机器学习实验（十二）：深度学习之图像分类模型AlexNet结构分析和tensorflow实现

声明：版权所有，转载请联系作者并注明出处  http://blog.csdn.net/u013719780?viewmode=contents

在ImageNet上的图像分类challenge上，Hinton和他的学生Alex Krizhevsky提出的AlexNet网络结构模型赢得了2012届的冠军，刷新了Image Classification的几率。因此，要研究CNN类型深度学习模型在图像分类上的应用，AlexNet就不得不谈，这是CNN在图像分类上的经典模型。下面先看看AlexNet的结构图：下面对其结构进行详细的分析，具体分析过程请看下面的流程图：conv1阶段DFD（data flow diagram）：In [13]:

Image(filename="/Users/youwei.tan/Documents/1.png")

Out[13]:conv2阶段DFD（data flow diagram）：In [6]:

Image(filename="/Users/youwei.tan/Documents/2.png")

Out[6]:conv3阶段DFD（data flow diagram）：In [7]:

Image(filename="/Users/youwei.tan/Documents/3.png")

Out[7]:conv4阶段DFD（data flow diagram）：In [8]:

Image(filename="/Users/youwei.tan/Documents/4.png")

Out[8]:conv5阶段DFD（data flow diagram）：In [9]:

Image(filename="/Users/youwei.tan/Documents/5.png")

Out[9]:fc6阶段DFD（data flow diagram）：In [10]:

Image(filename="/Users/youwei.tan/Documents/6.png")

Out[10]:fc7阶段DFD（data flow diagram）：In [11]:

Image(filename="/Users/youwei.tan/Documents/7.png")

Out[11]:fc8阶段DFD（data flow diagram）：In [12]:

Image(filename="/Users/youwei.tan/Documents/8.png")

Out[12]:理解了AlexNet模型的结构，实现AlexNet的代码应该不难了，在网上看到了已经有大神用tensorflow实现了AlexNet（代码出处：http://blog.csdn.net/chenriwei2/article/details/50615753 ）下面直接搬过来，具体代码如下：In [6]:

import tensorflow as tffrom tensorflow.examples.tutorials.mnist import input_data# load数据# import input_data# mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)mnist = input_data.read_data_sets("MNIST_data", one_hot=True)# 定义网络超参数learning_rate = 0.001training_iters = 200000batch_size = 64display_step = 20# 定义网络参数n_input = 784 # 输入的维度n_classes = 10 # 标签的维度dropout = 0.8 # Dropout 的概率# 占位符输入x = tf.placeholder(tf.float32, [None, n_input])y = tf.placeholder(tf.float32, [None, n_classes])keep_prob = tf.placeholder(tf.float32)# 卷积操作def conv2d(name, l_input, w, b):return tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(l_input, w, strides=[1, 1, 1, 1], padding='SAME'),b), name=name)# 最大下采样操作def max_pool(name, l_input, k):return tf.nn.max_pool(l_input, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME', name=name)# 归一化操作def norm(name, l_input, lsize=4):return tf.nn.lrn(l_input, lsize, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name=name)# 定义整个网络def alex_net(_X, _weights, _biases, _dropout):# 向量转为矩阵_X = tf.reshape(_X, shape=[-1, 28, 28, 1])# 卷积层conv1 = conv2d('conv1', _X, _weights['wc1'], _biases['bc1'])# 下采样层pool1 = max_pool('pool1', conv1, k=2)# 归一化层norm1 = norm('norm1', pool1, lsize=4)# Dropoutnorm1 = tf.nn.dropout(norm1, _dropout)# 卷积conv2 = conv2d('conv2', norm1, _weights['wc2'], _biases['bc2'])# 下采样pool2 = max_pool('pool2', conv2, k=2)# 归一化norm2 = norm('norm2', pool2, lsize=4)# Dropoutnorm2 = tf.nn.dropout(norm2, _dropout)# 卷积conv3 = conv2d('conv3', norm2, _weights['wc3'], _biases['bc3'])# 下采样pool3 = max_pool('pool3', conv3, k=2)# 归一化norm3 = norm('norm3', pool3, lsize=4)# Dropoutnorm3 = tf.nn.dropout(norm3, _dropout)# 全连接层，先把特征图转为向量dense1 = tf.reshape(norm3, [-1, _weights['wd1'].get_shape().as_list()[0]])dense1 = tf.nn.relu(tf.matmul(dense1, _weights['wd1']) + _biases['bd1'], name='fc1')# 全连接层dense2 = tf.nn.relu(tf.matmul(dense1, _weights['wd2']) + _biases['bd2'], name='fc2') # Relu activation# 网络输出层out = tf.matmul(dense2, _weights['out']) + _biases['out']return out# 存储所有的网络参数weights = {'wc1': tf.Variable(tf.random_normal([3, 3, 1, 64])),'wc2': tf.Variable(tf.random_normal([3, 3, 64, 128])),'wc3': tf.Variable(tf.random_normal([3, 3, 128, 256])),'wd1': tf.Variable(tf.random_normal([4*4*256, 1024])),'wd2': tf.Variable(tf.random_normal([1024, 1024])),'out': tf.Variable(tf.random_normal([1024, 10]))}biases = {'bc1': tf.Variable(tf.random_normal([64])),'bc2': tf.Variable(tf.random_normal([128])),'bc3': tf.Variable(tf.random_normal([256])),'bd1': tf.Variable(tf.random_normal([1024])),'bd2': tf.Variable(tf.random_normal([1024])),'out': tf.Variable(tf.random_normal([n_classes]))}# 构建模型pred = alex_net(x, weights, biases, keep_prob)# 定义损失函数和学习步骤cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)# 测试网络correct_pred = tf.equal(tf.argmax(pred,1), tf.argmax(y,1))accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))# 初始化所有的共享变量init = tf.initialize_all_variables()# 开启一个训练with tf.Session() as sess:sess.run(init)step = 1# Keep training until reach max iterationswhile step * batch_size < training_iters:batch_xs, batch_ys = mnist.train.next_batch(batch_size)# 获取批数据sess.run(optimizer, feed_dict={x: batch_xs, y: batch_ys, keep_prob: dropout})if step % display_step == 0:# 计算精度acc = sess.run(accuracy, feed_dict={x: batch_xs, y: batch_ys, keep_prob: 1.})# 计算损失值loss = sess.run(cost, feed_dict={x: batch_xs, y: batch_ys, keep_prob: 1.})print ("Iter " + str(step*batch_size) + ", Minibatch Loss= " + "{:.6f}".format(loss) + ", Training Accuracy= " + "{:.5f}".format(acc))step += 1print ("Optimization Finished!")# 计算测试精度print ("Testing Accuracy:", sess.run(accuracy, feed_dict={x: mnist.test.images[:256], y: mnist.test.labels[:256], keep_prob: 1.}))

Extracting MNIST_data/train-images-idx3-ubyte.gzExtracting MNIST_data/train-labels-idx1-ubyte.gzExtracting MNIST_data/t10k-images-idx3-ubyte.gzExtracting MNIST_data/t10k-labels-idx1-ubyte.gzIter 1280, Minibatch Loss= 105135.500000, Training Accuracy= 0.20312Iter 2560, Minibatch Loss= 53449.058594, Training Accuracy= 0.26562Iter 3840, Minibatch Loss= 46481.890625, Training Accuracy= 0.45312Iter 5120, Minibatch Loss= 27159.894531, Training Accuracy= 0.46875Iter 6400, Minibatch Loss= 24144.722656, Training Accuracy= 0.59375Iter 7680, Minibatch Loss= 34462.066406, Training Accuracy= 0.51562Iter 8960, Minibatch Loss= 29766.037109, Training Accuracy= 0.54688Iter 10240, Minibatch Loss= 18833.246094, Training Accuracy= 0.62500Iter 11520, Minibatch Loss= 14885.839844, Training Accuracy= 0.67188Iter 12800, Minibatch Loss= 29879.960938, Training Accuracy= 0.56250Iter 14080, Minibatch Loss= 24039.386719, Training Accuracy= 0.50000Iter 15360, Minibatch Loss= 6645.172852, Training Accuracy= 0.76562Iter 16640, Minibatch Loss= 18668.121094, Training Accuracy= 0.64062Iter 17920, Minibatch Loss= 12958.103516, Training Accuracy= 0.78125Iter 19200, Minibatch Loss= 11467.254883, Training Accuracy= 0.75000Iter 20480, Minibatch Loss= 11780.928711, Training Accuracy= 0.78125Iter 21760, Minibatch Loss= 17288.359375, Training Accuracy= 0.68750Iter 23040, Minibatch Loss= 10288.304688, Training Accuracy= 0.79688Iter 24320, Minibatch Loss= 10797.081055, Training Accuracy= 0.73438Iter 25600, Minibatch Loss= 15526.394531, Training Accuracy= 0.68750Iter 26880, Minibatch Loss= 5896.966309, Training Accuracy= 0.78125Iter 28160, Minibatch Loss= 5986.314941, Training Accuracy= 0.78125Iter 29440, Minibatch Loss= 14352.033203, Training Accuracy= 0.70312Iter 30720, Minibatch Loss= 9299.230469, Training Accuracy= 0.79688Iter 32000, Minibatch Loss= 12997.154297, Training Accuracy= 0.73438Iter 33280, Minibatch Loss= 8444.686523, Training Accuracy= 0.78125Iter 34560, Minibatch Loss= 8713.619141, Training Accuracy= 0.79688Iter 35840, Minibatch Loss= 11031.250000, Training Accuracy= 0.76562Iter 37120, Minibatch Loss= 14087.941406, Training Accuracy= 0.78125Iter 38400, Minibatch Loss= 6499.219727, Training Accuracy= 0.78125Iter 39680, Minibatch Loss= 7266.004883, Training Accuracy= 0.79688Iter 40960, Minibatch Loss= 13199.544922, Training Accuracy= 0.65625Iter 42240, Minibatch Loss= 7316.147949, Training Accuracy= 0.82812Iter 43520, Minibatch Loss= 5919.269531, Training Accuracy= 0.84375Iter 44800, Minibatch Loss= 4456.823242, Training Accuracy= 0.81250Iter 46080, Minibatch Loss= 2113.104492, Training Accuracy= 0.93750Iter 47360, Minibatch Loss= 4923.742188, Training Accuracy= 0.85938Iter 48640, Minibatch Loss= 7970.073730, Training Accuracy= 0.75000Iter 49920, Minibatch Loss= 5625.089844, Training Accuracy= 0.79688Iter 51200, Minibatch Loss= 8557.619141, Training Accuracy= 0.82812Iter 52480, Minibatch Loss= 4743.790039, Training Accuracy= 0.78125Iter 53760, Minibatch Loss= 972.031982, Training Accuracy= 0.90625Iter 55040, Minibatch Loss= 6338.499023, Training Accuracy= 0.81250Iter 56320, Minibatch Loss= 9248.547852, Training Accuracy= 0.76562Iter 57600, Minibatch Loss= 5703.125000, Training Accuracy= 0.85938Iter 58880, Minibatch Loss= 10426.742188, Training Accuracy= 0.78125Iter 60160, Minibatch Loss= 4342.888672, Training Accuracy= 0.84375Iter 61440, Minibatch Loss= 7261.451660, Training Accuracy= 0.84375Iter 62720, Minibatch Loss= 3899.342285, Training Accuracy= 0.89062Iter 64000, Minibatch Loss= 3670.834229, Training Accuracy= 0.85938Iter 65280, Minibatch Loss= 5787.583008, Training Accuracy= 0.82812Iter 66560, Minibatch Loss= 6649.801758, Training Accuracy= 0.78125Iter 67840, Minibatch Loss= 2080.366455, Training Accuracy= 0.92188Iter 69120, Minibatch Loss= 7153.665039, Training Accuracy= 0.84375Iter 70400, Minibatch Loss= 5228.269043, Training Accuracy= 0.82812Iter 71680, Minibatch Loss= 9432.683594, Training Accuracy= 0.78125Iter 72960, Minibatch Loss= 1656.271851, Training Accuracy= 0.92188Iter 74240, Minibatch Loss= 2362.667236, Training Accuracy= 0.89062Iter 75520, Minibatch Loss= 4994.963867, Training Accuracy= 0.85938Iter 76800, Minibatch Loss= 691.407837, Training Accuracy= 0.92188Iter 78080, Minibatch Loss= 3261.704102, Training Accuracy= 0.89062Iter 79360, Minibatch Loss= 4440.437500, Training Accuracy= 0.84375Iter 80640, Minibatch Loss= 5165.506348, Training Accuracy= 0.85938Iter 81920, Minibatch Loss= 2232.877441, Training Accuracy= 0.92188Iter 83200, Minibatch Loss= 2232.522949, Training Accuracy= 0.85938Iter 84480, Minibatch Loss= 1921.781006, Training Accuracy= 0.93750Iter 85760, Minibatch Loss= 5770.326172, Training Accuracy= 0.79688Iter 87040, Minibatch Loss= 2022.952881, Training Accuracy= 0.92188Iter 88320, Minibatch Loss= 1921.537109, Training Accuracy= 0.90625Iter 89600, Minibatch Loss= 8427.068359, Training Accuracy= 0.75000Iter 90880, Minibatch Loss= 3777.687012, Training Accuracy= 0.85938Iter 92160, Minibatch Loss= 1407.593140, Training Accuracy= 0.90625Iter 93440, Minibatch Loss= 3493.047363, Training Accuracy= 0.82812Iter 94720, Minibatch Loss= 3549.630371, Training Accuracy= 0.90625Iter 96000, Minibatch Loss= 4050.996338, Training Accuracy= 0.89062Iter 97280, Minibatch Loss= 2689.620605, Training Accuracy= 0.87500Iter 98560, Minibatch Loss= 775.571289, Training Accuracy= 0.92188Iter 99840, Minibatch Loss= 4966.501465, Training Accuracy= 0.81250Iter 101120, Minibatch Loss= 468.303955, Training Accuracy= 0.96875Iter 102400, Minibatch Loss= 4239.565430, Training Accuracy= 0.87500Iter 103680, Minibatch Loss= 2601.228760, Training Accuracy= 0.90625Iter 104960, Minibatch Loss= 2539.465820, Training Accuracy= 0.84375Iter 106240, Minibatch Loss= 3445.990234, Training Accuracy= 0.92188Iter 107520, Minibatch Loss= 2020.942261, Training Accuracy= 0.89062Iter 108800, Minibatch Loss= 2290.115479, Training Accuracy= 0.84375Iter 110080, Minibatch Loss= 5105.082520, Training Accuracy= 0.84375Iter 111360, Minibatch Loss= 2792.384521, Training Accuracy= 0.90625Iter 112640, Minibatch Loss= 3714.771973, Training Accuracy= 0.84375Iter 113920, Minibatch Loss= 2331.506348, Training Accuracy= 0.82812Iter 115200, Minibatch Loss= 5542.223633, Training Accuracy= 0.87500Iter 116480, Minibatch Loss= 2068.789795, Training Accuracy= 0.87500Iter 117760, Minibatch Loss= 3032.279541, Training Accuracy= 0.85938Iter 119040, Minibatch Loss= 2303.545166, Training Accuracy= 0.89062Iter 120320, Minibatch Loss= 1151.952393, Training Accuracy= 0.93750Iter 121600, Minibatch Loss= 2172.850342, Training Accuracy= 0.92188Iter 122880, Minibatch Loss= 1365.023438, Training Accuracy= 0.96875Iter 124160, Minibatch Loss= 2074.203613, Training Accuracy= 0.90625Iter 125440, Minibatch Loss= 3134.413330, Training Accuracy= 0.90625Iter 126720, Minibatch Loss= 1457.720703, Training Accuracy= 0.92188Iter 128000, Minibatch Loss= 7626.540039, Training Accuracy= 0.87500Iter 129280, Minibatch Loss= 332.535400, Training Accuracy= 0.95312Iter 130560, Minibatch Loss= 3173.010010, Training Accuracy= 0.81250Iter 131840, Minibatch Loss= 2052.775879, Training Accuracy= 0.89062Iter 133120, Minibatch Loss= 915.511108, Training Accuracy= 0.93750Iter 134400, Minibatch Loss= 2391.861572, Training Accuracy= 0.89062Iter 135680, Minibatch Loss= 1909.811890, Training Accuracy= 0.87500Iter 136960, Minibatch Loss= 1870.513428, Training Accuracy= 0.87500Iter 138240, Minibatch Loss= 3669.521973, Training Accuracy= 0.87500Iter 139520, Minibatch Loss= 3938.454834, Training Accuracy= 0.87500Iter 140800, Minibatch Loss= 3682.827393, Training Accuracy= 0.85938Iter 142080, Minibatch Loss= 773.881226, Training Accuracy= 0.89062Iter 143360, Minibatch Loss= 2542.113770, Training Accuracy= 0.82812Iter 144640, Minibatch Loss= 1390.634399, Training Accuracy= 0.95312Iter 145920, Minibatch Loss= 1816.718994, Training Accuracy= 0.92188Iter 147200, Minibatch Loss= 2383.392822, Training Accuracy= 0.84375Iter 148480, Minibatch Loss= 2805.485352, Training Accuracy= 0.90625Iter 149760, Minibatch Loss= 4273.811035, Training Accuracy= 0.85938Iter 151040, Minibatch Loss= 2829.233154, Training Accuracy= 0.90625Iter 152320, Minibatch Loss= 2350.943848, Training Accuracy= 0.87500Iter 153600, Minibatch Loss= 1390.092407, Training Accuracy= 0.92188Iter 154880, Minibatch Loss= 1447.536255, Training Accuracy= 0.85938Iter 156160, Minibatch Loss= 2814.929688, Training Accuracy= 0.92188Iter 157440, Minibatch Loss= 1454.184570, Training Accuracy= 0.92188Iter 158720, Minibatch Loss= 1053.621826, Training Accuracy= 0.90625Iter 160000, Minibatch Loss= 268.273071, Training Accuracy= 0.96875Iter 161280, Minibatch Loss= 421.640625, Training Accuracy= 0.93750Iter 162560, Minibatch Loss= 554.997803, Training Accuracy= 0.95312Iter 163840, Minibatch Loss= 756.904907, Training Accuracy= 0.92188Iter 165120, Minibatch Loss= 2533.083496, Training Accuracy= 0.90625Iter 166400, Minibatch Loss= 1262.960327, Training Accuracy= 0.90625Iter 167680, Minibatch Loss= 987.682190, Training Accuracy= 0.93750Iter 168960, Minibatch Loss= 2693.651367, Training Accuracy= 0.87500Iter 170240, Minibatch Loss= 2195.717285, Training Accuracy= 0.92188Iter 171520, Minibatch Loss= 2571.887451, Training Accuracy= 0.87500Iter 172800, Minibatch Loss= 632.802551, Training Accuracy= 0.96875Iter 174080, Minibatch Loss= 1144.768799, Training Accuracy= 0.90625Iter 175360, Minibatch Loss= 1107.609863, Training Accuracy= 0.89062Iter 176640, Minibatch Loss= 962.998108, Training Accuracy= 0.93750Iter 177920, Minibatch Loss= 475.736450, Training Accuracy= 0.95312Iter 179200, Minibatch Loss= 454.031738, Training Accuracy= 0.96875Iter 180480, Minibatch Loss= 1643.504272, Training Accuracy= 0.93750Iter 181760, Minibatch Loss= 520.336853, Training Accuracy= 0.96875Iter 183040, Minibatch Loss= 5099.615723, Training Accuracy= 0.82812Iter 184320, Minibatch Loss= 1832.333374, Training Accuracy= 0.92188Iter 185600, Minibatch Loss= 5085.391602, Training Accuracy= 0.87500Iter 186880, Minibatch Loss= 1165.275635, Training Accuracy= 0.90625Iter 188160, Minibatch Loss= 436.611694, Training Accuracy= 0.95312Iter 189440, Minibatch Loss= 729.550781, Training Accuracy= 0.92188Iter 190720, Minibatch Loss= 631.992798, Training Accuracy= 0.92188Iter 192000, Minibatch Loss= 254.609497, Training Accuracy= 0.95312Iter 193280, Minibatch Loss= 1927.740479, Training Accuracy= 0.90625Iter 194560, Minibatch Loss= 0.000000, Training Accuracy= 1.00000Iter 195840, Minibatch Loss= 735.920166, Training Accuracy= 0.95312Iter 197120, Minibatch Loss= 93.257446, Training Accuracy= 0.98438Iter 198400, Minibatch Loss= 328.502441, Training Accuracy= 0.93750Iter 199680, Minibatch Loss= 1295.930298, Training Accuracy= 0.92188Optimization Finished!('Testing Accuracy:', 0.96875)