Keras_Advanced_Coding_Style_Challenge.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "# 各種好用套件\n",
    "import os, random, sys, keras, math\n",
    "import numpy as np\n",
    "\n",
    "# 大畫家\n",
    "import matplotlib.patches as mpatches\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "\n",
    "# k~k~k~~~~Keras core sturcture\n",
    "from keras.models import Sequential, Model\n",
    "from keras.layers import Activation, Input, Dense, Lambda, Dropout, concatenate, Flatten, Reshape\n",
    "from keras.initializers import Ones\n",
    "\n",
    "# some\n",
    "from keras import initializers\n",
    "from keras.constraints import max_norm, non_neg, unit_norm\n",
    "from keras import backend as K\n",
    "from keras import metrics\n",
    "\n",
    "# used for RGB to Lab transformation\n",
    "from skimage import color\n",
    "color.colorconv.lab_ref_white = np.array([95.047, 100.0, 108.883])\n",
    "\n",
    "# MNIST:\n",
    "#     Dataset of 60,000 28x28 grayscale images of the 10 digits, along with a test set of 10,000 images.\n",
    "from keras.datasets import mnist\n",
    "# CIFAR10: \n",
    "#     Dataset of 50,000 32x32 color training images, labeled over 10 categories, and 10,000 test images.\n",
    "from keras.datasets import cifar10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "(x_train, y_train), (x_test, y_test) = cifar10.load_data()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_train = x_train/255\n",
    "x_test = x_test/255"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.text.Text at 0xf9e2240>"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAP8AAAEICAYAAACQ6CLfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x4e78550>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "imgNumber = 100\n",
    "\n",
    "rgb_data = x_train[imgNumber]\n",
    "plt.imshow(rgb_data)\n",
    "plt.title('RGB Image')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_train_R = x_train[:,:,:,0]\n",
    "x_train_G = x_train[:,:,:,1]\n",
    "x_train_B = x_train[:,:,:,2]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "x_R_img = Input(shape=(32, 32,), name='R_part')\n",
    "x_R = Flatten()(x_R_img)\n",
    "\n",
    "x_G_img = Input(shape=(32, 32,), name='G_part')\n",
    "x_G = Flatten()(x_G_img)\n",
    "\n",
    "x_B_img = Input(shape=(32, 32,), name='B_part')\n",
    "x_B = Flatten()(x_B_img)\n",
    "\n",
    "enc_R = Dense(100, activation='relu')(x_R)\n",
    "h_R = Dense(2, activation='relu')(enc_R)\n",
    "dec_R = Dense(100, activation='relu')(h_R)\n",
    "dec_R_img = Dense(32*32, activation='sigmoid')(dec_R)\n",
    "x_G_reconstructed = Reshape((32, 32))(dec_R_img)\n",
    "\n",
    "enc_G = Dense(100, activation='relu')(x_G)\n",
    "h_G = Dense(2, activation='relu')(enc_G)\n",
    "dec_G = Dense(100, activation='relu')(h_G)\n",
    "dec_G_img = Dense(32*32, activation='sigmoid')(dec_G)\n",
    "x_B_reconstructed = Reshape((32, 32))(dec_G_img)\n",
    "\n",
    "enc_B = Dense(100, activation='relu')(x_B)\n",
    "h_B = Dense(2, activation='relu')(enc_B)\n",
    "dec_B = Dense(100, activation='relu')(h_B)\n",
    "dec_B_img = Dense(32*32, activation='sigmoid')(dec_B)\n",
    "x_R_reconstructed = Reshape((32, 32))(dec_B_img)\n",
    "\n",
    "model = Model([x_R_img, x_G_img, x_B_img], [x_G_reconstructed, x_B_reconstructed, x_R_reconstructed])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def my_loss(y_true, y_pred):\n",
    "    error_G = x_G_img - x_G_reconstructed\n",
    "    error_B = x_B_img - x_B_reconstructed\n",
    "    error_R = x_R_img - x_R_reconstructed\n",
    "    \n",
    "    mse_G = K.mean(K.square(error_G))\n",
    "    mse_B = K.mean(K.square(error_B))\n",
    "    mse_R = K.mean(K.square(error_R))\n",
    "    return mse_R + mse_G + mse_B"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#model.compile(loss='mse', optimizer='rmsprop')\n",
    "model.compile(loss=my_loss, optimizer='rmsprop')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.fit([x_train_R, x_train_G, x_train_B], [x_train_G, x_train_B, x_train_R], verbose=2,\n",
    "          epochs=10, batch_size=500)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "imgNumber = 2\n",
    "\n",
    "test = x_train[imgNumber : imgNumber+1, :, :, :]\n",
    "test_R = test[:,:,:, 0]\n",
    "test_G = test[:,:,:, 1]\n",
    "test_B = test[:,:,:, 2]\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "G, B, R = model.predict([test_R, test_G, test_B])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "G.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# raw image\n",
    "plt.subplot(251)\n",
    "plt.imshow(test[0])\n",
    "plt.title('raw RGB Image')\n",
    "\n",
    "plt.subplot(252)\n",
    "plt.imshow(test_R[0], cmap='Reds')\n",
    "plt.title('reconstructed R Image')\n",
    "plt.axis('off')\n",
    "\n",
    "plt.subplot(253)\n",
    "plt.imshow(test_G[0], cmap='Greens')\n",
    "plt.title('reconstructed G Image')\n",
    "plt.axis('off')\n",
    "\n",
    "plt.subplot(254)\n",
    "plt.imshow(test_B[0], cmap='Blues')\n",
    "plt.title('reconstructed G Image')\n",
    "plt.axis('off')\n",
    "\n",
    "# reconstructed image\n",
    "plt.subplot(259)\n",
    "plt.imshow(R[0], cmap='Reds')\n",
    "#plt.title('reconstructed R Image')\n",
    "plt.axis('off')\n",
    "\n",
    "plt.subplot(257)\n",
    "plt.imshow(G[0], cmap='Greens')\n",
    "#plt.title('reconstructed G Image')\n",
    "plt.axis('off')\n",
    "\n",
    "plt.subplot(258)\n",
    "plt.imshow(B[0], cmap='Blues')\n",
    "#plt.title('reconstructed G Image')\n",
    "plt.axis('off')\n",
    "\n",
    "plt.subplot(2,5,10)\n",
    "plt.imshow(np.array([R[0], G[0], B[0]]).T)\n",
    "#plt.title('reconstructed Image')\n",
    "plt.axis('off')\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.array([R[0], G[0], B[0]]).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.5.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}