使用tensorflow实现AlexNet

时间:2022-09-15 21:40:49

AlexNet是2012年ImageNet比赛的冠军,虽然过去了很长时间,但是作为深度学习中的经典模型,AlexNet不但有助于我们理解其中所使用的很多技巧,而且非常有助于提升我们使用深度学习工具箱的熟练度。尤其是我刚入门深度学习,迫切需要一个能让自己熟悉tensorflow的小练习,于是就有了这个小玩意儿......

先放上我的代码:https://github.com/hjptriplebee/AlexNet_with_tensorflow

如果想运行代码,详细的配置要求都在上面链接的readme文件中了。本文建立在一定的tensorflow基础上,不会对太细的点进行说明。

模型结构

使用tensorflow实现AlexNet

关于模型结构网上的文献很多,我这里不赘述,一会儿都在代码里解释。

有一点需要注意,AlexNet将网络分成了上下两个部分,在论文中两部分结构完全相同,唯一不同的是他们放在不同GPU上训练,因为每一层的feature map之间都是独立的(除了全连接层),所以这相当于是提升训练速度的一种方法。很多AlexNet的复现都将上下两部分合并了,因为他们都是在单个GPU上运行的。虽然我也是在单个GPU上运行,但是我还是很想将最原始的网络结构还原出来,所以我的代码里也是分开的。

模型定义

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
def maxPoolLayer(x, kHeight, kWidth, strideX, strideY, name, padding = "SAME"):
  """max-pooling"""
  return tf.nn.max_pool(x, ksize = [1, kHeight, kWidth, 1],
             strides = [1, strideX, strideY, 1], padding = padding, name = name)
 
def dropout(x, keepPro, name = None):
  """dropout"""
  return tf.nn.dropout(x, keepPro, name)
 
def LRN(x, R, alpha, beta, name = None, bias = 1.0):
  """LRN"""
  return tf.nn.local_response_normalization(x, depth_radius = R, alpha = alpha,
                       beta = beta, bias = bias, name = name)
 
def fcLayer(x, inputD, outputD, reluFlag, name):
  """fully-connect"""
  with tf.variable_scope(name) as scope:
    w = tf.get_variable("w", shape = [inputD, outputD], dtype = "float")
    b = tf.get_variable("b", [outputD], dtype = "float")
    out = tf.nn.xw_plus_b(x, w, b, name = scope.name)
    if reluFlag:
      return tf.nn.relu(out)
    else:
      return out
 
def convLayer(x, kHeight, kWidth, strideX, strideY,
       featureNum, name, padding = "SAME", groups = 1):#group为2时等于AlexNet中分上下两部分
  """convlutional"""
  channel = int(x.get_shape()[-1])#获取channel
  conv = lambda a, b: tf.nn.conv2d(a, b, strides = [1, strideY, strideX, 1], padding = padding)#定义卷积的匿名函数
  with tf.variable_scope(name) as scope:
    w = tf.get_variable("w", shape = [kHeight, kWidth, channel/groups, featureNum])
    b = tf.get_variable("b", shape = [featureNum])
 
    xNew = tf.split(value = x, num_or_size_splits = groups, axis = 3)#划分后的输入和权重
    wNew = tf.split(value = w, num_or_size_splits = groups, axis = 3)
 
    featureMap = [conv(t1, t2) for t1, t2 in zip(xNew, wNew)] #分别提取feature map
    mergeFeatureMap = tf.concat(axis = 3, values = featureMap) #feature map整合
    # print mergeFeatureMap.shape
    out = tf.nn.bias_add(mergeFeatureMap, b)
    return tf.nn.relu(tf.reshape(out, mergeFeatureMap.get_shape().as_list()), name = scope.name) #relu后的结果

定义了卷积、pooling、LRN、dropout、全连接五个模块,其中卷积模块因为将网络的上下两部分分开了,所以比较复杂。接下来定义AlexNet。

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
class alexNet(object):
  """alexNet model"""
  def __init__(self, x, keepPro, classNum, skip, modelPath = "bvlc_alexnet.npy"):
    self.X = x
    self.KEEPPRO = keepPro
    self.CLASSNUM = classNum
    self.SKIP = skip
    self.MODELPATH = modelPath
    #build CNN
    self.buildCNN()
 
  def buildCNN(self):
    """build model"""
    conv1 = convLayer(self.X, 11, 11, 4, 4, 96, "conv1", "VALID")
    pool1 = maxPoolLayer(conv1, 3, 3, 2, 2, "pool1", "VALID")
    lrn1 = LRN(pool1, 2, 2e-05, 0.75, "norm1")
 
    conv2 = convLayer(lrn1, 5, 5, 1, 1, 256, "conv2", groups = 2)
    pool2 = maxPoolLayer(conv2, 3, 3, 2, 2, "pool2", "VALID")
    lrn2 = LRN(pool2, 2, 2e-05, 0.75, "lrn2")
 
    conv3 = convLayer(lrn2, 3, 3, 1, 1, 384, "conv3")
 
    conv4 = convLayer(conv3, 3, 3, 1, 1, 384, "conv4", groups = 2)
 
    conv5 = convLayer(conv4, 3, 3, 1, 1, 256, "conv5", groups = 2)
    pool5 = maxPoolLayer(conv5, 3, 3, 2, 2, "pool5", "VALID")
 
    fcIn = tf.reshape(pool5, [-1, 256 * 6 * 6])
    fc1 = fcLayer(fcIn, 256 * 6 * 6, 4096, True, "fc6")
    dropout1 = dropout(fc1, self.KEEPPRO)
 
    fc2 = fcLayer(dropout1, 4096, 4096, True, "fc7")
    dropout2 = dropout(fc2, self.KEEPPRO)
 
    self.fc3 = fcLayer(dropout2, 4096, self.CLASSNUM, True, "fc8")
 
  def loadModel(self, sess):
    """load model"""
    wDict = np.load(self.MODELPATH, encoding = "bytes").item()
    #for layers in model
    for name in wDict:
      if name not in self.SKIP:
        with tf.variable_scope(name, reuse = True):
          for p in wDict[name]:
            if len(p.shape) == 1
              #bias 只有一维
              sess.run(tf.get_variable('b', trainable = False).assign(p))
            else:
              #weights 
              sess.run(tf.get_variable('w', trainable = False).assign(p))

buildCNN函数完全按照alexnet的结构搭建网络。
loadModel函数从模型文件中读取参数,采用的模型文件见github上的readme说明。
至此,我们定义了完整的模型,下面开始测试模型。

模型测试

ImageNet训练的AlexNet有很多类,几乎包含所有常见的物体,因此我们随便从网上找几张图片测试。比如我直接用了之前做项目的渣土车图片:

使用tensorflow实现AlexNet

然后编写测试代码:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
#some params
dropoutPro = 1
classNum = 1000
skip = []
#get testImage
testPath = "testModel"
testImg = []
for f in os.listdir(testPath):
  testImg.append(cv2.imread(testPath + "/" + f))
 
imgMean = np.array([104, 117, 124], np.float)
x = tf.placeholder("float", [1, 227, 227, 3])
 
model = alexnet.alexNet(x, dropoutPro, classNum, skip)
score = model.fc3
softmax = tf.nn.softmax(score)
 
with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  model.loadModel(sess) #加载模型
 
  for i, img in enumerate(testImg):
    #img preprocess
    test = cv2.resize(img.astype(np.float), (227, 227)) #resize成网络输入大小
    test -= imgMean #去均值
    test = test.reshape((1, 227, 227, 3)) #拉成tensor
    maxx = np.argmax(sess.run(softmax, feed_dict = {x: test}))
    res = caffe_classes.class_names[maxx] #取概率最大类的下标
    #print(res)
    font = cv2.FONT_HERSHEY_SIMPLEX
    cv2.putText(img, res, (int(img.shape[0]/3), int(img.shape[1]/3)), font, 1, (0, 255, 0), 2)#绘制类的名字
    cv2.imshow("demo", img) 
    cv2.waitKey(5000) #显示5秒

如上代码所示,首先需要设置一些参数,然后读取指定路径下的测试图像,再对模型做一个初始化,最后是真正测试代码。测试结果如下:

使用tensorflow实现AlexNet

以上就是本文的全部内容,希望对大家的学习有所帮助,也希望大家多多支持服务器之家。

原文链接:http://blog.csdn.net/accepthjp/article/details/69999309