├── .DS_Store ├── .ipynb_checkpoints ├── LV3-checkpoint.ipynb ├── LV3_0.12-checkpoint.ipynb ├── XOR_gate-checkpoint.ipynb └── core_code-checkpoint.ipynb ├── JupyterNotebook_version ├── FNN.py ├── LICENSE ├── LV3.ipynb ├── LV3_0.12.ipynb ├── RNN │ ├── .DS_Store │ ├── .ipynb_checkpoints │ │ ├── BGRU_BLSTM_scan_LV2-checkpoint.ipynb │ │ └── LSTM_scan_LV1-checkpoint.ipynb │ ├── BGRU_BLSTM_scan_LV2.ipynb │ ├── LSTM_scan_LV1.ipynb │ ├── X.npy │ └── Y.npy ├── X.npy ├── XOR_gate.ipynb ├── Y.npy └── core_code.ipynb ├── Py_version ├── FNNs_Demo │ ├── X.npy │ ├── Y.npy │ ├── demoLV1.py │ ├── demoLV2.py │ └── demoLV3.py └── RNNs_Demo │ ├── LSTMLV1.py │ ├── Variant_LSTMLV2.py │ ├── X.npy │ └── Y.npy ├── README.md └── 深层学习设计理念加文字.pdf /.DS_Store: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/YJango/tensorflow_basic_tutorial/6a6110b91f4a035e0db2b316cf353d27338c4f0e/.DS_Store -------------------------------------------------------------------------------- /.ipynb_checkpoints/LV3-checkpoint.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 1, 6 | "metadata": { 7 | "collapsed": true 8 | }, 9 | "outputs": [], 10 | "source": [ 11 | "import tensorflow as tf\n", 12 | "import numpy as np\n", 13 | "import matplotlib.pyplot as plt\n", 14 | "%matplotlib inline\n", 15 | "tf.reset_default_graph()" 16 | ] 17 | }, 18 | { 19 | "cell_type": "code", 20 | "execution_count": 2, 21 | "metadata": {}, 22 | "outputs": [], 23 | "source": [ 24 | "class FNN(object):\n", 25 | " \"\"\"Build a general FeedForward neural network\n", 26 | " Parameters\n", 27 | " ----------\n", 28 | " learning_rate : float\n", 29 | " drop_out : float\n", 30 | " Layers : list\n", 31 | " The number of layers\n", 32 | " N_hidden : list\n", 33 | " The numbers of nodes in layers\n", 34 | " D_input : int\n", 35 | " Input dimension\n", 36 | " D_label : int\n", 37 | " Label dimension\n", 38 | " Task_type : string\n", 39 | " 'regression' or 'classification'\n", 40 | " L2_lambda : float\n", 41 | " Auther : YJango; 2016/11/25\n", 42 | " \"\"\"\n", 43 | " def __init__(self, learning_rate, Layers, N_hidden, D_input, D_label, Task_type='regression', L2_lambda=0.0):\n", 44 | " \n", 45 | " #var\n", 46 | " self.learning_rate = learning_rate\n", 47 | " self.Layers = Layers\n", 48 | " self.N_hidden = N_hidden\n", 49 | " self.D_input = D_input\n", 50 | " self.D_label = D_label\n", 51 | " # 类型控制loss函数的选择\n", 52 | " self.Task_type = Task_type\n", 53 | " # l2 regularization的惩罚强弱，过高会使得输出都拉向0\n", 54 | " self.L2_lambda = L2_lambda\n", 55 | " # 用于存放所累积的每层l2 regularization\n", 56 | " self.l2_penalty = tf.constant(0.0)\n", 57 | " \n", 58 | " # 用于生成tensorflow缩放图的,括号里起名字\n", 59 | " with tf.name_scope('Input'):\n", 60 | " self.inputs = tf.placeholder(tf.float32, [None, D_input], name=\"inputs\")\n", 61 | " with tf.name_scope('Label'):\n", 62 | " self.labels = tf.placeholder(tf.float32, [None, D_label], name=\"labels\")\n", 63 | " with tf.name_scope('keep_rate'):\n", 64 | " self.drop_keep_rate = tf.placeholder(tf.float32, name=\"dropout_keep\")\n", 65 | " \n", 66 | "\n", 67 | " # 初始化的时候直接生成，build方法是后面会建立的\n", 68 | " self.build('F')\n", 69 | " \n", 70 | " def weight_init(self,shape):\n", 71 | " # shape : list [in_dim, out_dim]\n", 72 | " # 在这里更改初始化方法\n", 73 | " # 方式1：下面的权重初始化若用ReLU激活函数，可以使用带有6个隐藏层的神经网络\n", 74 | " # 若过深，则使用dropout会难以拟合。\n", 75 | " #initial = tf.truncated_normal(shape, stddev=0.1)/ np.sqrt(shape[1])\n", 76 | " # 方式2：下面的权重初始化若用ReLU激活函数，可以扩展到15个隐藏层以上（通常不会用那么多）\n", 77 | " initial = tf.random_uniform(shape,minval=-np.sqrt(5)*np.sqrt(1.0/shape[0]), maxval=np.sqrt(5)*np.sqrt(1.0/shape[0]))\n", 78 | " return tf.Variable(initial)\n", 79 | "\n", 80 | " def bias_init(self,shape):\n", 81 | " # can change initialization here\n", 82 | " initial = tf.constant(0.1, shape=shape)\n", 83 | " return tf.Variable(initial)\n", 84 | " \n", 85 | " def variable_summaries(self, var, name):\n", 86 | " with tf.name_scope(name+'_summaries'):\n", 87 | " mean = tf.reduce_mean(var)\n", 88 | " tf.summary.scalar('mean/' + name, mean)\n", 89 | " with tf.name_scope(name+'_stddev'):\n", 90 | " stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))\n", 91 | " # 记录每次训练后变量的数值变化\n", 92 | " tf.summary.scalar('_stddev/' + name, stddev)\n", 93 | " tf.summary.scalar_('_max/' + name, tf.reduce_max(var))\n", 94 | " tf.summary.scalar_('_min/' + name, tf.reduce_min(var))\n", 95 | " tf.summary.histogram(name, var)\n", 96 | "\n", 97 | " def layer(self,in_tensor, in_dim, out_dim, layer_name, act=tf.nn.relu):\n", 98 | " with tf.name_scope(layer_name):\n", 99 | " with tf.name_scope(layer_name+'_weights'):\n", 100 | " # 用所建立的weight_init函数进行初始化。\n", 101 | " weights = self.weight_init([in_dim, out_dim])\n", 102 | " # 存放着每一个权重W\n", 103 | " self.W.append(weights)\n", 104 | " # 对权重进行统计\n", 105 | " self.variable_summaries(weights, layer_name + '/weights')\n", 106 | " with tf.name_scope(layer_name+'_biases'):\n", 107 | " biases = self.bias_init([out_dim])\n", 108 | " # 存放着每一个偏移b\n", 109 | " self.b.append(biases)\n", 110 | " self.variable_summaries(biases, layer_name + '/biases')\n", 111 | " with tf.name_scope(layer_name+'_Wx_plus_b'):\n", 112 | " # 计算Wx+b\n", 113 | " pre_activate = tf.matmul(in_tensor, weights) + biases\n", 114 | " # 记录直方图\n", 115 | " tf.summary.histogram(layer_name + '/pre_activations', pre_activate)\n", 116 | " # 计算a(Wx+b)\n", 117 | " activations = act(pre_activate, name='activation')\n", 118 | " tf.summary.histogram(layer_name + '/activations', activations)\n", 119 | " # 最终返回该层的输出，以及权重W的L2\n", 120 | " return activations, tf.nn.l2_loss(weights)\n", 121 | "\n", 122 | " def drop_layer(self,in_tensor):\n", 123 | " #tf.scalar_summary('dropout_keep', self.drop_keep_rate)\n", 124 | " dropped = tf.nn.dropout(in_tensor, self.drop_keep_rate)\n", 125 | " return dropped\n", 126 | "\n", 127 | " def build(self, prefix):\n", 128 | " # 建立网络 \n", 129 | " # incoming也代表当前tensor的流动位置\n", 130 | " incoming = self.inputs\n", 131 | " # 如果没有隐藏层\n", 132 | " if self.Layers!=0:\n", 133 | " layer_nodes = [self.D_input] + self.N_hidden\n", 134 | " else:\n", 135 | " layer_nodes = [self.D_input]\n", 136 | " \n", 137 | " # hid_layers用于存储所有隐藏层的输出\n", 138 | " self.hid_layers=[]\n", 139 | " # W用于存储所有层的权重\n", 140 | " self.W=[]\n", 141 | " # b用于存储所有层的偏移\n", 142 | " self.b=[]\n", 143 | " # total_l2用于存储所有层的L2\n", 144 | " self.total_l2=[]\n", 145 | " \n", 146 | " # 开始叠加隐藏层。这跟千层饼没什么区别。\n", 147 | " for l in range(self.Layers):\n", 148 | " # 使用刚才编写的函数来建立层，并更新incoming的位置\n", 149 | " incoming,l2_loss= self.layer(incoming,layer_nodes[l],layer_nodes[l+1],prefix+'_hid_'+str(l+1),act=tf.nn.relu)\n", 150 | " # 累计l2\n", 151 | " self.total_l2.append(l2_loss)\n", 152 | " # 输出一些信息，让我们知道网络在建造中做了什么\n", 153 | " print('Add dense layer: relu')\n", 154 | " print(' %sD --> %sD' %(layer_nodes[l],layer_nodes[l+1]))\n", 155 | " # 存储所有隐藏层的输出\n", 156 | " self.hid_layers.append(incoming)\n", 157 | " # 加入dropout层\n", 158 | " incoming = self.drop_layer(incoming)\n", 159 | " \n", 160 | " # 输出层的建立。输出层需要特别对待的原因是输出层的activation function要根据任务来变。\n", 161 | " # 回归任务的话，下面用的是tf.identity，也就是没有activation function\n", 162 | " if self.Task_type=='regression':\n", 163 | " out_act=tf.identity\n", 164 | " else:\n", 165 | " # 分类任务使用softmax来拟合概率\n", 166 | " out_act=tf.nn.softmax\n", 167 | " self.output,l2_loss= self.layer(incoming,layer_nodes[-1],self.D_label, layer_name='output',act=out_act)\n", 168 | " self.total_l2.append(l2_loss)\n", 169 | " print('Add output layer: linear')\n", 170 | " print(' %sD --> %sD' %(layer_nodes[-1],self.D_label))\n", 171 | " \n", 172 | " # l2 loss的缩放图\n", 173 | " with tf.name_scope('total_l2'):\n", 174 | " for l2 in self.total_l2:\n", 175 | " self.l2_penalty+=l2\n", 176 | " tf.summary.scalar('l2_penalty', self.l2_penalty)\n", 177 | " \n", 178 | " # 不同任务的loss\n", 179 | " # 若为回归，则loss是用于判断所有预测值和实际值差别的函数。\n", 180 | " if self.Task_type=='regression':\n", 181 | " with tf.name_scope('SSE'):\n", 182 | " self.loss=tf.reduce_mean((self.output-self.labels)**2)\n", 183 | " self.loss2=tf.nn.l2_loss(self.output-self.labels)\n", 184 | " \n", 185 | " tf.summary.scalar('loss', self.loss)\n", 186 | " else:\n", 187 | " # 若为分类，cross entropy的loss function\n", 188 | " entropy = tf.nn.softmax_cross_entropy_with_logits(self.output, self.labels)\n", 189 | " with tf.name_scope('cross entropy'):\n", 190 | " self.loss = tf.reduce_mean(entropy)\n", 191 | " tf.summary.scalar('loss', self.loss)\n", 192 | " with tf.name_scope('accuracy'):\n", 193 | " correct_prediction = tf.equal(tf.argmax(self.output, 1), tf.argmax(self.labels, 1))\n", 194 | " self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n", 195 | " tf.summary.scalar('accuracy', self.accuracy)\n", 196 | " \n", 197 | " # 整合所有loss，形成最终loss\n", 198 | " with tf.name_scope('total_loss'):\n", 199 | " self.total_loss=self.loss + self.l2_penalty*self.L2_lambda\n", 200 | " tf.summary.scalar('total_loss', self.total_loss)\n", 201 | " \n", 202 | " # 训练操作\n", 203 | " with tf.name_scope('train'):\n", 204 | " self.train_step = tf.train.AdamOptimizer(self.learning_rate).minimize(self.total_loss)\n", 205 | "\n", 206 | " # 洗牌功能\n", 207 | " def shufflelists(self,lists):\n", 208 | " ri=np.random.permutation(len(lists[1]))\n", 209 | " out=[]\n", 210 | " for l in lists:\n", 211 | " out.append(l[ri])\n", 212 | " return out" 213 | ] 214 | }, 215 | { 216 | "cell_type": "code", 217 | "execution_count": 3, 218 | "metadata": { 219 | "collapsed": true 220 | }, 221 | "outputs": [], 222 | "source": [ 223 | "def Standardize(seq):\n", 224 | " #subtract mean\n", 225 | " centerized=seq-np.mean(seq, axis = 0)\n", 226 | " #divide standard deviation\n", 227 | " normalized=centerized/np.std(centerized, axis = 0)\n", 228 | " return normalized\n", 229 | "def Makewindows(indata,window_size=41):\n", 230 | " outdata=[]\n", 231 | " mid=int(window_size/2)\n", 232 | " indata=np.vstack((np.zeros((mid,indata.shape[1])),indata,np.zeros((mid,indata.shape[1]))))\n", 233 | " for i in range(indata.shape[0]-window_size+1):\n", 234 | " outdata.append(np.hstack(indata[i:i+window_size]))\n", 235 | " return np.array(outdata)" 236 | ] 237 | }, 238 | { 239 | "cell_type": "code", 240 | "execution_count": 4, 241 | "metadata": { 242 | "collapsed": true 243 | }, 244 | "outputs": [], 245 | "source": [ 246 | "mfc=np.load('X.npy')\n", 247 | "art=np.load('Y.npy')\n", 248 | "x=[]\n", 249 | "y=[]\n", 250 | "for i in range(len(mfc)):\n", 251 | " x.append(Makewindows(Standardize(mfc[i])))\n", 252 | " y.append(Standardize(art[i]))\n", 253 | "vali_size=20\n", 254 | "totalsamples=len(np.vstack(x))\n", 255 | "X_train=np.vstack(x)[int(totalsamples/vali_size):].astype(\"float32\")\n", 256 | "Y_train=np.vstack(y)[int(totalsamples/vali_size):].astype(\"float32\")\n", 257 | "\n", 258 | "X_test=np.vstack(x)[:int(totalsamples/vali_size)].astype(\"float32\")\n", 259 | "Y_test=np.vstack(y)[:int(totalsamples/vali_size)].astype(\"float32\")" 260 | ] 261 | }, 262 | { 263 | "cell_type": "code", 264 | "execution_count": 5, 265 | "metadata": {}, 266 | "outputs": [ 267 | { 268 | "name": "stdout", 269 | "output_type": "stream", 270 | "text": [ 271 | "((37500, 1599), (37500, 24), (1973, 1599), (1973, 24))\n" 272 | ] 273 | } 274 | ], 275 | "source": [ 276 | "print(X_train.shape,Y_train.shape,X_test.shape,Y_test.shape)" 277 | ] 278 | }, 279 | { 280 | "cell_type": "code", 281 | "execution_count": 6, 282 | "metadata": {}, 283 | "outputs": [ 284 | { 285 | "name": "stdout", 286 | "output_type": "stream", 287 | "text": [ 288 | "Add dense layer: relu\n", 289 | " 1599D --> 2048D\n", 290 | "Add dense layer: relu\n", 291 | " 2048D --> 1024D\n", 292 | "Add dense layer: relu\n", 293 | " 1024D --> 512D\n", 294 | "Add dense layer: relu\n", 295 | " 512D --> 256D\n", 296 | "Add dense layer: relu\n", 297 | " 256D --> 128D\n", 298 | "Add output layer: linear\n", 299 | " 128D --> 24D\n" 300 | ] 301 | } 302 | ], 303 | "source": [ 304 | "# 生成网络实例\n", 305 | "ff=FNN(learning_rate=7e-5, Layers=5, N_hidden=[2048,1024,512,256,128], D_input=1599, D_label=24, L2_lambda=1e-4)" 306 | ] 307 | }, 308 | { 309 | "cell_type": "code", 310 | "execution_count": 7, 311 | "metadata": { 312 | "collapsed": true 313 | }, 314 | "outputs": [], 315 | "source": [ 316 | "sess = tf.InteractiveSession()\n", 317 | "sess.run(tf.global_variables_initializer())\n", 318 | "merged = tf.summary.merge_all()\n", 319 | "train_writer = tf.summary.FileWriter('log3' + '/train',sess.graph)\n", 320 | "test_writer = tf.summary.FileWriter('log3' + '/test')" 321 | ] 322 | }, 323 | { 324 | "cell_type": "code", 325 | "execution_count": 8, 326 | "metadata": {}, 327 | "outputs": [], 328 | "source": [ 329 | "def plots(T,P,i, n=21,length=400):\n", 330 | " m=0\n", 331 | " plt.figure(figsize=(20,16))\n", 332 | " plt.subplot(411)\n", 333 | " plt.plot(T[m:m+length,7],'--')\n", 334 | " plt.plot(P[m:m+length,7])\n", 335 | "\n", 336 | " plt.subplot(412)\n", 337 | " plt.plot(T[m:m+length,8],'--')\n", 338 | " plt.plot(P[m:m+length,8])\n", 339 | " \n", 340 | " plt.subplot(413)\n", 341 | " plt.plot(T[m:m+length,15],'--')\n", 342 | " plt.plot(P[m:m+length,15])\n", 343 | " \n", 344 | " plt.subplot(414)\n", 345 | " plt.plot(T[m:m+length,16],'--')\n", 346 | " plt.plot(P[m:m+length,16])\n", 347 | " plt.legend(['True','Predicted'])\n", 348 | " plt.savefig('epoch'+str(i)+'.png')\n", 349 | " plt.close()" 350 | ] 351 | }, 352 | { 353 | "cell_type": "code", 354 | "execution_count": 9, 355 | "metadata": { 356 | "scrolled": true 357 | }, 358 | "outputs": [ 359 | { 360 | "name": "stdout", 361 | "output_type": "stream", 362 | "text": [ 363 | "epoch0 | train_loss:0.693693 |test_loss:0.695782\n", 364 | "epoch1 | train_loss:0.561398 |test_loss:0.583177\n", 365 | "epoch2 | train_loss:0.478702 |test_loss:0.54134\n", 366 | "epoch3 | train_loss:0.389257 |test_loss:0.482584\n", 367 | "epoch4 | train_loss:0.320454 |test_loss:0.453059\n", 368 | "epoch5 | train_loss:0.27317 |test_loss:0.441251\n", 369 | "epoch6 | train_loss:0.234799 |test_loss:0.436898\n", 370 | "epoch7 | train_loss:0.203948 |test_loss:0.436239\n", 371 | "epoch8 | train_loss:0.176413 |test_loss:0.426064\n", 372 | "epoch9 | train_loss:0.156074 |test_loss:0.425965\n", 373 | "epoch10 | train_loss:0.132211 |test_loss:0.414978\n", 374 | "epoch11 | train_loss:0.120031 |test_loss:0.40392\n", 375 | "epoch12 | train_loss:0.105933 |test_loss:0.407968\n", 376 | "epoch13 | train_loss:0.100558 |test_loss:0.402615\n", 377 | "epoch14 | train_loss:0.0894171 |test_loss:0.39814\n", 378 | "epoch15 | train_loss:0.0805371 |test_loss:0.395325\n", 379 | "epoch16 | train_loss:0.0755967 |test_loss:0.396785\n", 380 | "epoch17 | train_loss:0.0714638 |test_loss:0.399197\n", 381 | "epoch18 | train_loss:0.0633848 |test_loss:0.399693\n", 382 | "epoch19 | train_loss:0.0614749 |test_loss:0.393158\n", 383 | "epoch20 | train_loss:0.0591606 |test_loss:0.394801\n", 384 | "epoch21 | train_loss:0.055984 |test_loss:0.38638\n", 385 | "epoch22 | train_loss:0.0540141 |test_loss:0.390381\n", 386 | "epoch23 | train_loss:0.0496812 |test_loss:0.384337\n", 387 | "epoch24 | train_loss:0.049253 |test_loss:0.387584\n", 388 | "epoch25 | train_loss:0.0445422 |test_loss:0.383592\n", 389 | "epoch26 | train_loss:0.0432337 |test_loss:0.382902\n", 390 | "epoch27 | train_loss:0.0451354 |test_loss:0.382096\n", 391 | "epoch28 | train_loss:0.041636 |test_loss:0.385363\n", 392 | "epoch29 | train_loss:0.0392896 |test_loss:0.380636\n", 393 | "epoch30 | train_loss:0.0386417 |test_loss:0.386117\n", 394 | "epoch31 | train_loss:0.0378463 |test_loss:0.38561\n", 395 | "epoch32 | train_loss:0.036046 |test_loss:0.383354\n", 396 | "epoch33 | train_loss:0.0359106 |test_loss:0.381586\n", 397 | "epoch34 | train_loss:0.0344725 |test_loss:0.379078\n", 398 | "epoch35 | train_loss:0.0337404 |test_loss:0.376501\n", 399 | "epoch36 | train_loss:0.0340016 |test_loss:0.379459\n", 400 | "epoch37 | train_loss:0.032026 |test_loss:0.37459\n", 401 | "epoch38 | train_loss:0.0317964 |test_loss:0.37812\n", 402 | "epoch39 | train_loss:0.033109 |test_loss:0.37789\n", 403 | "epoch40 | train_loss:0.0311886 |test_loss:0.372208\n", 404 | "epoch41 | train_loss:0.0303357 |test_loss:0.370058\n", 405 | "epoch42 | train_loss:0.0300332 |test_loss:0.37759\n", 406 | "epoch43 | train_loss:0.0293583 |test_loss:0.36993\n", 407 | "epoch44 | train_loss:0.0287083 |test_loss:0.373513\n", 408 | "epoch45 | train_loss:0.0296841 |test_loss:0.371504\n", 409 | "epoch46 | train_loss:0.0279586 |test_loss:0.370936\n", 410 | "epoch47 | train_loss:0.0285195 |test_loss:0.373591\n", 411 | "epoch48 | train_loss:0.0281992 |test_loss:0.37003\n", 412 | "epoch49 | train_loss:0.0267218 |test_loss:0.368262\n" 413 | ] 414 | } 415 | ], 416 | "source": [ 417 | "# 训练并记录\n", 418 | "k=0\n", 419 | "Batch=32\n", 420 | "for i in range(50):\n", 421 | " idx=0\n", 422 | " X0,Y0=ff.shufflelists([X_train,Y_train])\n", 423 | " while idxlen(layer_nodes):\n", 150 | " nodes_in=layer_nodes[-1]\n", 151 | " nodes_out=layer_nodes[-1]\n", 152 | " else:\n", 153 | " nodes_in=layer_nodes[l]\n", 154 | " nodes_out=layer_nodes[l+1]\n", 155 | " incoming,l2_loss= self.layer(incoming,nodes_in,nodes_out,prefix+'_hid_'+str(l+1),act=tf.nn.relu)\n", 156 | " # 累计l2\n", 157 | " self.total_l2.append(l2_loss)\n", 158 | " # 输出一些信息，让我们知道网络在建造中做了什么\n", 159 | " print('Add dense layer: relu')\n", 160 | " print(' %sD --> %sD' %(nodes_in,nodes_out))\n", 161 | " # 存储所有隐藏层的输出\n", 162 | " self.hid_layers.append(incoming)\n", 163 | " # 加入dropout层\n", 164 | " incoming = self.drop_layer(incoming)\n", 165 | " \n", 166 | " # 输出层的建立。输出层需要特别对待的原因是输出层的activation function要根据任务来变。\n", 167 | " # 回归任务的话，下面用的是tf.identity，也就是没有activation function\n", 168 | " if self.Task_type=='regression':\n", 169 | " out_act=tf.identity\n", 170 | " else:\n", 171 | " # 分类任务使用softmax来拟合概率\n", 172 | " out_act=tf.nn.softmax\n", 173 | " self.output,l2_loss= self.layer(incoming,layer_nodes[-1],self.D_label, layer_name='output',act=out_act)\n", 174 | " self.total_l2.append(l2_loss)\n", 175 | " print('Add output layer: linear')\n", 176 | " print(' %sD --> %sD' %(layer_nodes[-1],self.D_label))\n", 177 | " \n", 178 | " # l2 loss的缩放图\n", 179 | " with tf.name_scope('total_l2'):\n", 180 | " for l2 in self.total_l2:\n", 181 | " self.l2_penalty+=l2\n", 182 | " tf.summary.histogram('l2_penalty', self.l2_penalty)\n", 183 | " \n", 184 | " # 不同任务的loss\n", 185 | " # 若为回归，则loss是用于判断所有预测值和实际值差别的函数。\n", 186 | " if self.Task_type=='regression':\n", 187 | " with tf.name_scope('SSE'):\n", 188 | " self.loss=tf.reduce_mean((self.output-self.labels)**2)\n", 189 | " self.loss2=tf.nn.l2_loss(self.output-self.labels)\n", 190 | " \n", 191 | " tf.summary.histogram('loss', self.loss)\n", 192 | " else:\n", 193 | " # 若为分类，cross entropy的loss function\n", 194 | " entropy = tf.nn.softmax_cross_entropy_with_logits(self.output, self.labels)\n", 195 | " with tf.name_scope('cross entropy'):\n", 196 | " self.loss = tf.reduce_mean(entropy)\n", 197 | " tf.summary.histogram('loss', self.loss)\n", 198 | " with tf.name_scope('accuracy'):\n", 199 | " correct_prediction = tf.equal(tf.argmax(self.output, 1), tf.argmax(self.labels, 1))\n", 200 | " self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n", 201 | " tf.summary.histogram('accuracy', self.accuracy)\n", 202 | " \n", 203 | " # 整合所有loss，形成最终loss\n", 204 | " with tf.name_scope('total_loss'):\n", 205 | " self.total_loss=self.loss + self.l2_penalty*self.L2_lambda\n", 206 | " tf.summary.histogram('total_loss', self.total_loss)\n", 207 | " \n", 208 | " # 训练操作\n", 209 | " with tf.name_scope('train'):\n", 210 | " self.train_step = tf.train.AdamOptimizer(self.learning_rate).minimize(self.total_loss)\n", 211 | "\n", 212 | " # 洗牌功能\n", 213 | " def shufflelists(self,lists):\n", 214 | " ri=np.random.permutation(len(lists[1]))\n", 215 | " out=[]\n", 216 | " for l in lists:\n", 217 | " out.append(l[ri])\n", 218 | " return out" 219 | ] 220 | }, 221 | { 222 | "cell_type": "code", 223 | "execution_count": 3, 224 | "metadata": { 225 | "collapsed": true 226 | }, 227 | "outputs": [], 228 | "source": [ 229 | "def Standardize(seq):\n", 230 | " #subtract mean\n", 231 | " centerized=seq-np.mean(seq, axis = 0)\n", 232 | " #divide standard deviation\n", 233 | " normalized=centerized/np.std(centerized, axis = 0)\n", 234 | " return normalized\n", 235 | "def Makewindows(indata,window_size=41):\n", 236 | " outdata=[]\n", 237 | " mid=int(window_size/2)\n", 238 | " indata=np.vstack((np.zeros((mid,indata.shape[1])),indata,np.zeros((mid,indata.shape[1]))))\n", 239 | " for i in range(indata.shape[0]-window_size+1):\n", 240 | " outdata.append(np.hstack(indata[i:i+window_size]))\n", 241 | " return np.array(outdata)" 242 | ] 243 | }, 244 | { 245 | "cell_type": "code", 246 | "execution_count": 4, 247 | "metadata": {}, 248 | "outputs": [], 249 | "source": [ 250 | "mfc=np.load('X.npy')\n", 251 | "art=np.load('Y.npy')\n", 252 | "x=[]\n", 253 | "y=[]\n", 254 | "for i in range(len(mfc)):\n", 255 | " x.append(Makewindows(Standardize(mfc[i])))\n", 256 | " y.append(Standardize(art[i]))\n", 257 | "vali_size=20\n", 258 | "totalsamples=len(np.vstack(x))\n", 259 | "X_train=np.vstack(x)[int(totalsamples/vali_size):].astype(\"float32\")\n", 260 | "Y_train=np.vstack(y)[int(totalsamples/vali_size):].astype(\"float32\")\n", 261 | "\n", 262 | "X_test=np.vstack(x)[:int(totalsamples/vali_size)].astype(\"float32\")\n", 263 | "Y_test=np.vstack(y)[:int(totalsamples/vali_size)].astype(\"float32\")" 264 | ] 265 | }, 266 | { 267 | "cell_type": "code", 268 | "execution_count": 5, 269 | "metadata": {}, 270 | "outputs": [ 271 | { 272 | "name": "stdout", 273 | "output_type": "stream", 274 | "text": [ 275 | "((37500, 1599), (37500, 24), (1973, 1599), (1973, 24))\n" 276 | ] 277 | } 278 | ], 279 | "source": [ 280 | "print(X_train.shape,Y_train.shape,X_test.shape,Y_test.shape)" 281 | ] 282 | }, 283 | { 284 | "cell_type": "code", 285 | "execution_count": 9, 286 | "metadata": { 287 | "scrolled": true 288 | }, 289 | "outputs": [ 290 | { 291 | "name": "stdout", 292 | "output_type": "stream", 293 | "text": [ 294 | "Add dense layer: relu\n", 295 | " 1599D --> 1024D\n", 296 | "Add dense layer: relu\n", 297 | " 1024D --> 1024D\n", 298 | "Add dense layer: relu\n", 299 | " 1024D --> 1024D\n", 300 | "Add dense layer: relu\n", 301 | " 1024D --> 1024D\n", 302 | "Add dense layer: relu\n", 303 | " 1024D --> 1024D\n", 304 | "Add dense layer: relu\n", 305 | " 1024D --> 1024D\n", 306 | "Add dense layer: relu\n", 307 | " 1024D --> 1024D\n", 308 | "Add dense layer: relu\n", 309 | " 1024D --> 1024D\n", 310 | "Add dense layer: relu\n", 311 | " 1024D --> 1024D\n", 312 | "Add dense layer: relu\n", 313 | " 1024D --> 1024D\n", 314 | "Add dense layer: relu\n", 315 | " 1024D --> 1024D\n", 316 | "Add dense layer: relu\n", 317 | " 1024D --> 1024D\n", 318 | "Add dense layer: relu\n", 319 | " 1024D --> 1024D\n", 320 | "Add dense layer: relu\n", 321 | " 1024D --> 1024D\n", 322 | "Add dense layer: relu\n", 323 | " 1024D --> 1024D\n", 324 | "Add dense layer: relu\n", 325 | " 1024D --> 1024D\n", 326 | "Add dense layer: relu\n", 327 | " 1024D --> 1024D\n", 328 | "Add dense layer: relu\n", 329 | " 1024D --> 1024D\n", 330 | "Add dense layer: relu\n", 331 | " 1024D --> 1024D\n", 332 | "Add dense layer: relu\n", 333 | " 1024D --> 1024D\n", 334 | "Add dense layer: relu\n", 335 | " 1024D --> 1024D\n", 336 | "Add dense layer: relu\n", 337 | " 1024D --> 1024D\n", 338 | "Add dense layer: relu\n", 339 | " 1024D --> 1024D\n", 340 | "Add dense layer: relu\n", 341 | " 1024D --> 1024D\n", 342 | "Add dense layer: relu\n", 343 | " 1024D --> 1024D\n", 344 | "Add dense layer: relu\n", 345 | " 1024D --> 1024D\n", 346 | "Add dense layer: relu\n", 347 | " 1024D --> 1024D\n", 348 | "Add dense layer: relu\n", 349 | " 1024D --> 1024D\n", 350 | "Add dense layer: relu\n", 351 | " 1024D --> 1024D\n", 352 | "Add dense layer: relu\n", 353 | " 1024D --> 1024D\n", 354 | "Add dense layer: relu\n", 355 | " 1024D --> 1024D\n", 356 | "Add dense layer: relu\n", 357 | " 1024D --> 1024D\n", 358 | "Add dense layer: relu\n", 359 | " 1024D --> 1024D\n", 360 | "Add dense layer: relu\n", 361 | " 1024D --> 1024D\n", 362 | "Add dense layer: relu\n", 363 | " 1024D --> 1024D\n", 364 | "Add dense layer: relu\n", 365 | " 1024D --> 1024D\n", 366 | "Add dense layer: relu\n", 367 | " 1024D --> 1024D\n", 368 | "Add dense layer: relu\n", 369 | " 1024D --> 1024D\n", 370 | "Add dense layer: relu\n", 371 | " 1024D --> 1024D\n", 372 | "Add dense layer: relu\n", 373 | " 1024D --> 1024D\n", 374 | "Add dense layer: relu\n", 375 | " 1024D --> 1024D\n", 376 | "Add dense layer: relu\n", 377 | " 1024D --> 1024D\n", 378 | "Add dense layer: relu\n", 379 | " 1024D --> 1024D\n", 380 | "Add dense layer: relu\n", 381 | " 1024D --> 1024D\n", 382 | "Add dense layer: relu\n", 383 | " 1024D --> 1024D\n", 384 | "Add dense layer: relu\n", 385 | " 1024D --> 1024D\n", 386 | "Add dense layer: relu\n", 387 | " 1024D --> 1024D\n", 388 | "Add dense layer: relu\n", 389 | " 1024D --> 1024D\n", 390 | "Add dense layer: relu\n", 391 | " 1024D --> 1024D\n", 392 | "Add dense layer: relu\n", 393 | " 1024D --> 1024D\n", 394 | "Add output layer: linear\n", 395 | " 1024D --> 24D\n" 396 | ] 397 | } 398 | ], 399 | "source": [ 400 | "# 生成网络实例\n", 401 | "# 如果不指定全部隐藏层节点，则多出的隐藏层全部按最后一个指定的隐藏层节点数设定\n", 402 | "ff=FNN(learning_rate=7e-5, Layers=50,N_hidden=[1024], D_input=1599, D_label=24, L2_lambda=1e-4)" 403 | ] 404 | }, 405 | { 406 | "cell_type": "code", 407 | "execution_count": 10, 408 | "metadata": {}, 409 | "outputs": [], 410 | "source": [ 411 | "sess = tf.InteractiveSession()\n", 412 | "sess.run(tf.global_variables_initializer())\n", 413 | "merged = tf.summary.merge_all()\n", 414 | "train_writer = tf.summary.FileWriter('log' + '/train',sess.graph)\n", 415 | "test_writer = tf.summary.FileWriter('log' + '/test')" 416 | ] 417 | }, 418 | { 419 | "cell_type": "code", 420 | "execution_count": 11, 421 | "metadata": {}, 422 | "outputs": [], 423 | "source": [ 424 | "def plots(T,P,i, n=21,length=400):\n", 425 | " m=0\n", 426 | " plt.figure(figsize=(20,16))\n", 427 | " plt.subplot(411)\n", 428 | " plt.plot(T[m:m+length,7],'--')\n", 429 | " plt.plot(P[m:m+length,7])\n", 430 | "\n", 431 | " plt.subplot(412)\n", 432 | " plt.plot(T[m:m+length,8],'--')\n", 433 | " plt.plot(P[m:m+length,8])\n", 434 | " \n", 435 | " plt.subplot(413)\n", 436 | " plt.plot(T[m:m+length,15],'--')\n", 437 | " plt.plot(P[m:m+length,15])\n", 438 | " \n", 439 | " plt.subplot(414)\n", 440 | " plt.plot(T[m:m+length,16],'--')\n", 441 | " plt.plot(P[m:m+length,16])\n", 442 | " plt.legend(['True','Predicted'])\n", 443 | " plt.savefig('epoch'+str(i)+'.png')\n", 444 | " plt.close()" 445 | ] 446 | }, 447 | { 448 | "cell_type": "code", 449 | "execution_count": 14, 450 | "metadata": { 451 | "scrolled": true 452 | }, 453 | "outputs": [ 454 | { 455 | "name": "stdout", 456 | "output_type": "stream", 457 | "text": [ 458 | "epoch0 | train_loss:[0.60085243] |test_loss:0.659336\n", 459 | "epoch1 | train_loss:[0.56255817] |test_loss:0.653993\n", 460 | "epoch2 | train_loss:[0.53966987] |test_loss:0.644001\n", 461 | "epoch3 | train_loss:[0.47888839] |test_loss:0.609307\n", 462 | "epoch4 | train_loss:[0.4274255] |test_loss:0.580894\n", 463 | "epoch5 | train_loss:[0.41663638] |test_loss:0.58645\n", 464 | "epoch6 | train_loss:[0.39932683] |test_loss:0.594072\n", 465 | "epoch7 | train_loss:[0.35524291] |test_loss:0.57222\n", 466 | "epoch8 | train_loss:[0.35336751] |test_loss:0.565518\n", 467 | "epoch9 | train_loss:[0.30439419] |test_loss:0.545433\n" 468 | ] 469 | } 470 | ], 471 | "source": [ 472 | "# 训练并记录\n", 473 | "k=0\n", 474 | "EPOCH=10\n", 475 | "Batch=256\n", 476 | "for i in range(EPOCH):\n", 477 | " idx=0\n", 478 | " X0,Y0=ff.shufflelists([X_train,Y_train])\n", 479 | " while idx %sD' %(layer_nodes[l],layer_nodes[l+1])) 117 | self.hid_layers.append(incoming) 118 | #drop out layer 119 | incoming = self.drop_layer(incoming) 120 | 121 | #output layer 122 | self.output,l2_loss= self.layer(incoming,layer_nodes[-1],self.D_label, layer_name='output',act=tf.identity) 123 | self.total_l2.append(l2_loss) 124 | print('Add output layer: linear') 125 | print(' %sD --> %sD' %(layer_nodes[-1],self.D_label)) 126 | 127 | #loss 128 | with tf.name_scope('total_l2'): 129 | for l2 in self.total_l2: 130 | self.l2_penalty+=l2 131 | tf.summary.scalar_('l2_penalty', self.l2_penalty) 132 | 133 | if self.Task_type=='regression': 134 | with tf.name_scope('SSE'): 135 | self.loss=tf.reduce_mean(tf.nn.l2_loss((self.output - self.labels))) 136 | tf.summary.scalar('loss', self.loss) 137 | else: 138 | entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.output, labels=self.labels) 139 | with tf.name_scope('cross entropy'): 140 | self.loss = tf.reduce_mean(entropy) 141 | tf.summary.scalar('loss', self.loss) 142 | with tf.name_scope('accuracy'): 143 | correct_prediction = tf.equal(tf.argmax(self.output, 1), tf.argmax(self.labels, 1)) 144 | self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) 145 | tf.summary.scalar('accuracy', self.accuracy) 146 | 147 | with tf.name_scope('total_loss'): 148 | self.total_loss=self.loss + self.l2_penalty*self.L2_lambda 149 | tf.summary.scalar_('total_loss', self.total_loss) 150 | 151 | #train 152 | with tf.name_scope('train'): 153 | self.train_step = tf.train.AdamOptimizer(self.learning_rate).minimize(self.total_loss) 154 | 155 | 156 | def shufflelists(self,lists): 157 | ri=np.random.permutation(len(lists[1])) 158 | out=[] 159 | for l in lists: 160 | out.append(l[ri]) 161 | return out 162 | -------------------------------------------------------------------------------- /JupyterNotebook_version/LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | {project} Copyright (C) {year} {fullname} 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /JupyterNotebook_version/LV3.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 1, 6 | "metadata": { 7 | "collapsed": true 8 | }, 9 | "outputs": [], 10 | "source": [ 11 | "import tensorflow as tf\n", 12 | "import numpy as np\n", 13 | "import matplotlib.pyplot as plt\n", 14 | "%matplotlib inline\n", 15 | "tf.reset_default_graph()" 16 | ] 17 | }, 18 | { 19 | "cell_type": "code", 20 | "execution_count": 2, 21 | "metadata": {}, 22 | "outputs": [], 23 | "source": [ 24 | "class FNN(object):\n", 25 | " \"\"\"Build a general FeedForward neural network\n", 26 | " Parameters\n", 27 | " ----------\n", 28 | " learning_rate : float\n", 29 | " drop_out : float\n", 30 | " Layers : list\n", 31 | " The number of layers\n", 32 | " N_hidden : list\n", 33 | " The numbers of nodes in layers\n", 34 | " D_input : int\n", 35 | " Input dimension\n", 36 | " D_label : int\n", 37 | " Label dimension\n", 38 | " Task_type : string\n", 39 | " 'regression' or 'classification'\n", 40 | " L2_lambda : float\n", 41 | " Auther : YJango; 2016/11/25\n", 42 | " \"\"\"\n", 43 | " def __init__(self, learning_rate, Layers, N_hidden, D_input, D_label, Task_type='regression', L2_lambda=0.0):\n", 44 | " \n", 45 | " #var\n", 46 | " self.learning_rate = learning_rate\n", 47 | " self.Layers = Layers\n", 48 | " self.N_hidden = N_hidden\n", 49 | " self.D_input = D_input\n", 50 | " self.D_label = D_label\n", 51 | " # 类型控制loss函数的选择\n", 52 | " self.Task_type = Task_type\n", 53 | " # l2 regularization的惩罚强弱，过高会使得输出都拉向0\n", 54 | " self.L2_lambda = L2_lambda\n", 55 | " # 用于存放所累积的每层l2 regularization\n", 56 | " self.l2_penalty = tf.constant(0.0)\n", 57 | " \n", 58 | " # 用于生成tensorflow缩放图的,括号里起名字\n", 59 | " with tf.name_scope('Input'):\n", 60 | " self.inputs = tf.placeholder(tf.float32, [None, D_input], name=\"inputs\")\n", 61 | " with tf.name_scope('Label'):\n", 62 | " self.labels = tf.placeholder(tf.float32, [None, D_label], name=\"labels\")\n", 63 | " with tf.name_scope('keep_rate'):\n", 64 | " self.drop_keep_rate = tf.placeholder(tf.float32, name=\"dropout_keep\")\n", 65 | " \n", 66 | "\n", 67 | " # 初始化的时候直接生成，build方法是后面会建立的\n", 68 | " self.build('F')\n", 69 | " \n", 70 | " def weight_init(self,shape):\n", 71 | " # shape : list [in_dim, out_dim]\n", 72 | " # 在这里更改初始化方法\n", 73 | " # 方式1：下面的权重初始化若用ReLU激活函数，可以使用带有6个隐藏层的神经网络\n", 74 | " # 若过深，则使用dropout会难以拟合。\n", 75 | " #initial = tf.truncated_normal(shape, stddev=0.1)/ np.sqrt(shape[1])\n", 76 | " # 方式2：下面的权重初始化若用ReLU激活函数，可以扩展到15个隐藏层以上（通常不会用那么多）\n", 77 | " initial = tf.random_uniform(shape,minval=-np.sqrt(5)*np.sqrt(1.0/shape[0]), maxval=np.sqrt(5)*np.sqrt(1.0/shape[0]))\n", 78 | " return tf.Variable(initial)\n", 79 | "\n", 80 | " def bias_init(self,shape):\n", 81 | " # can change initialization here\n", 82 | " initial = tf.constant(0.1, shape=shape)\n", 83 | " return tf.Variable(initial)\n", 84 | " \n", 85 | " def variable_summaries(self, var, name):\n", 86 | " with tf.name_scope(name+'_summaries'):\n", 87 | " mean = tf.reduce_mean(var)\n", 88 | " tf.summary.scalar('mean/' + name, mean)\n", 89 | " with tf.name_scope(name+'_stddev'):\n", 90 | " stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))\n", 91 | " # 记录每次训练后变量的数值变化\n", 92 | " tf.summary.scalar('_stddev/' + name, stddev)\n", 93 | " tf.summary.scalar_('_max/' + name, tf.reduce_max(var))\n", 94 | " tf.summary.scalar_('_min/' + name, tf.reduce_min(var))\n", 95 | " tf.summary.histogram(name, var)\n", 96 | "\n", 97 | " def layer(self,in_tensor, in_dim, out_dim, layer_name, act=tf.nn.relu):\n", 98 | " with tf.name_scope(layer_name):\n", 99 | " with tf.name_scope(layer_name+'_weights'):\n", 100 | " # 用所建立的weight_init函数进行初始化。\n", 101 | " weights = self.weight_init([in_dim, out_dim])\n", 102 | " # 存放着每一个权重W\n", 103 | " self.W.append(weights)\n", 104 | " # 对权重进行统计\n", 105 | " self.variable_summaries(weights, layer_name + '/weights')\n", 106 | " with tf.name_scope(layer_name+'_biases'):\n", 107 | " biases = self.bias_init([out_dim])\n", 108 | " # 存放着每一个偏移b\n", 109 | " self.b.append(biases)\n", 110 | " self.variable_summaries(biases, layer_name + '/biases')\n", 111 | " with tf.name_scope(layer_name+'_Wx_plus_b'):\n", 112 | " # 计算Wx+b\n", 113 | " pre_activate = tf.matmul(in_tensor, weights) + biases\n", 114 | " # 记录直方图\n", 115 | " tf.summary.histogram(layer_name + '/pre_activations', pre_activate)\n", 116 | " # 计算a(Wx+b)\n", 117 | " activations = act(pre_activate, name='activation')\n", 118 | " tf.summary.histogram(layer_name + '/activations', activations)\n", 119 | " # 最终返回该层的输出，以及权重W的L2\n", 120 | " return activations, tf.nn.l2_loss(weights)\n", 121 | "\n", 122 | " def drop_layer(self,in_tensor):\n", 123 | " #tf.scalar_summary('dropout_keep', self.drop_keep_rate)\n", 124 | " dropped = tf.nn.dropout(in_tensor, self.drop_keep_rate)\n", 125 | " return dropped\n", 126 | "\n", 127 | " def build(self, prefix):\n", 128 | " # 建立网络 \n", 129 | " # incoming也代表当前tensor的流动位置\n", 130 | " incoming = self.inputs\n", 131 | " # 如果没有隐藏层\n", 132 | " if self.Layers!=0:\n", 133 | " layer_nodes = [self.D_input] + self.N_hidden\n", 134 | " else:\n", 135 | " layer_nodes = [self.D_input]\n", 136 | " \n", 137 | " # hid_layers用于存储所有隐藏层的输出\n", 138 | " self.hid_layers=[]\n", 139 | " # W用于存储所有层的权重\n", 140 | " self.W=[]\n", 141 | " # b用于存储所有层的偏移\n", 142 | " self.b=[]\n", 143 | " # total_l2用于存储所有层的L2\n", 144 | " self.total_l2=[]\n", 145 | " \n", 146 | " # 开始叠加隐藏层。这跟千层饼没什么区别。\n", 147 | " for l in range(self.Layers):\n", 148 | " # 使用刚才编写的函数来建立层，并更新incoming的位置\n", 149 | " incoming,l2_loss= self.layer(incoming,layer_nodes[l],layer_nodes[l+1],prefix+'_hid_'+str(l+1),act=tf.nn.relu)\n", 150 | " # 累计l2\n", 151 | " self.total_l2.append(l2_loss)\n", 152 | " # 输出一些信息，让我们知道网络在建造中做了什么\n", 153 | " print('Add dense layer: relu')\n", 154 | " print(' %sD --> %sD' %(layer_nodes[l],layer_nodes[l+1]))\n", 155 | " # 存储所有隐藏层的输出\n", 156 | " self.hid_layers.append(incoming)\n", 157 | " # 加入dropout层\n", 158 | " incoming = self.drop_layer(incoming)\n", 159 | " \n", 160 | " # 输出层的建立。输出层需要特别对待的原因是输出层的activation function要根据任务来变。\n", 161 | " # 回归任务的话，下面用的是tf.identity，也就是没有activation function\n", 162 | " if self.Task_type=='regression':\n", 163 | " out_act=tf.identity\n", 164 | " else:\n", 165 | " # 分类任务使用softmax来拟合概率\n", 166 | " out_act=tf.nn.softmax\n", 167 | " self.output,l2_loss= self.layer(incoming,layer_nodes[-1],self.D_label, layer_name='output',act=out_act)\n", 168 | " self.total_l2.append(l2_loss)\n", 169 | " print('Add output layer: linear')\n", 170 | " print(' %sD --> %sD' %(layer_nodes[-1],self.D_label))\n", 171 | " \n", 172 | " # l2 loss的缩放图\n", 173 | " with tf.name_scope('total_l2'):\n", 174 | " for l2 in self.total_l2:\n", 175 | " self.l2_penalty+=l2\n", 176 | " tf.summary.scalar('l2_penalty', self.l2_penalty)\n", 177 | " \n", 178 | " # 不同任务的loss\n", 179 | " # 若为回归，则loss是用于判断所有预测值和实际值差别的函数。\n", 180 | " if self.Task_type=='regression':\n", 181 | " with tf.name_scope('SSE'):\n", 182 | " self.loss=tf.reduce_mean((self.output-self.labels)**2)\n", 183 | " self.loss2=tf.nn.l2_loss(self.output-self.labels)\n", 184 | " \n", 185 | " tf.summary.scalar('loss', self.loss)\n", 186 | " else:\n", 187 | " # 若为分类，cross entropy的loss function\n", 188 | " entropy = tf.nn.softmax_cross_entropy_with_logits(self.output, self.labels)\n", 189 | " with tf.name_scope('cross entropy'):\n", 190 | " self.loss = tf.reduce_mean(entropy)\n", 191 | " tf.summary.scalar('loss', self.loss)\n", 192 | " with tf.name_scope('accuracy'):\n", 193 | " correct_prediction = tf.equal(tf.argmax(self.output, 1), tf.argmax(self.labels, 1))\n", 194 | " self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n", 195 | " tf.summary.scalar('accuracy', self.accuracy)\n", 196 | " \n", 197 | " # 整合所有loss，形成最终loss\n", 198 | " with tf.name_scope('total_loss'):\n", 199 | " self.total_loss=self.loss + self.l2_penalty*self.L2_lambda\n", 200 | " tf.summary.scalar('total_loss', self.total_loss)\n", 201 | " \n", 202 | " # 训练操作\n", 203 | " with tf.name_scope('train'):\n", 204 | " self.train_step = tf.train.AdamOptimizer(self.learning_rate).minimize(self.total_loss)\n", 205 | "\n", 206 | " # 洗牌功能\n", 207 | " def shufflelists(self,lists):\n", 208 | " ri=np.random.permutation(len(lists[1]))\n", 209 | " out=[]\n", 210 | " for l in lists:\n", 211 | " out.append(l[ri])\n", 212 | " return out" 213 | ] 214 | }, 215 | { 216 | "cell_type": "code", 217 | "execution_count": 3, 218 | "metadata": { 219 | "collapsed": true 220 | }, 221 | "outputs": [], 222 | "source": [ 223 | "def Standardize(seq):\n", 224 | " #subtract mean\n", 225 | " centerized=seq-np.mean(seq, axis = 0)\n", 226 | " #divide standard deviation\n", 227 | " normalized=centerized/np.std(centerized, axis = 0)\n", 228 | " return normalized\n", 229 | "def Makewindows(indata,window_size=41):\n", 230 | " outdata=[]\n", 231 | " mid=int(window_size/2)\n", 232 | " indata=np.vstack((np.zeros((mid,indata.shape[1])),indata,np.zeros((mid,indata.shape[1]))))\n", 233 | " for i in range(indata.shape[0]-window_size+1):\n", 234 | " outdata.append(np.hstack(indata[i:i+window_size]))\n", 235 | " return np.array(outdata)" 236 | ] 237 | }, 238 | { 239 | "cell_type": "code", 240 | "execution_count": 4, 241 | "metadata": { 242 | "collapsed": true 243 | }, 244 | "outputs": [], 245 | "source": [ 246 | "mfc=np.load('X.npy')\n", 247 | "art=np.load('Y.npy')\n", 248 | "x=[]\n", 249 | "y=[]\n", 250 | "for i in range(len(mfc)):\n", 251 | " x.append(Makewindows(Standardize(mfc[i])))\n", 252 | " y.append(Standardize(art[i]))\n", 253 | "vali_size=20\n", 254 | "totalsamples=len(np.vstack(x))\n", 255 | "X_train=np.vstack(x)[int(totalsamples/vali_size):].astype(\"float32\")\n", 256 | "Y_train=np.vstack(y)[int(totalsamples/vali_size):].astype(\"float32\")\n", 257 | "\n", 258 | "X_test=np.vstack(x)[:int(totalsamples/vali_size)].astype(\"float32\")\n", 259 | "Y_test=np.vstack(y)[:int(totalsamples/vali_size)].astype(\"float32\")" 260 | ] 261 | }, 262 | { 263 | "cell_type": "code", 264 | "execution_count": 5, 265 | "metadata": {}, 266 | "outputs": [ 267 | { 268 | "name": "stdout", 269 | "output_type": "stream", 270 | "text": [ 271 | "((37500, 1599), (37500, 24), (1973, 1599), (1973, 24))\n" 272 | ] 273 | } 274 | ], 275 | "source": [ 276 | "print(X_train.shape,Y_train.shape,X_test.shape,Y_test.shape)" 277 | ] 278 | }, 279 | { 280 | "cell_type": "code", 281 | "execution_count": 6, 282 | "metadata": {}, 283 | "outputs": [ 284 | { 285 | "name": "stdout", 286 | "output_type": "stream", 287 | "text": [ 288 | "Add dense layer: relu\n", 289 | " 1599D --> 2048D\n", 290 | "Add dense layer: relu\n", 291 | " 2048D --> 1024D\n", 292 | "Add dense layer: relu\n", 293 | " 1024D --> 512D\n", 294 | "Add dense layer: relu\n", 295 | " 512D --> 256D\n", 296 | "Add dense layer: relu\n", 297 | " 256D --> 128D\n", 298 | "Add output layer: linear\n", 299 | " 128D --> 24D\n" 300 | ] 301 | } 302 | ], 303 | "source": [ 304 | "# 生成网络实例\n", 305 | "ff=FNN(learning_rate=7e-5, Layers=5, N_hidden=[2048,1024,512,256,128], D_input=1599, D_label=24, L2_lambda=1e-4)" 306 | ] 307 | }, 308 | { 309 | "cell_type": "code", 310 | "execution_count": 7, 311 | "metadata": { 312 | "collapsed": true 313 | }, 314 | "outputs": [], 315 | "source": [ 316 | "sess = tf.InteractiveSession()\n", 317 | "sess.run(tf.global_variables_initializer())\n", 318 | "merged = tf.summary.merge_all()\n", 319 | "train_writer = tf.summary.FileWriter('log3' + '/train',sess.graph)\n", 320 | "test_writer = tf.summary.FileWriter('log3' + '/test')" 321 | ] 322 | }, 323 | { 324 | "cell_type": "code", 325 | "execution_count": 8, 326 | "metadata": {}, 327 | "outputs": [], 328 | "source": [ 329 | "def plots(T,P,i, n=21,length=400):\n", 330 | " m=0\n", 331 | " plt.figure(figsize=(20,16))\n", 332 | " plt.subplot(411)\n", 333 | " plt.plot(T[m:m+length,7],'--')\n", 334 | " plt.plot(P[m:m+length,7])\n", 335 | "\n", 336 | " plt.subplot(412)\n", 337 | " plt.plot(T[m:m+length,8],'--')\n", 338 | " plt.plot(P[m:m+length,8])\n", 339 | " \n", 340 | " plt.subplot(413)\n", 341 | " plt.plot(T[m:m+length,15],'--')\n", 342 | " plt.plot(P[m:m+length,15])\n", 343 | " \n", 344 | " plt.subplot(414)\n", 345 | " plt.plot(T[m:m+length,16],'--')\n", 346 | " plt.plot(P[m:m+length,16])\n", 347 | " plt.legend(['True','Predicted'])\n", 348 | " plt.savefig('epoch'+str(i)+'.png')\n", 349 | " plt.close()" 350 | ] 351 | }, 352 | { 353 | "cell_type": "code", 354 | "execution_count": 9, 355 | "metadata": { 356 | "scrolled": true 357 | }, 358 | "outputs": [ 359 | { 360 | "name": "stdout", 361 | "output_type": "stream", 362 | "text": [ 363 | "epoch0 | train_loss:0.693693 |test_loss:0.695782\n", 364 | "epoch1 | train_loss:0.561398 |test_loss:0.583177\n", 365 | "epoch2 | train_loss:0.478702 |test_loss:0.54134\n", 366 | "epoch3 | train_loss:0.389257 |test_loss:0.482584\n", 367 | "epoch4 | train_loss:0.320454 |test_loss:0.453059\n", 368 | "epoch5 | train_loss:0.27317 |test_loss:0.441251\n", 369 | "epoch6 | train_loss:0.234799 |test_loss:0.436898\n", 370 | "epoch7 | train_loss:0.203948 |test_loss:0.436239\n", 371 | "epoch8 | train_loss:0.176413 |test_loss:0.426064\n", 372 | "epoch9 | train_loss:0.156074 |test_loss:0.425965\n", 373 | "epoch10 | train_loss:0.132211 |test_loss:0.414978\n", 374 | "epoch11 | train_loss:0.120031 |test_loss:0.40392\n", 375 | "epoch12 | train_loss:0.105933 |test_loss:0.407968\n", 376 | "epoch13 | train_loss:0.100558 |test_loss:0.402615\n", 377 | "epoch14 | train_loss:0.0894171 |test_loss:0.39814\n", 378 | "epoch15 | train_loss:0.0805371 |test_loss:0.395325\n", 379 | "epoch16 | train_loss:0.0755967 |test_loss:0.396785\n", 380 | "epoch17 | train_loss:0.0714638 |test_loss:0.399197\n", 381 | "epoch18 | train_loss:0.0633848 |test_loss:0.399693\n", 382 | "epoch19 | train_loss:0.0614749 |test_loss:0.393158\n", 383 | "epoch20 | train_loss:0.0591606 |test_loss:0.394801\n", 384 | "epoch21 | train_loss:0.055984 |test_loss:0.38638\n", 385 | "epoch22 | train_loss:0.0540141 |test_loss:0.390381\n", 386 | "epoch23 | train_loss:0.0496812 |test_loss:0.384337\n", 387 | "epoch24 | train_loss:0.049253 |test_loss:0.387584\n", 388 | "epoch25 | train_loss:0.0445422 |test_loss:0.383592\n", 389 | "epoch26 | train_loss:0.0432337 |test_loss:0.382902\n", 390 | "epoch27 | train_loss:0.0451354 |test_loss:0.382096\n", 391 | "epoch28 | train_loss:0.041636 |test_loss:0.385363\n", 392 | "epoch29 | train_loss:0.0392896 |test_loss:0.380636\n", 393 | "epoch30 | train_loss:0.0386417 |test_loss:0.386117\n", 394 | "epoch31 | train_loss:0.0378463 |test_loss:0.38561\n", 395 | "epoch32 | train_loss:0.036046 |test_loss:0.383354\n", 396 | "epoch33 | train_loss:0.0359106 |test_loss:0.381586\n", 397 | "epoch34 | train_loss:0.0344725 |test_loss:0.379078\n", 398 | "epoch35 | train_loss:0.0337404 |test_loss:0.376501\n", 399 | "epoch36 | train_loss:0.0340016 |test_loss:0.379459\n", 400 | "epoch37 | train_loss:0.032026 |test_loss:0.37459\n", 401 | "epoch38 | train_loss:0.0317964 |test_loss:0.37812\n", 402 | "epoch39 | train_loss:0.033109 |test_loss:0.37789\n", 403 | "epoch40 | train_loss:0.0311886 |test_loss:0.372208\n", 404 | "epoch41 | train_loss:0.0303357 |test_loss:0.370058\n", 405 | "epoch42 | train_loss:0.0300332 |test_loss:0.37759\n", 406 | "epoch43 | train_loss:0.0293583 |test_loss:0.36993\n", 407 | "epoch44 | train_loss:0.0287083 |test_loss:0.373513\n", 408 | "epoch45 | train_loss:0.0296841 |test_loss:0.371504\n", 409 | "epoch46 | train_loss:0.0279586 |test_loss:0.370936\n", 410 | "epoch47 | train_loss:0.0285195 |test_loss:0.373591\n", 411 | "epoch48 | train_loss:0.0281992 |test_loss:0.37003\n", 412 | "epoch49 | train_loss:0.0267218 |test_loss:0.368262\n" 413 | ] 414 | } 415 | ], 416 | "source": [ 417 | "# 训练并记录\n", 418 | "k=0\n", 419 | "Batch=32\n", 420 | "for i in range(50):\n", 421 | " idx=0\n", 422 | " X0,Y0=ff.shufflelists([X_train,Y_train])\n", 423 | " while idxlen(layer_nodes):\n", 148 | " nodes_in=layer_nodes[-1]\n", 149 | " nodes_out=layer_nodes[-1]\n", 150 | " else:\n", 151 | " nodes_in=layer_nodes[l]\n", 152 | " nodes_out=layer_nodes[l+1]\n", 153 | " incoming,l2_loss= self.layer(incoming,nodes_in,nodes_out,prefix+'_hid_'+str(l+1),act=tf.nn.relu)\n", 154 | " # 累计l2\n", 155 | " self.total_l2.append(l2_loss)\n", 156 | " # 输出一些信息，让我们知道网络在建造中做了什么\n", 157 | " print('Add dense layer: relu')\n", 158 | " print(' %sD --> %sD' %(nodes_in,nodes_out))\n", 159 | " # 存储所有隐藏层的输出\n", 160 | " self.hid_layers.append(incoming)\n", 161 | " # 加入dropout层\n", 162 | " incoming = self.drop_layer(incoming)\n", 163 | " \n", 164 | " # 输出层的建立。输出层需要特别对待的原因是输出层的activation function要根据任务来变。\n", 165 | " # 回归任务的话，下面用的是tf.identity，也就是没有activation function\n", 166 | " if self.Task_type=='regression':\n", 167 | " out_act=tf.identity\n", 168 | " else:\n", 169 | " # 分类任务使用softmax来拟合概率\n", 170 | " out_act=tf.nn.softmax\n", 171 | " self.output,l2_loss= self.layer(incoming,layer_nodes[-1],self.D_label, layer_name='output',act=out_act)\n", 172 | " self.total_l2.append(l2_loss)\n", 173 | " print('Add output layer: linear')\n", 174 | " print(' %sD --> %sD' %(layer_nodes[-1],self.D_label))\n", 175 | " \n", 176 | " # l2 loss的缩放图\n", 177 | " with tf.name_scope('total_l2'):\n", 178 | " for l2 in self.total_l2:\n", 179 | " self.l2_penalty+=l2\n", 180 | " tf.summary.histogram('l2_penalty', self.l2_penalty)\n", 181 | " \n", 182 | " # 不同任务的loss\n", 183 | " # 若为回归，则loss是用于判断所有预测值和实际值差别的函数。\n", 184 | " if self.Task_type=='regression':\n", 185 | " with tf.name_scope('SSE'):\n", 186 | " self.loss=tf.reduce_mean((self.output-self.labels)**2)\n", 187 | " self.loss2=tf.nn.l2_loss(self.output-self.labels)\n", 188 | " \n", 189 | " tf.summary.histogram('loss', self.loss)\n", 190 | " else:\n", 191 | " # 若为分类，cross entropy的loss function\n", 192 | " entropy = tf.nn.softmax_cross_entropy_with_logits(self.output, self.labels)\n", 193 | " with tf.name_scope('cross entropy'):\n", 194 | " self.loss = tf.reduce_mean(entropy)\n", 195 | " tf.summary.histogram('loss', self.loss)\n", 196 | " with tf.name_scope('accuracy'):\n", 197 | " correct_prediction = tf.equal(tf.argmax(self.output, 1), tf.argmax(self.labels, 1))\n", 198 | " self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n", 199 | " tf.summary.histogram('accuracy', self.accuracy)\n", 200 | " \n", 201 | " # 整合所有loss，形成最终loss\n", 202 | " with tf.name_scope('total_loss'):\n", 203 | " self.total_loss=self.loss + self.l2_penalty*self.L2_lambda\n", 204 | " tf.summary.histogram('total_loss', self.total_loss)\n", 205 | " \n", 206 | " # 训练操作\n", 207 | " with tf.name_scope('train'):\n", 208 | " self.train_step = tf.train.AdamOptimizer(self.learning_rate).minimize(self.total_loss)\n", 209 | "\n", 210 | " # 洗牌功能\n", 211 | " def shufflelists(self,lists):\n", 212 | " ri=np.random.permutation(len(lists[1]))\n", 213 | " out=[]\n", 214 | " for l in lists:\n", 215 | " out.append(l[ri])\n", 216 | " return out" 217 | ] 218 | }, 219 | { 220 | "cell_type": "code", 221 | "execution_count": 3, 222 | "metadata": {}, 223 | "outputs": [], 224 | "source": [ 225 | "def Standardize(seq):\n", 226 | " #subtract mean\n", 227 | " centerized=seq-np.mean(seq, axis = 0)\n", 228 | " #divide standard deviation\n", 229 | " normalized=centerized/np.std(centerized, axis = 0)\n", 230 | " return normalized\n", 231 | "def Makewindows(indata,window_size=41):\n", 232 | " outdata=[]\n", 233 | " mid=int(window_size/2)\n", 234 | " indata=np.vstack((np.zeros((mid,indata.shape[1])),indata,np.zeros((mid,indata.shape[1]))))\n", 235 | " for i in range(indata.shape[0]-window_size+1):\n", 236 | " outdata.append(np.hstack(indata[i:i+window_size]))\n", 237 | " return np.array(outdata)" 238 | ] 239 | }, 240 | { 241 | "cell_type": "code", 242 | "execution_count": 4, 243 | "metadata": {}, 244 | "outputs": [], 245 | "source": [ 246 | "mfc=np.load('X.npy')\n", 247 | "art=np.load('Y.npy')\n", 248 | "x=[]\n", 249 | "y=[]\n", 250 | "for i in range(len(mfc)):\n", 251 | " x.append(Makewindows(Standardize(mfc[i])))\n", 252 | " y.append(Standardize(art[i]))\n", 253 | "vali_size=20\n", 254 | "totalsamples=len(np.vstack(x))\n", 255 | "X_train=np.vstack(x)[int(totalsamples/vali_size):].astype(\"float32\")\n", 256 | "Y_train=np.vstack(y)[int(totalsamples/vali_size):].astype(\"float32\")\n", 257 | "\n", 258 | "X_test=np.vstack(x)[:int(totalsamples/vali_size)].astype(\"float32\")\n", 259 | "Y_test=np.vstack(y)[:int(totalsamples/vali_size)].astype(\"float32\")" 260 | ] 261 | }, 262 | { 263 | "cell_type": "code", 264 | "execution_count": 5, 265 | "metadata": {}, 266 | "outputs": [ 267 | { 268 | "name": "stdout", 269 | "output_type": "stream", 270 | "text": [ 271 | "((37500, 1599), (37500, 24), (1973, 1599), (1973, 24))\n" 272 | ] 273 | } 274 | ], 275 | "source": [ 276 | "print(X_train.shape,Y_train.shape,X_test.shape,Y_test.shape)" 277 | ] 278 | }, 279 | { 280 | "cell_type": "code", 281 | "execution_count": 6, 282 | "metadata": { 283 | "scrolled": true 284 | }, 285 | "outputs": [ 286 | { 287 | "name": "stdout", 288 | "output_type": "stream", 289 | "text": [ 290 | "Add dense layer: relu\n", 291 | " 1599D --> 1024D\n", 292 | "Add dense layer: relu\n", 293 | " 1024D --> 1024D\n", 294 | "Add dense layer: relu\n", 295 | " 1024D --> 1024D\n", 296 | "Add dense layer: relu\n", 297 | " 1024D --> 1024D\n", 298 | "Add dense layer: relu\n", 299 | " 1024D --> 1024D\n", 300 | "Add dense layer: relu\n", 301 | " 1024D --> 1024D\n", 302 | "Add dense layer: relu\n", 303 | " 1024D --> 1024D\n", 304 | "Add dense layer: relu\n", 305 | " 1024D --> 1024D\n", 306 | "Add dense layer: relu\n", 307 | " 1024D --> 1024D\n", 308 | "Add dense layer: relu\n", 309 | " 1024D --> 1024D\n", 310 | "Add dense layer: relu\n", 311 | " 1024D --> 1024D\n", 312 | "Add dense layer: relu\n", 313 | " 1024D --> 1024D\n", 314 | "Add dense layer: relu\n", 315 | " 1024D --> 1024D\n", 316 | "Add dense layer: relu\n", 317 | " 1024D --> 1024D\n", 318 | "Add dense layer: relu\n", 319 | " 1024D --> 1024D\n", 320 | "Add dense layer: relu\n", 321 | " 1024D --> 1024D\n", 322 | "Add dense layer: relu\n", 323 | " 1024D --> 1024D\n", 324 | "Add dense layer: relu\n", 325 | " 1024D --> 1024D\n", 326 | "Add dense layer: relu\n", 327 | " 1024D --> 1024D\n", 328 | "Add dense layer: relu\n", 329 | " 1024D --> 1024D\n", 330 | "Add dense layer: relu\n", 331 | " 1024D --> 1024D\n", 332 | "Add dense layer: relu\n", 333 | " 1024D --> 1024D\n", 334 | "Add dense layer: relu\n", 335 | " 1024D --> 1024D\n", 336 | "Add dense layer: relu\n", 337 | " 1024D --> 1024D\n", 338 | "Add dense layer: relu\n", 339 | " 1024D --> 1024D\n", 340 | "Add dense layer: relu\n", 341 | " 1024D --> 1024D\n", 342 | "Add dense layer: relu\n", 343 | " 1024D --> 1024D\n", 344 | "Add dense layer: relu\n", 345 | " 1024D --> 1024D\n", 346 | "Add dense layer: relu\n", 347 | " 1024D --> 1024D\n", 348 | "Add dense layer: relu\n", 349 | " 1024D --> 1024D\n", 350 | "Add dense layer: relu\n", 351 | " 1024D --> 1024D\n", 352 | "Add dense layer: relu\n", 353 | " 1024D --> 1024D\n", 354 | "Add dense layer: relu\n", 355 | " 1024D --> 1024D\n", 356 | "Add dense layer: relu\n", 357 | " 1024D --> 1024D\n", 358 | "Add dense layer: relu\n", 359 | " 1024D --> 1024D\n", 360 | "Add dense layer: relu\n", 361 | " 1024D --> 1024D\n", 362 | "Add dense layer: relu\n", 363 | " 1024D --> 1024D\n", 364 | "Add dense layer: relu\n", 365 | " 1024D --> 1024D\n", 366 | "Add dense layer: relu\n", 367 | " 1024D --> 1024D\n", 368 | "Add dense layer: relu\n", 369 | " 1024D --> 1024D\n", 370 | "Add dense layer: relu\n", 371 | " 1024D --> 1024D\n", 372 | "Add dense layer: relu\n", 373 | " 1024D --> 1024D\n", 374 | "Add dense layer: relu\n", 375 | " 1024D --> 1024D\n", 376 | "Add dense layer: relu\n", 377 | " 1024D --> 1024D\n", 378 | "Add dense layer: relu\n", 379 | " 1024D --> 1024D\n", 380 | "Add dense layer: relu\n", 381 | " 1024D --> 1024D\n", 382 | "Add dense layer: relu\n", 383 | " 1024D --> 1024D\n", 384 | "Add dense layer: relu\n", 385 | " 1024D --> 1024D\n", 386 | "Add dense layer: relu\n", 387 | " 1024D --> 1024D\n", 388 | "Add dense layer: relu\n", 389 | " 1024D --> 1024D\n", 390 | "Add output layer: linear\n", 391 | " 1024D --> 24D\n" 392 | ] 393 | } 394 | ], 395 | "source": [ 396 | "# 生成网络实例\n", 397 | "# 如果不指定全部隐藏层节点，则多出的隐藏层全部按最后一个指定的隐藏层节点数设定\n", 398 | "ff=FNN(learning_rate=7e-5, Layers=50,N_hidden=[1024], D_input=1599, D_label=24, L2_lambda=1e-4)" 399 | ] 400 | }, 401 | { 402 | "cell_type": "code", 403 | "execution_count": 7, 404 | "metadata": {}, 405 | "outputs": [], 406 | "source": [ 407 | "sess = tf.InteractiveSession()\n", 408 | "sess.run(tf.global_variables_initializer())\n", 409 | "merged = tf.summary.merge_all()\n", 410 | "train_writer = tf.summary.FileWriter('log' + '/train',sess.graph)\n", 411 | "test_writer = tf.summary.FileWriter('log' + '/test')" 412 | ] 413 | }, 414 | { 415 | "cell_type": "code", 416 | "execution_count": 8, 417 | "metadata": {}, 418 | "outputs": [], 419 | "source": [ 420 | "def plots(T,P,i, n=21,length=400):\n", 421 | " m=0\n", 422 | " plt.figure(figsize=(20,16))\n", 423 | " plt.subplot(411)\n", 424 | " plt.plot(T[m:m+length,7],'--')\n", 425 | " plt.plot(P[m:m+length,7])\n", 426 | "\n", 427 | " plt.subplot(412)\n", 428 | " plt.plot(T[m:m+length,8],'--')\n", 429 | " plt.plot(P[m:m+length,8])\n", 430 | " \n", 431 | " plt.subplot(413)\n", 432 | " plt.plot(T[m:m+length,15],'--')\n", 433 | " plt.plot(P[m:m+length,15])\n", 434 | " \n", 435 | " plt.subplot(414)\n", 436 | " plt.plot(T[m:m+length,16],'--')\n", 437 | " plt.plot(P[m:m+length,16])\n", 438 | " plt.legend(['True','Predicted'])\n", 439 | " plt.savefig('epoch'+str(i)+'.png')\n", 440 | " plt.close()" 441 | ] 442 | }, 443 | { 444 | "cell_type": "code", 445 | "execution_count": 9, 446 | "metadata": { 447 | "scrolled": true 448 | }, 449 | "outputs": [ 450 | { 451 | "name": "stdout", 452 | "output_type": "stream", 453 | "text": [ 454 | "epoch0 | train_loss:[0.85698944] |test_loss:0.849185\n", 455 | "epoch1 | train_loss:[0.75586188] |test_loss:0.74935\n", 456 | "epoch2 | train_loss:[0.66804069] |test_loss:0.697456\n", 457 | "epoch3 | train_loss:[0.59767956] |test_loss:0.649171\n", 458 | "epoch4 | train_loss:[0.57731116] |test_loss:0.652371\n", 459 | "epoch5 | train_loss:[0.5061782] |test_loss:0.596573\n", 460 | "epoch6 | train_loss:[0.46441522] |test_loss:0.595333\n", 461 | "epoch7 | train_loss:[0.40969208] |test_loss:0.555003\n", 462 | "epoch8 | train_loss:[0.40397146] |test_loss:0.573359\n", 463 | "epoch9 | train_loss:[0.37393975] |test_loss:0.553064\n" 464 | ] 465 | } 466 | ], 467 | "source": [ 468 | "# 训练并记录\n", 469 | "k=0\n", 470 | "EPOCH=10\n", 471 | "Batch=256\n", 472 | "for i in range(EPOCH):\n", 473 | " idx=0\n", 474 | " X0,Y0=ff.shufflelists([X_train,Y_train])\n", 475 | " while idx