Use Tensorflow or Keras When the GPU Memory limit

The original blog —— Use Tensorflow or Keras When the GPU Memory limit

run Keras perhaps Tensorflow By default, it takes up all GPU Memory , At this time, if you want to start another process , Or others can't squeeze in when they want to start a process , therefore must Limit GPU Memory

The best information is Official documents

• visible_device_list Specify which graphics card to use
• per_process_gpu_memory_fraction The proportion of allocated memory in the total memory
• allow_growth Allocate according to the needs of the runtime GPU Memory , Allocate very little memory at first , With Session Start running and need more GPU Memory will expand automatically , But then Don't release Memory
• per_process_gpu_memory_fraction And allow_growth Choose one of the two methods . It's usually used allow_growth that will do , If you can accurately estimate how much video memory your program needs , recommend per_process_gpu_memory_fraction

#  Add the following code to .py Just start the file 
import tensorflow as tf
config = tf.ConfigProto()
config.gpu_options.visible_device_list = '0'
# config.gpu_options.per_process_gpu_memory_fraction = 0.2
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# from keras import backend as K
# K.set_session(sess)

In terms of personal habits , Each project has a configuration file config.py, Write the basic configuration function inside , Such as GPU To configure , Path configuration, etc .
set_gpu() function , Automatically identify whether there is a graphics card , If there is a graphics card, specify more free memory or GPU Use a low graphics card , And set the distribution proportion

The code is for two graphics cards , It's too lazy to change. , Many graphics cards are similar ~

def set_gpu(ratio=0, target='memory'):
    """
     To configure GPU,0 It means adaptive ,(0, 1] Indicates the proportion of video memory 
    :param ratio:
    :param target:  Choose a card with large video memory or GPU Cards with low utilization 
    :return:
    """
    command1 = "nvidia-smi -q -d Memory | grep -A4 GPU | grep Free | awk '{print $3}'"
    command2 = "nvidia-smi -q | grep Gpu | awk '{print $3}'"
    memory = list(map(int, os.popen(command1).readlines()))
    gpu = list(map(int, os.popen(command2).readlines()))
    if memory and gpu:  #  If you don't have a graphics card ,memory,gpu Are all []
        if target == 'memory':
            num = (1, 0)[memory[0] > memory[1]]
        else:
            num = (0, 1)[gpu[0] > gpu[1]]
        print('>>> Free Memory       : GPU0 %6d MiB | GPU1 %6d MiB' % (memory[0], memory[1]))
        print('>>> Volatile GPU-Util : GPU0 %6d %%   | GPU1 %6d %% ' % (gpu[0], gpu[1]))
        print('>>> Using GPU%d' % num)
        import tensorflow as tf
        config = tf.ConfigProto()
        config.gpu_options.visible_device_list = str(num)  #  choice GPU
        if ratio == 0:
            config.gpu_options.allow_growth = True
        else:
            config.gpu_options.per_process_gpu_memory_fraction = ratio
        sess = tf.Session(config=config)
        from keras import backend as K
        K.set_session(sess)
    else:
        print('>>> Could not find the GPU')

There are two other ways to specify a graphics card

• Add environment variables

export CUDA_VISIBLE_DEVICES=0

• Use os

import os
os.environ["CUDA_VISIBLE_DEVICES"] = '0'

see GPU usage

• watch -n 1 nvidia-smi Refresh per second
• nvidia-smi