一，直接指定GPU:
tf.ConfigProto一般用在创建session的时候。用来对session进行参数配置
with tf.Session(config = tf.ConfigProto(...),...)#tf.ConfigProto()的参数
log_device_placement=True : 是否打印设备分配日志
allow_soft_placement=True ： 如果你指定的设备不存在，允许TF自动分配设备#coding:utf-8
import tensorflow as tf
a=tf.constant([1.0,2.0,3.0],shape=[3],name='a')
b=tf.constant([1.0,2.0,3.0],shape=[3],name='b')
c=a+b
#log_device_placement=True 输出运行每一个运算的设备
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:print(sess.run(c))可看出每个运算的设备是GPU:0
可以把运算放在不同的device上
#coding:utf-8
import tensorflow as tf
with tf.device('/cpu:0'):a=tf.constant([1.0,2.0,3.0],shape=[3],name='a')b=tf.constant([1.0,2.0,3.0],shape=[3],name='b')
with tf.device('/gpu:0'):c=a+b
#log_device_placement=True 输出运行每一个运算的设备
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:print(sess.run(c))
直接指定的缺点是，降低程序的可移植性，同时有些运算不能在GPU上执行，bi比如tf.Variable操作只支持实数型（float16,float32,和double的参数），对于整数不支持，故可用allow_soft_placement参数
报错：
a_cpu=tf.Variable(0,name='a_cpu')
with tf.device('/gpu:0'):a_gpu=tf.Variable(0,name='a_gpu')
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:print(sess.run(tf.global_variables_initializer()))
a_cpu=tf.Variable(0,name='a_cpu')
with tf.device('/gpu:0'):a_gpu=tf.Variable(0,name='a_gpu')
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,log_device_placement=True)) as sess:print(sess.run(tf.global_variables_initializer()))打印信息：
2018-08-30 03:40:26.905925: I tensorflow/core/common_runtime/placer.cc:874] a_gpu: (VariableV2)/job:localhost/replica:0/task:0/device:CPU:0
a_gpu/read: (Identity): /job:localhost/replica:0/task:0/device:CPU:0
2018-08-30 03:40:26.905952: I tensorflow/core/common_runtime/placer.cc:874] a_gpu/read: (Identity)/job:localhost/replica:0/task:0/device:CPU:0
a_gpu/Assign: (Assign): /job:localhost/replica:0/task:0/device:CPU:0
import tensorflow as tf
a=tf.constant([1.0,2.0,3.0],shape=[3],name='a')
b=tf.constant([1.0,2.0,3.0],shape=[3],name='b')c=a+b
#log_device_placement=True 输出运行每一个运算的设备
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:print(sess.run(c))
二，命令行用gpu
打印信息
2018-08-30 03:46:25.130842: I tensorflow/core/common_runtime/placer.cc:874] add: (Add)/job:localhost/replica:0/task:0/device:GPU:0
b: (Const): /job:localhost/replica:0/task:0/device:GPU:0
2018-08-30 03:46:25.130872: I tensorflow/core/common_runtime/placer.cc:874] b: (Const)/job:localhost/replica:0/task:0/device:GPU:0
a: (Const): /job:localhost/replica:0/task:0/device:GPU:0
2018-08-30 03:46:25.130892: I tensorflow/core/common_runtime/placer.cc:874] a: (Const)/job:localhost/replica:0/task:0/device:GPU:0
运算在第二块gpu
或者直接在程序里
import tensorflow as tf
import os
os.environ['CUDA_VISIBLE_DEVICES']='1'
a=tf.constant([1.0,2.0,3.0],shape=[3],name='a')
b=tf.constant([1.0,2.0,3.0],shape=[3],name='b')c=a+b
#log_device_placement=True 输出运行每一个运算的设备
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:print(sess.run(c))三，动态分配GPU
上面的方式会占用整个GPU，可以在一块GPU上同时运行多个任务
import tensorflow as tf
import os
def set_config():# 控制使用率os.environ['CUDA_VISIBLE_DEVICES'] = '1'# 假如有16GB的显存并使用其中的8GB:gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)config = tf.ConfigProto(gpu_options=gpu_options)# session = tf.Session(config=config)return configif __name__ == '__main__':a = tf.constant([1.0, 2.0, 3.0], shape=[3], name='a')b = tf.constant([1.0, 2.0, 3.0], shape=[3], name='b')c = a + bcfg=set_config()with tf.Session(config=cfg) as sess:print(sess.run(c))
