[J36          0.134893    0.43196        0.956921       0.887362           999.282       \n",
      "\u001b[J37          0.118279    0.462323       0.961037       0.883703           1026.7        \n",
      "\u001b[J38          0.115384    0.472403       0.961939       0.888548           1054.32       \n",
      "\u001b[J39          0.112095    0.467267       0.964154       0.888647           1081.62       \n",
      "\u001b[J40          0.113962    0.457077       0.964183       0.885384           1109.56       \n",
      "\u001b[J41          0.110753    0.505515       0.965393       0.889537           1138.57       \n",
      "\u001b[J42          0.103556    0.498732       0.966352       0.887955           1168.16       \n",
      "\u001b[J43          0.0962798   0.498892       0.969111       0.890724           1194.84       \n",
      "\u001b[J44          0.0960623   0.472915       0.970189       0.893592           1222.28       \n",
      "\u001b[J45          0.102829    0.51535        0.96859        0.885977           1249.02       \n",
      "\u001b[J46          0.0906494   0.476635       0.970032       0.890724           1276.07       \n",
      "\u001b[J47          0.0943576   0.490182       0.970428       0.890823           1302.99       \n",
      "\u001b[J48          0.0864711   0.51007        0.971394       0.892405           1330.72       \n",
      "\u001b[J49          0.0802152   0.457499       0.974265       0.891317           1358.93       \n",
      "\u001b[J50          0.0816278   0.509579       0.974744       0.8929             1386.67       \n",
      "\u001b[J51          0.0871677   0.485501       0.971595       0.891021           1414.12       \n",
      "\u001b[J52          0.0874276   0.523145       0.973186       0.889537           1441.15       \n",
      "\u001b[J53          0.0777248   0.502344       0.975264       0.891614           1468.35       \n",
      "\u001b[J54          0.0758293   0.53078        0.976603       0.887362           1496.67       \n",
      "\u001b[J55          0.0687765   0.519517       0.977342       0.892801           1523.47       \n",
      "\u001b[J56          0.0739846   0.541304       0.9752         0.895471           1551.22       \n",
      "\u001b[J57          0.0765782   0.515796       0.975863       0.895273           1578.83       \n",
      "\u001b[J58          0.0709365   0.537792       0.978101       0.884197           1606.45       \n",
      "\u001b[J59          0.0685968   0.522859       0.977845       0.890823           1634.61       \n",
      "\u001b[J60          0.0711584   0.566463       0.977022       0.895075           1662.33       \n",
      "\u001b[J61          0.0685315   0.543004       0.97898        0.889438           1691.69       \n",
      "\u001b[J62          0.0657762   0.571561       0.978846       0.893888           1718.57       \n",
      "\u001b[J63          0.062869    0.581711       0.980579       0.889834           1746.83       \n",
      "\u001b[J64          0.0678531   0.539453       0.978766       0.894877           1774.47       \n",
      "\u001b[J65          0.0686105   0.530505       0.979779       0.893394           1802.22       \n",
      "\u001b[J66          0.0655598   0.47872        0.979979       0.895273           1830.27       \n",
      "\u001b[J67          0.061014    0.555248       0.981851       0.893691           1857.79       \n",
      "\u001b[J68          0.0654965   0.572162       0.97958        0.896262           1886.21       \n",
      "\u001b[J69          0.0628353   0.543463       0.981658       0.891119           1912.95       \n",
      "\u001b[J70          0.0624894   0.586859       0.980769       0.88568            1940.62       \n",
      "\u001b[J71          0.058222    0.558202       0.982936       0.892108           1968.19       \n",
      "\u001b[J72          0.062677    0.597039       0.980649       0.892009           1995.34       \n",
      "\u001b[J73          0.0616648   0.596771       0.981578       0.887559           2022.19       \n",
      "\u001b[J74          0.0543586   0.552333       0.983056       0.893394           2049.87       \n",
      "\u001b[J75          0.0604921   0.576755       0.980649       0.892108           2077.87       \n",
      "\u001b[J76          0.053782    0.588882       0.983656       0.888252           2105.94       \n",
      "\u001b[J77          0.0544463   0.582021       0.982737       0.890724           2133.67       \n",
      "\u001b[J78          0.055034    0.588683       0.982412       0.893888           2161.91       \n",
      "\u001b[J79          0.0508388   0.600886       0.983176       0.899031           2190.12       \n",
      "\u001b[J80          0.0524239   0.597491       0.983614       0.891119           2217.18       \n",
      "\u001b[J81          0.0568873   0.613139       0.982537       0.896262           2244.21       \n",
      "\u001b[J82          0.0507426   0.603448       0.983935       0.890427           2271.56       \n",
      "\u001b[J83          0.0531336   0.583005       0.984255       0.896064           2299.25       \n",
      "\u001b[J84          0.0498179   0.559626       0.984015       0.895669           2326.75       \n",
      "\u001b[J85          0.0474439   0.558293       0.986173       0.89646            2355.46       \n",
      "\u001b[J86          0.0470474   0.615457       0.985978       0.893295           2382.93       \n",
      "\u001b[J87          0.053192    0.617062       0.984255       0.893888           2410.82       \n",
      "\u001b[J88          0.048872    0.686229       0.984335       0.886768           2438.42       \n",
      "\u001b[J89          0.0521594   0.702471       0.984375       0.892306           2465.61       \n",
      "\u001b[J90          0.0481439   0.598124       0.984974       0.890032           2494.1        \n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\u001b[J91          0.0514463   0.611787       0.984816       0.887362           2521.67       \n",
      "\u001b[J92          0.0474498   0.632413       0.985374       0.8929             2549.76       \n",
      "\u001b[J93          0.0378744   0.709121       0.987572       0.880142           2577.05       \n",
      "\u001b[J94          0.0438717   0.621942       0.986378       0.889142           2604.65       \n",
      "\u001b[J95          0.0428542   0.655592       0.987012       0.89735            2632.81       \n",
      "\u001b[J96          0.0460428   0.54094        0.987179       0.898734           2659.57       \n",
      "\u001b[J97          0.0507341   0.587891       0.984495       0.893888           2687.69       \n",
      "\u001b[J98          0.0407599   0.594971       0.987012       0.899624           2715.2        \n",
      "\u001b[J99          0.0460981   0.658251       0.986418       0.896361           2742.91       \n",
      "\u001b[J100         0.0444968   0.602549       0.986253       0.898536           2770.95       \n"
     ]
    }
   ],
   "source": [
    "from chainercv import transforms\n",
    "\n",
    "def transform(inputs):\n",
    "    img, label = inputs\n",
    "    img = transforms.random_crop(img, (28, 28))\n",
    "    img = transforms.random_flip(img)\n",
    "    img = transforms.pca_lighting(img, 0.1)\n",
    "    return img, label\n",
    "    \n",
    "batchsize = 128\n",
    "\n",
    "train, test = get_cifar10()\n",
    "\n",
    "train = datasets.TransformDataset(train, transform)\n",
    "\n",
    "train_iter = iterators.MultiprocessIterator(train, batchsize)\n",
    "test_iter = iterators.SerialIterator(test, batchsize, repeat=False, shuffle=False)\n",
    "\n",
    "model = L.Classifier(VGG(n_class=10))\n",
    "optimizer = optimizers.Adam()\n",
    "optimizer.setup(model)\n",
    "\n",
    "devices = {\n",
    "    'main': 0,\n",
    "    'gpu1': 1,\n",
    "}\n",
    "updater = training.ParallelUpdater(\n",
    "    train_iter, optimizer, devices=devices)\n",
    "\n",
    "trainer = training.Trainer(\n",
    "    updater, (100, 'epoch'), out='multi_gpu_result_2')\n",
    "\n",
    "trainer.extend(extensions.LogReport())\n",
    "trainer.extend(extensions.PrintReport(['epoch', 'main/loss', 'val/main/loss', 'main/accuracy', 'val/main/accuracy', 'elapsed_time']))\n",
    "trainer.extend(extensions.Evaluator(test_iter, model, device=devices['main']), name='val')\n",
    "trainer.extend(extensions.PlotReport(['main/loss', 'val/main/loss'], x_key='epoch', file_name='loss.png'))\n",
    "trainer.extend(extensions.PlotReport(['main/accuracy', 'val/main/accuracy'], x_key='epoch', file_name='accuracy.png'))\n",
    "\n",
    "trainer.run()"
   ]
  }
 ],
 "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": "3.6.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}