转载注明出处：http://blog.csdn.net/accepthjp/article/details/70170217
接上一篇AlexNet，本文讲述使用tensorflow实现VGG19网络。
VGG网络与AlexNet类似，也是一种CNN，VGG在2014年的 ILSVRC localization and classification 两个问题上分别取得了第一名和第二名。VGG网络非常深，通常有16－19层，卷积核大小为 3 x 3，16和19层的区别主要在于后面三个卷积部分卷积层的数量。第二个用tensorflow独立完成的小玩意儿......
同样先放上我的代码，由AlexNet的代码改过来的：https://github.com/hjptriplebee/VGG19_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
可以看到VGG的前几层为卷积和maxpool的交替，每个卷积包含多个卷积层，后面紧跟三个全连接层。激活函数采用Relu，训练采用了dropout，但并没有像AlexNet一样采用LRN（论文给出的理由是加LRN实验效果不好）。
模型定义
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 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"): """convlutional""" channel = int(x.get_shape()[-1]) #获取channel数 with tf.variable_scope(name) as scope: w = tf.get_variable("w", shape = [kHeight, kWidth, channel, featureNum]) b = tf.get_variable("b", shape = [featureNum]) featureMap = tf.nn.conv2d(x, w, strides = [1, strideY, strideX, 1], padding = padding) out = tf.nn.bias_add(featureMap, b) return tf.nn.relu(tf.reshape(out, featureMap.get_shape().as_list()), name = scope.name)
定义了卷积、pooling、dropout、全连接五个模块，使用了上一篇AlexNet中的代码，其中卷积模块去除了group参数，因为网络没有像AlexNet一样分成两部分。接下来定义VGG19。
class VGG19(object): """VGG model""" def __init__(self, x, keepPro, classNum, skip, modelPath = "vgg19.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_1 = convLayer(self.X, 3, 3, 1, 1, 64, "conv1_1" ) conv1_2 = convLayer(conv1_1, 3, 3, 1, 1, 64, "conv1_2") pool1 = maxPoolLayer(conv1_2, 2, 2, 2, 2, "pool1") conv2_1 = convLayer(pool1, 3, 3, 1, 1, 128, "conv2_1") conv2_2 = convLayer(conv2_1, 3, 3, 1, 1, 128, "conv2_2") pool2 = maxPoolLayer(conv2_2, 2, 2, 2, 2, "pool2") conv3_1 = convLayer(pool2, 3, 3, 1, 1, 256, "conv3_1") conv3_2 = convLayer(conv3_1, 3, 3, 1, 1, 256, "conv3_2") conv3_3 = convLayer(conv3_2, 3, 3, 1, 1, 256, "conv3_3") conv3_4 = convLayer(conv3_3, 3, 3, 1, 1, 256, "conv3_4") pool3 = maxPoolLayer(conv3_4, 2, 2, 2, 2, "pool3") conv4_1 = convLayer(pool3, 3, 3, 1, 1, 512, "conv4_1") conv4_2 = convLayer(conv4_1, 3, 3, 1, 1, 512, "conv4_2") conv4_3 = convLayer(conv4_2, 3, 3, 1, 1, 512, "conv4_3") conv4_4 = convLayer(conv4_3, 3, 3, 1, 1, 512, "conv4_4") pool4 = maxPoolLayer(conv4_4, 2, 2, 2, 2, "pool4") conv5_1 = convLayer(pool4, 3, 3, 1, 1, 512, "conv5_1") conv5_2 = convLayer(conv5_1, 3, 3, 1, 1, 512, "conv5_2") conv5_3 = convLayer(conv5_2, 3, 3, 1, 1, 512, "conv5_3") conv5_4 = convLayer(conv5_3, 3, 3, 1, 1, 512, "conv5_4") pool5 = maxPoolLayer(conv5_4, 2, 2, 2, 2, "pool5") fcIn = tf.reshape(pool5, [-1, 7*7*512]) fc6 = fcLayer(fcIn, 7*7*512, 4096, True, "fc6") dropout1 = dropout(fc6, self.KEEPPRO) fc7 = fcLayer(dropout1, 4096, 4096, True, "fc7") dropout2 = dropout(fc7, self.KEEPPRO) self.fc8 = 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VGG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VGG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图片，为了避免审美疲劳，我们不只用渣土车，还要用挖掘机、采沙船：
然后编写测试代码：
parser = argparse.ArgumentParser(description='Classify some images.') parser.add_argument('mode', choices=['folder', 'url'], default='folder') parser.add_argument('path', help='Specify a path [e.g. testModel]') args = parser.parse_args(sys.argv[1:]) if args.mode == 'folder': #测试方式为本地文件夹 #get testImage withPath = lambda f: '{}/{}'.format(args.path,f) testImg = dict((f,cv2.imread(withPath(f))) for f in os.listdir(args.path) if os.path.isfile(withPath(f))) elif args.mode == 'url': #测试方式为URL def url2img(url): #获取URL图像 '''url to image''' resp = urllib.request.urlopen(url) image = np.asarray(bytearray(resp.read()), dtype="uint8") image = cv2.imdecode(image, cv2.IMREAD_COLOR) return image testImg = {args.path:url2img(args.path)} if testImg.values(): #some params dropoutPro = 1 classNum = 1000 skip = [] imgMean = np.array([104, 117, 124], np.float) x = tf.placeholder("float", [1, 224, 224, 3]) model = vgg19.VGG19(x, dropoutPro, classNum, skip) score = model.fc8 softmax = tf.nn.softmax(score) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) model.loadModel(sess) #加载模型 for key,img in testImg.items(): #img preprocess resized = cv2.resize(img.astype(np.float), (224, 224)) - imgMean #去均值 maxx = np.argmax(sess.run(softmax, feed_dict = {x: resized.reshape((1, 224, 224, 3))})) #网络输入为224*224 res = caffe_classes.class_names[maxx] 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) #在图像上绘制结果 print("{}: {}\n----".format(key,res)) #输出测试结果 cv2.imshow("demo", img) cv2.waitKey(0)
如果你看完了我AlexNet的博客，那么一定会发现我这里的测试代码做了一些小的修改，增加了URL测试的功能，可以测试网上的图像 ，测试结果如下：
