将keras后端从Tensorflow CPU更改为GPU

Question

Is there a difference (in code) between keras tensorflow-cpu backend and tensorflow-gpu backend? If I want to change tensorflow from cpu to gpu, what code do I need to add or what environmental variables do I need to set?

From keras link I know that I can use tf.devices - something like the code below. But what if I want the whole code, not just some part to run on GPU?

with tf.device('/gpu:0'):
    x = tf.placeholder(tf.float32, shape=(None, 20, 64))
    y = LSTM(32)(x)  # all ops in the LSTM layer will live on GPU:0

with tf.device('/cpu:0'):
    x = tf.placeholder(tf.float32, shape=(None, 20, 64))
    y = LSTM(32)(x)  # all ops in the LSTM layer will live on CPU:0

Answer 1

Just uninstall tensorflow-cpu ( pip uninstall tensorflow ) and install tensorflow-gpu ( pip install tensorflow-gpu ). Now tensorflow will always use your gpu(s).

If you only want to use cpu in tensorflow-gpu set the environmental variable CUDA_VISIBLE_DEVICES so that the gpus are invisible. Before loading tensorflow do this in your script:

import os
os.environ["CUDA_VISIBLE_DEVICES"]=""
import tensorflow