I discovered the same issue. It happens in WSL. I’m currently investigating.
fastai and Pytorch do not work in Windows or in WSL.
I will try again tonight to do dual boot with Ubuntu on my system.
Do you know of any virtual system (like VMWare or something ) that may take the cards in with Ubuntu)?
Pretty sure there are none, unless you have a Tesla card.
Hello @gerardo, I’ve got the same issue.
I’m trying to setup my own local DL server on my computer with the following configuration :
- Windows 10 Family (last version of October 2017 : “Fall Creators Update” (1709))
- NVIDIA GeForce GTX 1070, CUDA 8.0.61 installed + CuDNN 5.1 installed through Windows x86_64 installation (cf Windows 10 CUDA tutorial)
- Ubuntu installed as WSL through Windows store (cf the Workshop video)
- Anaconda for python 3.6 installed for Linux as showed in the Workshop video
- fastai virtual environment activated thanks to environment.yml (after getting fastai folder through git)
BUT… I read from @jeremy :
Then, I conclude that it is not possible to use Windows as a local DL server for this course (Part 1 v2). I’m right ?
It looks like that the unique solution would be to create a linux partition on my computer, boot on it and install Ubuntu, Anaconda, fastai, CUDA, CuDNN, Pytorch for Linux. I’m right ?
Thank you.
Looks like that’s the way to go.
I created my dual boot with Ubuntu and Windows 10 Pro on that computer but now I’m encountering other issues.
@jeremy I think you need to add on your pre-requisites Master in Linux with Administration 
25 years after I programmed one of the first NeXT computers. My knowledge of linux is coming back slowly but coming back. 
1 Like
When I did dual boot with the latest Ubuntu 17.04…
It messesed up with my partitioning…(even though I have used the Install Ubuntu Along Windows Boot Manager...
So far I find that Windows is not working with Lesson 1 notebook. I’ve listed the issues in the original post. I’ll try a fix. It’s a big deal for some if we can get Windows working.
Windows seems to work well with most machine learning software (Tensorflow, Scipy, etc.) including GPU support but not with the issues listed in the original post.
Hello @bsalita,
I agree with you. It is a big deal as it looks like Microsoft has no plans at this time :
- PyTorch & Windows 10 : PyTorch has no “Get Started” procedure for Windows at http://pytorch.org/
-
CUDA & WSL : from Rich Turner [MSFT] on October 16, 2017
We are looking into how we might be able to support CUDA/OpenCL in future versions of Windows, but have no plans at this time to support graphical/UI technologies.
Then, I think @jeremy should tell in the next video that only a linux OS can be setup as a local server for this course. Great if I’m wrong but for the moment, I stop trying to setup my Windows 10 with WSL Ubuntu as a local server for this course and switch to Crestle/AWS/PaperSpace.
1 Like
Will it work if I use ubuntu through Hyper-v on win 10 pro? Or dual boot is what you suggest?
Setting up AWS is relatively much easier…
@vrajjshah I’m still having troubles even with the dual boot Ubuntu seems finicky to certain things specially with Anaconda running.
Windows 10 with WSL does not work and looks like
I’m going to continue this weekend with the installation of Ubuntu and the tools required by the course but If cannot have it ready by Monday I will try AWS that I know it works 
@Gerardo Did you successfully setup your Linux on other partition? If yes what all did you install and did you find any tutorial/guide for setting it up?
@vrajjshah I thought that I had the two partitions until I restarted and Windows 10 is now dead on my PC.
My Ubuntu configuration is up and running 
Two cards are showing up.
and the pytorch seems to be responsive to the GPUs.
Maybe during Thanksgiving break I will go back and re-install the whole thing again and see if I can get the dual boot that I need.
Thanks @bsalita
Looking at this for the first time. In the fast.ai system we are loading an environment file. But this instruction makes no mention of it that I could see. Can you clarify please, if these installation requirements should be fulfilled before or after loading the fastai environment, or supersedes the loading of it?
Also, you specify conda install pytorch first, then suggest conda install -c peterjc123 pytorch. Can we clarify please - I suspect we only need the peterjc123 version for GPU support directly under Windows 10. On the other hand if we are running from the WSL subsystem we can’t access our GPU from there so standard pytorch should be sufficient. Is this correct?
Heh What you’re doing is great practice - but note if you use Crestle there’s nothing to set up at all!
1 Like
I get “execstack: cannot open ELF file: invalid file descriptor” when I try this step…
I have not explored the environment file. The instructions show how to manually set up requirements for lesson1 on Windows or WSL and is helpful for Linux too. Please note the showstopper issues. I believe getting all of Lesson 1 to run on Windows or WSL can be done but it requires changes to the fastai library. I haven’t had time to work on Windows pull requests for fastai library. There’s possibly some non-fastai issues too.
Yes, peterjc123 is specifically for Windows prompt, others can use pytorch.
Hi Robert
Thought I would give this a try. You are right, I can get Pytorch working directly under Windows 10, and even started running the code arch=resnet34, but after a number of epochs (87, then 360, then on the 32nd of 32), I got a memory issue - see snapshot below.
On the process of from torch._C import * the error comes as DLL load failed: The paging file is too small for this operation to complete.
Perhaps you could let me know if this makes sense to you? I have 16GB RAM in my PC, but my GPU is rather inadequate - an NVDIA GeForce GTX 650 Ti.
Do you have a link to the model that you’re running? I can test on my systems. Are you able to run CPU only?