5 ms·

GPU Accelerated Theano and Keras with Windows 10

64 points by efavdb 10y ago

naveen99 10y ago
https://github.com/philferriere/dlwin https://github.com/philferriere/dlwin Worked for me on windows7 also.
- efavdb 10y ago
  His guide helped a lot, but it didn't for me. I like his sections for the optional add-ons though.
  - alok-g 10y ago
    Could you please expand on what went wrong? :-) Thanks.
gtani 10y ago
I'll try this later, seems to be targetting pre-Pascal cards (need more info on whether this will work with Cuda 8). "Cuda for Engineers" (excellent book by Storti/Yurtoglu) uses VS 2013 and the VC++ that comes with it, all my other cuda books (Handbook by Wilt, Wrox Pro Cuda Prog'g) run on linux. Also worth noting that OS X seems to have gotten feature parity with Cuda 8's unified memory (my mac doesn't have recent Nvidia card to test).
- efavdb 10y ago
  I suspect that it will work with CUDA 8, will have to give that a try.
- psyklic 10y ago
  I got a 1080 working with VS 2015 (via the CUDA 8 beta). However, I strongly recommend using VS 2013. Frustratingly, CUDA VS 2015 support does not extend to the latest Visual Studio update level. So, to use CUDA I had to forgo the latest versions of useful extensions (e.g. Python Tools) and even forgo MS software updates (e.g. for SQL Studio, which apparently auto-installed the latest VS updates).
2bitencryption 10y ago
Any hope getting this to work with Windows Subsystem for Linux? I got my model up and training on wsl, which really surprised me -- it was super easy and just worked. But I didn't even try venturing into gpu acceleration... perhaps it would "just work" as well? Though I suspect not.
- efavdb 10y ago
  If you want to try on the Linux subsystem I would give a Linux guide a try. I would be surprised though.
  - kenrick95 10y ago
    Curiously, I tried and it seems that it couldn't communicate to the GPU. Installation is successful, but upon running `nvidia-smi` command, it returned the following: NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
- akhilcacharya 10y ago
  It does not yet, one of the program managers (I think?) said that GPU acceleration in WSL is a low priority issue but one that they're considering.
DTE 10y ago
This is great. We do a lot of work with GPUs in Win10 at Paperspace (https://www.paperspace.com https://www.paperspace.com) but we mostly use GPUs for video acceleration not compute. That said, all of our cloud VMs have GPUs and we are adding new instance types designed for ML/ HPC. Reach out directly if you are in the space and want early access to CUDA-backed Windows instances in the cloud.

64 points by efavdb 10y ago

naveen99 10y ago

https://github.com/philferriere/dlwin https://github.com/philferriere/dlwin Worked for me on windows7 also.

efavdb 10y ago
His guide helped a lot, but it didn't for me. I like his sections for the optional add-ons though.
- alok-g 10y ago
  Could you please expand on what went wrong? :-) Thanks.

gtani 10y ago

I'll try this later, seems to be targetting pre-Pascal cards (need more info on whether this will work with Cuda 8). "Cuda for Engineers" (excellent book by Storti/Yurtoglu) uses VS 2013 and the VC++ that comes with it, all my other cuda books (Handbook by Wilt, Wrox Pro Cuda Prog'g) run on linux. Also worth noting that OS X seems to have gotten feature parity with Cuda 8's unified memory (my mac doesn't have recent Nvidia card to test).

efavdb 10y ago
I suspect that it will work with CUDA 8, will have to give that a try.
psyklic 10y ago
I got a 1080 working with VS 2015 (via the CUDA 8 beta). However, I strongly recommend using VS 2013. Frustratingly, CUDA VS 2015 support does not extend to the latest Visual Studio update level. So, to use CUDA I had to forgo the latest versions of useful extensions (e.g. Python Tools) and even forgo MS software updates (e.g. for SQL Studio, which apparently auto-installed the latest VS updates).

2bitencryption 10y ago

Any hope getting this to work with Windows Subsystem for Linux? I got my model up and training on wsl, which really surprised me -- it was super easy and just worked. But I didn't even try venturing into gpu acceleration... perhaps it would "just work" as well? Though I suspect not.

efavdb 10y ago
If you want to try on the Linux subsystem I would give a Linux guide a try. I would be surprised though.
- kenrick95 10y ago
  Curiously, I tried and it seems that it couldn't communicate to the GPU. Installation is successful, but upon running `nvidia-smi` command, it returned the following: NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
akhilcacharya 10y ago
It does not yet, one of the program managers (I think?) said that GPU acceleration in WSL is a low priority issue but one that they're considering.

DTE 10y ago

This is great. We do a lot of work with GPUs in Win10 at Paperspace (https://www.paperspace.com https://www.paperspace.com) but we mostly use GPUs for video acceleration not compute. That said, all of our cloud VMs have GPUs and we are adding new instance types designed for ML/ HPC. Reach out directly if you are in the space and want early access to CUDA-backed Windows instances in the cloud.