1-Write-the-training-loop_ja.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "# 学習ループを書いてみよう\n",
    "\n",
    "ここでは、\n",
    "\n",
    "1. データセットからデータを取り出す\n",
    "2. モデルに入力する\n",
    "3. Optimizerを使ってモデルのパラメータを更新して学習を行うループを回す\n",
    "\n",
    "ことをやってみます。このノートから得られるものは、`Trainer`を使わない学習ループの書き方です。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 1. データセットの準備\n",
    "\n",
    "ここでは、Chainerが用意しているMNISTデータセットを使うための便利なメソッドを利用します。これを使うと、データのダウンロードから、一つ一つのデータを取り出せるようにするところまでが隠蔽されます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "slideshow": {
     "slide_type": "-"
    }
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAWEAAAFfCAYAAACfj30KAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAAPYQAAD2EBqD+naQAAIABJREFUeJztnX2MvVtV3797Zs7L/Ob3+92bQnMx0lT02ihpNBZbSpRy\nW5po+QPxH4w1QWpMgy+NMakSElIQTIwaDI1KYxoLmlYTEm1RA1wr4gsioigqGiXQi6hwryAv9/cy\nZ86cmd0/Zta566xZa+99zpwzzzlnvp9k59nPPm/P85yZ71nP2mutnXLOIIQQ0g07XR8AIYRcZyjC\nhBDSIRRhQgjpEIowIYR0CEWYEEI6hCJMCCEdQhEmhJAOoQgTQkiHUIQJIaRD9ro+gJTS0wB8HYCP\nARh1ezSEELIUhgC+CMCjOee/Lz1xZSKcUvouAP8ZwDMA/DGA/5Rz/n3nqV8H4H+t6jgIIaRDvgXA\nz5WesBJ3RErpmwC8AcBrAHwVzkT40ZTS052nf2wVx0AIIWvAx2pPWJVP+HsB/FTO+Wdzzn8B4BUA\n7gP4Nue5dEEQQraVqr4tXYRTSj0AzwHwLhnLZ6Xafg3A85b9eYQQssmswhJ+OoBdAE+Y8Sdw5h8m\nhBByDkPUCCGkQ1Yhwp8GcALgITP+EIDHV/B5hBCysSxdhHPOxwA+AOCFMpZSSuf771325xFCyCaz\nqjjhHwPwlpTSBwC8H2fREjcAvGVFn0cIIRvJSkQ45/zW85jg1+HMDfFBAF+Xc/7UKj6PEEI2ldT1\nQp8ppX+GM/cFIYRsG8/JOf9h6QmMjiCEkA6hCBNCSIdQhAkhpEMowoQQ0iEUYUII6RCKMCGEdAhF\nmBBCOoQiTAghHUIRJoSQDqEIE0JIh1CECSGkQyjChBDSIRRhQgjpEIowIYR0CEWYEEI6hCJMCCEd\nQhEmhJAOoQgTQkiHUIQJIaRDKMKEENIhFGFCCOkQijAhhHQIRZgQQjqEIkwIIR1CESaEkA6hCBNC\nSIdQhAkhpEMowoQQ0iEUYUII6RCKMCGEdAhFmBBCOoQiTAghHUIRJoSQDqEIE0JIh1CECSGkQyjC\nhBDSIRRhQgjpEIowIYR0yF7XB0DIJpJScrfR2Lzv2/JYzjnc2rHSMeutdx6l/rx4x1U6j5btpkMR\nJhtLJAaXEYlWdnZ2pqKl+95+6zHNI+wpJZyens60nHO4L8e0s7NT7XvnELUW7I+DHFdpGz3mjW06\nFGGycZSssmVZbDV2d3enwmWbfax0/HbMWqVWyPX+yclJU5PjlSbHGI3pY9cC7e17RBaqiOjJycnM\n1htr2Z6cnCCltPEWMUWYbBStVuKqj8ETsr29PVfcvGOK9ksWtW2TyQTHx8eYTCYXmozL+8rxtTQt\nyrWtxoqh3dc/DJPJpNj3tpPJBDs7O5hMJtP3166XTYUiTDaCFtGNfJurOJaaqGlRrvlW9Zi2Nq0Q\n27HxeIzj4+PpVvf1NTg9PcXe3h56vV5Ts8df+qGRz/B8vHZfxFT/cNT6cl4i/FqAd3Z2ppb+JkMR\nJmtNze/r3a6XXrcMxAoUYbMCp/f39vYuHE/tx6PFHSAifHR0hKOjo2lfW9/AmQBrAe33+xfaYDCY\n2bdCXGrWHVDqn5ycuD8Yti/74/F4ehzywyLvJS6Jq/D/r5qli3BK6TUAXmOG/yLn/OxlfxbZbiLf\n6Tx+01WgXREibL1eb0bEpK9FuMWCj/zMXhuNRhiNRuj1ehcEWG7TxfrUPwyDwWDahsPhhX70Y+L1\ntQjXIh4mk8lUXOWHw9s/Ojqafs54PJ7+6Mj7yA9L5JPeNFZlCX8IwAsByH/CZEWfQ7aUmoiWZutX\nLcTWEhZh01albEWoSi4TvdUTY7X+4eHhjNXqWYraf6t/HIbDIYbDIfb39y9s5bjl+Xpr+/oa10LJ\njo+PMRqNpta7NDtmf1S88xJ/Ny3hmEnO+VMrem9yDbG38y1tVYgYyi25vq0Xa1K2/X5/5pijvnZF\nRBELNnpBC7AVK+1/tVa7HN/+/j5u3LiBGzduTPv7+/sz7gnrqrDN+oRLMcvj8XhqvY9GIxweHs7s\ny5hcV+0P1yFqk8mElnADX5pS+lsAIwC/C+BVOee/XtFnkS2jFElQclFclRB7lrC2LrWVKUIVWen2\neO1EWNREhMUC1pNwJycnU9+qiJm12rUIHxwc4ODgYNrXPyDadeGNaRGuJY2Mx2McHh7i/v37ODw8\nvNDkR6Vk2YtrRfvIN51ViPD7ALwcwF8C+AIArwXwWymlf5pzvreCzyPXkJLwrvqf0/MJa7/q/v7+\ntIlQ1dwmWoR1BIKNRoiiEwDMCJX4V+1rtSUsVvDBwQFu3rw5bfIj4ln29jHrEy71j46OcP/+/QtN\n/NAtAiyTdvbHZ5NZugjnnB9Vux9KKb0fwF8BeCmANy/788jyWOQPOnrNPOOtlq8ntFH21yr/Qff2\n9qYia2/l7Zi2FmsWuxX3Un9nZ2cmO07H0npRBvpHQv9QaEtYRFj8w/r5JREGZmN2vQy5nDP6/f6F\nH0gtsjpWWI7di19e9fd71aw8RC3n/PmU0ocBPLzqzyKzlMSs9vzS4zX/5qJ973PsWGuW2iqt4d3d\nXXdCy9sOBoOZcy25JvS5RNlt+lytL1oSG3R6r2CtXRFecUPI8WpXhA1Xs5Z3KR3ZttPT05kfB/0D\nYUPV9JhOQLHntulJGsLKRTildBNnAvyzq/4s8hQlofNEzr4ues+WmgLRc2qvLR2jZy3W+qsU4dpt\nuhazSHS9fulHxY6JL3owGOD4+HhqSUYiLEJsxVc3HbJmRdj+uMn71+pAyNaKcCTINmZYZ9TJ+W1L\nthywmjjhHwXwyzhzQXwhgB8AcAzg55f9WcQnuv2NxqLX27FWN4CXaFDqRyLs9b1JqshvukoRthNU\n0b4XolY619Zrl1KaWqvaUjw5OZkRJ3lv7XYQAdaREdYS1jHBNgLDuiBqxYOklUTYE2VrBesfmG0o\n3COswhJ+JoCfA/A0AJ8C8B4A/zLn/Pcr+CwSULNWS0Jsx2qugFb3QPRY6Rbd9nUWl83osvurFGEd\nc+s1HU+rr2HN6p/n7kE+w96uawGW672/vz8TAaG3VoglycQTYR0Wpt0RthCP1zyfdc094bkjtskK\nBlYzMffNy35PshjRP7MnfPZ13v48hV0890DkPihZ6nZcBFDEodRflQiLK6C1DoN3nVvcL6UxANNw\nMyu++rlahLX1W7KEJfRNC7D3XWlLuKU6Wk18o326I8hG4gmwt9XPr/VbY1hr7gIv5rXmNpEW1Wvw\n2ipFWFvcWvy9Vrqu0bVuIXJByDHq669D0qKoDhFh/X15SSIppQuTblp0bblJ2dei2iLGWoDtxJz9\n0dlkKMJbiCdm1rdYsoYjsbCiOs+2NNYy2adFuJRWqx/b3Z0ttbjM69saxyvHULrbsI/V0n91yJcV\nJH3HIK3X612IX9bNCnHJlaSP1Qpxra6xJ76lSnCeJczoCLJReP7F0j9V7bY5Ktc4b7OWY0mE7WNa\nfG0lMDu+ShG2/m0tXJ7P23sPbz8K8YrCvjwB1pa6zebTYmu30uwdU+TGinzBXt3gki+4JMrWH+y5\nXjYdivCWUhJgO8lS809KX4uu54u1Y3Y8qsZV+4fXfR2WZbd2bFUiDMR1f70ohuj78Whd0scTYOuC\n0CI8GAwuiK3XBoPBzHGXfqRL7gidPFITYSu8tYk5+oTJ2tPiipB/WPv80tYKrFfacJFm3SOeJSb9\nUilG21YpwvYal/q19xC8sC4rcF6Ilr5G2gXR7/engtYS0yxb63LQWztWEmAtxF6h9kiISxNz9gdp\nG6AIbxFaAPQ/ZMuEmH29bHW/JKTaF1sa9/y5kVXpjdl4XCsgenyVItxKznk6kQVg2tdjmmhiy76n\nvJf+QdXWob6rkLhmG0Jns+G8iTf5vCgbLhJab0yXq7QF6UuJGqXoiG2AIrxBtFis4rttCeOy4VNR\nX7aRiyEaK1nJejxyPXiCrP2++lxqPtiusGLrWZYaETt7G+6tIzfPUkFW2G1EwtHR0fQHTo6jJR25\ndGy2jUYj3Lt3b6ZJBTUpZSnCHAkx3RGkMzyRjCayrFBFKz/oMK7abbWIe2toVkn4rWjPMzFnoyOs\nJdfiBrhK5hEJG2VQ8qOWFvn0WlTkp9frTQVYWjQJ6O17roeoPxqN3CpqIsLWMmbGHFlLalbjzs6O\nWwc2qgsLxFlqdsyLipgngy0aK/2o2HOVW2trVVshXhcit0OEtoS9mgradxpZy17fCvB4PA7dSPo4\nvNTkKCzN+zy9lVKWYv3ausKj0Sgs3uP5hbfBCgYowhuDJ0h2Jl58g6XZcFsjIBJcr9XiYUsxs15f\nttHneccik07WurY+zXUR4nmFQoubFky7JNDR0dGMMNX68l6liBXZ1oTXWsLe53pjspSRFl27uoa2\nhL33ZJww6RTP6rUhZ3K7LkIbZUjduHFjWg+2VYi9LKqW/dp49LnecdhstXV3R8xjCZcE2C4DNBqN\nqskRusmP3vHxcTXOuyVEzlrCLckadnkjK77aHaGrwtnP2DYhpghvCCXx1VsRYbGEpXCLraAl2xZX\ngP1sm4xQ2q8lM4gI63OUvjcmIuxZ2Os4MTePAMtW11rQImyXA7Li56UKWxH2fgTtuBf/6wlyKVXZ\nO5bIoo9EOKpHsU2TcgBFeKPQYhhZljo4XyxhXdD71q1b074W4ZbUYRuB4GXfeVl5tcfk3Fq2+py9\n67CJlrAnwNZ/KyKsJ7SiYjle37te0daLV47GrEh6xyDb4+PjmfC0aLl78XeXYqa3RYABivBGEAmi\nZ9HYDCkR4Vu3buHWrVu4ffv2tH/z5k3X11yaGJu33zIm56jPtzQWifkm+oTt45ElLBbj/fv3p+Fd\nJZGyrfUHcWdnJxS9eRJJvP3JZOJmx3n9yWQSukCsn3rToQhvEN4/j01T9XzCYvnevn0bDzzwwHR7\n69atqgiXhFmOqdRanyPPs+cbXYPaD8Qm4AmIZwlbd4QnwpG7QPql79Tue+6GeQS/JNQt9SNkMs4L\nlbNtG6AIbwiRb9YKcMkSvn37Nm7fvo0HH3wQDz74IG7fvl0NebMWqxyL3raO1Z7vnXPpWtj3aXnP\ndcETEO2LtSFq1h1x9+7dYtSCZy22XLOUUtXKLbkISvst2XUyfnp6OnOd5kl42TQowhuEZ6V6VrH1\nC3ti/MADD+CBBx4oug2sC2HTWMY/afQelxH6SIBlq8O7tDtCRPjevXu4e/du0UK8zARWiwh7dRys\n+HqP1WKJpb8tAtsCRXiD8P7RvH8ML67Ss1rkPcQPKEJrLadNx1pOeuv5ZEvvIUTWZGT9t3yO/k69\nyTkbrtZyu74MEW6xfktWuPRL8b7b5mKYB4rwBhFZPFaI7Qy1FWP9Oi3AWpzENyj9TWUekfJueaP9\nFp+5vqYt7zuPCB8eHl74MWk5rxZKLoWS1VuyxEt/j9Ztct2gCG8I9h/Oiqj9A2/14wGYisU2ia9Q\nssw88ZDX6NfbMWC2nnDU19fTHlPUt3cyesLKxtbaY4725xW3kjuhdP0iy1Y/Xku6uI5CTBHeICIh\nTinh5ORkOqlSitn0rA/btkF8gfh6RRadfb59D9lG/nh5nRd2p98j6luxanVHeMfs9ee5bpEQtwpu\nJMI199h1hCK8YXh/2HL761nCNYtYo63gy1hS64QnAN6PU2RJen0RYZ00knOeqV8s19JawqW+tYR1\ntTMbL3x4eOgem90u8t2VhLYmvvr1djzyJW/T39siUIQ3gJJlYf2QrS4J/Xr7ftv0j2AtuSittlVQ\nhL29vem19IrHaxePvI8+Jm9M/1i0WsKl97Nj81yzeZt+nX0Pe352S0uYbBRWMLUQA7ggNCUxlvcB\nLlrB8lmbTmQJe8VhWgVHrrsUu9HXSb4L8dfLMXhbOzavNTyPwC963Wo/SN7fS2lb8iMveqybDkV4\ng4gsYe1WqEVGeO6IyA0hn7mp2H9+66bRMaott8hWhO21slln8wiUPs553BHee3jXYNHrZrfzno/3\n+pqwXzcowhtCJAjWr1uLEbZWoU5T1f7LbfmnsNfKipxeeaLVB+qJrJ6sq8361yzFVneEfV973pe5\nZtGxRmO1/VZB34a/uXmhCG8gnvgKJavXExrtzthGi6QkwFrorEui1LchfV4G4zxWoxWhFmtYW8It\n12AZ13Ge9235cViG1b4NUIQ3CO9WTotxztn1d5Ya8JQVV7sNXxXzhMTNcxw1C1jH4MrknDdxZMVY\n3zHozDgdrlayhqOtd5yRO+Lo6Kj5OpD1hiK8IdjbYBmTrVho+lbWW5lB6g8MBoMLywyVliDSEQAt\nfsB5qnYtej2i45G+1F7wml1qXeoVtAjxzo6/mKrXtytal7YnJyczxxWtNEy2C4rwBmKtsZolJaJz\neHg4XeRTlpovLcJp9z3LOBrTVqHe2rFFRLjl83POoQXp7UeTc5El7K1k7W31atK12sqnp6euCHsR\nHGR7oAhvENrK1P+I4tcVl8Lp6alrCR8eHl4Qh9py9Z4Il8KMpNkym9qqzjlPLcRFq7N57hJ7HPYH\nSN8N2EUmpXxizScsIiyLjerVir2+XYS01BcRPjo6mqmt600cku2BIryBRP+EdmbdirC9TdYiXFuF\nVy8AWQq4l619j16vN42tBbCwAEfC6x2DXhpIVqXQS63rJdfH43H1B8azhEut3+9fWIy0tCI1gKol\nTHfE9kER3jBqVlAkwiLEvV5vaoEBmFk+XguIt1+LuND72jdqi7XYWgvzuiTshFt0PNYSllUp9BJB\n0hcRbrH0tSUcNX39vLsLbx94SoS9lSZoCW8nFOENwc7G239EEQcrwtoa1LfG8j41S05uh3u9XnER\nRzs2GAwwGAwuCEhKaWr5XcaqsyLpHYu9E9CrUty9e3daHP3u3bs4Ojqqujf0xFzLj5Z3PaOxyWSC\nlNLMisN6Uo4ivL1QhDcMzy+srUkb6C8iZH2T8no7kWSFQYtxVHfBG7Nrhckx7+7uYjKZTN0bi14D\nawl7x6HLP4oPWIT3zp070+2dO3dwdHTkiq4nyGIJey4Xb0zuCGRSVLfj4+Npf3d3N4yOsFY+2R4o\nwhtI9E8oAqFFaDwez/gdtQCfnp5eCKvSwjCZTKbuBC3CNuHBG/MEWPugF/VveuLo/Sjo89eheeKG\nuHv3Lp588slp06UhS2Is19i6FDy3g4iw3BVIGw6H7p2CFmHritCWMNkuKMIbhLZ4S+4IGx2hBVie\nm/NZYodYaNoHqd0QVmC9tcGi9cJEZHXpRxF8LdCLXovIGpbP96JDxB1x584dPPnkk/j85z+Pz33u\nc8XlgiKfcC2qRERYVr6WNh6PMRwOL/h79/b2muKEKcTbBUV4wyj9A8pj2iesXRDyHO0zHQwGGI/H\n01tlEWQrrrVmnyciqwVYhGkwGFzKv1mygvXn24k5vUimFuDPfe5z0+WCalawtoSjZq3h/f39mWYj\nH+Qze73ejE9YZ/PRJ7y9UIQ3HPsPaSfmrA/YC2MT4ZXbY1tdTARAi7P0ozEdu6wF2IrPoudcsoRF\ngG2ImvUJixB/9rOfnYpwSxMRtskstu3u7qLf7+PGjRszwmrFF8DUjeS5I/QdCQV4+6AIbyGe2NpJ\nOXle6ZZeRKDf78/M1lvR9VqUCOJNXEmMrE1ltgXr9TF7Vca0S+X4+HgaB6wz42q3+brpz7PjEiVh\nr50OtxPBti4h7/qfnJyg1+vhzp0709C5w8PDmUgJCvF2QhHeQrzbdRFGKxLe82xxG5m8i1wT3riI\nsGct2kks+wNR2kbWvFcfQmKDxRK2mWj2Ft/zs5du/+31E2tWk1LC8fHx1N9rf1DEfy8RIzp+WY5b\n/3AwYWP7oAhvIS0CHD1PJuX07bwOWatNxsm21RLWt+5ROi+Amey6SIBtXQhrCddqMsg1iQTZ7lu3\nSJR0Ir55uwKK/eHb29ubxjJbS1j7hmkJbxcU4S3E3ubu7OxgMpm4j1shkNjgXq83FeCjo6MZS7gU\npraICItLIkrlBRDewosQ60pxNk1ZW8KRX9aKsO3b62uPxRNgeZ6eGAXi9PK9vb2p9W5/PLQlTBHe\nLijCW4oWqsgHLMK5t7c344bQFqqIw97enhuLGyVttLghZCvHYGtU6EkrXe84ygrUoWi2NkQUf+uJ\nsN62jMkxRc/3XBD6+HU9D/2DEbkjKMLbBUV4y7BCZf2UkQiLEEcz/pJm7GWmef15LGGZlJIiP6en\nZ7UnrPjacxAhs0v/iABHE3M24sC6I/TntF5r7zE5dj1mJz7ljkOusa51rLd0R2wvc4twSun5AL4P\nwHMAfAGAl+Scf8k853UAvh3AgwB+B8B35Jw/cvnDJS1YIfbGxFoVX6TnCtATa7u7uzORBFpwvTEr\n5qX0XpuSK1EF2i+sRdKbRIxE2E5w1dwR8hm2H7kldJSEvs7e4qnWAtY/cHt7Z1XtbISHtd45Mbd9\nLGIJHwD4IICfBvCL9sGU0isBfDeAlwH4GIAfBPBoSunLc87jxQ+VtOLdHutVN2xhdZk4sk0e1wV3\nbEiWHZMWWdSeCFv3A/CUH1WO27v91/HA3uohniXsuSO0gOprVrvGgvYJi/jqZe+9cEHvGovvvpYU\nQ0t4u5hbhHPO7wTwTgBI/nTw9wB4fc75V86f8zIATwB4CYC3Ln6opBUtVLqvQ70mk8m0oE7LKhha\nDK3gemNahGu1iuWYBVtpzVqrNp655o7QPmGvME4pFK0kePYxcf/YWOednZ3pj1+piesocvVY1wnZ\nDpbqE04pPQvAMwC8S8Zyzk+mlH4PwPNAEb4S9Iy9WGRe4oNdIViLtDeuhb221VlyVoyl2LlYxXbh\nTG0dirvCuiMkvtZOzOlJuZboiJJPuHaNdd+uF6f70dp6pWSU2vUl28OyJ+aeASDjzPLVPHH+GLki\nIssOuCgU3tpnXt9ajdba9kQ4qqegRViw4muz2fR52QkuW7LS+oNboyMWuc56G1EqXO/Fb0efQ7YP\nRkdcQ+w/tLaUxX+sfZtiTVsBLrXITWF9yp5f2b6XRv8o6GZ9zrpam5581C4WbYFe9TVvfYxsP8sW\n4ccBJAAPYdYafgjAHy35s8iKiERhHhHW7zPP66L30niFgUR0pU7v/v7+hR8CPcElFvdVCjEhHksV\n4ZzzYymlxwG8EMCfAEBK6TaA5wL4yWV+Flke8/hBS+JYeu9IkFusXz2mLWEtwnr1iuFwODPppl0X\nMpFnS3wS0hWLxAkfAHgYZxYvAHxxSukrAXwm5/zXAN4I4NUppY/gLETt9QD+BsDblnLEZKnoSSU7\n7o1ZS7clqsB7fuRLts/10JEb2g2hi9PbVGovNveqXRKEeCxiCX81gHfjbAIuA3jD+fjPAPi2nPOP\npJRuAPgpnCVr/DaAf5cZI7z2aNHTEQv6sUgoIwtW92uTeiWXhj6uyBLWNZFtGJut0UARJuvCInHC\nvwmgeA+Xc34tgNcudkikC7RFXJtE8lwMtefXJvBqPmXBC2Gzq0Pr+FodxhZVbZP35QQZ6QJGR5AZ\nItdCyTL2trXPmFeINZ4QSySEDjuTjDpJ0tDZet7EHIWYdAFFmFwQ2UiMPBG2ffv8RSIhSr7myB3h\nLRmkRfjo6GimWlk0MUchJlcNRZgA8N0RWpBK/t7W/VZ3RHR88pidmJNax7YA0MnJyVSAB4PBBSs4\n8glTiMlVQhEmU0oTc7Xnl57jWbgtE3OeMHuWsJTC1M+T+hiSRSchbNYSjibmKMTkqqAIE5dlCFDk\nQ/bEuNUiBi6mN9t6CrpI0Wg0wmAwmAqwZw0T0iUUYbIStMDaxUPFRSCRCtq9YKMeZCv1HgBc8P1a\ny1jGxuPxVIC1Jey1yWRy4YdC+naMkGVCESYrQQTYrqM2Go2mwmtdAbqimLV2RYxTSu4knJS/1EKp\nRVcLsde8lZettS7nRcgyoQiTlRCJsA0N89wI2qq1scBSm1eLpl4GSRYHTSldsIBLIiyWsFcjWY6R\nQkxWAUWYrIRIhL36xF5lNG81jn6/P7NShV49RF6v+y0WsBZhW+HNVpDjZB1ZBRRhshI8EbYuCBG8\nmgtCxFQsac9nq+sfixhLllyLNWwz7STETZ8PBZisAoowWQmeCFsBlscAXJiYsxNng8FgWvdBiPzJ\nQqsVrEVYln2yYWu25jIhy4IiTFaGFlodCqaL64g4lyzgwWAwTbjQPmC7PJAd8yzgyBqWCI5oqSG9\nojIhy4QiTFaCtoS1BSk1HUSAJeLB+oC1QEqh9vH4rBCfXYDUW2lD4oTnmZizFrCNY2a1NbIKKMJk\nJWgRtvt7e3vT2r4SKWEF2FrBsj6cPFfQxXe8NepaJuek/KW1gkV8rYVNyDKhCJOVoEXXCpq4J7Q1\na9eG06tkyCKe4/G4ukip+JNFhFt9wmJlW+tXzoGZdWRVUITJStAJDnIrL64JK5xahLX7YTAYYDQa\nYTgcTstR2tfa6AhtUYsIW+s6anLcWoRFgGkFk1VBESYrQSc4aJ+wxN3q23udyqyXrR8MBtOtiKie\n1PPSlqNUZqk5rC3s4XCIGzduYDweI+c8U/Rd15ew+xLRoc+ztt+6JdcPijC5MiKh0VEUWowPDw8v\nVD2TcLfBYDAjxiLoIpQ2nVmvQyerMeu16HZ3d6efLVZ3tC8i7FWI8/o6HVpn4zElmgAUYXIFaEG0\nYwAuxBNLjQkrwACmoWRe2rK4I8Rils/U4/1+H8PhcFoQSER8b2/PFV4rwuPxeGYl51qzvmVvX64B\nRfh6QhEmK0VEUvoeniXs1f0FcKFuBIALk3Keq8KuyGxFPBJhb0yLcMtWzk2vAG2z8hiDfH2hCJMr\nwYqxzjwTodKlLkej0YwAa0HzqqdFyxzZx8US1rUnZGKwJMLWHaEtWtvsxJ5Y77rZUDhO+l1fKMJk\nZWixLaX72prDR0dHzQJsXQ3aStXP2d3dvVCyEsDMYyVfsG7Hx8cX3Aulvvy46PoZ+hpRhK83FGFy\nJZQE2boj9AKcWoBF/LQfWFwQ/X5/xs+rIzDsWnRagGW8JLqeCGu3iHUz2DEpPGRD3SR70AozuV5Q\nhMlKKbn/igOWAAAgAElEQVQhrJUrlqJnAYuVnHOeyYyzk21exISIrRVwHVMsadHzinDk79UFgfQ5\nCXJeEvJGEb6+UITJyqlNztlCP5EAHx0dTQu3awGW1OaaJWxdEHrpJPFFtwjweDy+ILSlvrXsdf0M\nW9yIXD8owuRK0OJrBVmLbSTA4/EYvV4PAGaSL2xtCesT1qFrsi8CrEtYlkTYPiafo/3Y3uSbLuEp\n52sn7Gz0B7l+UITJlWOtYa/Qj0zW6UI/IqTaBTEcDrG/vz8jjjr+VixhaxXbiAbxR9cEWMbks7QA\n2zF5rT4vbQFHa+2R6wVFmHSOjhDwxFj7Tnu9Hg4PD6dtf38fo9HIbTZxQhAx1rUnRBDtis9SSF7X\nOPbE1tvqiUYrslEihz7OlhRnJnhsPhRhshZ4mWaehahdB7rOxP3792dqTOzt7WEwGMyUutTv5/W1\n68KuX2cflx+Emkvi+Ph4Kupa5HXdY1uQyKY4R2Ny3fQ19PpkvaEIk7XBCrAXP6vFTQvx/fv3ZwR4\nd3cXw+HQLfauV3TWReF3d3enIif+Z+DiKtAtE3Lan61FV7tGvJVBSkkgtsk1k60Xi00xXn8owqRz\nvOI32hLVQuIJ8OHh4QUB3tnZmboSxLVgt8BTk4RaJDVWtK3QRqFpevUQK/5a1K1FbmOMo3177bQA\n15JjyHpBESZrg73t9h73RNgTYCmRqX25ugFPxSnbovCCtYBlVZCaQFpRtuKrt/I5AKZV4qw1rd/H\nltH0fN7yGNkMKMJkLbACrBfW1I95IhwV+zk+Pp6GsUnKso0V1skbOolEi7IW2JY0ZRuG5lnCkW/a\n8yfLexwfH4fXq3ZtyfpCESZrgxVgLU6RCEcCnHPG8fExhsPhTMlKeT+bwGFD2XS8soh1S8EeL/TN\niq9tgljv0nZ3d4v1JvT18sSY4rsZUIRJ53hxw1oItaVqRTgSYF0UKBJg8a3q97dRCl6EghXB0tYu\nIOqJr0bOza4mLeel4421D9i7e6BfeDOgCJO1wvNtajGR8DA96eYJsPanagGWELN+v38hs64Wl2sn\nD0tN3lvqBltLODpvXUEuEmAJcxPBtXcPFODNgiJM1oJIfIHZFTl0Ft3R0dGMuNUE2KYr64I+IpQe\nkWhaUfb6IsLW7WDPW7YisPY1+tyk8ppG3z1oAaYYrz8UYbI2tIRY2dhbLVQ6IkEqnQEXBdjWmLCW\nqrymlODhHbe3b90RtXOPSl7alGcrwvrHxBNjsr5QhMlaEomHFll7i299psBTlqVNPe73+zg6Oppu\nbeyu3gK4MLkG+BayHRPfs/inS5N70efI9bATgDpqQ8cY29dRjNcbijDZOKzY6qQILV5S7rLmOz45\nOcFwOHRTi71+yUqOLGn9QyBhczq+WF7jJZZEzYsn1payTuqQc671ydVDESYbhbUIbSyuft7Ozk6T\nCE8mEwyHwxnxk6b3tTBqP7IW2ygJQ68CIkXkrQDrMpsl8ZUWldOUpA59rnbrWccU426gCJONQwuo\nCI4XSyuREJ742nrFw+FwKtg6s87uS9acl4KsXRfyWQBcS1jXf9DWcmuT5A0dU2zdNPp8o8lDfb1I\nN1CEycZhhVRnk+nHrGWqHxcXhsTlDofDafpzv9+faTJ2cnIyFVBdAEhn3wkixNpCFktYi6IWcW0B\nW2vYc43oGseS2BGdqw6fs9fSi9YgVwdFmGwUWkxs+JcVZ2B2FQ95jViMIsCj0QjD4RCDwWCmie9W\nuw50rK6O2RW0K0KLrC4mb+OW9VJL1hViJxb1mHa1eNEi+lpEiSe2hjG5euYW4ZTS8wF8H4DnAPgC\nAC/JOf+SevzNAL7VvOydOecXXeZACRFEPFIqF4HXz5fHo1KYskqHbXryTN/Oa0G1Imb9rTpJJBrX\n7o+SC0I37Rax/nAbU2xdEdoVol9Hrp5FLOEDAB8E8NMAfjF4zjsAvByAfMNHC3wOIRewImLHtIvC\nWnw1Ad7f3y8ulQQ8JVqnp6dTUbXCKhayHrMuC+0nluLw/X7/gouj1KxvWrAxxTq7Tq6Frc9BAe6O\nuUU45/xOAO8EgBRHoB/lnD91mQMjJMIrVCPCIgIsYlgSYO1yGA6HFxYMtRawCKe2LGVchNbe8mtB\nledqF4Qte1mzfksV2bykDnkPfa3kGjK7bj1YlU/4kZTSEwA+C+DXAbw65/yZFX0WuUZoEZF9m7Ir\nTURUC7BO0pAJt16vF1rAnv+2JMLaAteP68gJsUxtsoYnwiVBtv5uK8B6kVQhulYU4O5YhQi/A8Av\nAHgMwJcA+CEAb08pPS/zWyZLQP6MdHaYvimTvl1mqNfrTctf6vCzvb09jEajJgG2omajH7So6iw4\neY71zern6Ym+SHz1vjfpKOdr46Oj69iSUk1Wy9JFOOf8VrX7ZymlPwXwUQCPAHj3sj+PXE9sAoKH\n547Y29ubWog6AeP4+HhGEG0Chs2Y0xN2VnS97DlBJ3DYMW+STz/HHpN+jfWJ61RpqaMh7hJ5XF9L\na92Tq2PlIWo558dSSp8G8DAowuSK8UKybI2FlNK0PKaErHmTXvJ6cWvoJv5lL8bYS+zw+nIsWvDF\nWtfWuWdNC96PhvzIeKtBa6G2vnZyNaxchFNKzwTwNACfXPVnEaLxwrJ0WJtGEj5qheJPTk4wHo+r\nAixjvV6vKQPOs3R10SFb6Ed8uNb6tq4LEWHxd8vEpGx1tAdFuBsWiRM+wJlVK/dYX5xS+koAnzlv\nr8GZT/jx8+f9MIAPA3h0GQdMyDxYK9i6CuQ5Ozs7TYXiRai18FoRtvteIobsW7+xtYS9DDttOQMX\nBdhL8pCIEPmRkcpxcn7ivuAE3dWziCX81ThzK+Tz9obz8Z8B8J0AvgLAywA8COATOBPf/5JzPr74\nVoSsFivAXmysiJstFC+P2TTnkgjrsDcv9Vm20aoesi+WsLVQtUjrscgNIZOQdkFUe35ejWJyNSwS\nJ/ybAErf1tcvfjiELJ9SgodYwSkljMfjqgCPx+MLQtuyHQwGmEwmGAwGF4rJi9CKOOpIC7GWPSu5\nJL62frIIsD0/HU/MKIluYO0IstVYd0T0mIhZTYDnEV2dCCIRClaARTQ9365k2FmhFYH1hFk/rq3v\nkgBL5TVawt1AESZbiwiYTdXVwiuPyb68zhNgW2EtEl09NhgMLggwgBlBjWKS9b4VbR11Ebkg9PFG\nLgid0MGkjW6gCJOtRguKTdXVPmKdMKEFuNfrTSfsdIKHV2ktsoJLFrCUyLQiLGgr2C5t5FnHdgmn\n0WiEfr8fCrCNBiFXD0WYbD12Ak7HB+sQLW0BizDpiAkdseAJr7c/Ho+LAqwXHJVj0jHD8pi4Jmw5\nTR3KpovG6zYej6fnbwVYojfoE+4OijDZauyttRY7QcROF2sXP6mXNry3txeKrh0fj8cXXBBipeo0\naXtcIro6M0620vd8wGK5y+dLUSItwDoppSW9mawWijC5FkRirBHfq14uyW4llE2ETCxN3SSEbTwe\nT2NvRXxFgCNf8Ty1HLSrIoqIkOPzSnjq8DmvKDy5GijChJxjJ+esr1gsRV1/QS806iWBpJRmLFKd\nPiyvt6nIAJrEMMqSs5N9Oeep2NrJRW85J22Ze2nS+jjJ5aEIE6KwQmwfk60IsRVgnU4sIXEyQadT\nh3W1Nm+JIX0cJWz4mrzf3t7sv3ZJeK04TyYTt+aGrVVBIV4OFGFCDJG46OgKHaEgLgcd8qbXuZN0\nYXFXaCs4EmE5jpIQ21oTOrbYvp8V2pJVbM9DR5LoHxeK8HKgCBMSUBJjvYKHfr4NIwOA0Wg0447Q\nlrBdCdmKbkmIPVeEV9QnpeRWeIuanqiUHwr94+OJPFkcijAhCi0uUeKCFVk7Lm0ymQDAjBVsfcLW\nEl5EiG2GnYzLdmdnJ7SAPWsYwIz46rKfErFBEV4eFGFCzvH8wVaIRYSsCFs3hNRjyDnPWMLREkr2\nM+admJP3sGUxxUouia8VYuAsUkTEVyJGShY7WRyKMCEGGxlhx20lNivAOrY45zzjE/aiI1om5iLh\n02Jrx7SFXBNf3Qcw4+fW56l9wmQ5UIQJUWgBLgmx9LVlqFOJtQiPRqMZK1jig7U7wr63t2/xBBjA\n1FLXqc4t4it9m1loBdhGgpDLQREmxBAJsEZHCthkDi2Op6enRUvYW7Jonlt9/VlaGOUHQN6/1QqW\nIvKRAHvx0ORyUIQJcSgJsBZKb6v7ngjb6IgoRM1+lkXGtU9Y93Wbxx1hkzO0ADOrbvlQhAlpIIqa\naLkdj8TXhqfpz/KEzoq/92PgHVfOeSadOepL0z8M2s+tJ+ooxMuDIky2HluspzRmLVlv33ufaNvv\n9/HAAw/g1q1buHnzJm7cuIH9/X0Mh8OZCmZ2Yq3Wt9aq7nv73urK8qOgayZbl0mUWEJf8PKgCJOt\npkVUrR/X+li9sajZ9+33+7h9+zZu3749FeHhcHhhNWbvx6GG53aImhVeK8ZR01a7l7pMLg9FmFwL\nIgG1YuuVr/RaJNS2ry3hg4ODqSUsIuyt7iy0CrL229q+bFuE10Zw2AnEmv+aLAZFmGw1nuh6ZSq9\n5eJFHL0x77XevljC4o44ODhw3RE1AY4E2RNd7TawCSSREGvxLbkjrIVNLg9FmGwtkQBHTQTRbr0x\n+9porNfruT5h647Qq2l45+FhBdGmTWvr1QpwZBGLNazdEC3FhsjiUITJ1uMJsBVNu0SQLpDujev3\nsFazHuv1elNLuOSOiCIOai4JT4St+0CLsBZgXejdWsO2ZrK1sCnEy4MiTLYazxL2RNML1YrGPEs5\nar1eDzdv3py2aGLOW1poHp+wZwFrMZ7HEhYRtkJu/cxkOVCEydYTuSSsWErCQm3rCXHkwuj1ejg4\nOJi2WojaZaMjrAjr1T/m8QmXfMwU4OVCESZbS018PcHUa8DpymJ6P3JR2KXn5T1FePf392fcEfJ5\nJXdEDc8frC1gT4BLQixF5220hReBQZYDRZhsNVaIvYgHLZi11ZO1L1eLcak/HA4vNC9EbVG0QHoC\nrIXYRkB4Quyte+ftk+VAESZbjyfE0XLx2vIVsbRbsWKtzzja1yJuLWubrCHHG6VJ6/3IHVFzRZQy\n5iRBwxNZCu9qoAiTtSJK09X7LZlqKaUZN4O1Uu02soCt+C5iCXtuDPkR8M7Bil1pP3IVeOFqXoRD\nVFTeQgFeHRRhshbU6jNIvzVJQkc8tBSvsRaqdUPoscgnHPmFIwH2qpLNI3aeu8BL3PDqPpRa9Dlk\nNVCEydpQq8tQCjPz4nS9ELOoRYte2gk67UKoRUjYCTotxPZHY95JuVYBrrXWzyGrgyJMOqdUz8Hu\nRxls1sr0BNgudFkKSyuNiQi3tMha1vUiFomKAOICPq2uCFrA6wFFmKwNXsUy625o8fNaN0Or0HoC\n7Y3rjLmWrLkojngRSzgSxtpEnRfr60U91D6HLB+KMFkLSkV2bHpxi3B6qwhbl4Ldr/mNtdC31o7w\nCgAtYglHlmpr7Yh5hJgCfLVQhMla4Pl/PWHTouj5a0u+3NK4zoSr1Y3Y29trnhz0BNmO1YS45CpY\nhktCvw+5eijCpHNsFETJorRhXzaMzAs18yIf7FZPtnmRDrpv3Qi1fkmcbXSEpUUYl+ETJt1BESZr\nQWsUhM1si7LRbGyvJ9A27MxLZ/bG7MRh1OS8ShON9vktlKxgr5awFxvM2sDrA0WYrBXWSvQm5Gx6\n8XA4nNZmkOI4so0SMGzTImz9uPbHQFuv82xLYx4tEQulULVSJTQvUYNC3A0UYbIW1Cxh7RLQ7ggR\nXCmMo7faMrbWse3rGg61ybaa1bpoyFkNL1NO9+eJD45cEhTiq4ciTC5FJDg1y0/3o8kvrw0GgwsV\nyWxfW8SRn9irERxZ4Z5Pt5Uo7CsKCytNtNmx8XiMo6OjafUzb1/anTt3cPfuXdy/fx+Hh4cYjUbT\n5+p15MjVQxEmc6OF1+uXki9s/G9KKazp4I2JCGuXg+eG0KtXeEVz5D09a9dOvC3it7WUwsEWsWRP\nTk5C4dX70r9z5w6efPJJ3L17F/fu3cPh4SFFeE2gCJO5aPV5tliV0mrpxHo/mozzWpT5VqrjEEU6\n2POfh5pFq8d0+cla0yLrCa/u3717d9ru37+P+/fvYzQaVSunkdVDESZz47kVvDAzLznBG/My2KJ+\nyb/rVTuzoh7VcdAVzawVLOe1KFZsvQgG6UcF2L1SlDUR1mP37t3DvXv3cP/+/QuWsIgwV83oBoow\nWZhSeFZrgZvd3d1iIoXnUogqnUUCXPIxt4adyfkuSktW2+npqbvuW9Ss2Ja2h4eH06b9wuKOmEwm\ndEd0xFwinFJ6FYBvBPBlAA4BvBfAK3POHzbPex2AbwfwIIDfAfAdOeePLOWISWdEAuUJmC2kU8pE\nKwmptz9vtbPSD4BXw6EkxItgBViLrl0Jw5tUs01bwdIi8ZX+aDSaTsZJn+6I9WBeS/j5AH4cwB+c\nv/aHAPxqSunLc86HAJBSeiWA7wbwMgAfA/CDAB49f854WQdOuqM26aYtYc8VoP2y2s/bEkpWKshj\nt7bQTsk1YmN/lyXAgifE3lJEVjy10EYteo3tRxN52h1Brp65RDjn/CK9n1J6OYC/A/AcAO85H/4e\nAK/POf/K+XNeBuAJAC8B8NZLHi/pmFoEhGy9kpKRrzfKeIsm21rrBJeK6nhNn5fXX5RSEoVdgkgL\npbZYbauJsd0vrSunoyNoDV89l/UJPwggA/gMAKSUngXgGQDeJU/IOT+ZUvo9AM8DRXgr8CbhdESE\n9LX7Iarf0O/3p2FlNsTM21o/b9QXd4P3AxFt7bl557soLQIsAikCrP244sPV+yUBtuM64iLq0xLu\nhoVFOJ39Rb4RwHtyzn9+PvwMnInyE+bpT5w/RrYEzxK2FqdX68FrNsmi1Gx4WakfRTnUJt2s2C4z\nPljEOBJgbQXriTTbrAjXWqmmhE1fJlfLZSzhNwF4NoCvWdKxkDWnNnEV1XooFd0ZDAZuyrE3JiJc\nKq6jWy2+dxm+3la8iTkrxFaER6PRNKTMhpjdu3dvLhFuydYj3bCQCKeUfgLAiwA8P+f8SfXQ4wAS\ngIcwaw0/BOCPFj1Ishwi0SlZfrrfuqTP7u7uzGRa5Ov1iu94qcd2qXmvulpUaKeVWnqx7nsZb9G+\nDTvzfLIyLvG82uLV+7qvLeEoNli/N1lf5hbhcwH+BgAvyDl/XD+Wc34spfQ4gBcC+JPz598G8FwA\nP3n5wyXzEIlp5O/0/KK631LbQSzTeSIeonKUOt5XpxhHGW6XnUiL0om9sZbECxv7q4XRa+PxeCaO\n1/a1b1gnWojQWv8urdvNYN444TcB+GYALwZwL6X00PlDn885j877bwTw6pTSR3AWovZ6AH8D4G1L\nOWLSRDTRFAltSytFItiIhcj/K5XP9ASdzXTTj9s6Dzau1wqxPrdFsGnFUZqxDTGLws5OTk5cEfb6\nOp5XxDbqS8SDvFaLsPX1kvVmXkv4FTibePsNM/4fAPwsAOScfySldAPAT+EseuK3Afw7xghfPZ7Q\n6n0AYZyvFwPcul5bbRWLlq0tuDNPcsWieKFkXt9GFuitN6ZFVve9rQ4/08kVdswLO9OWsLXkyfoy\nb5xwk5Mt5/xaAK9d4HjIEqhFA3gTal6Ime3XMtrsY7U13bTI1rZeuUnPEl6GCNdSi0VcrQB6omiF\n1gqxt99aD8LWk9CWOK3gzYG1I7aYeYQ3Ejcd7WCL5cxTwSyymFtXOdY+YS/Gd9n+YOtm0Ftr3Zbc\nC1ZsW1OSo/eyLojIKqcQbw4U4S0mEt4opKzWF+u1JZ7XCqzej6xcG+/rxf9Gbgiv7OS81LLatI93\nXqvV2/fGdbiat/VcD/Y4Gfu7WVCEtxTrjogKq9uIg1Lolw4pk1jeqLWkFev04lrJSx37W2qXIarv\nYC1NPYnm+Wq9tOKW2r8iwlFWW2Tt2sk4WsGbBUV4CylFQHguh3lif7UIHxwchM1bFSNaMaNUz8Fz\nlbSG1S2CJ8SlWr5eBIPdjwrreNuSG8Ru9WSht6UlvBlQhLeUVgG2IqyjEGzfE+GbN2+6zavlEJWz\nbHExlPy+UX8eakV2vPoOOr3YxvPatdxaC+6URDUS2VI8M0V4/aEIbzGlSblIgEs1GSIRvnXr1ky7\nefPmjJVb8/la69UK6TJEtoVSfQfrirA1Hrz6DnoFi9aSlF5Wnj42PWb7pTGyvlCEN4BSyJk3Virb\naMdaJ8X29vamAnvz5s2p20HXedAVz7zi6VFh9XmEtSW9WO9HImYf10kVLStaeKJbKrRTKjOp3RG1\ncybbB0V4TahZf60+Uy+yoTTWUo1Mmli+2vd748aNmbA0O9HmHdey/La635JeXNp6NR2iNOOjo6ML\nqcSeS0L7g6MU49okGgV4+6EIrwElP6cIVs1na/dLEQ+RH9iKrxXhKBJCLy/vLaAZRTUsSiS0UV2H\n0sSWF/vrCbDetwXX7aScTS/Wsb5RUgXdCtcXinDHeLP7spW+iJkXWWC3XllHLwJCl5uMtrav3Q22\n0tm8xXb0uc6DrVBWKp4ThZh5Bc1LiRZRjYdaeJpew81byaJUaIcCfH2gCK8BpbArcUV4ywR5WWhe\n3G2tlYRYb2tlKbUIt4aZLcI86cVekkPUr9V2qKUXe1vrgmhJMaYAXy8owh3iRQVEIixiWKrXIELo\nuSpqYy0TaV6xHTsWre3mhZstinU5RFXMbO2GmquhVtOhVISnJN5RwoV1R1B8rycU4TWglgWm3RF2\ndQq7PttgMAhFtDZWe24t/Vi7S1ZZ4yGK59UWpk0vbkkbLtV5sGJri/V4+9L3UotZ95cIFOE1oiTC\n4pfVIuwtATQcDqv+XTtWm+TzJulKiRh7e3tNiReLEmW1eSUkS3G52n87b5EdW7bSa9bijdKLKcTX\nG4pwx8wT86sLpsukmA4Vu3nz5lSES/G+niiX3Bbec2qTf6VMt2UKcbRqsV25uNZKQmwfE+u2ZOHq\n5k0c2jGK8PWFItwhkT/YWo+eT1gsYZs6fOPGjaKF6glxq9/Yxit78cv6uPU5XsYFoYkm5bQQ24k2\nu3qx16xFXHJdtFq3ns83auT6QhFeAzwr0YqcuAAiEb516xZu376Ng4OD0GXguRDmsXhrwloS2mWm\nG0fVzmrLx5cy2yScrKXa2fHxcTVWORLZKLOPXF8ownPSUtegZasFtmRRSh3fUsUy20qWr1e/t7WS\n2jzMIy5eGnG0jQqqe2OyZLxkspVSi2suCVtuspSpZ8+JkBIU4QZKMbyesNqJKC9SYJ7W7/dnXA62\nhoNMzEm8rucH9ibjagkVy7ReNZ416KUSe2PzrFRRc0F45SZtpIOXXuxZt4zzJYtCEW4kKqsYiasV\nOC10pTHv8V6v59ZtsEV0rAi3TLLZz4x8upellBUWJVt4+6VECTsWpRJ7rSTAUWIFxZcsA4pwhSha\nwbNwW2Jtbb9U40H6/X4/rNtgLeHhcNjsXmhJqpBrcBlKflBt8VrR89KLS2FnUe3elhRjG3oWpReX\nXCelcyYkgiLcgBXgqEkYmed/tf1WkZT39NZx0yUktTuipXiPZ/0uO7MNqJee1FlvpZhbad5S8Nql\nYLfzpBdHsb+RFVzaEtIKRbgBL2bXcyHY+g5eX/bnidPt9XpuhpxtssCm5+LwfjAii34VfuGSWHkh\nZqVavq1uBi8TLupHhX1s5bVSxAMFmCwCRbgRT4g98ZQaCrp5Y7bqWWkrCRq1NhwO0ev1mmN5rdiu\nygpu2Yo7orZUvNTy1ZNuUV/7eSNxt64HrwaFF/Orj7/UJ6QGRbiC54qIrFYtuiKKpUI7LZltYglb\nEffaYDBAr9dritSIRDdqy8Kb1PIsYev71TG/tnh6VFRdLNyWNm/ShXdOtk9ICxThBiIhtuIpYhm5\nC2zpx9ZYXu3msG4NO+bVbaiJa7S/DDzR9fpiCXuJFnZVYy24Nt5X96WCWeRmiNKLa1t9Xt65EjIP\nFOEG7C27N3GmU4r1gphRE6s1msCLJvVa0pCjuObSmD7X6LFFiWJqtQDb4uo6gkH7faOEi3v37l3Y\nPz4+DlfS8Pa9xIt5st8IWQSKcAUtRnZSzrNUdVqxRDB4Tfy3+rXW4tUCPE/ImSecpUy/lvHLEiU4\n6PRjzxLW67lZkZUW7YsIR4kfdquPrdYnZFlQhBuIfMJWNLUVrC1hG9N7cHBQFWHdjxIrvH5JRCMR\nsZEKLa+pYTPfoky4nPOMn9dzK9gxLbRagK11fHx8PD2WVuuWkKuGItyAdUVYy1d8sy0RDLrV3A0t\nacXR5FkpJKx02x2NzytUpWWHbIuiG6Ixr96DuDG0f3dZ50LIKqEIV/B8wZ4LQrshWrZahKOoiFJW\nW+R2AGKR9cZaJ6PmvS2v1djVrbXmr/YN69hgne0W1XnwIhooxmQdoAg34MUHewJcit/1ti1pzSUr\nuJRUEbkBPF9s1Kz16hEJWbTqhDdWSz/2Uo11QoZYw/J+kQBTeMk6QhGu4CVpaEvYiwv2rF6vtdZ1\nEF/vPBltLVaurcdrV4mwj7X4lAVvrbWo35paHBXqidwRkQBzso2sExThBrwCPV5ExLzuiFpNBy3E\nteQK4KnIBk+AvcQDWzTH6+sxoSVGNlp5eJ7VjWv7duXkyB2hj88TYkK6hCLcgGcJz+OOiFpUzyGq\nalbLcLN4AmxdDK0ZZZPJ5MJ7e58neCtSRCtWlGpF1FKNvfKTrPFANgmKcAUvMsJmyM1rBQ+Hw5lC\nO9bVEBXU0cfT4oqwvl9bD0GSIzxRs2OymoT9HO+zATSVmpR9LaD686J9b2VlW/+3xf1AMSbrAEW4\ngZIlbIW4tYkIW1FtcTm0ZLVFk29erV57W6+bHtfv7fX1/jzVziJhjdwiUaEdr9hO7TgJ6RqKcIVa\niFpLaJrX+v1+KKylOg4tfS8SIiqWXqtaph+T99afo9H7Ucyv10oVzKKKZp7P2/YpvGQToAg34Lkk\ntEH2YBgAAAiaSURBVAjXhDiyhPX7e31vPxqz1CzhaGXi0vJB+r1r/VJxHTuul5D3BLVUSrLWJ2Td\noQgvSM0nK9vIJ6ujDZZJztmNyY36tRCwVhG2Y7Ul5nWzBdOtpesljBCyLVCEGxDh1FbjaDSaiWaQ\n52nr0lYBkzoHN27cQK/XW8mxiuuhFnom/daVi8UdIZ9R69u0Y5terF0QnvuEE2jkukARriBWmBbh\no6OjqX9Yx+aKCNtSjIeHhxeWJVq1CHvN+ll1NIQXd2vH5P3t53n90lJDVoBLIWUUYLLtUIQr6Nth\nEa2SAOvb+NFoVKwbsSq8yawWIa4tAyTnaq+Ptx+tiKzTi1stYUK2GYpwBWsJj8fjmbhd/bhYwIPB\nYCrAdp05vbTRqo83StLQfS8GN4rNlfe3n+cdg1wLO9Gn04ujQjsUY3KdoAg3oEVY+4CtAI/HY/T7\nfYxGo2l5S2/V5X6/j93d3ZUcqxe6VWrRpJ23jQTRGy+5NvRn2Uk3uiXIdWMuEU4pvQrANwL4MgCH\nAN4L4JU55w+r57wZwLeal74z5/yiSx5rJ1hLOLKARWC8JYq8rQj5KqjFz+pta7nJRQv4WKu6ZAnL\ne3kRF4RsK/Naws8H8OMA/uD8tT8E4FdTSl+ecz5Uz3sHgJcDkBiuI2woWqw8H7COmPCWq4+Wsl+l\nCNuwLm/yywpxzW1RC6mzgtma/WYt32hLyLYylwhbazal9HIAfwfgOQDeox46yjl/6tJHtwaIQIgI\nedESXglKvfX6VyHCLc1zXURWdMvnClGtiij7zb6eIWrkunBZn/CDADKAz5jxR1JKTwD4LIBfB/Dq\nnLN9zsYgAqQF2FY58yqflfqrWlBTjlO2nmXp+V8jcV5kgiwS+Jq4t0z6EbJtLCzC6UxF3gjgPTnn\nP1cPvQPALwB4DMCX4Mxl8faU0vPyBv5XWTGJqpnVaj9EtSBWfewt+63beT87mmyLJt8IuY5cxhJ+\nE4BnA/gaPZhzfqva/bOU0p8C+CiARwC8+xKf1xmcnSeErIqFHJMppZ8A8CIAj+ScP1l6bs75MQCf\nBvDwIp9FCCHbzNyW8LkAfwOAF+ScP97w/GcCeBqAolgTQsh1ZC5LOKX0JgDfAuDfA7iXUnrovA3P\nHz9IKf1ISum5KaV/nFJ6IYD/A+DDAB5d9sETQsimM6874hUAbgP4DQCfUO2l54+fAPgKAG8D8JcA\n/juA3wfwr3LOx/bNCCHkujNvnHBRtHPOIwBff6kjIoSQa8TqMgYIIYRUoQgTQkiHUIQJIaRDKMKE\nENIhFGFCCOkQijAhhHQIRZgQQjqEIkwIIR1CESaEkA6hCBNCSIdQhAkhpEMowoQQ0iEUYUII6RCK\nMCGEdAhFmBBCOoQiTAghHUIRJoSQDqEIE0JIh1CECSGkQyjChBDSIRRhQgjpkHUQ4WHXB0AIISui\nqm/rIMJf1PUBEELIivii2hNSzvkKjqNwACk9DcDXAfgYgFGnB0MIIcthiDMBfjTn/PelJ3YuwoQQ\ncp1ZB3cEIYRcWyjChBDSIRRhQgjpEIowIYR0yFqKcErpu1JKj6WUDlNK70sp/fOuj2kZpJRek1I6\nNe3Puz6uRUgpPT+l9Esppb89P48XO895XUrpEyml+yml/5tSeriLY12E2vmllN7sfJdv7+p4W0kp\nvSql9P6U0pMppSdSSv87pfRPnOdt5HfXcn7r9t2tnQinlL4JwBsAvAbAVwH4YwCPppSe3umBLY8P\nAXgIwDPO29d2ezgLcwDggwC+E8CFEJuU0isBfDeA/wjgXwC4h7PvsX+VB3kJiud3zjsw+11+89Uc\n2qV4PoAfB/BcAP8WQA/Ar6aU9uUJG/7dVc/vnPX57nLOa9UAvA/Af1X7CcDfAPj+ro9tCef2GgB/\n2PVxrOC8TgG82Ix9AsD3qv3bAA4BvLTr413S+b0ZwC92fWxLOLenn5/f127pd+ed31p9d2tlCaeU\negCeA+BdMpbPrtqvAXheV8e1ZL70/Bb3oyml/5lS+kddH9CySSk9C2fWhf4enwTwe9ie7xEAHjm/\n5f2LlNKbUkr/oOsDWoAHcWbpfwbYyu9u5vwUa/PdrZUI4+xXaxfAE2b8CZz9YWw67wPwcpxlCL4C\nwLMA/FZK6aDLg1oBz8DZH/62fo/A2e3sywD8GwDfD+AFAN6eUkqdHtUcnB/rGwG8J+cscxNb890F\n5wes2Xe318WHXldyzo+q3Q+llN4P4K8AvBRnt0hkQ8g5v1Xt/llK6U8BfBTAIwDe3clBzc+bADwb\nwNd0fSArwj2/dfvu1s0S/jSAE5w5zDUPAXj86g9nteScPw/gwwA2YuZ5Dh7HmS//WnyPAJBzfgxn\nf78b8V2mlH4CwIsAPJJz/qR6aCu+u8L5XaDr726tRDjnfAzgAwBeKGPntwgvBPDero5rVaSUbuLs\niy/+kWwa53/Uj2P2e7yNsxnrrfseASCl9EwAT8MGfJfnAvUNAP51zvnj+rFt+O5K5xc8v9Pvbh3d\nET8G4C0ppQ8AeD+A7wVwA8BbujyoZZBS+lEAv4wzF8QXAvgBAMcAfr7L41qEcz/2wzizmgDgi1NK\nXwngMznnv8aZL+7VKaWP4KxC3utxFuXytg4Od25K53feXgPgF3AmWA8D+GGc3dU8evHd1oeU0ptw\nFo71YgD3Ukpi8X4+5yxVDDf2u6ud3/n3ul7fXdfhGUFYyXfi7Ms/BPC7AL6662Na0nn9PM7+mA8B\nfBzAzwF4VtfHteC5vABnoT8npv0P9ZzX4izc6T7O/sAf7vq4l3F+OCtT+E6c/ROPAPw/AP8NwD/s\n+rgbzss7pxMALzPP28jvrnZ+6/jdsZQlIYR0yFr5hAkh5LpBESaEkA6hCBNCSIdQhAkhpEMowoQQ\n0iEUYUII6RCKMCGEdAhFmBBCOoQiTAghHUIRJoSQDqEIE0JIh1CECSGkQ/4/dWtzaJOhYVoAAAAA\nSUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f1de88868d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "label: 5\n"
     ]
    }
   ],
   "source": [
    "from chainer.datasets import mnist\n",
    "\n",
    "# データセットがダウンロード済みでなければ、ダウンロードも行う\n",
    "train, test = mnist.get_mnist(withlabel=True, ndim=1)\n",
    "\n",
    "# matplotlibを使ったグラフ描画結果がnotebook内に表示されるようにします。\n",
    "%matplotlib inline\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# データの例示\n",
    "x, t = train[0]\n",
    "plt.imshow(x.reshape(28, 28), cmap='gray')\n",
    "plt.show()\n",
    "print('label:', t)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 2. Iteratorの作成\n",
    "\n",
    "データセットから決まった数のデータを取り出し、それらを束ねてミニバッチを作成して返してくれる`Iterator`を作成しましょう。これをこの後の学習ループの中で使用していきます。イテレータは、`next()`メソッドで新しいミニバッチを返してくれます。内部ではデータセットを何周なめたか（`epoch`）、現在のイテレーションが新しいepochの最初のイテレーションか、などを管理するプロパティ（`is_new_epoch`）などを持っています。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from chainer import iterators\n",
    "\n",
    "batchsize = 128\n",
    "\n",
    "train_iter = iterators.SerialIterator(train, batchsize)\n",
    "test_iter = iterators.SerialIterator(test, batchsize,\n",
    "                                     repeat=False, shuffle=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "### Iteratorについて\n",
    "\n",
    "- Chainerがいくつか用意しているIteratorの一種である`SerialIterator`は、データセットの中のデータを順番に取り出してくる最もシンプルなIteratorです。\n",
    "- 引数にデータセットオブジェクトと、バッチサイズを取ります。\n",
    "- また、このとき渡したデータセットから、何周も何周もデータを繰り返し読み出す必要がある場合は`repeat`引数を`True`とし、1周が終わったらそれ以上データを取り出したくない場合はこれを`False`とします。デフォルトでは、`True`になっています。\n",
    "- `shuffle`引数に`True`を渡すと、データセットから取り出されてくるデータの順番をエポックごとにランダムに変更します。\n",
    "\n",
    "ここで、`batchsize = 128`としているので、ここで作成した訓練データ用の`Iterator`である`train_iter`およびテストデータ用の`Iterator`である`test_iter`は、それぞれ128枚の数字画像データを一括りにして返す`Iterator`ということになります。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 3. モデルの定義\n",
    "\n",
    "ここでは、シンプルな三層パーセプトロンを定義します。これは全結合層のみからなるネットワークです。中間層のユニット数は適当に100とし、出力は10クラスなので10とします。ここで用いるMNISTデータセットは10種のラベルを持つためです。では、モデルを定義するために必要な`Link`, `Function`, そして`Chain`について、簡単にここで説明を行います。\n",
    "\n",
    "### LinkとFunction\n",
    "\n",
    "- Chainerでは、ニューラルネットワークの各層を、`Link`と`Function`に区別します。\n",
    "- **`Link`は、パラメータを持つ関数です。**\n",
    "- **`Function`は、パラメータを持たない関数です。**\n",
    "- これらを組み合わせてモデルを記述します。\n",
    "- パラメータを持つ層は、`chainer.links`モジュール以下にたくさん用意されています。\n",
    "- パラメータを持たない層は、`chainer.functions`モジュール以下にたくさん用意されています。\n",
    "- これらを簡単に使うために、\n",
    "    ```\n",
    "    import chainer.links as L\n",
    "    import chainer.functions as F\n",
    "    ```\n",
    "    と別名を与えて、`L.Convolution2D(...)`や`F.relu(...)`のように用いる慣習があります。\n",
    "\n",
    "### Chain\n",
    "\n",
    "- Chainは、パラメータを持つ層＝**`Link`をまとめておくためのクラス**です。\n",
    "- パラメータを持つということは、基本的にモデルの学習の際にそれらを更新していく必要があるということです（例外はあります）。\n",
    "- そこで、学習中に`Optimizer`が更新すべき全てのパラメータを簡単に取得できるように、`Chain`で一箇所にまとめておきます。\n",
    "\n",
    "### Chainを継承して定義されるモデル\n",
    "\n",
    "- モデルは`Chain`クラスを継承したクラスとして定義されることが多いです。\n",
    "- その場合、モデルを表すクラスのコンストラクタで、親クラスのコンストラクタにキーワード引数の形で登録したい層の名前と、オブジェクトを渡しておくと、自動的に`Optimizer`から見つけられる形で保持しておいてくれます。\n",
    "- これは、別の場所で`add_link`メソッドを使っても行うことができます。 \n",
    "- また、関数呼び出しのようにしてモデルに`()`アクセサでデータを渡せるように、`__call__`メソッドを定義して、その中にforward処理を記述すると便利です。\n",
    "\n",
    "### GPUで実行するには\n",
    "\n",
    "- `Chain`クラスは`to_gpu`メソッドを持ち、この引数にGPU IDを指定すると、指定したGPU IDのメモリ上にモデルの全パラメータを転送します。\n",
    "- これはモデル内部でのforward/backward計算をその指定したGPU上で行うために必要になります。\n",
    "- これを行わない場合、それらの処理はCPU上で行われます。\n",
    "\n",
    "---\n",
    "\n",
    "それでは実際にモデルの定義を行い、オブジェクトを作って、GPUに送信してみましょう。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<__main__.MLP at 0x7f1df1710080>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import chainer\n",
    "import chainer.links as L\n",
    "import chainer.functions as F\n",
    "\n",
    "class MLP(chainer.Chain):\n",
    "\n",
    "    def __init__(self, n_mid_units=100, n_out=10):\n",
    "        # パラメータを持つ層の登録\n",
    "        super(MLP, self).__init__()\n",
    "        with self.init_scope():\n",
    "            self.l1=L.Linear(None, n_mid_units)\n",
    "            self.l2=L.Linear(None, n_mid_units)\n",
    "            self.l3=L.Linear(None, n_out)\n",
    "\n",
    "\n",
    "    def __call__(self, x):\n",
    "        # データを受け取った際のforward計算を書く\n",
    "        h1 = F.relu(self.l1(x))\n",
    "        h2 = F.relu(self.l2(h1))\n",
    "        return self.l3(h2)\n",
    "\n",
    "gpu_id = 0  # change to -1 if not using GPU\n",
    "\n",
    "model = MLP()\n",
    "if gpu_id >= 0:\n",
    "    model.to_gpu(gpu_id)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "### NOTE\n",
    "\n",
    "ここで、`L.Linear`クラスは全結合層を意味します。コンストラクタの第一引数に`None`を渡すと、実行時に、データがその層に入力された瞬間、必要な数の入力側ユニット数を自動的に計算し、`(n_input)` $\\times$ `n_mid_units`の大きさの行列を作成し、パラメータとして保持します。これは後々、畳み込み層を全結合層の前に配置する際などに便利な機能です。\n",
    "\n",
    "前述のように、`Link`はパラメータを持つので、そのパラメータの値にアクセスすることができます。例えば、上のモデル`MLP`は`l1`という名前の全結合相が登録されています。この全結合相は`W`と`b`という2つのパラメータを持ちます。これらは外からアクセスすることができます。例えば`b`へアクセスするには、以下のようにします。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1つ目の全結合相のバイアスパラメータの形は、 (100,)\n",
      "初期化直後のその値は、 [ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.\n",
      "  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.\n",
      "  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.\n",
      "  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.\n",
      "  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.\n",
      "  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.]\n"
     ]
    }
   ],
   "source": [
    "print('1つ目の全結合相のバイアスパラメータの形は、', model.l1.b.shape)\n",
    "print('初期化直後のその値は、', model.l1.b.data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 4. 最適化手法の選択\n",
    "\n",
    "Chainerは多くの最適化手法を提供しています。`chainer.optimizers`モジュール以下にそれらはあります。ここでは最もシンプルな勾配降下法の手法である`optimizers.SGD`を用います。Optimizerのオブジェクトには、`setup`メソッドを使ってモデル（`Chain`オブジェクト）を渡します。こうすることでOptimizerは、自身が更新すべきモデル内のパラメータを自動的にたどってくれます。\n",
    "\n",
    "他にもいろいろな最適化手法が手軽に試せるので、色々と試してみて結果の変化を見てみてください。例えば、下の`chainer.optimizers.SGD`のうち`SGD`の部分を`MomentumSGD`, `RMSprop`,  `Adam`などに変えて、結果の違いを見てみましょう。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "from chainer import optimizers\n",
    "\n",
    "optimizer = optimizers.SGD(lr=0.01)\n",
    "optimizer.setup(model)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### NOTE\n",
    "\n",
    "今回はSGDのコンストラクタの`lr`という引数に $0.01$ を与えました。この値は学習率として知られ、モデルをうまく訓練して良いパフォーマンスを発揮させるために調整する必要がある重要な**ハイパーパラメータ**として知られています。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 5. 学習ループ\n",
    "\n",
    "いよいよ学習ループです。今回は分類問題なので、`softmax_cross_entropy`というロス関数を使って最小化すべきロスの値を計算します。\n",
    "\n",
    "Chainerでは、`Function`や`Link`を使ってモデルのforward計算を行い、結果と正解ラベルを`Function`の一種でありスカラ値を返すロス関数に渡してやり、ロスの計算を行うと、それは他の`Link`や`Function`と同じく、`Variable`オブジェクトを返します。`Variable`オブジェクトはこれまでの計算過程をあとから逆向きに辿り返すための参照を保持しているため、`Variable.backward`メソッドを呼ぶだけで、自動的にそこからこれまでの計算過程を遡って、途中で施された計算に用いられたパラメータの勾配を計算してくれます。\n",
    "\n",
    "つまり、学習ループの1回の中で行うのは、以下の4項目です。\n",
    "\n",
    "1. モデルにデータを渡して出力`y`を得る\n",
    "2. `y`と正解ラベル`t`を使って、最小化すべきロスの値を`softmax_cross_entropy`関数で計算する\n",
    "3. `softmax_cross_entropy`関数の出力`Variable`の`backward`メソッドを呼んで、モデル内部のパラメータに`grad`プロパティ（これがパラメータ更新に使われる勾配）を持たせる\n",
    "4. Optimizerの`update`メソッドを呼び、3.で計算した`grad`を使って全パラメータを更新する\n",
    "\n",
    "以上です。では実際に訓練ループを書いていきます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch:01 train_loss:0.9434 val_loss:0.7707 val_accuracy:0.8273\n",
      "epoch:02 train_loss:0.4465 val_loss:0.4543 val_accuracy:0.8805\n",
      "epoch:03 train_loss:0.3872 val_loss:0.3693 val_accuracy:0.8978\n",
      "epoch:04 train_loss:0.3494 val_loss:0.3296 val_accuracy:0.9053\n",
      "epoch:05 train_loss:0.3283 val_loss:0.3057 val_accuracy:0.9160\n",
      "epoch:06 train_loss:0.3001 val_loss:0.2900 val_accuracy:0.9176\n",
      "epoch:07 train_loss:0.3056 val_loss:0.2781 val_accuracy:0.9198\n",
      "epoch:08 train_loss:0.4101 val_loss:0.2649 val_accuracy:0.9249\n",
      "epoch:09 train_loss:0.2721 val_loss:0.2546 val_accuracy:0.9254\n",
      "epoch:10 train_loss:0.2789 val_loss:0.2470 val_accuracy:0.9286\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "from chainer.dataset import concat_examples\n",
    "from chainer.cuda import to_cpu\n",
    "\n",
    "max_epoch = 10\n",
    "\n",
    "while train_iter.epoch < max_epoch:\n",
    "    \n",
    "    # ---------- 学習の1イテレーション ----------\n",
    "    train_batch = train_iter.next()\n",
    "    x, t = concat_examples(train_batch, gpu_id)\n",
    "    \n",
    "    # 予測値の計算\n",
    "    y = model(x)\n",
    "\n",
    "    # ロスの計算\n",
    "    loss = F.softmax_cross_entropy(y, t)\n",
    "\n",
    "    # 勾配の計算\n",
    "    model.cleargrads()\n",
    "    loss.backward()\n",
    "\n",
    "    # パラメータの更新\n",
    "    optimizer.update()\n",
    "    # --------------- ここまで ----------------\n",
    "\n",
    "    # 1エポック終了ごとにValidationデータに対する予測精度を測って、\n",
    "    # モデルの汎化性能が向上していることをチェックしよう\n",
    "    if train_iter.is_new_epoch:  # 1 epochが終わったら\n",
    "\n",
    "        # ロスの表示\n",
    "        print('epoch:{:02d} train_loss:{:.04f} '.format(\n",
    "            train_iter.epoch, float(to_cpu(loss.data))), end='')\n",
    "\n",
    "        test_losses = []\n",
    "        test_accuracies = []\n",
    "        while True:\n",
    "            test_batch = test_iter.next()\n",
    "            x_test, t_test = concat_examples(test_batch, gpu_id)\n",
    "\n",
    "            # テストデータをforward\n",
    "            y_test = model(x_test)\n",
    "\n",
    "            # ロスを計算\n",
    "            loss_test = F.softmax_cross_entropy(y_test, t_test)\n",
    "            test_losses.append(to_cpu(loss_test.data))\n",
    "\n",
    "            # 精度を計算\n",
    "            accuracy = F.accuracy(y_test, t_test)\n",
    "            accuracy.to_cpu()\n",
    "            test_accuracies.append(accuracy.data)\n",
    "            \n",
    "            if test_iter.is_new_epoch:\n",
    "                test_iter.epoch = 0\n",
    "                test_iter.current_position = 0\n",
    "                test_iter.is_new_epoch = False\n",
    "                test_iter._pushed_position = None\n",
    "                break\n",
    "\n",
    "        print('val_loss:{:.04f} val_accuracy:{:.04f}'.format(\n",
    "            np.mean(test_losses), np.mean(test_accuracies)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 6. 学習済みモデルを保存する\n",
    "\n",
    "Chainerには2つのシリアライズ機能が用意されています。一つはHDF5形式でモデルを保存するもので、もう一つはNumPyのNPZ形式でモデルを保存するものです。今回は、追加ライブラリのインストールが必要なHDF5ではなく、NumPy標準機能で提供されているシリアライズ機能を利用したNPZ形式でのモデルの保存を行います。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "-rw-rw-r-- 1 crissman crissman 333959 Jun 26 16:13 my_mnist.model\r\n"
     ]
    }
   ],
   "source": [
    "from chainer import serializers\n",
    "\n",
    "serializers.save_npz('my_mnist.model', model)\n",
    "\n",
    "# ちゃんと保存されていることを確認\n",
    "%ls -la my_mnist.model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## 7. 保存したモデルを読み込んで推論する\n",
    "\n",
    "今しがた保存したNPZファイルを読み込んで、テストデータに対して予測を計算させてみます。NPZファイルにはパラメータが保存されているので、forward計算のロジックを持つモデルのオブジェクトをまず作成し、そのパラメータを先程保存したNPZが持つ値で上書きすることで学習直後のモデルの状態を復元します。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<__main__.MLP at 0x7f1deea6a630>"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# まず同じモデルのオブジェクトを作る\n",
    "infer_model = MLP()\n",
    "\n",
    "# そのオブジェクトに保存済みパラメータをロードする\n",
    "serializers.load_npz('my_mnist.model', infer_model)\n",
    "\n",
    "# GPU上で計算させるために、モデルをGPUに送る\n",
    "if gpu_id >= 0:\n",
    "    infer_model.to_gpu(gpu_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAWEAAAFfCAYAAACfj30KAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nCGkRijAhhLQIRZgQQlqEIkwIIS3y/wFYJOCHHvLvRQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f1de8871f28>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "label: 7\n"
     ]
    }
   ],
   "source": [
    "# テストデータ\n",
    "x, t = test[0]\n",
    "plt.imshow(x.reshape(28, 28), cmap='gray')\n",
    "plt.show()\n",
    "print('label:', t)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(784,) -> (1, 784)\n",
      "predicted label: 7\n"
     ]
    }
   ],
   "source": [
    "from chainer.cuda import to_gpu\n",
    "\n",
    "# ミニバッチの形にする（ここではサイズ1のミニバッチにするが、\n",
    "# 複数まとめてサイズnのミニバッチにしてまとめて推論することもできる）\n",
    "print(x.shape, end=' -> ')\n",
    "x = x[None, ...]\n",
    "print(x.shape)\n",
    "\n",
    "# GPU上で計算させるため、データもGPU上に送る\n",
    "if gpu_id >= 0:\n",
    "    x = to_gpu(x, 0)\n",
    "\n",
    "# モデルのforward関数に渡す\n",
    "y = infer_model(x)\n",
    "\n",
    "# Variable形式で出てくるので中身を取り出す\n",
    "y = y.data\n",
    "\n",
    "# 結果をCPUに送る\n",
    "y = to_cpu(y)\n",
    "\n",
    "# 最大値のインデックスを見る\n",
    "pred_label = y.argmax(axis=1)\n",
    "\n",
    "print('predicted label:', pred_label[0])"
   ]
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [default]",
   "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": "3.5.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}