Reference article:
Install Keras on native Windows
Keras Installation and Configuration Guide (Windows)
Deep learning environment construction: win10+GTX1060 + tensorflow1.5+keras+cuda9.0+cudnn7
Install the latest Tensorflow1.6 (CUDA9.0+cuDNN7.0) under win10
One, anaconda3 installation
Use anaconda for library installation. Download address: https://repo.continuum.io/archive/index.html
The version I have installed is:
Second, visual studio2017 installation
Three, cuda installation
illustrate:
(1) cuda9.1 was installed at the beginning, and then import tensorflow reported an error.
(2) It was found that the official tensorflow supports up to cuda9.0. If you want to maintain cuda9.1, you can only compile the available tensorflow version yourself, which uses the online resource - the version compiled by Daniel.
(3) Since I am a tensorflow novice, I am very worried about using my own compiled version when installing tensorflow for the first time.
(4) Next, I searched for the version provided by Daniel on the Internet, but found that the requirement is that the graphics card computing power should be above 3, and my current computer Thinkpad T420 graphics card is nvdia 4200M, and the computing power is 2.1. So this way doesn't work either.
(5) Only cuda can be downgraded. Ready to install combination: cuda9.0+visual studio2017+cudnn7.0+tensorflow
1. Download and install
If using gpu acceleration: install cudn; otherwise skip step 2.
https://developer.nvidia.com/cuda-downloads
The following downloads and installations are: base installer and patch3.
If you run the above 3 .exes directly and an error is reported when decompressing, decompress the .exe first, and then run the setup.exe file.
2. Environment variable configuration
Configure the environment variables of cuda below. After installing cuda, two system environment variables will be automatically generated
CUDA_PATH C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1
CUDA_PATH_V6_5 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1
Add the following environment variables yourself
CUDA_BIN_PATH% CUDA_PATH% \ binCUDA_LIB_PATH %CUDA_PATH%\lib\x64
CUDA_SDK_PATH C:\ProgramData\NVIDIA Corporation\CUDA Samples\v6.5
CUDA_SDK_BIN %CUDA_SDK_PATH%\bin\win64
CUDA_SDK_LIB %CUDA_SDK_PATH%\common\lib\x64
After configuring the environment variables, restart the computer.
3. How to check the cuda version number supported by the computer
Control Panel---Search:nvidia---select NVIDIA Control Panel---Help---System Information---Components
4. Verify the installation of cudn: cmd-->nvcc -V
Four, cudnn installation
Download address: https://developer.nvidia.com/rdp/cudnn-download
The cudnn for cuda9.0 is:
To download from the official website, you need to register an Nvidia developer account. My account password is: [email protected], and the password is Xq881116. Download and decompress the folder named cuda, which contains bin, include, lib. Copy the three folders to the place where CUDA is installed to cover the corresponding folder. The default folder is:C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\9.0
Five, pip switch source address
Switch to a domestic mirror to improve the speed and success rate of installation. Create a pip directory in the user directory, such as :\Users\administrator\pip , create pip.ini , and write the following content:
[global]
index-url = http://mirrors.aliyun.com/pypi/simple/
[install]
trusted-host=mirrors.aliyun.com
The above is Alibaba Cloud, and other domestic mirrors include:
Tsinghua University: https://pypi.tuna.tsinghua.edu.cn/simple
Aliyun: http://mirrors.aliyun.com/pypi/simple/
University of Science and Technology of China https://pypi.mirrors.ustc.edu.cn/simple/
Huazhong University of Science and Technology: http://pypi.hustunique.com/
Shandong University of Technology: http://pypi.sdutlinux.org/
Douban: http://pypi.douban.com/simple/
Six, tensorflow
Open the anaconda prompt and enter pip install --upgrade tensorflow and press Enter. If it is the GPU version, enter pip install --upgrade tensorflow-gpu . After installing the test (in IPython), it is successful to output the result of 32.
>>>import tensorflow as tf
>>>sess = tf.Session()
>>>a = tf.constant(10)
>>>b = tf.constant(22)
>>>print(sess.run(a + b))
安装测试过程:
import tensorflow as tf一开始有如下报错:
报错原因:h5py和numpy不相容,修改numpy类型。tensorflow需要的numpy至少1.13.3以上,所以:pip install numpy==1.13.3
再进行验证,解决。
七、keras
打开anaconda prompt输入pip install keras --pre即可安装。装完了测试下,能正常运行就是成功,他给的示例数据下载的比较慢,需要开启科学上网才行:
>>> conda install git
>>> git clone https://github.com/fchollet/keras.git
>>> cd keras/examples/
>>> python mnist_mlp.py
测试安装过程:
没有报错
conda install git
git clone https://github.com/fchollet/keras.git
cd keras/examples/
python mnist_mlp.py
程序无错进行,至此,keras安装完成。