1.首先下载nccl:
cd /usr/local
https://github.com/NVIDIA/nccl/archive/master.zip2.解压:
sudo unzip master.zip
cd nccl-master/
3.安装:
sudo make CUDA_HOME=/usr/local/cuda-8.0/ test (按照自己的cuda地址)
make install (加sudo的话报错了,不加的话链接正常)
重新make caffe吧,然后就可以开启并行训练了!!!!
只需要在Makefile.comfig文件中把use_nccl=1的注释去掉,然后在/caffe/目录下
sudo make clean
make all -j20
make pycaffe -j20
make test -j20
- 1
- 2
- 3
sudo make -j20 runtest # take long time
sudo make py
sudo make distribute