04_VariationalAutoEncoder.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "VQ-VAEの実装の前に、VAEを実装しておく"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Variational Autoencoders\n",
    "変分自己符号化器(Variational AutoEncoder, VAE)は、観測値として得られたデータが何らかの(単純な)確率分布に従う低次元の確率変数から変換されたものであるという前提のもと、それらのデータの変換元である低次元表現の推定、および低次元表現から観測データへの変換を行うニューラルネットワークを学習させるものである。\n",
    "\n",
    "通常の自己符号化器と異なりネットワークの中間層の値(すなわち低次元表現)に特定の確率分布が想定されているため、その分布にしたがって低次元表現をサンプリングしデータへの変換を施すことで、元のデータ群に似通ったデータ群を新たに生成することができる。\n",
    "\n",
    "実装は[https://gluon.mxnet.io/chapter13_unsupervised-learning/vae-gluon.html](https://gluon.mxnet.io/chapter13_unsupervised-learning/vae-gluon.html) を参考に行った。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import mxnet as mx\n",
    "from mxnet import gluon\n",
    "from mxnet import nd\n",
    "import numpy as np\n",
    "\n",
    "ctx = mx.gpu(0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "class VAELoss(gluon.HybridBlock):\n",
    "    def __init__(self, **kwargs):\n",
    "        super(VAELoss, self).__init__(**kwargs)\n",
    "        with self.name_scope():\n",
    "            # エンコーダー部分は、2次元(独立)のガウス分布を出力。\n",
    "            # それぞれの次元で平均と分散(の対数)を推定するので、\n",
    "            # 出力次元は4。\n",
    "            encoder = gluon.nn.HybridSequential()\n",
    "            with encoder.name_scope():\n",
    "                encoder.add(gluon.nn.Dense(units=512, activation='tanh'))\n",
    "                encoder.add(gluon.nn.Dense(units=256, activation='tanh'))\n",
    "                encoder.add(gluon.nn.Dense(units=128, activation='tanh'))\n",
    "                encoder.add(gluon.nn.Dense(units=4, activation=None))\n",
    "            self.encoder = encoder\n",
    "\n",
    "            # デコーダー部分は、画像の各ピクセルにベルヌーイ分布を出力。\n",
    "            # パラメータは[0,1]区間のため、最後はシグモイド関数を通す。\n",
    "            decoder = gluon.nn.HybridSequential()\n",
    "            with decoder.name_scope():\n",
    "                decoder.add(gluon.nn.Dense(units=128, activation='tanh'))\n",
    "                decoder.add(gluon.nn.Dense(units=256, activation='tanh'))\n",
    "                decoder.add(gluon.nn.Dense(units=512, activation='tanh'))\n",
    "                decoder.add(gluon.nn.Dense(units=28*28, activation='sigmoid'))\n",
    "            self.decoder = decoder\n",
    "            \n",
    "    def hybrid_forward(self, F, x):\n",
    "        mean, lvar = F.split(self.encoder(F.flatten(x)), num_outputs=2, axis=-1)\n",
    "        kl = F.sum(.5*(F.exp(lvar)+mean**2-1-lvar), axis=1)\n",
    "        \n",
    "        L=10\n",
    "        mean = F.repeat(mean, repeats=L, axis=0)\n",
    "        lvar = F.repeat(lvar, repeats=L, axis=0)\n",
    "        eps = F.random_normal(loc=0, scale=1, shape=mean.shape, ctx=ctx)\n",
    "        z = mean +  eps * (F.exp(lvar/2))\n",
    "        y = self.decoder(z)\n",
    "        \n",
    "        repx = F.repeat(F.flatten(x), repeats=L, axis=0)\n",
    "        logp = F.sum(repx*F.log(y+1e-10)+ \\\n",
    "                    (1-repx)*F.log(1-y+1e-10), axis=1)\n",
    "        logp = F.mean(F.reshape(logp, (-1,L)), axis=1)\n",
    "        loss = kl - logp\n",
    "        return loss\n",
    "\n",
    "vaeLoss = VAELoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_loader = gluon.data.DataLoader(\n",
    "    gluon.data.vision.MNIST('.', transform=lambda x,y: (mx.nd.transpose(x, [2,0,1]).astype(np.float32)/255.0, y)), batch_size=24)\n",
    "test_loader = gluon.data.DataLoader(\n",
    "    gluon.data.vision.MNIST('.', train=False, transform=lambda x,y: (mx.nd.transpose(x, [2,0,1]).astype(np.float32)/255.0, y)), batch_size=24)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "initer = mx.init.Xavier()\n",
    "vaeLoss.initialize(initer, ctx=ctx, force_reinit=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "training loss: 195.40, test loss: 181.14\n",
      "training loss: 173.53, test loss: 169.70\n",
      "training loss: 163.32, test loss: 161.47\n",
      "training loss: 157.93, test loss: 157.41\n",
      "training loss: 154.42, test loss: 154.10\n",
      "training loss: 152.08, test loss: 152.10\n",
      "training loss: 150.66, test loss: 151.54\n",
      "training loss: 149.88, test loss: 150.42\n",
      "training loss: 147.95, test loss: 148.22\n",
      "training loss: 146.83, test loss: 147.30\n"
     ]
    }
   ],
   "source": [
    "trainer = gluon.Trainer(vaeLoss.collect_params(), 'adam', {'learning_rate': 1e-4})\n",
    "\n",
    "for _ in range(10):\n",
    "    cumulative_loss = 0\n",
    "    for x,y in train_loader:\n",
    "        x = mx.nd.array(x, ctx=ctx)\n",
    "        y = mx.nd.array(y, ctx=ctx)\n",
    "        with mx.autograd.record():\n",
    "            loss = vaeLoss(x)\n",
    "        loss.backward()\n",
    "        trainer.step(24)\n",
    "        cumulative_loss += mx.nd.mean(loss).asscalar()\n",
    "    test_loss = 0\n",
    "    for x, y in test_loader:\n",
    "        x = mx.nd.array(x, ctx=ctx)\n",
    "        y = mx.nd.array(y, ctx=ctx)\n",
    "        loss = vaeLoss(x)\n",
    "        test_loss += mx.nd.mean(loss).asscalar()\n",
    "    print('training loss: %.2f, test loss: %.2f'%(cumulative_loss/len(train_loader), test_loss/len(test_loader)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# テストデータの再構成"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "x, _ = list(test_loader)[0]\n",
    "x = mx.nd.array(x, ctx=ctx)\n",
    "_mean, _lvar = nd.split(vaeLoss.encoder(x), num_outputs=2)\n",
    "z = _mean + nd.exp(_lvar/2) * nd.random_normal(shape=_mean.shape, ctx=ctx)\n",
    "mean = vaeLoss.decoder(z)\n",
    "reconst = nd.reshape(mean, (-1,28,28))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1728x144 with 24 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, ax = plt.subplots(2,12, figsize=(24,2))\n",
    "for i in range(12):\n",
    "    a1, a2 = ax[:, i]\n",
    "    a1.set_xticks([])\n",
    "    a2.set_xticks([])\n",
    "    a1.imshow(x.asnumpy()[i,0,:,:])\n",
    "    #a1.plot([0,28], [14,14], 'r')\n",
    "    a2.imshow(reconst.asnumpy()[i,:,:])\n",
    "    #a2.plot([0,28], [14,14], 'r')\n",
    "    a1.set_xticks([])\n",
    "    a2.set_xticks([])\n",
    "    a1.set_xlim([0,28])\n",
    "    a2.set_xlim([0,28])\n",
    "    a1.set_yticks([])\n",
    "    a2.set_yticks([])"
   ]
  }
 ],
 "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": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}