08_Autoencoder/4-1_Convolutional_Variational_Autoencoder_Interpolation.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Convolutional Variational Autoencoder\n",
    "- MNIST dataset\n",
    "- Variational Inference \n",
    "- Convolutional \n",
    "- [Auto-Encoding Variational Bayes](https://arxiv.org/abs/1312.6114)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Settings\n",
    "### 1) Import required libraries"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "import torch.nn.init as init\n",
    "import torchvision.datasets as dset\n",
    "import torchvision.transforms as transforms\n",
    "from torch.utils.data import DataLoader\n",
    "from torch.autograd import Variable\n",
    "import torchvision.utils as v_utils\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2) Set hyperparameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "batch_size = 128\n",
    "learning_rate = 0.0005\n",
    "num_epoch = 1\n",
    "hidden_size = 2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. Data\n",
    "\n",
    "### 1) Download Data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "mnist_train = dset.MNIST(\"./\", train=True, transform=transforms.ToTensor(), target_transform=None, download=True)\n",
    "mnist_test = dset.MNIST(\"./\", train=False, transform=transforms.ToTensor(), target_transform=None, download=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2) Set DataLoader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "train_loader = torch.utils.data.DataLoader(mnist_train,batch_size=batch_size, shuffle=True,num_workers=2,drop_last=True)\n",
    "test_loader = torch.utils.data.DataLoader(mnist_test,batch_size=batch_size, shuffle=False,num_workers=2,drop_last=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4. Model & Optimizer\n",
    "### 1) Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "class Encoder(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(Encoder,self).__init__()\n",
    "        self.fc1 = nn.Sequential(\n",
    "                        nn.Conv2d(1,8,3,padding=1),   # batch x 8 x 28 x 28\n",
    "                        nn.BatchNorm2d(8),\n",
    "                        nn.ReLU(),\n",
    "                        nn.MaxPool2d(2,2), \n",
    "                        nn.Conv2d(8,16,3,padding=1),  # batch x 16 x 14 x 14\n",
    "                        nn.BatchNorm2d(16),\n",
    "                        nn.ReLU(),\n",
    "                        nn.MaxPool2d(2,2),\n",
    "                        nn.Conv2d(16,32,3,padding=1),  # batch x 32 x 7 x 7\n",
    "                        nn.ReLU(),\n",
    "        )\n",
    "        self.fc2_1 = nn.Sequential(\n",
    "                        nn.Linear(32*7*7, 800),\n",
    "                        nn.Linear(800, hidden_size),\n",
    "        )\n",
    "        self.fc2_2 = nn.Sequential(\n",
    "                        nn.Linear(32*7*7, 800),\n",
    "                        nn.Linear(800, hidden_size),\n",
    "        )\n",
    "        self.relu = nn.ReLU()\n",
    "                        \n",
    "    def encode(self,x):\n",
    "        #x = x.view(batch_size,-1)\n",
    "        out = self.fc1(x)\n",
    "        out = out.view(batch_size,-1)\n",
    "        out = self.relu(out)\n",
    "        mu = self.fc2_1(out)\n",
    "        log_var = self.fc2_2(out)\n",
    "                \n",
    "        return mu,log_var\n",
    "    \n",
    "    def reparametrize(self, mu, logvar):\n",
    "        std = logvar.mul(0.5).exp_()\n",
    "        \n",
    "        eps = torch.FloatTensor(std.size()).normal_()\n",
    "        eps = Variable(eps).cuda()\n",
    "        \n",
    "        return eps.mul(std).add_(mu)\n",
    "    \n",
    "    def forward(self,x):\n",
    "        mu, logvar = self.encode(x)\n",
    "        reparam = self.reparametrize(mu,logvar)\n",
    "        \n",
    "        return mu,logvar,reparam\n",
    "        \n",
    "encoder = Encoder().cuda()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "class Decoder(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(Decoder,self).__init__()\n",
    "        self.fc1 = nn.Sequential(\n",
    "                        nn.Linear(hidden_size,800),\n",
    "                        nn.BatchNorm1d(800),\n",
    "                        nn.ReLU(),\n",
    "                        nn.Linear(800,1568),\n",
    "                        nn.ReLU(),\n",
    "        )\n",
    "        self.fc2 = nn.Sequential(\n",
    "                        nn.ConvTranspose2d(32,16,3,2,1,1),\n",
    "                        nn.ReLU(),\n",
    "                        nn.BatchNorm2d(16),\n",
    "                        nn.ConvTranspose2d(16,8,3,2,1,1),\n",
    "                        nn.ReLU(),\n",
    "                        nn.BatchNorm2d(8),\n",
    "                        nn.ConvTranspose2d(8,1,3,1,1),\n",
    "                        nn.BatchNorm2d(1),\n",
    "        )\n",
    "        self.sigmoid = nn.Sigmoid()\n",
    "        self.relu = nn.ReLU()\n",
    "    \n",
    "    def forward(self,x):\n",
    "        out = self.fc1(x)\n",
    "        out = self.relu(out)\n",
    "        out = out.view(batch_size,32,7,7)\n",
    "        out = self.fc2(out)\n",
    "        out = self.sigmoid(out)\n",
    "        out = out.view(batch_size,28,28,1)\n",
    "        \n",
    "        return out\n",
    "        \n",
    "decoder = Decoder().cuda()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2) Loss func & Optimizer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "reconstruction_function = nn.BCELoss(size_average=False)\n",
    "\n",
    "def loss_function(recon_x, x, mu, logvar):\n",
    "    BCE = reconstruction_function(recon_x, x)\n",
    "\n",
    "    # see Appendix B from VAE paper:\n",
    "    # Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014\n",
    "    # https://arxiv.org/abs/1312.6114\n",
    "    # 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)\n",
    "    KLD_element = mu.pow(2).add_(logvar.exp()).mul_(-1).add_(1).add_(logvar)\n",
    "    KLD = torch.sum(KLD_element).mul_(-0.5)\n",
    "\n",
    "    return BCE + KLD\n",
    "\n",
    "parameters = list(encoder.parameters())+ list(decoder.parameters())\n",
    "optimizer = torch.optim.Adam(parameters, lr=learning_rate)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5. Train "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "--------model restored--------\n",
      "\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/gunho/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/nn/functional.py:767: UserWarning: Using a target size (torch.Size([128, 1, 28, 28])) that is different to the input size (torch.Size([128, 28, 28, 1])) is deprecated. Please ensure they have the same size.\n",
      "  \"Please ensure they have the same size.\".format(target.size(), input.size()))\n",
      "/home/gunho/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/serialization.py:147: UserWarning: Couldn't retrieve source code for container of type Encoder. It won't be checked for correctness upon loading.\n",
      "  \"type \" + obj.__name__ + \". It won't be checked \"\n",
      "/home/gunho/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/serialization.py:147: UserWarning: Couldn't retrieve source code for container of type Decoder. It won't be checked for correctness upon loading.\n",
      "  \"type \" + obj.__name__ + \". It won't be checked \"\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Variable containing:\n",
      " 25204.7480\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25778.5293\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24359.8262\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24821.7793\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24867.2637\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24461.1973\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 23822.7969\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24553.3086\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24723.6035\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25092.3594\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25412.0684\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24468.2949\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24313.4023\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24653.1680\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25153.8145\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24523.4102\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24593.1367\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24181.1992\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25287.2559\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24906.5742\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24553.4062\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24803.1758\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 23882.0488\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24268.8809\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 26207.7969\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24545.0625\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24963.0547\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25970.0273\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25052.2910\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25019.2578\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24450.9590\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24767.3105\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25527.9102\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24445.2246\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25131.3691\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25247.7051\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24961.4609\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24447.7344\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25085.9668\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24941.0898\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24094.2207\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25296.0078\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25018.1250\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24475.7246\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 25278.3711\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24111.0586\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n",
      "Variable containing:\n",
      " 24870.4648\n",
      "[torch.cuda.FloatTensor of size 1 (GPU 0)]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "try:\n",
    "    encoder, decoder = torch.load('./model/conv_variational_autoencoder_interpolation.pkl')\n",
    "    print(\"\\n--------model restored--------\\n\")\n",
    "except:\n",
    "    print(\"\\n--------model not restored--------\\n\")\n",
    "    pass\n",
    "\n",
    "for i in range(num_epoch):\n",
    "    for j,[image,label] in enumerate(train_loader):\n",
    "        optimizer.zero_grad()\n",
    "        \n",
    "        image = Variable(image).cuda()\n",
    "        mu,log_var,reparam = encoder(image)\n",
    "        output = decoder(reparam)\n",
    "        \n",
    "        loss = loss_function(output, image, mu, log_var)\n",
    "        \n",
    "        loss.backward()\n",
    "        optimizer.step()\n",
    "\n",
    "        if j % 10 == 0:\n",
    "            torch.save([encoder,decoder],'./model/conv_variational_autoencoder_interpolation.pkl')\n",
    "            print(loss)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6. Check with Train Image"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([128, 28, 28])\n"
     ]
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\ngg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f47f17ed400>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f4798555e80>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nMfxEQTH8REEx/ERB/T/nSfRpgIXxqQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f47957fdf60>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n1P8DWZs2Nya6jQAAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f479858f320>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAD8CAYAAAC4nHJkAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nfqKgmPxEQTH5iYJi8hMF9b+9kIyoRyroKgAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f4795804f60>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "out_img = torch.squeeze(output.cpu().data)\n",
    "print(out_img.size())\n",
    "\n",
    "for i in range(5):\n",
    "    #plt.imshow(torch.squeeze(image[i]).numpy(),cmap='gray')\n",
    "    #plt.show()\n",
    "    plt.imshow(out_img[i].numpy(),cmap='gray')\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch.Size([128, 1, 28, 28])\n"
     ]
    }
   ],
   "source": [
    "for i in range(1):\n",
    "    for j,[image,label] in enumerate(test_loader):\n",
    "       \n",
    "        image = Variable(image,volatile=True).cuda()\n",
    "        mu,log_var,reparam = encoder(image)\n",
    "        \n",
    "        for k in range(11):\n",
    "            for l in range(11):\n",
    "                reparam[10*k+l,0]=(k-5)/1.5\n",
    "                reparam[10*k+l,1]=(l-5)/1.5\n",
    "        \n",
    "        output = decoder(reparam)\n",
    "            \n",
    "output = output.data.unsqueeze_(1).squeeze_(4)\n",
    "print(output.size())\n",
    "v_utils.save_image(output[0:121],\"./result/gen.png\",nrow=11)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![alt text](./result/gen.png)"
   ]
  }
 ],
 "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": "3.6.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}