Learn machine learning in windows environment
At first, I want an environment that is as plain as possible.
Stable operation in the following environment
| item | version |
|---|---|
| python | 3.5.3 |
| CUDA | 8.0 |
| cuDNN | 5.1 for 8.0 |
| Visual Studio | community 2015 update1 |
Compare using the chainer sample (chainer / examples / mnist / train_mnist.py) as it is as a benchmark of processing time. The elapsed time (elapsed_time) when 20 epoch was rotated with the default was compared in the existing environment.
| Use | name | CPU | benchmark |
|---|---|---|---|
| mobile | GODWIN | Atom Z8700 | 1.5h |
| Note | Lenovo X250 | i5 | 15m |
| desktop | self made | i5 | 15m |
| 〃 | GPU usage | GT 730 | 3m |
| 〃 | GPU update | GTX 1080ti | 1m |
I bought 1080ti in a hurry, but GT730 was enough for studying for the time being orz. Also, I am grateful for the existence of GPD WIN, which allows GUI operation of GPU machines even on the go.
python Tensorflow 1.01 did not work under python 3.6 system environment Stable with 3.53 (x64) downloaded from the official website
For other general-purpose libraries, follow the steps below.
> python -m pip install --upgrade pip
> python -m pip install --upgrade -I setuptools
> python -m pip install ipython
> python -m pip install numpy
> python -m pip install pandas
> python -m pip install matplotlib
tensorflow No difficulty in gaining
> python -m pip install tensorflow
On GPU-enhanced machines
> python -m pip install tensorflow-gpu
chainer There is no particular difficulty in operating the CPU
> python -m pip install chainer
I used the MNIST sample to compare the processing performance of each machine. In order to run this on GPU, it is necessary to install VS and I am addicted to it.
sklearn I stumbled. Requires numpy + mkl and scipy Download .whl respectively as it was not available for online pip installation
> python -m pip install "numpy-1.12.1+mkl-cp35-cpm35-win_amd64.whl"
> python -m pip install "scipy-0.19.0-cp35-cp35m-win_amd64.whl"
> python -m pip install sklearn
sklearn is used for trial implementation of SVM.
CUDA I am using it because 8.0 has been released. cuda_8.0.61_win10.exe
cuDNN You need to create an account to download. As soon as I declared the purpose of use for business use ticks, he issued an account.
v6.0 for CUDA 8.0 was released, but NG. cuDNN v5.1 (Jan 20, 2017), for CUDA 8.0 is used.
The entity is just like a DLL module used via CUDA, so you can put it where you can see it. He passed me through the path when installing CUDA "\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0" I put it in.
If you try running mmist_with_sumaries.py etc. from the tensorflow tutorial using GPU If you don't see "cupti64_80.dll", you may get angry. "\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\extras\CUPTI\libx64" I had to give it through an additional path.
VisualStudio I installed it because I went to use the compiler from CUDA or cuDNN for the operation of the chainer sample. Stable with VS 2015 community update 1.
I was super addicted. It cannot be used from CUDA after 2015 update2. VS 2013 community may be enough. In the first place, VS official is already 2017. I can't find a link to a previous version
Visual Studio Community 2015 with update 1(.iso)
Git DL and use git for windows 2.12.2 (2) from github.io
> git clone https://github.com/tensorflow/tensorflow.git
> cd tensorflow\tensorflow\examples\tutorials\mnist
> python mnist_softmax.py
> git clone https://github.com/pfnet/chainer.git
> cd chainer/examples/mnist/train_mnist.py
> python train_mnist.py
octave v4.2.1(x64)
Useful for trial vector / matrix operations and visualization of results. After taking Coursera Machine Learning, I experienced the convenience. I can't write any code to check the operating principle. Instead of MATLAB.
v4.0.0 is told to avoid.