安装caffe踩雷经历
1 编译make all 时出现,cuda.h 找不到,因为cuda没安装,或安装不正确
安装cuda时,cuda_10.0.130_410.48_linux.run,driver一向选择no,否则安装不成功
安装cudnn时,是把cudnn-10.0-linux-x64-v7.4.2.24.tgz解压缩,按照https://www.cnblogs.com/pertor/p/8733010.html 文章参考
2 ./train_xxxxx.sh时,出现_ZN5caffexxxx错误是库路径不正确,将编译好的caffe/build/lib 输出到~/.bashrc中,
export LD_LIBRARY_PATH=/home/youname/caffe-master/build/lib:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
具体方法见网页
3 解析log时,log重定向出现makers错误,采用list进行包装,如下所示
list(markers.keys())[idx]
4 draw_net画出网络结构图时,安装graphviz,pydot,如文章https://www.jianshu.com/p/0abc1488e331,出现caffe._caffe错误时,原因在caffe-master目录下没有进行,make pycaffe操作,完成后出现 ZN5boost错误,
在make.config中注意
PYTHON_INCLUDE := /usr/include/python3.6 \
/usr/lib/python3/dist-packages/numpy/core/include
# Uncomment to use Python 3 (default is Python 2)
PYTHON_LIBRARIES := boost_python3 python3.6m
#PYTHON_INCLUDE := /usr/include/python3.6m \
# /usr/lib/python3.6/dist-packages/numpy/core/include
boost库路径错误,重新make all
再make pycaffe,
再出现错误一般是pip3 install --upgrade scikit-image scikit-learn,将caffe组件升级即可按照https://www.jianshu.com/p/0abc1488e331画出网络结构
以上是我为期两周的踩雷经历,也是一天搞不定,放了几天,再次拿起来,重新一鼓作气搞定了!
That's all,Enjoy your caffe life! Goog Job!