1 #coding = utf-8 2 from datetime import datetime 3 import tensorflow as tf 4 import time 5 import math 6 7 def conv_op(input_op, name, kh, kw, n_out, dh, dw, p): 8 n_in = input_op.get_shape()[-1].value 9 10 '''创建卷积层并把本层的参数存入参数列表 11 input_op：tensor 12 name:层名 13 kh:kernel height 14 kw:kernel width 15 n_out:卷积核数量即输出通道数 16 dh：步长的高 17 dw: 步长的宽 18 p： 参数列表''' 19 with tf.name_scope(name) as scope: 20 21 kernel = tf.get_variable(scope+"w", 22 shape=[kh, kw, n_in, n_out], dtype=tf.float32, 23 initializer=tf.contrib.layers.xavier_initializer_conv2d()) 24 25 conv = tf.nn.conv2d(input_op, kernel, (1, dh, dw, 1), 26 padding = 'SAME') 27 bias_init_val = tf.constant(0.0, shape=[n_out], dtype=tf.float32) 28 biases = tf.Variable(bias_init_val, trainable=True, name='b') 29 z = tf.nn.bias_add(conv, biases) 30 activation = tf.nn.relu(z, name=scope) 31 p += [kernel, biases] 32 return activation 33 34 35 def fc_op(input_op, name, n_out, p): 36 '''定义创建全连接层函数：''' 37 n_in = input_op.get_shape()[-1].value#获取输入input_op的通道数 38 39 with tf.name_scope(name) as scope: 40 #创建全连接层参数，维度为2：n_in, n_out 41 kernel = tf.get_variable(scope+"w", 42 shape=[n_in, n_out], dtype=tf.float32, 43 initializer=tf.contrib.layers.xavier_initializer()) 44 #偏置初始化为0.1，避免dead neoron 45 biases = tf.Variable(tf.constant(0.1, shape=[n_out], 46 dtype=tf.float32), name="b") 47 #激活：relu非线性变换得到 48 activation = tf.nn.relu_layer(input_op, kernel, biases, name=scope) 49 #添加全连接参数kernel和biases到参数列表p 50 p += [kernel, biases] 51 52 return activation 53 54 55 def mpool_op(input_op, name, kh, kw, dh, dw): 56 '''定义maxpool层 57 input：input_op 58 池化尺寸：khxkw 59 步长：dhxdw 60 padding：SAME 61 padding="SAME" 输出尺寸为W/S(其中W为输入尺寸,S为stride) 62 padding="VALID" 输出尺寸为(W-F+1)/S(其中F为卷积核尺寸) 63 ''' 64 65 return tf.nn.max_pool(input_op, 66 ksize=[1, kh, kw, 1], 67 strides=[1, dh, dw, 1], 68 padding='SAME', 69 name=name) 70 71 def inference_op(input_op, keep_prob): 72 73 p = [] 74 '''第一层卷积神经网络：2个卷积层和一个最大池化层, 75 2个卷积层卷积核大小为3x3，卷积核数量(输出通道数)：64，步长为1x1 76 输出尺寸：112x112x64''' 77 conv1_1 = conv_op(input_op, name="conv1_1", kh=3, kw=3, n_out=64, dh=1, 78 dw=1, p=p) 79 conv1_2 = conv_op(conv1_1, name="conv1_2", kh=3, kw=3, n_out=64, dh=1, 80 dw=1, p=p) 81 pool1 = mpool_op(conv1_2, name="pool1", kh=2, kw=2, dw=2, dh=2) 82 83 '''第二层卷积神经网络:2个卷积层和一个最大池化层， 84 2个卷积层卷积核大小为：3x3，输出通道数为：128 85 输出尺寸为56x56x128''' 86 conv2_1 = conv_op(pool1, name="conv2_1", kh=3, kw=3, n_out=128, dh=1, 87 dw=1, p=p) 88 conv2_2 = conv_op(conv2_1, name="conv2_2", kh=3, kw=3, n_out=128, dh=1, 89 dw=1, p=p) 90 pool2 = mpool_op(conv2_2, name="pool2", kh=2, kw=2, dh=2, dw=2) 91 92 '''第三层卷积神经网络：3个卷积层和一个最大池化层。 93 3个卷积层的卷积核大小均为：3X3，输出通道数为:256 94 输出尺寸为28x28x256''' 95 conv3_1 = conv_op(pool2, name="conv3_1", kh=3, kw=3, n_out=256, dh=1, 96 dw=1, p=p) 97 conv3_2 = conv_op(conv3_1, name="conv3_2", kh=3, kw=3, n_out=256, dh=1, 98 dw=1, p=p) 99 conv3_3 = conv_op(conv3_2, name="conv3_3", kh=3, kw=3, n_out=256, dh=1, 100 dw=1, p=p) 101 pool3 = mpool_op(conv3_3, name="pool3", kh=2, kw=2, dh=2, dw=2) 102 '''第四层卷积神经网络：3个卷积层和一个最大池化层。 103 3个卷积层的卷积核大小均为3x3，输出通道为512， 104 输出尺寸为14x14x512''' 105 conv4_1 = conv_op(pool3, name="conv3_1", kh=3, kw=3, n_out=512, dh=1, 106 dw=1, p=p) 107 conv4_2 = conv_op(conv4_1, name="conv4_2", kh=3, kw=3, n_out=512, dh=1, 108 dw=1, p=p) 109 conv4_3 = conv_op(conv4_2, name="conv4_3", kh=3, kw=3, n_out=512, dh=1, 110 dw=1, p=p) 111 pool4 =mpool_op(conv4_3, name="pool4", kh=2, kw=2, dh=2, dw=2) 112 113 '''第五层卷积神经网络：3个卷积层和一个最大池化层。 114 3个卷积层的卷积核大小均为3x3，输出通道仍然为512， 115 输出尺寸为7x7x512''' 116 conv5_1 = conv_op(pool4, name="conv5_1", kh=3, kw=3, n_out=512, dh=1, 117 dw=1, p=p) 118 conv5_2 = conv_op(conv5_1, name="conv4_2", kh=3, kw=3, n_out=512, dh=1, 119 dw=1, p=p) 120 conv5_3 = conv_op(conv5_2, name="conv4_3", kh=3, kw=3, n_out=512, dh=1, 121 dw=1, p=p) 122 pool5 =mpool_op(conv5_3, name="pool5", kh=2, kw=2, dh=2, dw=2) 123 124 125 '''输出结果扁平化，用tf.reshape将各个样本转化为7x7x512=25088的一维向量''' 126 shp = pool5.get_shape() 127 flattened_shape = shp[1].value*shp[2].value*shp[3].value 128 resh1 = tf.reshape(pool5, [-1, flattened_shape], name="resh1") 129 130 '''隐含节点数为4096的Fully connection，activation function：ReLu, 131 再连接Dropout层，训练时节点保留率为0.5，预测时为1.0''' 132 fc6 = fc_op(resh1, name="fc6", n_out=4096, p=p) 133 fc6_drop = tf.nn.dropout(fc6, keep_prob, name="fc_drop") 134 135 fc7 = fc_op(fc6_drop, name="fc7", n_out=4096, p=p) 136 fc7_drop = tf.nn.dropout(fc7, keep_prob, name="fc7_drop") 137 138 '''输出节点为1000的全连接层，并使用softmax进行分类得到分类输出概率''' 139 fc8 = fc_op(fc7_drop, name="fc8", n_out=1000, p=p) 140 softmax = tf.nn.softmax(fc8) 141 #tf.argmax()求出输出概率的最大类别 142 predictions = tf.argmax(softmax, 1) 143 return predictions, softmax, fc8, p 144 145 146 def time_tensorflow_run(session, target, feed, info_string): 147 148 n_steps_burn_in = 10 149 total_duration = 0.0 150 total_duration_squared = 0.0 151 for i in range(num_batches+n_steps_burn_in): 152 start_time = time.time() 153 _ = session.run(target, feed_dict=feed) 154 duration = time.time()-start_time 155 156 if i>=n_steps_burn_in: 157 if not i%10: 158 print('%s: step %d, duration = %.3f' % 159 (datetime.now(), i-n_steps_burn_in, duration)) 160 total_duration += duration 161 total_duration_squared +=duration*duration 162 163 mn = total_duration/num_batches 164 vr = total_duration_squared/num_batches-mn*mn 165 sd = math.sqrt(vr) 166 167 print('%s: %s across %d steps, %.3f +/- %3.3f sec/batch' % 168 (datetime.now(), info_string, num_batches, mn, sd)) 169 170 def run_benchmark(): 171 172 with tf.Graph().as_default(): 173 '''定义图片尺寸224，利用tf.random_normal函数生成标准差为0.1的正态分布 174 的随机数来构建224x224的随机图片''' 175 image_size = 64#GPU空间不够,改为64 176 images = tf.Variable(tf.random_normal([batch_size, 177 image_size, 178 image_size, 3], 179 dtype=tf.float32, 180 stddev=1e-1)) 181 #构建keep_prob的placeholder 182 keep_prob = tf.placeholder(tf.float32) 183 #调用inference_op构建VGGNET-16网络结构，赋值给相应参数 184 predictions, softmax, fc8, p = inference_op(images, keep_prob) 185 init = tf.global_variables_initializer() 186 sess = tf.Session() 187 sess.run(init) 188 #设置keep_prob为1.0，运用time_tensorflow_run来评测forward运算随机 189 time_tensorflow_run(sess, predictions, {keep_prob:1.0}, "Forward") 190 #计算fc8输出的loss 191 objective = tf.nn.l2_loss(fc8) 192 #求相对于loss的所有模型的梯度 193 grad = tf.gradients(objective, p) 194 #评测backward运算时间 195 time_tensorflow_run(sess, grad, {keep_prob:0.5}, "Forward-backward") 196 197 198 batch_size = 32 199 num_batches = 100 200 run_benchmark()
1 2017-12-21 20:27:00.788000: step 0, duration = 0.055 2 2017-12-21 20:27:01.343000: step 10, duration = 0.055 3 2017-12-21 20:27:01.897000: step 20, duration = 0.055 4 2017-12-21 20:27:02.451000: step 30, duration = 0.055 5 2017-12-21 20:27:03.008000: step 40, duration = 0.055 6 2017-12-21 20:27:03.566000: step 50, duration = 0.056 7 2017-12-21 20:27:04.123000: step 60, duration = 0.056 8 2017-12-21 20:27:04.679000: step 70, duration = 0.056 9 2017-12-21 20:27:05.236000: step 80, duration = 0.056 10 2017-12-21 20:27:05.793000: step 90, duration = 0.056 11 2017-12-21 20:27:06.295000: Forward across 100 steps, 0.006 +/- 0.017 sec/batch 12 13 2017-12-21 20:27:09.294000: step 0, duration = 0.188 14 2017-12-21 20:27:11.173000: step 10, duration = 0.187 15 2017-12-21 20:27:13.053000: step 20, duration = 0.188 16 2017-12-21 20:27:14.934000: step 30, duration = 0.188 17 2017-12-21 20:27:16.814000: step 40, duration = 0.188 18 2017-12-21 20:27:18.695000: step 50, duration = 0.188 19 2017-12-21 20:27:20.575000: step 60, duration = 0.188 20 2017-12-21 20:27:22.456000: step 70, duration = 0.188 21 2017-12-21 20:27:24.337000: step 80, duration = 0.188 22 2017-12-21 20:27:26.217000: step 90, duration = 0.187 23 2017-12-21 20:27:27.911000: Forward-backward across 100 steps, 0.019 +/- 0.056 sec/batch