###以下均基于Anaconda3 for Python 3.6.5,安装文件为Anaconda3-5.2.0-Linux-x86_64.sh
####安装tensorflow-gpu,依赖包及其顺序
termcolor
grpcio
protobuf
gast
astor
absl-py
markdown
tensorboard
tensorflow_gpu
##tensorflow-gpu文件为tensorflow_gpu-1.10.0-cp36-cp36m-manylinux1_x86_64.whl
其他依赖包版本如下:
Collecting termcolor>=1.1.0 (from tensorflow-gpu==1.10.0)
Collecting grpcio>=1.8.6 (from tensorflow-gpu==1.10.0)
Collecting protobuf>=3.6.0 (from tensorflow-gpu==1.10.0)
Collecting gast>=0.2.0 (from tensorflow-gpu==1.10.0)
Collecting astor>=0.6.0 (from tensorflow-gpu==1.10.0)
Collecting absl-py>=0.1.6 (from tensorflow-gpu==1.10.0)
Collecting markdown>=2.6.8 (from tensorboard==1.10.0)
Collecting tensorboard<1.11.0,>=1.10.0 (from tensorflow-gpu==1.10.0)
有些未明确指定是用到tensorflow的,如果也想使用GPU资源,加入如下代码即可
os.environ["CUDA_VISIBLE_DEVICES"] = "0" ###单卡
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1" ###多卡
先运行export CUDA_VISIBLE_DEVICES=0,再运行py文件
或者直接将CUDA_VISIBLE_DEVICES=0与py一起运行,如CUDA_VISIBLE_DEVICES=0 python train.py ##单卡操作
多卡操作:CUDA_VISIBLE_DEVICES=0,1 python train.py
参考资料:https://www.cnblogs.com/darkknightzh/p/6591923.html
##tensorflow单GPU多GPU使用
参考资料:https://www.cnblogs.com/greentomlee/archive/2018/07/27/9379139.html
在TensorFlow中,不是所有的操作都可以被放在GPU上,如果强行将无法放在GPU上的操作指定到GPU上,那么程序将会报错。