我们知道,再网络训练好之后,只需要forward过程就能做预测,当然,我们也可以直接把这个网络当成一个feature extractor来用,可以直接用任何一层的输出作为特征,根据R-CNN论文对Alexnet的实验结果,如果不做fine-tuning,pool5和fc6和fc7的特征效果并没有很强的提升,所以,如果直接用作feature extractor,直接用pool的最后一层输出就OK.
这里是一个简单的演示,提取VGG的pool5层特征,存储为.mat文件;
import scipy.io as sio
from scipy.misc import imread, imresize
sess = tf.Session()
imgs = tf.placeholder(tf.float32, [None, 224, 224, 3])
vgg = vgg16(imgs, '/aa/data/vgg16_weights.npz', sess)
img1 = imread('/aa/data/laska.png', mode='RGB')
img1 = imresize(img1, (224, 224))
path = '/aa/data/AllSample/'
for i in range(1,211):
img = imread(path+str(i)+'.jpg',mode='RGB')
print(path+str(i)+'.jpg')
img = imresize(img, (224, 224))
feature = sess.run(vgg.pool5, feed_dict={vgg.imgs: [img]})
feature = np.reshape(feature,[7,7,512])
dic = {'features':feature}
sio.savemat('/aa/data/features/'+str(i)+'.mat',dic)
# features = feature.eval(session=sess)
# features = np.reshape(features,[7,7,512])