AlexNet卷积神经网络【前向反馈】
1.代码实现
1 # -*- coding: utf-8 -*- 2 """ 3 Created on Wed Nov 14 17:13:05 2018 4 5 @author: zhen 6 """ 7 8 from datetime import datetime 9 import math 10 import time 11 import tensorflow as tf 12 13 batch_size = 32 14 num_batchs = 100 15 16 def print_activations(t): 17 print(t.op.name, " ", t.get_shape().as_list()) 18 19 def inference(images): 20 parameters = [] 21 with tf.name_scope('conv1') as scope: 22 kernel = tf.Variable(tf.truncated_normal([11,11, 3, 64],dtype=tf.float32, stddev=1e-1), name='weights') 23 conv = tf.nn.conv2d(images, kernel, [1, 4, 4, 1], padding='SAME') 24 biases = tf.Variable(tf.constant(0.0, shape=[64], dtype=tf.float32), trainable=True, name='biases') 25 bias = tf.nn.bias_add(conv, biases) 26 conv1 = tf.nn.relu(bias, name=scope) 27 print_activations(conv1) 28 parameters += [kernel, biases] 29 lrn1 = tf.nn.lrn(conv1, depth_radius=4, bias=1.0, alpha=0.001/9, beta=0.75, name='lrn1') 30 pool1 = tf.nn.max_pool(lrn1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool1') 31 print_activations(pool1) 32 33 with tf.name_scope('conv2') as scope: 34 kernel = tf.Variable(tf.truncated_normal([5, 5, 64, 128], dtype=tf.float32, stddev=1e-1, name='weights')) 35 conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME') 36 biases = tf.Variable(tf.constant(0.0, shape=[128], dtype=tf.float32), trainable=True, name='biases') 37 bias = tf.nn.bias_add(conv, biases) 38 conv2 = tf.nn.relu(bias, name=scope) 39 parameters += [kernel, biases] 40 print_activations(conv2) 41 42 lrn2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.001/9, beta=0.75, name='lrn2') 43 pool2 = tf.nn.max_pool(lrn2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool2') 44 print_activations(pool2) 45 46 with tf.name_scope('conv3') as scope: 47 kernel = tf.Variable(tf.truncated_normal([3, 3, 128, 256], dtype=tf.float32, stddev=1e-1, name='weights')) 48 conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME') 49 biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32), trainable=True, name='biases') 50 bias = tf.nn.bias_add(conv, biases) 51 conv3 = tf.nn.relu(bias, name=scope) 52 parameters += [kernel, biases] 53 print_activations(conv3) 54 55 with tf.name_scope('conv4') as scope: 56 kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 128], dtype=tf.float32, stddev=1e-1, name='weights')) 57 conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME') 58 biases = tf.Variable(tf.constant(0.0, shape=[128], dtype=tf.float32), trainable=True, name='biases') 59 bias = tf.nn.bias_add(conv, biases) 60 conv4 = tf.nn.relu(bias, name=scope) 61 parameters += [kernel, biases] 62 print_activations(conv4) 63 64 with tf.name_scope('conv5') as scope: 65 kernel = tf.Variable(tf.truncated_normal([3, 3, 128, 128], dtype=tf.float32, stddev=1e-1, name='weights')) 66 conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME') 67 biases = tf.Variable(tf.constant(0.0, shape=[128], dtype=tf.float32), trainable=True, name='biases') 68 bias = tf.nn.bias_add(conv, biases) 69 conv5 = tf.nn.relu(bias, name=scope) 70 parameters += [kernel, biases] 71 print_activations(conv5) 72 73 pool5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool5') 74 print_activations(pool5) 75 76 return pool5, parameters 77 78 # 评估AlexNet每轮计算时间 79 def fit_date(session, target, info_string): 80 num_steps_burn_in = 10 # 初始计算轮数 81 total_duration = 0.0 82 total_duration_squared = 0.0 83 84 for i in range(num_batchs + num_steps_burn_in): 85 start_time = time.time() 86 session.run(target) 87 duration = time.time() - start_time 88 if i >= num_steps_burn_in: 89 if not i % 10: 90 print('%s:step %d, duration=%.3f'%(datetime.now(), i - num_steps_burn_in, duration)) 91 total_duration += duration 92 total_duration_squared += duration * duration 93 mn = total_duration / num_batchs 94 vr = total_duration_squared / num_batchs - mn * mn 95 sd = math.sqrt(vr) 96 print('%s:%s across %d steps,%.3f +/- %.3f sec / batch'%(datetime.now(), info_string, num_batchs, mn, sd)) 97 98 def fit_benchmark(): 99 with tf.Graph().as_default(): 100 image_size = 224 101 images = tf.Variable(tf.random_normal([batch_size, image_size, image_size, 3], dtype=tf.float32, stddev=1e-1)) 102 pool5, parameters = inference(images) 103 init = tf.global_variables_initializer() 104 sess = tf.Session() 105 sess.run(init) 106 107 fit_date(sess, pool5, "Forward") 108 objective = tf.nn.l2_loss(pool5) 109 grad = tf.gradients(objective, parameters) 110 fit_date(sess, grad, "Forward-backward") 111 112 fit_benchmark()
2.结果
conv1 [32, 56, 56, 64] pool1 [32, 27, 27, 64] conv2 [32, 27, 27, 128] pool2 [32, 13, 13, 128] conv3 [32, 13, 13, 256] conv4 [32, 13, 13, 128] conv5 [32, 13, 13, 128] pool5 [32, 6, 6, 128] 2019-01-27 10:51:37.551617:step 0, duration=1.625 2019-01-27 10:51:54.082824:step 10, duration=1.766 2019-01-27 10:52:10.582787:step 20, duration=1.641 2019-01-27 10:52:27.051502:step 30, duration=1.672 2019-01-27 10:52:43.507558:step 40, duration=1.625 2019-01-27 10:52:59.913772:step 50, duration=1.625 2019-01-27 10:53:16.245750:step 60, duration=1.672 2019-01-27 10:53:32.511337:step 70, duration=1.625 2019-01-27 10:53:48.901938:step 80, duration=1.609 2019-01-27 10:54:05.183145:step 90, duration=1.625 2019-01-27 10:54:19.917492:Forward across 100 steps,1.640 +/- 0.031 sec / batch 2019-01-27 10:55:47.146016:step 0, duration=7.719 2019-01-27 10:57:04.602639:step 10, duration=7.766 2019-01-27 10:58:26.594245:step 20, duration=9.842 2019-01-27 11:00:01.957195:step 30, duration=8.391 2019-01-27 11:01:35.103007:step 40, duration=10.073 2019-01-27 11:03:07.656318:step 50, duration=8.988 2019-01-27 11:04:31.844207:step 60, duration=8.590 2019-01-27 11:06:01.173490:step 70, duration=9.422 2019-01-27 11:07:28.737373:step 80, duration=10.635 2019-01-27 11:09:03.830375:step 90, duration=8.653 2019-01-27 11:10:19.836018:Forward-backward across 100 steps,8.804 +/- 0.817 sec / batch
3.分析
1、AlexNet是比赛分类项目的2012年冠军，top5错误率16.4%，8层神经网络。
2、AlexNet中包含了几个比较新的技术点，首次在CNN中成功应用了Relu、Dropout、 Lrn等Trick。
3、运用Relu，解决Sigmoid在网络层次较深时的梯度弥散。
4、训练Dropout，随机忽略一些神经元，避免过拟合。
5、使用重叠的最大池化，此前CNN普遍平均池化，最大池化避免平均池化的模糊化效果。
6、提出了Lrn层，局部神经元活动创建竞争机制，响应比较大的值变得更大，抑制其他反馈小的神经元，增强泛化能力。
7、数据增强，随机地从256*256的原始图像中截取224*224大小的区域，以及水平翻转的镜像，相当于增加了【（256-224）^2】*2=2048倍的数据量。
注意：没有数据增强，仅靠原始的数据量，参数众多的CNN会陷入过拟合中。
