去年提出的DenseNet获得今年CVPR2017最佳论文
在github上已经发布了代码,在此贴一下链接:https://github.com/liuzhuang13/DenseNetCaffe
在readMe里面详细的讲了如何构建DenseNet的prototxt
0. Get the CIFAR data prepared following the [Caffe's official CIFAR tutorial](http://caffe.berkeleyvision.org/gathered/examples/cifar10.html).
1. make\_densenet.py contains the code to generate the network and solver prototxt file. First change the data path in function make\_net() and preprocessing mean file in function densenet() to your own path of corresponding data file.
2. By default make\_densenet.py generates a DenseNet with Depth L=40, Growth rate k=12 and Dropout=0.2. To experiment with different settings, change the code accordingly (see the comments in the code). Example prototxt files are already included. Use ```python densenet_make.py``` to generate new prototxt files.
3. Change the caffe path in train.sh. Then use ```sh train.sh``` to train a DenseNet.
1)先根据官网的cifar10使用,下载并且用cifar10的shell,具体见我另一篇博客:
http://blog.csdn.net/mdjxy63/article/details/76359262
2)运行make\_densenet.py文件,构造prototxt的train_test神经网络
3)运行train.sh,这里需要修改train.sh里面关于caffe_home的配置路径,才能进行训练