add Conditional GANs section

Author: demon386

imgs/mirza-arxiv-14.png

@@ -34,7 +34,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 4,
+   "execution_count": 30,
    "metadata": {
     "autoscroll": false,
     "collapsed": true,
@@ -50,7 +50,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 31,
    "metadata": {
     "autoscroll": false,
     "collapsed": true,
@@ -335,7 +335,6 @@
    ]
   },
   {
-   "attachments": {},
    "cell_type": "markdown",
    "metadata": {},
    "source": [
@@ -353,7 +352,6 @@
    ]
   },
   {
-   "attachments": {},
    "cell_type": "markdown",
    "metadata": {},
    "source": [
@@ -396,7 +394,9 @@
   {
    "cell_type": "code",
    "execution_count": 31,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "# 产生一组数据，模型不知道数据怎么产生的。\n",
@@ -432,7 +432,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 59,
+   "execution_count": 8,
    "metadata": {
     "collapsed": true
    },
@@ -491,7 +491,9 @@
   {
    "cell_type": "code",
    "execution_count": 56,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "dm = Discriminator(2)\n",
@@ -584,7 +586,9 @@
   {
    "cell_type": "code",
    "execution_count": 101,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "criterion = nn.BCELoss() # binary cross entropy loss\n",
@@ -602,7 +606,9 @@
   {
    "cell_type": "code",
    "execution_count": 146,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "num_epochs = 2000\n",
@@ -688,7 +694,9 @@
   {
    "cell_type": "code",
    "execution_count": 157,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "gm_init = Generator(2, 2)\n",
@@ -698,7 +706,9 @@
   {
    "cell_type": "code",
    "execution_count": 158,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "g_input_data = g_input.data.numpy()\n",
@@ -732,13 +742,141 @@
     "    ax.legend()"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "collapsed": true
+   },
+   "source": [
+    "## Conditional GANs\n",
+    "以 mnist dataset 为例，解释 Conditional GANs"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### MNIST 数据集读取\n",
+    "- 一个著名的手写数字数据集"
+   ]
+  },
   {
    "cell_type": "code",
-   "execution_count": null,
+   "execution_count": 77,
    "metadata": {
     "collapsed": true
    },
    "outputs": [],
+   "source": [
+    "import torchvision.datasets as dset\n",
+    "import torchvision.transforms as transforms\n",
+    "import os"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 78,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "if not os.path.exists('./data'):\n",
+    "    os.mkdir('./data')\n",
+    "    \n",
+    "train_set = dset.MNIST(root='./data', train=True, \n",
+    "                       transform=transforms.Compose([transforms.ToTensor(),                                        \n",
+    "                                                     transforms.Normalize((0.5,), (1.0,))]),\n",
+    "                       download=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 70,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "dataloader = torch.utils.data.DataLoader(train_set, batch_size=200, shuffle=True)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Take a look at the data"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 71,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "data_sample = next(iter(dataloader))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "- dataloader 迭代返回的每个元素是一个 list, 其第一个元素是这批 batch 的图像，第二个元素是这批 batch 的 label\n",
+    "- 第一个元素是一个 FloatTensor, 大小为 (200 x 1 x 28 x 28)，即 (batch_size x channel x height x width)。因为是灰度图像，所以只有一个 channel\n",
+    "- 因为 channel 这个维度没用，所以我们用 view 把它消解"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 72,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "torch.Size([200, 1, 28, 28])\n",
+      "torch.Size([200, 28, 28])\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(data_sample[0].size())\n",
+    "print(data_sample[0].view(-1, 28, 28).size())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 74,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQmCC\n",
+      "text/plain": [
+       "<matplotlib.figure.Figure at 0x1114ba128>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "plt.imshow(data_sample[0].view(-1, 28, 28)[0].numpy(), cmap='gray');"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "对于这样一个数据集，我们想用 GANs 的 Generator 自己产生对应的数字。如果用原始的 GANs 形式，需要针对 0-9 这 10 个数字产生对应 10 组模型。对更多类的问题，需要的模型会更多。而且各个模型之间互相独立，没有办法「共享」一部分他们学习的知识，这不是一个很好的解决方案。\n",
+    "\n",
+    "针对这个问题，Conditional GANs [1] 给出的方案是除了噪音输入 Z，给将每一个 $y$ 的信息也输入给 DM 及 GM（比如 embedding 形式）。其模型示意图（取自 [1]）及问题形式如下：\n",
+    "\n",
+    "![](./imgs/mirza-arxiv-14.png)\n",
+    "\n",
+    "$$ \\min_G\\max_D V(D, G) = E_{x \\sim P_{data}(x|y)}[\\log D(x)] + E_{z \\sim P_z(z)}[\\log (1 - D(G(z|y))]$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
    "source": []
   }
  ],