这是一份你们需要的Windows版深度学习软件安装指南

百家 作者:机器之心 2017-10-06 14:03:41

选自Github

机器之心编译

参与:蒋思源、刘晓坤


本文从最基本的依赖项开始,依次配置了 VS 2015、Anaconda 4.4.0、CUDA 8.0.61 和 cuDNN v5.1 等基本环境,然后再从 Keras 出发安装 Theano、TensorFlow 和 CNTK 以作为其后端。在完成配置深度学习框架后,本文分别利用这三个框架作为 Keras 后端在 CPU 和 GPU 上训练了一个标准的卷积神经网络,完成该简单的卷积网络也就意味着我们完成了深度学习环境的配置。



该配置版本最后更新的日期是今年七月,该更新版本允许本地使用 3 个不同的 GPU 加速后端,并添加对 MKL BLAS 库的支持。


目前有很多帮助我们在 Linux 或 Mac OS 上构建深度学习(DL)环境的指导文章,但很少有文章完整地叙述如何高效地在 Windows 10 上配置深度学习开发环境。此外,很多开发者安装 Windows 和 Ubuntu 双系统或在 Windows 上安装虚拟机以配置深度学习环境,但对于入门者来说,我们更希望还是直接使用 Windows 直接配置深度学习环境。因此,本文作者 Phil Ferriere 在 GitHub 上发布了该教程,他希望能从最基本的环境变量配置开始一步步搭建 Keras 深度学习开发环境。


如果读者希望在 Windows 10 上配置深度学习环境,那么本文将为大家提供很多有利的信息。


依赖项


下面是我们将在 Windows 10(Version 1607 OS Build 14393.222)上配置深度学习环境所需要的工具和软件包:


  1. Visual Studio 2015 Community Edition Update 3 w. Windows Kit 10.0.10240.0:用于其 C/C++编译器(而不是 IDE)和 SDK,选择该确定的版本是因为 CUDA 8.0.61 所支持的 Windows 编译器。

  2. Anaconda (64-bit) w. Python 3.6 (Anaconda3-4.4.0) [for Tensorflow support] or Python 2.7 (Anaconda2-4.4.0) [no Tensorflow support] with MKL:Anaconda 是一个开源的 Python 发行版本,其包含了 conda、Python、NumPy、SciPy 等 180 多个科学包及其依赖项,是一个集成开发环境。MKL 可以利用 CPU 加速许多线性代数运算。

  3. CUDA 8.0.61 (64-bit):CUDA 是一种由 NVIDIA 推出的通用并行计算架构,该架构使 GPU 能够解决复杂的计算问题,该软件包能提供 GPU 数学库、显卡驱动和 CUDA 编译器等。

  4. cuDNN v5.1 (Jan 20, 2017) for CUDA 8.0:用于加速卷积神经网络的运算。

  5. Keras 2.0.5 with three different backends: Theano 0.9.0, Tensorflow-gpu 1.2.0, and CNTK 2.0:Keras 以 Theano、Tensorflow 或 CNTK 等框架为后端,并提供深度学习高级 API。使用不同的后端在张量数学计算等方面会有不同的效果。


硬件


  • Dell Precision T7900, 64GB RAM:Intel Xeon E5-2630 v4 @ 2.20 GHz (1 processor, 10 cores total, 20 logical processors)

  • NVIDIA GeForce Titan X, 12GB RAM:Driver version: 372.90 / Win 10 64


安装步骤


我们可能喜欢让所有的工具包和软件包在一个根目录下(如 e:toolkits.win),所以在下文只要看到以 e:toolkits.win 开头的路径,那么我们可能就需要小心不要覆盖或随意更改必要的软件包目录。


  • Visual Studio 2015 Community Edition Update 3 w. Windows Kit 10.0.10240.0

  • 下载地址:https://www.visualstudio.com/vs/older-downloads


运行下载的软件包以安装 Visual Studio,可能我们还需要做一些额外的配置:



  • 基于我们安装 VS 2015 的地址,需要将 C:Program Files (x86)Microsoft Visual Studio 14.0VCbin 添加到 PATH 中。

  • 定义系统环境变量(sysenv variable)INCLUDE 的值为 C:Program Files (x86)Windows Kits10Include10.0.10240.0ucrt

  • 定义系统环境变量(sysenv variable)LIB 的值为 C:Program Files (x86)Windows Kits10Lib10.0.10240.0umx64;C:Program Files (x86)Windows Kits10Lib10.0.10240.0ucrtx64


Anaconda 4.4.0 (64-bit) (Python 3.6 TF support / Python 2.7 no TF support))


本教程最初使用的是 Python 2.7,而随着 TensorFlow 可作为 Keras 的后端,我们决定使用 Python 3.6 作为默认配置。因此,根据我们配置的偏好,可以设置 e:toolkits.winanaconda3-4.4.0 或 e:toolkits.winanaconda2-4.4.0 为安装 Anaconda 的文件夹名。


  • Python 3.6 版本的 Anaconda 下载地址:https://repo.continuum.io/archive/Anaconda3-4.4.0-Windows-x86_64.exe

  • Python 2.7 版本的 Anaconda 下载地址:https://repo.continuum.io/archive/Anaconda2-4.4.0-Windows-x86_64.exe



运行安装程序完成安装:



如上,本教程选择了第二个选项,但不一定是最好的。


定义一下变量并更新 PATH:


  • 定义系统环境(sysenv variable)变量 PYTHON_HOME 的值为 e:toolkits.winanaconda3-4.4.0

  • 添加 %PYTHON_HOME%, %PYTHON_HOME%Scripts 和 %PYTHON_HOME%Librarybin 到 PATH 中


创建 dlwin36 conda 环境


在安装 Anaconda 后,打开 Windows 命令窗口并执行:


  1. #使用以下命令行创建环境

  2. $ conda create --yes -n dlwin36 numpy scipy mkl-service m2w64-toolchain libpython jupyter



  3. # 使用以下命令行激活环境:

  4. # > activate dlwin36

  5. #

  6. # 使用以下命令行关闭环境:

  7. # > deactivate dlwin36

  8. #

  9. # * for power-users using bash, you must source

  10. #


如上所示,使用 active dlwin36 命令激活这个新的环境。如果已经有了旧的 dlwin36 环境,可以先用 conda env remove -n dlwin36 命令删除。既然打算使用 GPU,为什么还要安装 CPU 优化的线性代数库如 MKL 呢?在我们的设置中,大多数深度学习都是由 GPU 承担的,这并没错,但 CPU 也不是无所事事。基于图像的 Kaggle 竞赛一个重要部分是数据增强。如此看来,数据增强是通过转换原始训练样本(利用图像处理算子)获得额外输入样本(即更多的训练图像)的过程。基本的转换比如下采样和均值归 0 的归一化也是必需的。如果你觉得这样太冒险,可以试试额外的预处理增强(噪声消除、直方图均化等等)。当然也可以用 GPU 处理并把结果保存到文件中。然而在实践过程中,这些计算通常都是在 CPU 上平行执行的,而 GPU 正忙于学习深度神经网络的权重,况且增强数据是用完即弃的。因此,我们强烈推荐安装 MKL,而 Theanos 用 BLAS 库更好。


CUDA 8.0.61 (64-bit)


从英伟达网站下载 CUDA 8.0 (64-bit):https://developer.nvidia.com/cuda-downloads


选择合适的操作系统:



下载安装包:



运行安装包,安装文件到 e:toolkits.wincuda-8.0.61 中:



完成安装后,安装包应该创建了一个名为 CUDA_PATH 的系统环境变量(sysenv variable),并且已经添加了%CUDA_PATH%bin 和 %CUDA_PATH%libnvvp 到 PATH 中。检查是否真正添加了,若 CUDA 环境变量因为一些原因出错了,那么完成下面两个步骤:


  • 定义名为 CUDA_PATH 的系统环境变量的值为 e:toolkits.wincuda-8.0.61

  • 添加%CUDA_PATH%bin 和 %CUDA_PATH%libnvvp 到 PATH 中


cuDNN v5.1 (Jan 20, 2017) for CUDA 8.0


根据英伟达官网「cuDNN 为标准的运算如前向和反向卷积、池化、归一化和激活层等提供高度调优的实现」,它是为卷积神经网络和深度学习设计的一款加速方案。


cuDNN 的下载地址:https://developer.nvidia.com/rdp/cudnn-download


我们需要选择符合 CUDA 版本和 Window 10 编译器的 cuDNN 软件包,一般来说,cuDNN 5.1 可以支持 CUDA 8.0 和 Windows 10。



下载的 ZIP 文件包含三个目录(bin、include、lib),抽取这三个的文件夹到%CUDA_PATH% 中。


安装 Keras 2.0.5 和 Theano0.9.0 与 libgpuarray


运行以下命令安装 libgpuarray 0.6.2,即 Theano 0.9.0 唯一的稳定版:


  1. (dlwin36) $ conda install pygpu==0.6.2 nose

  2. #下面是该命令行安装的效果

  3. Fetching package metadata ...........

  4. Solving package specifications: .

  5. Package plan for installation in environment e:toolkits.winanaconda3-4.4.0envsdlwin36:

  6. The following NEW packages will be INSTALLED:

  7.    libgpuarray: 0.6.2-vc14_0 [vc14]

  8.    nose:        1.3.7-py36_1

  9.    pygpu:       0.6.2-py36_0

  10. Proceed ([y]/n)? y


输入以下命令安装 Keras 和 Theano:


  1. (dlwin36) $ pip install keras==2.0.5

  2. #下面是该命令行安装的效果

  3. Collecting keras==2.0.5

  4. Requirement already satisfied: six in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from keras==2.0.5)

  5. Collecting pyyaml (from keras==2.0.5)

  6. Collecting theano (from keras==2.0.5)

  7. Requirement already satisfied: scipy>=0.14 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from theano->keras==2.0.5)

  8. Requirement already satisfied: numpy>=1.9.1 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from theano->keras==2.0.5)

  9. Installing collected packages: pyyaml, theano, keras

  10. Successfully installed keras-2.0.5 pyyaml-3.12 theano-0.9.0


安装 CNTK 2.0 后端


根据 CNTK 安装文档,我们可以使用以下 pip 命令行安装 CNTK:


  1. (dlwin36) $ pip install https://cntk.ai/PythonWheel/GPU/cntk-2.0-cp36-cp36m-win_amd64.whl

  2. #下面是该命令行安装的效果

  3. Collecting cntk==2.0 from https://cntk.ai/PythonWheel/GPU/cntk-2.0-cp36-cp36m-win_amd64.whl

  4.  Using cached https://cntk.ai/PythonWheel/GPU/cntk-2.0-cp36-cp36m-win_amd64.whl

  5. Requirement already satisfied: numpy>=1.11 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from cntk==2.0)

  6. Requirement already satisfied: scipy>=0.17 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from cntk==2.0)

  7. Installing collected packages: cntk

  8. Successfully installed cntk-2.0


该安装将导致在 conda 环境目录下额外安装 CUDA 和 cuDNN DLLs:


  1. (dlwin36) $ cd E:toolkits.winanaconda3-4.4.0envsdlwin36

  2. (dlwin36) $ dir cu*.dll

  3. #下面是该命令行安装的效果

  4. Volume in drive E is datasets

  5. Volume Serial Number is 1ED0-657B

  6. Directory of E:toolkits.winanaconda3-4.4.0envsdlwin36

  7. 06/30/2017  02:47 PM        40,744,896 cublas64_80.dll

  8. 06/30/2017  02:47 PM           366,016 cudart64_80.dll

  9. 06/30/2017  02:47 PM        78,389,760 cudnn64_5.dll

  10. 06/30/2017  02:47 PM        47,985,208 curand64_80.dll

  11. 06/30/2017  02:47 PM        41,780,280 cusparse64_80.dll

  12.               5 File(s)    209,266,160 bytes

  13.               0 Dir(s)  400,471,019,520 bytes free


这个问题并不是因为浪费硬盘空间,而是安装的 cuDNN 版本和我们安装在 c:toolkitscuda-8.0.61 下的 cuDNN 版本不同,因为在 conda 环境目录下的 DLL 将首先加载,所以我们需要这些 DLL 移除出%PATH% 目录:


  1. (dlwin36) $ md discard & move cu*.dll discard

  2. #下面是该命令行安装的效果

  3. E:toolkits.winanaconda3-4.4.0envsdlwin36cublas64_80.dll

  4. E:toolkits.winanaconda3-4.4.0envsdlwin36cudart64_80.dll

  5. E:toolkits.winanaconda3-4.4.0envsdlwin36cudnn64_5.dll

  6. E:toolkits.winanaconda3-4.4.0envsdlwin36curand64_80.dll

  7. E:toolkits.winanaconda3-4.4.0envsdlwin36cusparse64_80.dll

  8.        5 file(s) moved.


安装 TensorFlow-GPU 1.2.0 后端


运行以下命令行使用 pip 安装 TensorFlow:


  1. (dlwin36) $ pip install tensorflow-gpu==1.2.0

  2. #以下是安装效果

  3. Collecting tensorflow-gpu==1.2.0

  4.  Using cached tensorflow_gpu-1.2.0-cp36-cp36m-win_amd64.whl

  5. Requirement already satisfied: bleach==1.5.0 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from tensorflow-gpu==1.2.0)

  6. Requirement already satisfied: numpy>=1.11.0 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from tensorflow-gpu==1.2.0)

  7. Collecting markdown==2.2.0 (from tensorflow-gpu==1.2.0)

  8. Requirement already satisfied: wheel>=0.26 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from tensorflow-gpu==1.2.0)

  9. Collecting protobuf>=3.2.0 (from tensorflow-gpu==1.2.0)

  10. Collecting backports.weakref==1.0rc1 (from tensorflow-gpu==1.2.0)

  11.  Using cached backports.weakref-1.0rc1-py3-none-any.whl

  12. Collecting html5lib==0.9999999 (from tensorflow-gpu==1.2.0)

  13. Collecting werkzeug>=0.11.10 (from tensorflow-gpu==1.2.0)

  14.  Using cached Werkzeug-0.12.2-py2.py3-none-any.whl

  15. Requirement already satisfied: six>=1.10.0 in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packages (from tensorflow-gpu==1.2.0)

  16. Requirement already satisfied: setuptools in e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagessetuptools-27.2.0-py3.6.egg (from protobuf>=3.2.0->tensorflow-gpu==1.2.0)

  17. Installing collected packages: markdown, protobuf, backports.weakref, html5lib, werkzeug, tensorflow-gpu

  18.  Found existing installation: html5lib 0.999

  19.    DEPRECATION: Uninstalling a distutils installed project (html5lib) has been deprecated and will be removed in a future version. This is due to the fact that uninstalling a distutils project will only partially uninstall the project.

  20.    Uninstalling html5lib-0.999:

  21.      Successfully uninstalled html5lib-0.999

  22. Successfully installed backports.weakref-1.0rc1 html5lib-0.9999999 markdown-2.2.0 protobuf-3.3.0 tensorflow-gpu-1.2.0 werkzeug-0.12.2


使用 conda 检查安装的软件包


完成以上安装和配置后,我们应该在 dlwin36 conda 环境中看到以下软件包列表:



为了快速检查上述三个后端安装的效果,依次运行一下命令行分别检查 Theano、TensorFlow 和 CNTK 导入情况:


  1. (dlwin36) $ python -c "import theano; print('theano: %s, %s' % (theano.__version__, theano.__file__))"

  2. theano: 0.9.0, E:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagestheano__init__.py


  3. (dlwin36) $ python -c "import pygpu; print('pygpu: %s, %s' % (pygpu.__version__, pygpu.__file__))"

  4. pygpu: 0.6.2, e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagespygpu__init__.py


  5. (dlwin36) $ python -c "import tensorflow; print('tensorflow: %s, %s' % (tensorflow.__version__, tensorflow.__file__))"

  6. tensorflow: 1.2.0, E:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagestensorflow__init__.py


  7. (dlwin36) $ python -c "import cntk; print('cntk: %s, %s' % (cntk.__version__, cntk.__file__))"

  8. cntk: 2.0, E:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagescntk__init__.py



验证 Theano 的安装


因为 Theano 是安装 Keras 时自动安装的,为了快速地在 CPU 模式、GPU 模式和带 cuDNN 的 GPU 模式之间转换,我们需要创建以下三个系统环境变量(sysenv variable):


  • 系统环境变量 THEANO_FLAGS_CPU 的值定义为:floatX=float32,device=cpu

  • 系统环境变量 THEANO_FLAGS_GPU 的值定义为:floatX=float32,device=cuda0,dnn.enabled=False,gpuarray.preallocate=0.8

  • 系统环境变量 THEANO_FLAGS_GPU_DNN 的值定义为:floatX=float32,device=cuda0,optimizer_including=cudnn,gpuarray.preallocate=0.8,dnn.conv.algo_bwd_filter=deterministic,dnn.conv.algo_bwd_data=deterministic,dnn.include_path=e:/toolkits.win/cuda-8.0.61/include,dnn.library_path=e:/toolkits.win/cuda-8.0.61/lib/x64


现在,我们能直接使用 THEANO_FLAGS_CPU、THEANO_FLAGS_GPU 或 THEANO_FLAGS_GPU_DNN 直接设置 Theano 使用 CPU、GPU 还是 GPU+cuDNN。我们可以使用以下命令行验证这些变量是否成功加入环境中:


  1. (dlwin36) $ set KERAS_BACKEND=theano

  2. (dlwin36) $ set | findstr /i theano


  3. KERAS_BACKEND=theano

  4. THEANO_FLAGS=floatX=float32,device=cuda0,optimizer_including=cudnn,gpuarray.preallocate=0.8,dnn.conv.algo_bwd_filter=deterministic,dnn.conv.algo_bwd_data=deterministic,dnn.include_path=e:/toolkits.win/cuda-8.0.61/include,dnn.library_path=e:/toolkits.win/cuda-8.0.61/lib/x64

  5. THEANO_FLAGS_CPU=floatX=float32,device=cpu

  6. THEANO_FLAGS_GPU=floatX=float32,device=cuda0,dnn.enabled=False,gpuarray.preallocate=0.8

  7. THEANO_FLAGS_GPU_DNN=floatX=float32,device=cuda0,optimizer_including=cudnn,gpuarray.preallocate=0.8,dnn.conv.algo_bwd_filter=deterministic,dnn.conv.algo_bwd_data=deterministic,dnn.include_path=e:/toolkits.win/cuda-8.0.61/include,dnn.library_path=e:/toolkits.win/cuda-8.0.61/lib/x64



更多具体的 Theano 验证代码与命令请查看原文。


检查系统环境变量


现在,不论 dlwin36 conda 环境什么时候激活,PATH 环境变量应该需要看起来如下面列表一样:



使用 Keras 验证 GPU+cuDNN 的安装


我们可以使用 Keras 在 MNIST 数据集上训练简单的卷积神经网络(convnet)而验证 GPU 的 cuDNN 是否正确安装,该文件名为 mnist_cnn.py,其可以在 Keras 案例中找到。该卷积神经网络的代码如下:


Keras案例地址:https://github.com/fchollet/keras/blob/2.0.5/examples/mnist_cnn.py


  1. (dlwin36) $ set KERAS_BACKEND=cntk

  2. (dlwin36) $ python mnist_cnn.py

  3. Using CNTK backend

  4. Selected GPU[0] GeForce GTX TITAN X as the process wide default device.

  5. x_train shape: (60000, 28, 28, 1)

  6. 60000 train samples

  7. 10000 test samples

  8. Train on 60000 samples, validate on 10000 samples

  9. Epoch 1/12

  10. e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagescntkcore.py:351: UserWarning: your data is of type "float64", but your input variable (uid "Input113") expects "". Please convert your data beforehand to speed up training.

  11.  (sample.dtype, var.uid, str(var.dtype)))

  12. 60000/60000 [==============================] - 8s - loss: 0.3275 - acc: 0.8991 - val_loss: 0.0754 - val_acc: 0.9749

  13. Epoch 2/12

  14. 60000/60000 [==============================] - 7s - loss: 0.1114 - acc: 0.9662 - val_loss: 0.0513 - val_acc: 0.9841

  15. Epoch 3/12

  16. 60000/60000 [==============================] - 7s - loss: 0.0862 - acc: 0.9750 - val_loss: 0.0429 - val_acc: 0.9859

  17. Epoch 4/12

  18. 60000/60000 [==============================] - 7s - loss: 0.0721 - acc: 0.9784 - val_loss: 0.0373 - val_acc: 0.9868

  19. Epoch 5/12

  20. 60000/60000 [==============================] - 7s - loss: 0.0649 - acc: 0.9803 - val_loss: 0.0339 - val_acc: 0.9878

  21. Epoch 6/12

  22. 60000/60000 [==============================] - 8s - loss: 0.0580 - acc: 0.9831 - val_loss: 0.0337 - val_acc: 0.9890

  23. Epoch 7/12

  24. 60000/60000 [==============================] - 8s - loss: 0.0529 - acc: 0.9846 - val_loss: 0.0326 - val_acc: 0.9895

  25. Epoch 8/12

  26. 60000/60000 [==============================] - 8s - loss: 0.0483 - acc: 0.9858 - val_loss: 0.0307 - val_acc: 0.9897

  27. Epoch 9/12

  28. 60000/60000 [==============================] - 8s - loss: 0.0456 - acc: 0.9864 - val_loss: 0.0299 - val_acc: 0.9898

  29. Epoch 10/12

  30. 60000/60000 [==============================] - 8s - loss: 0.0407 - acc: 0.9875 - val_loss: 0.0274 - val_acc: 0.9906

  31. Epoch 11/12

  32. 60000/60000 [==============================] - 8s - loss: 0.0405 - acc: 0.9883 - val_loss: 0.0276 - val_acc: 0.9904

  33. Epoch 12/12

  34. 60000/60000 [==============================] - 8s - loss: 0.0372 - acc: 0.9889 - val_loss: 0.0274 - val_acc: 0.9906

  35. Test loss: 0.0274011099327

  36. Test accuracy: 0.9906


1. 使用带 Theano 后端的 Keras


为了有一个能进行对比的基线模型,首先我们使用 Theano 后端和 CPU 训练简单的卷积神经网络:


  1. (dlwin36) $ set KERAS_BACKEND=theano

  2. (dlwin36) $ set THEANO_FLAGS=%THEANO_FLAGS_CPU%

  3. (dlwin36) $ python mnist_cnn.py


  4. #以下为训练过程和结果

  5. Using Theano backend.

  6. x_train shape: (60000, 28, 28, 1)

  7. 60000 train samples

  8. 10000 test samples

  9. Train on 60000 samples, validate on 10000 samples

  10. Epoch 1/12

  11. 60000/60000 [==============================] - 233s - loss: 0.3344 - acc: 0.8972 - val_loss: 0.0743 - val_acc: 0.9777

  12. Epoch 2/12

  13. 60000/60000 [==============================] - 234s - loss: 0.1106 - acc: 0.9674 - val_loss: 0.0504 - val_acc: 0.9837

  14. Epoch 3/12

  15. 60000/60000 [==============================] - 237s - loss: 0.0865 - acc: 0.9741 - val_loss: 0.0402 - val_acc: 0.9865

  16. Epoch 4/12

  17. 60000/60000 [==============================] - 238s - loss: 0.0692 - acc: 0.9792 - val_loss: 0.0362 - val_acc: 0.9874

  18. Epoch 5/12

  19. 60000/60000 [==============================] - 241s - loss: 0.0614 - acc: 0.9821 - val_loss: 0.0370 - val_acc: 0.9879

  20. Epoch 6/12

  21. 60000/60000 [==============================] - 245s - loss: 0.0547 - acc: 0.9839 - val_loss: 0.0319 - val_acc: 0.9885

  22. Epoch 7/12

  23. 60000/60000 [==============================] - 248s - loss: 0.0517 - acc: 0.9840 - val_loss: 0.0293 - val_acc: 0.9900

  24. Epoch 8/12

  25. 60000/60000 [==============================] - 256s - loss: 0.0465 - acc: 0.9863 - val_loss: 0.0294 - val_acc: 0.9905

  26. Epoch 9/12

  27. 60000/60000 [==============================] - 264s - loss: 0.0422 - acc: 0.9870 - val_loss: 0.0276 - val_acc: 0.9902

  28. Epoch 10/12

  29. 60000/60000 [==============================] - 263s - loss: 0.0423 - acc: 0.9875 - val_loss: 0.0287 - val_acc: 0.9902

  30. Epoch 11/12

  31. 60000/60000 [==============================] - 262s - loss: 0.0389 - acc: 0.9884 - val_loss: 0.0291 - val_acc: 0.9898

  32. Epoch 12/12

  33. 60000/60000 [==============================] - 270s - loss: 0.0377 - acc: 0.9885 - val_loss: 0.0272 - val_acc: 0.9910

  34. Test loss: 0.0271551907005

  35. Test accuracy: 0.991



我们现在使用以下命令行利用带 Theano 的后端的 Keras 在 GPU 和 cuDNN 环境下训练卷积神经网络:


  1. (dlwin36) $ set THEANO_FLAGS=%THEANO_FLAGS_GPU_DNN%

  2. (dlwin36) $ python mnist_cnn.py

  3. Using Theano backend.

  4. Using cuDNN version 5110 on context None

  5. Preallocating 9830/12288 Mb (0.800000) on cuda0

  6. Mapped name None to device cuda0: GeForce GTX TITAN X (0000:03:00.0)

  7. x_train shape: (60000, 28, 28, 1)

  8. 60000 train samples

  9. 10000 test samples

  10. Train on 60000 samples, validate on 10000 samples

  11. Epoch 1/12

  12. 60000/60000 [==============================] - 17s - loss: 0.3219 - acc: 0.9003 - val_loss: 0.0774 - val_acc: 0.9743

  13. Epoch 2/12

  14. 60000/60000 [==============================] - 16s - loss: 0.1108 - acc: 0.9674 - val_loss: 0.0536 - val_acc: 0.9822

  15. Epoch 3/12

  16. 60000/60000 [==============================] - 16s - loss: 0.0832 - acc: 0.9766 - val_loss: 0.0434 - val_acc: 0.9862

  17. Epoch 4/12

  18. 60000/60000 [==============================] - 16s - loss: 0.0694 - acc: 0.9795 - val_loss: 0.0382 - val_acc: 0.9876

  19. Epoch 5/12

  20. 60000/60000 [==============================] - 16s - loss: 0.0605 - acc: 0.9819 - val_loss: 0.0353 - val_acc: 0.9884

  21. Epoch 6/12

  22. 60000/60000 [==============================] - 16s - loss: 0.0533 - acc: 0.9836 - val_loss: 0.0360 - val_acc: 0.9883

  23. Epoch 7/12

  24. 60000/60000 [==============================] - 16s - loss: 0.0482 - acc: 0.9859 - val_loss: 0.0305 - val_acc: 0.9897

  25. Epoch 8/12

  26. 60000/60000 [==============================] - 16s - loss: 0.0452 - acc: 0.9865 - val_loss: 0.0295 - val_acc: 0.9911

  27. Epoch 9/12

  28. 60000/60000 [==============================] - 16s - loss: 0.0414 - acc: 0.9878 - val_loss: 0.0315 - val_acc: 0.9898

  29. Epoch 10/12

  30. 60000/60000 [==============================] - 16s - loss: 0.0386 - acc: 0.9886 - val_loss: 0.0282 - val_acc: 0.9911

  31. Epoch 11/12

  32. 60000/60000 [==============================] - 16s - loss: 0.0378 - acc: 0.9887 - val_loss: 0.0306 - val_acc: 0.9904

  33. Epoch 12/12

  34. 60000/60000 [==============================] - 16s - loss: 0.0354 - acc: 0.9893 - val_loss: 0.0296 - val_acc: 0.9898

  35. Test loss: 0.0296215178292

  36. Test accuracy: 0.9898



我们看到每一个 Epoch 的训练时间只需要 16 秒,相对于使用 CPU 要 250 秒左右取得了很大的提高(在同一个批量大小的情况下)。


2. 使用 TensorFlow 后端的 Keras


为了激活和测试 TensorFlow 后端,我们需要使用以下命令行:


  1. (dlwin36) $ set KERAS_BACKEND=tensorflow

  2. (dlwin36) $ python mnist_cnn.py

  3. Using TensorFlow backend.

  4. x_train shape: (60000, 28, 28, 1)

  5. 60000 train samples

  6. 10000 test samples

  7. Train on 60000 samples, validate on 10000 samples

  8. Epoch 1/12

  9. 2017-06-30 12:49:22.005585: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.

  10. 2017-06-30 12:49:22.005767: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.

  11. 2017-06-30 12:49:22.005996: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.

  12. 2017-06-30 12:49:22.006181: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.

  13. 2017-06-30 12:49:22.006361: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.

  14. 2017-06-30 12:49:22.006539: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.

  15. 2017-06-30 12:49:22.006717: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.

  16. 2017-06-30 12:49:22.006897: W c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.

  17. 2017-06-30 12:49:22.453483: I c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcorecommon_runtimegpugpu_device.cc:940] Found device 0 with properties:

  18. name: GeForce GTX TITAN X

  19. major: 5 minor: 2 memoryClockRate (GHz) 1.076

  20. pciBusID 0000:03:00.0

  21. Total memory: 12.00GiB

  22. Free memory: 10.06GiB

  23. 2017-06-30 12:49:22.454375: I c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcorecommon_runtimegpugpu_device.cc:961] DMA: 0

  24. 2017-06-30 12:49:22.454489: I c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcorecommon_runtimegpugpu_device.cc:971] 0:   Y

  25. 2017-06-30 12:49:22.454624: I c:tf_jenkinshomeworkspacerelease-winmwindows-gpupy36tensorflowcorecommon_runtimegpugpu_device.cc:1030] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce G

  26. TX TITAN X, pci bus id: 0000:03:00.0)

  27. 60000/60000 [==============================] - 8s - loss: 0.3355 - acc: 0.8979 - val_loss: 0.0749 - val_acc: 0.9760

  28. Epoch 2/12

  29. 60000/60000 [==============================] - 5s - loss: 0.1134 - acc: 0.9667 - val_loss: 0.0521 - val_acc: 0.9825

  30. Epoch 3/12

  31. 60000/60000 [==============================] - 5s - loss: 0.0863 - acc: 0.9745 - val_loss: 0.0436 - val_acc: 0.9854

  32. Epoch 4/12

  33. 60000/60000 [==============================] - 5s - loss: 0.0722 - acc: 0.9787 - val_loss: 0.0381 - val_acc: 0.9872

  34. Epoch 5/12

  35. 60000/60000 [==============================] - 5s - loss: 0.0636 - acc: 0.9811 - val_loss: 0.0339 - val_acc: 0.9880

  36. Epoch 6/12

  37. 60000/60000 [==============================] - 5s - loss: 0.0552 - acc: 0.9838 - val_loss: 0.0328 - val_acc: 0.9888

  38. Epoch 7/12

  39. 60000/60000 [==============================] - 5s - loss: 0.0515 - acc: 0.9851 - val_loss: 0.0318 - val_acc: 0.9893

  40. Epoch 8/12

  41. 60000/60000 [==============================] - 5s - loss: 0.0479 - acc: 0.9862 - val_loss: 0.0311 - val_acc: 0.9891

  42. Epoch 9/12

  43. 60000/60000 [==============================] - 5s - loss: 0.0441 - acc: 0.9870 - val_loss: 0.0310 - val_acc: 0.9898

  44. Epoch 10/12

  45. 60000/60000 [==============================] - 5s - loss: 0.0407 - acc: 0.9871 - val_loss: 0.0302 - val_acc: 0.9903

  46. Epoch 11/12

  47. 60000/60000 [==============================] - 5s - loss: 0.0405 - acc: 0.9877 - val_loss: 0.0309 - val_acc: 0.9892

  48. Epoch 12/12

  49. 60000/60000 [==============================] - 5s - loss: 0.0373 - acc: 0.9886 - val_loss: 0.0309 - val_acc: 0.9898

  50. Test loss: 0.0308696583555

  51. Test accuracy: 0.9898



我们看到使用 TensorFlow 后端要比 Theano 后端在该任务上快 3 倍左右,它们都是用了 GPU 和 cuDNN 加速。这可能是因为在该测试中它们有相同的通道等级(channel ordering),但实际上两个平台在这一点是不一样的。因此,程序可能强制 Theano 后端重新排序数据而造成性能上的差异。但在该案例下,TensorFlow 在 GPU 上的负载一直没有超过 70%。



3. 使用 CNTK 后端的 Keras


为了激活和测试 CNTK 后算,我们需要使用以下命令行:


  1. (dlwin36) $ set KERAS_BACKEND=cntk

  2. (dlwin36) $ python mnist_cnn.py

  3. Using CNTK backend

  4. Selected GPU[0] GeForce GTX TITAN X as the process wide default device.

  5. x_train shape: (60000, 28, 28, 1)

  6. 60000 train samples

  7. 10000 test samples

  8. Train on 60000 samples, validate on 10000 samples

  9. Epoch 1/12

  10. e:toolkits.winanaconda3-4.4.0envsdlwin36libsite-packagescntkcore.py:351: UserWarning: your data is of type "float64", but your input variable (uid "Input113") expects "". Please convert your data beforehand to speed up training.

  11.  (sample.dtype, var.uid, str(var.dtype)))

  12. 60000/60000 [==============================] - 8s - loss: 0.3275 - acc: 0.8991 - val_loss: 0.0754 - val_acc: 0.9749

  13. Epoch 2/12

  14. 60000/60000 [==============================] - 7s - loss: 0.1114 - acc: 0.9662 - val_loss: 0.0513 - val_acc: 0.9841

  15. Epoch 3/12

  16. 60000/60000 [==============================] - 7s - loss: 0.0862 - acc: 0.9750 - val_loss: 0.0429 - val_acc: 0.9859

  17. Epoch 4/12

  18. 60000/60000 [==============================] - 7s - loss: 0.0721 - acc: 0.9784 - val_loss: 0.0373 - val_acc: 0.9868

  19. Epoch 5/12

  20. 60000/60000 [==============================] - 7s - loss: 0.0649 - acc: 0.9803 - val_loss: 0.0339 - val_acc: 0.9878

  21. Epoch 6/12

  22. 60000/60000 [==============================] - 8s - loss: 0.0580 - acc: 0.9831 - val_loss: 0.0337 - val_acc: 0.9890

  23. Epoch 7/12

  24. 60000/60000 [==============================] - 8s - loss: 0.0529 - acc: 0.9846 - val_loss: 0.0326 - val_acc: 0.9895

  25. Epoch 8/12

  26. 60000/60000 [==============================] - 8s - loss: 0.0483 - acc: 0.9858 - val_loss: 0.0307 - val_acc: 0.9897

  27. Epoch 9/12

  28. 60000/60000 [==============================] - 8s - loss: 0.0456 - acc: 0.9864 - val_loss: 0.0299 - val_acc: 0.9898

  29. Epoch 10/12

  30. 60000/60000 [==============================] - 8s - loss: 0.0407 - acc: 0.9875 - val_loss: 0.0274 - val_acc: 0.9906

  31. Epoch 11/12

  32. 60000/60000 [==============================] - 8s - loss: 0.0405 - acc: 0.9883 - val_loss: 0.0276 - val_acc: 0.9904

  33. Epoch 12/12

  34. 60000/60000 [==============================] - 8s - loss: 0.0372 - acc: 0.9889 - val_loss: 0.0274 - val_acc: 0.9906

  35. Test loss: 0.0274011099327

  36. Test accuracy: 0.9906



在具体的试验中,CNTK 同样也十分快速,并且 GPU 负载达到了 80%。


原文链接:https://github.com/philferriere/dlwin



本文为机器之心编译,转载请联系本公众号获得授权

✄------------------------------------------------

加入机器之心(全职记者/实习生):hr@jiqizhixin.com

投稿或寻求报道:content@jiqizhixin.com

广告&商务合作:bd@jiqizhixin.com

关注公众号:拾黑(shiheibook)了解更多

[广告]赞助链接:

四季很好,只要有你,文娱排行榜:https://www.yaopaiming.com/
让资讯触达的更精准有趣:https://www.0xu.cn/

公众号 关注网络尖刀微信公众号
随时掌握互联网精彩
赞助链接