参考
http://blog.csdn.net/jsond/article/details/72667829
资源：
1.相关的vgg模型下载网址
http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat
2.ImageNet 1000种分类以及排列
https://github.com/sh1r0/caffe-Android-demo/blob/master/app/src/main/assets/synset_words.txt（如果下载单个txt格式不对的话就整包下载）
这里以E网络为测试模型VGG19
#coding=utf-8 import numpy as np import scipy.misc import scipy.io as sio import tensorflow as tf import os ##卷积层 def _conv_layer(input, weight, bias): conv = tf.nn.conv2d(input, tf.constant(weight), strides=(1, 1, 1, 1), padding='SAME') return tf.nn.bias_add(conv, bias) ##池化层 def _pool_layer(input): return tf.nn.max_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1), padding='SAME') ##全链接层 def _fc_layer(input, weights, bias): shape = input.get_shape().as_list() dim = 1 for d in shape[1:]: dim *= d x = tf.reshape(input, [-1, dim]) fc = tf.nn.bias_add(tf.matmul(x, weights), bias) return fc ##softmax输出层 def _softmax_preds(input): preds = tf.nn.softmax(input, name='prediction') return preds ##图片处里前减去均值 def _preprocess(image, mean_pixel): return image - mean_pixel ##加均值 显示图片 def _unprocess(image, mean_pixel): return image + mean_pixel ##读取图片 并压缩 def _get_img(src, img_size=False): img = scipy.misc.imread(src, mode='RGB') if not (len(img.shape) == 3 and img.shape[2] == 3): img = np.dstack((img, img, img)) if img_size != False: img = scipy.misc.imresize(img, img_size) return img.astype(np.float32) ##获取名列表 def list_files(in_path): files = [] for (dirpath, dirnames, filenames) in os.walk(in_path): # print("dirpath=%s, dirnames=%s, filenames=%s"%(dirpath, dirnames, filenames)) files.extend(filenames) break return files ##获取文件路径列表dir+filename def _get_files(img_dir): files = list_files(img_dir) return [os.path.join(img_dir, x) for x in files] ##获得图片lable列表 def _get_allClassificationName(file_path): f = open(file_path, 'r') lines = f.readlines() f.close() return lines ##构建cnn前向传播网络 def net(data, input_image): layers = ( 'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2', 'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'conv3_4', 'relu3_4', 'pool3', 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 'relu4_3', 'conv4_4', 'relu4_4', 'pool4', 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3', 'conv5_4', 'relu5_4', 'pool5', 'fc6', 'relu6', 'fc7', 'relu7', 'fc8', 'softmax' ) weights = data['layers'][0] net = {} current = input_image for i, name in enumerate(layers): kind = name[:4] if kind == 'conv': kernels, bias = weights[i][0][0][0][0] kernels = np.transpose(kernels, (1, 0, 2, 3)) bias = bias.reshape(-1) current = _conv_layer(current, kernels, bias) elif kind == 'relu': current = tf.nn.relu(current) elif kind == 'pool': current = _pool_layer(current) elif kind == 'soft': current = _softmax_preds(current) kind2 = name[:2] if kind2 == 'fc': kernels1, bias1 = weights[i][0][0][0][0] kernels1 = kernels1.reshape(-1, kernels1.shape[-1]) bias1 = bias1.reshape(-1) current = _fc_layer(current, kernels1, bias1) net[name] = current assert len(net) == len(layers) return net, mean_pixel, layers if __name__ == '__main__': imagenet_path = 'imagenet-vgg-verydeep-19.mat' image_dir = 'images/' data = sio.loadmat(imagenet_path) ##加载ImageNet mat模型 mean = data['normalization'][0][0][0] mean_pixel = np.mean(mean, axis=(0, 1)) ##获取图片像素均值 lines = _get_allClassificationName('synset_words.txt') ##加载ImageNet mat标签 images = _get_files(image_dir) ##获取图片路径列表 with tf.Session() as sess: for i, imgPath in enumerate(images): image = _get_img(imgPath, (224, 224, 3)); ##加载图片并压缩到标准格式=>224 224 image_pre = _preprocess(image, mean_pixel) # image_pre = image_pre.transpose((2, 0, 1)) image_pre = np.expand_dims(image_pre, axis=0) image_preTensor = tf.convert_to_tensor(image_pre) image_preTensor = tf.to_float(image_preTensor) # Test pretrained model nets, mean_pixel, layers = net(data, image_preTensor) preds = nets['softmax'] predsSortIndex = np.argsort(-preds[0].eval()) print(' #####%s#######' % imgPath) for i in range(3): ##输出前3种分类 nIndex = predsSortIndex classificationName = lines[nIndex[i]] ##分类名称 problity = preds[0][nIndex[i]] ##某一类型概率 print('%d.ClassificationName=%s Problity=%f' % ((i + 1), classificationName, problity.eval())) sess.close()
分类结果
#####images/airplay.jpg####### 1.ClassificationName=n04228054 ski Problity=0.177715 2.ClassificationName=n04286575 spotlight, spot Problity=0.108483 3.ClassificationName=n04127249 safety pin Problity=0.026277 #####images/bird.jpg####### 1.ClassificationName=n01608432 kite Problity=0.096818 2.ClassificationName=n01833805 hummingbird Problity=0.072687 3.ClassificationName=n02231487 walking stick, walkingstick, stick insect Problity=0.069186 #####images/cat1.jpg####### 1.ClassificationName=n02123045 tabby, tabby cat Problity=0.232015 2.ClassificationName=n02123159 tiger cat Problity=0.094694 3.ClassificationName=n02124075 Egyptian cat Problity=0.030673 #####images/cat2.jpg####### 1.ClassificationName=n02123045 tabby, tabby cat Problity=0.333797 2.ClassificationName=n02123159 tiger cat Problity=0.164726 3.ClassificationName=n02124075 Egyptian cat Problity=0.057272 #####images/cat3.jpg####### 1.ClassificationName=n03887697 paper towel Problity=0.086723 2.ClassificationName=n02111889 Samoyed, Samoyede Problity=0.055845 3.ClassificationName=n03131574 crib, cot Problity=0.052640 #####images/dog1.jpg####### 1.ClassificationName=n02096585 Boston bull, Boston terrier Problity=0.429622 2.ClassificationName=n02108089 boxer Problity=0.199422 3.ClassificationName=n02093256 Staffordshire bullterrier, Staffordshire bull terrier Problity=0.093615 #####images/dog2.jpg####### 1.ClassificationName=n02085936 Maltese dog, Maltese terrier, Maltese Problity=0.172208 2.ClassificationName=n03445777 golf ball Problity=0.139949 3.ClassificationName=n02259212 leafhopper Problity=0.118109 #####images/lena.jpg####### 1.ClassificationName=n02869837 bonnet, poke bonnet Problity=0.130357 2.ClassificationName=n04356056 sunglasses, dark glasses, shades Problity=0.066170 3.ClassificationName=n04355933 sunglass Problity=0.043199 #####images/sky.jpg####### 1.ClassificationName=n03733281 maze, labyrinth Problity=0.711163 2.ClassificationName=n03065424 coil, spiral, volute, whorl, helix Problity=0.181123 3.ClassificationName=n04259630 sombrero Problity=0.010005