In Chainer2.0, the program structure has changed due to the independence of cupy. When I updated it, I got stuck a little, so I wrote a procedure at the memorandum level.
Please refer to the following page for how to create a basic environment. Install Chainer + CUDA 8.0 on Windows 10
Environment before update
pip uninstall chainer
The new Chainer has been separated from cupy (a numpy library with CUDA functionality), so it will be compiled against CUDNN during CUPY installation. The current cupy (1.0.0.1) requires cudnn v6.0. So get v6 from NVIDIA. https://developer.nvidia.com/cudnn Note: Member registration is required
Extract the downloaded zip file and select bin, include, lib under the cuda folder for each folder. C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0 Copy to.
Install using pip
pip install cupy
It takes about 1 minute. If it finishes smoothly, review the environment variables because cudnn is not recognized.
Confirmation work
python
>>> import cupy
>>> import cupy.cudnn
If there are no errors, it is successful.
#### 4. Chainer installation
`pip install cahiner --no-cache-dir`
## Summary
I was already running it in an environment with cudnn, so this time it was relatively easy,
If you make it from the initial environment, it seems that you need to review the environment variable settings of path and include again.
(Maybe it's a little easier !?)
Recommended Posts