API tools faq
paste
Advertisement
Guest User

Untitled

a guest
Sep 29th, 2016
90
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 9.26 KB | None | 0 0
raw download clone embed print report
  1. {
  2. "cells": [
  3. {
  4. "cell_type": "markdown",
  5. "metadata": {},
  6. "source": [
  7. "## 加载包"
  8. ]
  9. },
  10. {
  11. "cell_type": "code",
  12. "execution_count": 1,
  13. "metadata": {
  14. "collapsed": false
  15. },
  16. "outputs": [
  17. {
  18. "name": "stdout",
  19. "output_type": "stream",
  20. "text": [
  21. "Packages loaded\n"
  22. ]
  23. }
  24. ],
  25. "source": [
  26. "import scipy.io\n",
  27. "import numpy as np \n",
  28. "import os \n",
  29. "import scipy.misc \n",
  30. "import matplotlib.pyplot as plt \n",
  31. "import tensorflow as tf\n",
  32. "%matplotlib inline \n",
  33. "print (\"Packages loaded\")"
  34. ]
  35. },
  36. {
  37. "cell_type": "markdown",
  38. "metadata": {},
  39. "source": [
  40. "## 定义网络结构"
  41. ]
  42. },
  43. {
  44. "cell_type": "code",
  45. "execution_count": 2,
  46. "metadata": {
  47. "collapsed": true
  48. },
  49. "outputs": [],
  50. "source": [
  51. "IMAGE_W = 800 \n",
  52. "IMAGE_H = 600 \n",
  53. "cwd = os.getcwd()\n",
  54. "# 内容图片文档\n",
  55. "CONTENT_IMG = cwd + \"/images/Taipei101.jpg\"\n",
  56. "# 风格图片文档\n",
  57. "STYLE_IMG = cwd + \"/images/StarryNight.jpg\"\n",
  58. "# 输出结果的目录和文档名\n",
  59. "OUTOUT_DIR = './images'\n",
  60. "OUTPUT_IMG = 'results.png'\n",
  61. "# VGG模型文件\n",
  62. "VGG_MODEL = cwd + \"/data/imagenet-vgg-verydeep-19.mat\"\n",
  63. "INI_NOISE_RATIO = 0.7\n",
  64. "STYLE_STRENGTH = 500\n",
  65. "ITERATION = 5000\n",
  66. "\n",
  67. "CONTENT_LAYERS =[('conv4_2',1.)]\n",
  68. "STYLE_LAYERS=[('conv1_1',1.),('conv2_1',1.5),('conv3_1',2.),('conv4_1',2.5),('conv5_1',3.)]\n",
  69. "\n",
  70. "\n",
  71. "MEAN_VALUES = np.array([123, 117, 104]).reshape((1,1,1,3))\n"
  72. ]
  73. },
  74. {
  75. "cell_type": "code",
  76. "execution_count": 3,
  77. "metadata": {
  78. "collapsed": true
  79. },
  80. "outputs": [],
  81. "source": [
  82. "# 定义前向计算函数，如果是conv层则计算卷积，如果是pool则进行池化\n",
  83. "def build_net(ntype, nin, nwb=None):\n",
  84. " if ntype == 'conv':\n",
  85. " return tf.nn.relu(tf.nn.conv2d(nin, nwb[0], strides=[1, 1, 1, 1], padding='SAME')+ nwb[1])\n",
  86. " elif ntype == 'pool':\n",
  87. " return tf.nn.avg_pool(nin, ksize=[1, 2, 2, 1],\n",
  88. " strides=[1, 2, 2, 1], padding='SAME')\n",
  89. "\n",
  90. "# 从VGG模型中提取参数\n",
  91. "def get_weight_bias(vgg_layers, i,):\n",
  92. " weights = vgg_layers[i][0][0][0][0][0]\n",
  93. " weights = tf.constant(weights)\n",
  94. " bias = vgg_layers[i][0][0][0][0][1]\n",
  95. " bias = tf.constant(np.reshape(bias, (bias.size)))\n",
  96. " return weights, bias\n",
  97. "\n",
  98. "# 构建VGG模型网络结构，从现成的VGG模型文档中读取参数\n",
  99. "# 以conv1_1层参数为例，长下面这个样子\n",
  100. "# (<tf.Tensor 'Const_83:0' shape=(3, 3, 3, 64) dtype=float32>,\n",
  101. "# <tf.Tensor 'Const_84:0' shape=(64,) dtype=float32>)\n",
  102. "# conv1_1层输出长下面这个样子\n",
  103. "# <tf.Tensor 'Relu_32:0' shape=(1, 600, 800, 64) dtype=float32>\n",
  104. "\n",
  105. "def build_vgg19(path):\n",
  106. " net = {}\n",
  107. " vgg_rawnet = scipy.io.loadmat(path)\n",
  108. " vgg_layers = vgg_rawnet['layers'][0]\n",
  109. " net['input'] = tf.Variable(np.zeros((1, IMAGE_H, IMAGE_W, 3)).astype('float32'))\n",
  110. " net['conv1_1'] = build_net('conv',net['input'],get_weight_bias(vgg_layers,0))\n",
  111. " net['conv1_2'] = build_net('conv',net['conv1_1'],get_weight_bias(vgg_layers,2))\n",
  112. " net['pool1'] = build_net('pool',net['conv1_2'])\n",
  113. " net['conv2_1'] = build_net('conv',net['pool1'],get_weight_bias(vgg_layers,5))\n",
  114. " net['conv2_2'] = build_net('conv',net['conv2_1'],get_weight_bias(vgg_layers,7))\n",
  115. " net['pool2'] = build_net('pool',net['conv2_2'])\n",
  116. " net['conv3_1'] = build_net('conv',net['pool2'],get_weight_bias(vgg_layers,10))\n",
  117. " net['conv3_2'] = build_net('conv',net['conv3_1'],get_weight_bias(vgg_layers,12))\n",
  118. " net['conv3_3'] = build_net('conv',net['conv3_2'],get_weight_bias(vgg_layers,14))\n",
  119. " net['conv3_4'] = build_net('conv',net['conv3_3'],get_weight_bias(vgg_layers,16))\n",
  120. " net['pool3'] = build_net('pool',net['conv3_4'])\n",
  121. " net['conv4_1'] = build_net('conv',net['pool3'],get_weight_bias(vgg_layers,19))\n",
  122. " net['conv4_2'] = build_net('conv',net['conv4_1'],get_weight_bias(vgg_layers,21))\n",
  123. " net['conv4_3'] = build_net('conv',net['conv4_2'],get_weight_bias(vgg_layers,23))\n",
  124. " net['conv4_4'] = build_net('conv',net['conv4_3'],get_weight_bias(vgg_layers,25))\n",
  125. " net['pool4'] = build_net('pool',net['conv4_4'])\n",
  126. " net['conv5_1'] = build_net('conv',net['pool4'],get_weight_bias(vgg_layers,28))\n",
  127. " net['conv5_2'] = build_net('conv',net['conv5_1'],get_weight_bias(vgg_layers,30))\n",
  128. " net['conv5_3'] = build_net('conv',net['conv5_2'],get_weight_bias(vgg_layers,32))\n",
  129. " net['conv5_4'] = build_net('conv',net['conv5_3'],get_weight_bias(vgg_layers,34))\n",
  130. " net['pool5'] = build_net('pool',net['conv5_4'])\n",
  131. " return net\n",
  132. "\n",
  133. "# 内容损失函数\n",
  134. "def build_content_loss(p, x):\n",
  135. " M = p.shape[1]*p.shape[2]\n",
  136. " N = p.shape[3]\n",
  137. " loss = (1./(2* N**0.5 * M**0.5 )) * tf.reduce_sum(tf.pow((x - p),2)) \n",
  138. " return loss\n",
  139. "\n",
  140. "\n",
  141. "def gram_matrix(x, area, depth):\n",
  142. " x1 = tf.reshape(x,(area,depth))\n",
  143. " g = tf.matmul(tf.transpose(x1), x1)\n",
  144. " return g\n",
  145. "\n",
  146. "def gram_matrix_val(x, area, depth):\n",
  147. " x1 = x.reshape(area,depth)\n",
  148. " g = np.dot(x1.T, x1)\n",
  149. " return g\n",
  150. "\n",
  151. "# 风格损失函数，A为风格标准图片，G为训练后的结果图片\n",
  152. "def build_style_loss(a, x):\n",
  153. " M = a.shape[1]*a.shape[2]\n",
  154. " N = a.shape[3]\n",
  155. " A = gram_matrix_val(a, M, N )\n",
  156. " G = gram_matrix(x, M, N )\n",
  157. " loss = (1./(4 * N**2 * M**2)) * tf.reduce_sum(tf.pow((G - A),2))\n",
  158. " return loss\n",
  159. "\n",
  160. "\n",
  161. "# 读取图片函数，同时做白化\n",
  162. "def read_image(path):\n",
  163. " image = scipy.misc.imread(path)\n",
  164. " image = image[np.newaxis,:IMAGE_H,:IMAGE_W,:] \n",
  165. " image = image - MEAN_VALUES\n",
  166. " return image\n",
  167. "\n",
  168. "# 写图片函数\n",
  169. "def write_image(path, image):\n",
  170. " image = image + MEAN_VALUES\n",
  171. " image = image[0]\n",
  172. " image = np.clip(image, 0, 255).astype('uint8')\n",
  173. " scipy.misc.imsave(path, image)\n"
  174. ]
  175. },
  176. {
  177. "cell_type": "markdown",
  178. "metadata": {},
  179. "source": [
  180. "## 定义主函数"
  181. ]
  182. },
  183. {
  184. "cell_type": "code",
  185. "execution_count": 18,
  186. "metadata": {
  187. "collapsed": true
  188. },
  189. "outputs": [],
  190. "source": [
  191. "def main():\n",
  192. " net = build_vgg19(VGG_MODEL)\n",
  193. " sess = tf.Session()\n",
  194. " sess.run(tf.initialize_all_variables())\n",
  195. "# 建立一个纯噪音图片做为训练参数，使内容符合内容图片，而风格符合风格图片\n",
  196. " noise_img = np.random.uniform(-20, 20, (1, IMAGE_H, IMAGE_W, 3)).astype('float32')\n",
  197. " content_img = read_image(CONTENT_IMG)\n",
  198. " style_img = read_image(STYLE_IMG)\n",
  199. "# 将内容图片输入到VGG网络中，取出conv4_2层输出结果，计算内容损失\n",
  200. " sess.run([net['input'].assign(content_img)])\n",
  201. " cost_content = sum(map(lambda l,: l[1]*build_content_loss(sess.run(net[l[0]]) , net[l[0]])\n",
  202. " , CONTENT_LAYERS))\n",
  203. "# 将风格图片输入到VGG网络中，取出conv1_1-conv5_1五个层的输出结果，计算风格损失\n",
  204. " sess.run([net['input'].assign(style_img)])\n",
  205. " cost_style = sum(map(lambda l: l[1]*build_style_loss(sess.run(net[l[0]]) , net[l[0]])\n",
  206. " , STYLE_LAYERS))\n",
  207. "# 加总两种损失做为最小化训练目标，用cost_style做为调整系数\n",
  208. " cost_total = cost_content + STYLE_STRENGTH * cost_style\n",
  209. " optimizer = tf.train.AdamOptimizer(2.0)\n",
  210. "\n",
  211. " train = optimizer.minimize(cost_total)\n",
  212. " sess.run(tf.initialize_all_variables())\n",
  213. "# 把内容图片加噪音后，做为VGG网络输入层，算法将学习去调整这个输入层，来使得训练目标最小\n",
  214. " sess.run(net['input'].assign( INI_NOISE_RATIO* noise_img + (1.-INI_NOISE_RATIO) * content_img))\n",
  215. "\n",
  216. " if not os.path.exists(OUTOUT_DIR):\n",
  217. " os.mkdir(OUTOUT_DIR)\n",
  218. "\n",
  219. " for i in range(500):\n",
  220. " sess.run(train)\n",
  221. " print i\n",
  222. " if i%100 ==0:\n",
  223. " result_img = sess.run(net['input'])\n",
  224. " print sess.run(cost_total)\n",
  225. " write_image(os.path.join(OUTOUT_DIR,'%s.png'%(str(i).zfill(4))),result_img)\n",
  226. " \n",
  227. " write_image(os.path.join(OUTOUT_DIR,OUTPUT_IMG),result_img)\n"
  228. ]
  229. },
  230. {
  231. "cell_type": "code",
  232. "execution_count": null,
  233. "metadata": {
  234. "collapsed": true
  235. },
  236. "outputs": [],
  237. "source": [
  238. "main()"
  239. ]
  240. },
  241. {
  242. "cell_type": "code",
  243. "execution_count": null,
  244. "metadata": {
  245. "collapsed": true
  246. },
  247. "outputs": [],
  248. "source": []
  249. }
  250. ],
  251. "metadata": {
  252. "kernelspec": {
  253. "display_name": "Python 2",
  254. "language": "python",
  255. "name": "python2"
  256. },
  257. "language_info": {
  258. "codemirror_mode": {
  259. "name": "ipython",
  260. "version": 2
  261. },
  262. "file_extension": ".py",
  263. "mimetype": "text/x-python",
  264. "name": "python",
  265. "nbconvert_exporter": "python",
  266. "pygments_lexer": "ipython2",
  267. "version": "2.7.11"
  268. }
  269. },
  270. "nbformat": 4,
  271. "nbformat_minor": 0
  272. }
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement
Public Pastes
Advertisement
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!