版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/chai_zheng/article/details/79779371
本文主要介绍CNN复兴之作，2012年Alex所作ImageNet Classification with Deep Convolutional
Neural Networks及其基于TensorFlow GPU 1.3.0 rc1版本的实现。
AlexNet采用上图架构。请注意上图，由于当时的GPU运算性能有限，文章将2、4、5三个卷积层分为两部分，分配至两个GPU并行计算，而2→3为一个GPU。具体参数我写到了下图中。
相比于前人工作，AlexNet主要采用了如下几点改进方法：
1. ReLU activation function.
2. Local Response Normalization.
3. Overlapping Pooling.
4. Dropout.
5. Data augmentation.
5. Multiple GPU.
代码分为两部分，AlexNet.py为AlexNet网络构建部分，train.py为模型验证部分。搭建好模型之后，我们应用由caffe训练好的权重w和偏置b转换为tensorflow的bvlc_alexnet.npy参数作为模型的参数，以省去多日训练之苦。参数可从我的百度云下载。链接：https://pan.baidu.com/s/1V6MPQvXv0ler5on0znlIXA 密码：n5im，若失效可从http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/bvlc_alexnet.npy 下载。除了两份代码与一份预训练参数之外，还需要的ImageNet 1000类名与测试图片可从这里下载。
代码部分
第一部分，AlexNet.py
# !/usr/bin/env python3
# coding=utf-8
"""
AlexNet Using TensorFlow
Author : Chai Zheng, Ph.D.@Zhejiang University, Hangzhou
Email : [email protected]
Blog : http://blog.csdn.net/chai_zheng/
Github : https://github.com/Chai-Zheng/
Date : 2018.4.1
"""
import tensorflow as tf
import numpy as np
class AlexNet(object):
def __init__(self, x, keep_prob, skip_layer, weights_path='bvlc_alexnet.npy'):
self.x = x
self.keep_prob = keep_prob
self.skip_layer = skip_layer
self.weights_path = weights_path
self.build_AlexNet()
def build_AlexNet(self):
conv1 = conv_layer(self.x, 96, 11, 11, 4, 4, 'conv1', groups=1, padding='VALID')
norm1 = LRN_layer(conv1, 2, 1e-4, 0.75, 'norm1')
pool1 = max_pool_layer(norm1, 3, 3, 2, 2, 'pool1')
conv2 = conv_layer(pool1, 256, 5, 5, 1, 1, 'conv2', groups=2)
norm2 = LRN_layer(conv2, 2, 1e-4, 0.75, 'norm2')
pool2 = max_pool_layer(norm2, 3, 3, 2, 2, 'pool2', padding='VALID')
conv3 = conv_layer(pool2, 384, 3, 3, 1, 1, 'conv3')
conv4 = conv_layer(conv3, 384, 3, 3, 1, 1, 'conv4', groups=2)
conv5 = conv_layer(conv4, 256, 3, 3, 1, 1, 'conv5', groups=2)
pool5 = max_pool_layer(conv5, 3, 3, 2, 2, 'pool5', 'VALID')
pool5_flatted = tf.reshape(pool5, [-1, 6*6*256], 'pool5_flatted')
fc6 = fc_layer(pool5_flatted, 6*6*256, 4096, name='fc6')
dropout6 = dropout(fc6, self.keep_prob)
fc7 = fc_layer(dropout6, 4096, 4096, name='fc7')
dropout7 = dropout(fc7, self.keep_prob)
self.fc8 = fc_output_layer(dropout7, 4096, 1000, name='fc8')
def load_weights(self, sess):
weights_dict = np.load(self.weights_path, encoding='bytes').item()
for name in weights_dict:
if name not in self.skip_layer:
with tf.variable_scope(name, reuse=True):
for p in weights_dict[name]:
if len(p.shape) == 1: # bias
var = tf.get_variable('b', trainable=False)
sess.run(var.assign(p))
else: # weights
var = tf.get_variable('w', trainable=False)
sess.run(var.assign(p))
def weights(shape):
return tf.get_variable('w', shape, trainable=True)
def bias(shape):
return tf.get_variable('b', shape, trainable=True)
def conv_layer(x, filter_num, filter_height, filter_width, stride_x, stride_y, name, groups=1, padding='SAME'):
channel = int(x.shape[-1])
conv2d = lambda a, b: tf.nn.conv2d(input=a, filter=b, strides=[1, stride_y, stride_x, 1], padding=padding)
with tf.variable_scope(name) as scope:
w = weights([filter_height, filter_width, int(channel/groups), filter_num])
b = bias([filter_num])
x_split = tf.split(value=x, num_or_size_splits=groups, axis=3)
w_split = tf.split(value=w, num_or_size_splits=groups, axis=3)
conv_split = [conv2d(m, n) for m, n in zip(x_split, w_split)]
conv_merge = tf.concat(conv_split, axis=3)
return tf.nn.relu(conv_merge+b, name='scope.name')
def LRN_layer(x, R, alpha, beta, name, bias=1.0):
return tf.nn.local_response_normalization(x, depth_radius=R, alpha=alpha, beta=beta, name=name, bias=bias)
def max_pool_layer(x, filter_height, filter_width, stride_x, stride_y, name, padding='SAME'):
return tf.nn.max_pool(x, ksize=[1, filter_height, filter_width, 1], strides=[1, stride_y, stride_x, 1],
padding=padding, name=name)
def dropout(x, keep_prob, name=None):
return tf.nn.dropout(x, keep_prob, name)
def fc_layer(x, input_num, output_num, name):
with tf.variable_scope(name) as scope:
w = weights([input_num, output_num])
b = bias([output_num])
return tf.nn.relu(tf.matmul(x, w)+b)
def fc_output_layer(x, input_num, output_num, name):
with tf.variable_scope(name) as scope:
w = weights([input_num, output_num])
b = bias([output_num])
return tf.nn.softmax(tf.matmul(x, w)+b)
第二部分，模型调用参数与验证 train.py
# !/usr/bin/env python3
# coding=utf-8
"""
AlexNet Using TensorFlow
Author : Chai Zheng, Ph.D.@Zhejiang University, Hangzhou
Email : [email protected]
Blog : http://blog.csdn.net/chai_zheng/
Github : https://github.com/Chai-Zheng/
Date : 2018.4.1
"""
import os
import cv2
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import AlexNet
import caffe_classes
path = 'test_images'
withPath = lambda f: '{}/{}'.format(path, f)
test_images = dict((f, cv2.imread(withPath(f))) for f in os.listdir(path) if os.path.isfile(withPath(f)))
if test_images.values():
dropout_prob = 1.0
skip_layer = []
image_mean = np.array([104, 117, 124], np.float)
x = tf.placeholder('float', [1, 227, 227, 3])
trained_model = AlexNet.AlexNet(x, dropout_prob, skip_layer)
y_predict = trained_model.fc8
fig = plt.figure()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
trained_model.load_weights(sess)
j = 0
for i, img in test_images.items():
image_resized = cv2.resize(img.astype(np.float), (227, 227))-image_mean
probs = sess.run(y_predict, feed_dict={x: image_resized.reshape(1, 227, 227, 3)})
max_prob = np.max(probs)
y_pre = caffe_classes.class_names[np.argmax(probs)]
fig.add_subplot(1, 3, j+1)
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.title('class:{} probability:{}'.format(y_pre, max_prob))
j += 1
plt.show()
最终效果如图：
参考：
[1] Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[C]//Advances in neural information processing systems. 2012: 1097-1105.
[2] https://kratzert.github.io/2017/02/24/finetuning-alexnet-with-tensorflow.html
[3] https://github.com/kratzert/finetune_alexnet_with_tensorflow
[4] https://blog.csdn.net/zyqdragon/article/details/72353420