使用Caffe过程中的若干问题

cannot find lopencv_xxxx:

apt-cache search opencv

sudo apt-get install yyy

一次解决烦恼?如果用的是opencv3+的话,要面临更多痛苦。

这里的问题应当是make之后没有install,在sudo gedit /etc/ld.so.conf.d/opencv.conf中添加/usr/local/lib和/usr/local/lib/x86_64-linux-gnu

Check failed: fd != -1 (-1 vs. -1):

文件路径不恰当,一般把[caffe]/当做项目根,其他文件以此做相对路径。

  

Check failed: net_->num_inputs() == 1 (0 vs. 1) Network should have exactly one input:

引错参数文件

  

Check failed: error == cudaSuccess (2 vs. 0) out of memory:

减小batch_size(每次迭代送入的样本数),最好为8的倍数

设定test_iter(测试时调入的batch数量) = TEST样本总量/batch_size(TEST的) (进一法)

增大snapshot(每迭代xx次生成一个模型)

Check failed: labels_.size() == output_layer->channels() (4 vs. 5) Number of labels is different from the output layer dimension.

train_val.prototxt和deploy.prototxt内,标签数据不一致

  

Check failed: status == CUDNN_STATUS_SUCCESS (8 vs. 0) CUDNN_STATUS_EXECUTION_FAILED

刷新一下nvidia信息之类的,再train就好了,不是很懂.jpg

Check failed: mdb_status == 0 (13 vs. 0) Permission denied

sudo 

Check failed: error == cudaSuccess (30 vs. 0) unknown error  或  Cannot create Cublas handle:

sudo

Check failed: error == cudaSuccess (73 vs. 0):

重运行几次就好了,原因未知

Check failed: error == cudaSuccess (74 vs. 0):

调大max_iter,建议保持为test_interval的倍数

Check failed: error == cudaSuccess (77 vs. 0) an illegal memory access was encountered:

nvidia-smi检查进程,sudo kill -9 [PID],然后无sudo前缀运行一次cuda程序,再加上sudo前缀运行cuda程序。

注意prototxt中需求的空间和文件夹是否存在

Segmentation fault (core dumped)  或  malloc: memory coruption :

修改源文件,排查出错误的行,换用安全的方法重写。(这一般是悲剧的开始)

corrupted size vs. prev_size:

???(悲剧达到高潮)

caffe/proto/caffe.pb.h not such file:

QT项目的.pro中 INCLUDEPATH += [caffe]/build/src 

或复制该文件到[caffe]/src/caffe/proto

Error parsing text-format caffe.SolverParameter:

看具体报错行号,参考 http://www.cnblogs.com/denny402/p/5074212.html 和 http://www.cnblogs.com/denny402/p/5074049.html 修正

猜你喜欢

转载自www.cnblogs.com/aimhabo/p/8721357.html
今日推荐