GPU模型

GPU支持
多GPU模型

硬件要求：
英伟达显卡
https://developer.nvidia.com/cuda-gpus
软件要求：
NVIDIA GPU驱动程序
https://www.nvidia.com/Download/index.aspx?lang=en-us
CUDA
对通用GPU计算操作做了封装，方便其他组件调用
https://developer.nvidia.com/cuda-zone


!pip install tensorflow-gpu


#查看gpu
import tensorflow as tf
gpu_device_name = tf.test.gpu_device_name()
print(gpu_device_name)

#查看可用的gpu数量
from __future__ import absolute_import, division, print_function, unicode_literals

import tensorflow as tf
print("Num GPUs Available: ", len(tf.config.experimental.list_physical_devices('GPU')))

from tensorflow.python.client import device_lib

# 列出所有的本地机器设备
local_device_protos = device_lib.list_local_devices()
print(local_device_protos)

#限制显存使用
import tensorflow as tf

# 设置显存使用上限
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.7) 8G*0.7=5.6G 
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) 

# 按需取用
gpu_options = tf.GPUOptions(allow_growth=True)  4G  6G 
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))


#keras
import tensorflow as tf
from keras.applications import Xception
from keras.utils import multi_gpu_model

#cpu
with tf.device('/device:CPU:0'):
  model= Xception(weights=None, input_shape=(w,h,3), classes=100)

#单个gpu
with tf.device('/device:GPU:0'):
  model= Xception(weights=None, input_shape=(w,h,3), classes=100)

#多个gpu
model1=multi_gpu_model(model , gpus=8)
model1.compile(loss='crossentropy',optimizer='adam')

#训练
model1.fit(x,y,batch_size=256)
#单个gpu的batch_size=256/8=32