tensorflow GPU版本配置加速环境

       提取码：2p9i

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\bin

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\lib\x64

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\libnvvp

echo %path%

import ctypes  

import imp  

import sys  

def main():  

    try:  

        import tensorflow as tf  

        print("TensorFlow successfully installed.")  

        if tf.test.is\_built\_with\_cuda():  

            print("The installed version of TensorFlow includes GPU support.")  

        else:  

            print("The installed version of TensorFlow does not include GPU support.")  

        sys.exit(0)  

    except ImportError:  

        print("ERROR: Failed to import the TensorFlow module.")  

  

    candidate\_explanation = False  

  

    python\_version = sys.version\_info.major, sys.version\_info.minor  

    print("\n- Python version is %d.%d." % python\_version)  

    if not (python\_version == (3, 5) or python\_version == (3, 6)):  

        candidate\_explanation = True  

        print("- The official distribution of TensorFlow for Windows requires "  

              "Python version 3.5 or 3.6.")  

  

    try:  

        \_, pathname, \_ = imp.find\_module("tensorflow")  

        print("\n- TensorFlow is installed at: %s" % pathname)  

    except ImportError:  

        candidate\_explanation = False  

        print(""" 

- No module named TensorFlow is installed in this Python environment. You may 

  install it using the command `pip install tensorflow`.""")  

  

    try:  

        msvcp140 = ctypes.WinDLL("msvcp140.dll")  

    except OSError:  

        candidate\_explanation = True  

        print(""" 

- Could not load 'msvcp140.dll'. TensorFlow requires that this DLL be 

  installed in a directory that is named in your %PATH% environment 

  variable. You may install this DLL by downloading Microsoft Visual 

  C++ 2015 Redistributable Update 3 from this URL: 

  https://www.microsoft.com/en-us/download/details.aspx?id=53587""")  

  

    try:  

        cudart64\_80 = ctypes.WinDLL("cudart64\_80.dll")  

    except OSError:  

        candidate\_explanation = True  

        print(""" 

- Could not load 'cudart64\_80.dll'. The GPU version of TensorFlow 

  requires that this DLL be installed in a directory that is named in 

  your %PATH% environment variable. Download and install CUDA 8.0 from 

  this URL: https://developer.nvidia.com/cuda-toolkit""")  

  

    try:  

        nvcuda = ctypes.WinDLL("nvcuda.dll")  

    except OSError:  

        candidate\_explanation = True  

        print(""" 

- Could not load 'nvcuda.dll'. The GPU version of TensorFlow requires that 

  this DLL be installed in a directory that is named in your %PATH% 

  environment variable. Typically it is installed in 'C:\Windows\System32'. 

  If it is not present, ensure that you have a CUDA-capable GPU with the 

  correct driver installed.""")  

  

    cudnn5\_found = False  

    try:  

        cudnn5 = ctypes.WinDLL("cudnn64\_5.dll")  

        cudnn5\_found = True  

    except OSError:  

        candidate\_explanation = True  

        print(""" 

- Could not load 'cudnn64\_5.dll'. The GPU version of TensorFlow 

  requires that this DLL be installed in a directory that is named in 

  your %PATH% environment variable. Note that installing cuDNN is a 

  separate step from installing CUDA, and it is often found in a 

  different directory from the CUDA DLLs. You may install the 

  necessary DLL by downloading cuDNN 5.1 from this URL: 

  https://developer.nvidia.com/cudnn""")  

  

    cudnn6\_found = False  

    try:  

        cudnn = ctypes.WinDLL("cudnn64\_6.dll")  

        cudnn6\_found = True  

    except OSError:  

        candidate\_explanation = True  

  

    if not cudnn5\_found or not cudnn6\_found:  

        print()  

        if not cudnn5\_found and not cudnn6\_found:  

            print("- Could not find cuDNN.")  

        elif not cudnn5\_found:  

            print("- Could not find cuDNN 5.1.")  

        else:  

            print("- Could not find cuDNN 6.")  

            print(""" 

  The GPU version of TensorFlow requires that the correct cuDNN DLL be installed 

  in a directory that is named in your %PATH% environment variable. Note that 

  installing cuDNN is a separate step from installing CUDA, and it is often 

  found in a different directory from the CUDA DLLs. The correct version of 

  cuDNN depends on your version of TensorFlow: 

 

  \* TensorFlow 1.2.1 or earlier requires cuDNN 5.1. ('cudnn64\_5.dll') 

  \* TensorFlow 1.3 or later requires cuDNN 6. ('cudnn64\_6.dll') 

 

  You may install the necessary DLL by downloading cuDNN from this URL: 

  https://developer.nvidia.com/cudnn""")  

  

    if not candidate\_explanation:  

        print(""" 

- All required DLLs appear to be present. Please open an issue on the 

  TensorFlow GitHub page: https://github.com/tensorflow/tensorflow/issues""")  

    sys.exit(-1)  

if \_\_name\_\_ == "\_\_main\_\_":  

    main()  

TensorFlow successfully installed.

The installed version of TensorFlow includes GPU support.

Could not load ‘cudart64\_80.dll’. The GPU version of TensorFlow

requires that this DLL be installed in a directory that is named in

your %PATH% environment variable. Download and install CUDA 8.0 from

this URL: https://developer.nvidia.com/cuda-toolkit

# 测试tensorflow\_gpu版本加速效果代码

from datetime import datetime

import math

import time

import tensorflow as tf

import os

#os.environ["CUDA\_DEVICE\_ORDER"] = "PCI\_BUS\_ID"

#os.environ["CUDA\_VISIBLE\_DEVICES"] = "-1"

batch\_size = 32

num\_batches = 100

# 该函数用来显示网络每一层的结构，展示tensor的尺寸



def print\_activations(t):

print(t.op.name, ' ', t.get\_shape().as\_list())



# with tf.name\_scope('conv1') as scope # 可以将scope之内的variable自动命名为conv1/xxx，便于区分不同组件



def inference(images):

parameters = []

# 第一个卷积层

with tf.name\_scope('conv1') as scope:

# 卷积核、截断正态分布

kernel = tf.Variable(tf.truncated\_normal([11, 11, 3, 64],

dtype=tf.float32, stddev=1e-1), name='weights')

conv = tf.nn.conv2d(images, kernel, [1, 4, 4, 1], padding='SAME')

# 可训练

biases = tf.Variable(tf.constant(0.0, shape=[64], dtype=tf.float32), trainable=True, name='biases')

bias = tf.nn.bias\_add(conv, biases)

conv1 = tf.nn.relu(bias, name=scope)

print\_activations(conv1)

parameters += [kernel, biases]

# 再加LRN和最大池化层，除了AlexNet，基本放弃了LRN，说是效果不明显，还会减速？

lrn1 = tf.nn.lrn(conv1, 4, bias=1.0, alpha=0.001 / 9, beta=0.75, name='lrn1')

pool1 = tf.nn.max\_pool(lrn1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool1')

print\_activations(pool1)

# 第二个卷积层，只有部分参数不同

with tf.name\_scope('conv2') as scope:

kernel = tf.Variable(tf.truncated\_normal([5, 5, 64, 192], dtype=tf.float32, stddev=1e-1), name='weights')

conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME')

biases = tf.Variable(tf.constant(0.0, shape=[192], dtype=tf.float32), trainable=True, name='biases')

bias = tf.nn.bias\_add(conv, biases)

conv2 = tf.nn.relu(bias, name=scope)

parameters += [kernel, biases]

print\_activations(conv2)

# 稍微处理一下

lrn2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.001 / 9, beta=0.75, name='lrn2')

pool2 = tf.nn.max\_pool(lrn2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool2')

print\_activations(pool2)

# 第三个

with tf.name\_scope('conv3') as scope:

kernel = tf.Variable(tf.truncated\_normal([3, 3, 192, 384], dtype=tf.float32, stddev=1e-1), name='weights')

conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')

biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32), trainable=True, name='biases')

bias = tf.nn.bias\_add(conv, biases)

conv3 = tf.nn.relu(bias, name=scope)

parameters += [kernel, biases]

print\_activations(conv3)

# 第四层

with tf.name\_scope('conv4') as scope:

kernel = tf.Variable(tf.truncated\_normal([3, 3, 384, 256], dtype=tf.float32, stddev=1e-1), name='weights')

conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME')

biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32), trainable=True, name='biases')

bias = tf.nn.bias\_add(conv, biases)

conv4 = tf.nn.relu(bias, name=scope)

parameters += [kernel, biases]

print\_activations(conv4)

# 第五个

with tf.name\_scope('conv5') as scope:

kernel = tf.Variable(tf.truncated\_normal([3, 3, 256, 256], dtype=tf.float32, stddev=1e-1), name='weights')

conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME')

biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32), trainable=True, name='biases')

bias = tf.nn.bias\_add(conv, biases)

conv5 = tf.nn.relu(bias, name=scope)

parameters += [kernel, biases]

print\_activations(conv5)

# 之后还有最大化池层

pool5 = tf.nn.max\_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool5')

print\_activations(pool5)

return pool5, parameters

# 全连接层

# 评估每轮计算时间，第一个输入是tf得Session，第二个是运算算子，第三个是测试名称

# 头几轮有显存加载，cache命中等问题，可以考虑只计算第10次以后的

def time\_tensorflow\_run(session, target, info\_string):

num\_steps\_burn\_in = 10

total\_duration = 0.0

total\_duration\_squared = 0.0

# 进行num\_batches+num\_steps\_burn\_in次迭代

# 用time.time()记录时间，热身过后，开始显示时间

for i in range(num\_batches + num\_steps\_burn\_in):

start\_time = time.time()

\_ = session.run(target)

duration = time.time() - start\_time

if i >= num\_steps\_burn\_in:

if not i % 10:

print('%s:step %d, duration = %.3f' % (datetime.now(), i - num\_steps\_burn\_in, duration))

total\_duration += duration

total\_duration\_squared += duration \* duration

# 计算每轮迭代品均耗时和标准差sd

mn = total\_duration / num\_batches

vr = total\_duration\_squared / num\_batches - mn \* mn

sd = math.sqrt(vr)

print('%s: %s across %d steps, %.3f +/- %.3f sec / batch' % (datetime.now(), info\_string, num\_batches, mn, sd))

def run\_benchmark():

# 首先定义默认的Graph

with tf.Graph().as\_default():

# 并不实用ImageNet训练，知识随机计算耗时

image\_size = 224

images = tf.Variable(tf.random\_normal([batch\_size, image\_size, image\_size, 3], dtype=tf.float32, stddev=1e-1))

pool5, parameters = inference(images)

init = tf.global\_variables\_initializer()

sess = tf.Session(config=tf.ConfigProto(allow\_soft\_placement=True, log\_device\_placement=False))

sess.run(init)

# 下面直接用pool5传入训练（没有全连接层）

# 只是做做样子，并不是真的计算

time\_tensorflow\_run(sess, pool5, "Forward")

# 瞎弄的，伪装

objective = tf.nn.l2\_loss(pool5)

grad = tf.gradients(objective, parameters)

time\_tensorflow\_run(sess, grad, "Forward-backward")

run\_benchmark()