Enet caffe训练错误

左手的ㄟ右手 2022-05-23 06:35 228阅读 0赞

转自：https://blog.csdn.net/hong\_\_fang/article/details/78530390

根据ENet说明https://github.com/TimoSaemann/ENet/tree/master/Tutorial ，进行训练遇到以下问题

1 No module named spatial\_dropout

新增加的Python层如果用caffe-root/caffe train -solver solver.prototxt方式可能会出现找不到新添加的Python层，此时需要将新添加的Python层文件的路径添加到环境变量中，通过终端python后输入import \*\*.py看是否添加，然后通过Python接口来训练，如下：

ENet通过python接口定义了新层spatial\_dropout，根据说明直接在终端启动训练，出现“No module named spatial\_dropout”，原因是caffe，python层所在文件夹路径，没有添加到系统PYTHONPATH，一个解决方法是：用Python接口来启动训练：

\[python\] [view plain][] [copy][view plain]

若采用微调的方式进行训练则如下，参考https://blog.csdn.net/lilai619/article/details/54343395

\[python\] [view plain][] [copy][view plain]

1.  **import** sys 
2.  caffe\_root = 'home/f/caffe/'
3.  sys.path.insert(0,caffe\_root+'python') 
4.  
5.  **import** caffe 
6.  caffe.set\_device(0) 
7.  caffe.set\_mode\_gpu() 
8.  solver = caffe.SGDSolver('/home/xxx/data/solver.prototxt') 
9.  solver.net.copy\_from('models/bvlc\_reference\_caffenet/bvlc\_reference\_caffenet.caffemodel')   
    
10. \# We run the solver for niter times, and record the training loss. 
11. for it in range(niter): 
12.  solver.step(1) \# SGD by Caffe

2 drop1\_0\_3 新建时， img\_height == height 验证失败

data层的new\_height new\_weight不能删掉，根据输入图像设定大小即可。

\[python\] [view plain][] [copy][view plain]

1.  layer \{ 
2.   name: "data"
3.   type: "DenseImageData"
4.   top: "data"
5.   top: "label"
6.   dense\_image\_data\_param \{ 
7.   source: "ENet/dataset/train\_fine\_2columns.txt"
8.   batch\_size: 4
9.   shuffle: true 
10.  new\_height: 512
11.  new\_width: 1024
12.  label\_divide\_factor: 8
13.  \} 
14. \}

[view plain]: https://blog.csdn.net/hong__fang/article/details/78530390#