Howto: installation on Windows

Hi David,
I confirmed. Your suggestion solved the error. Thanks for the post.

Regards,
Abhishek

Hello Jonas,
Thanks for sharing information. So pytorch cuda installation would be required on a windows machine supporting GPU… right? Machine that does not have GPU would not support it, I believe.

I am getting following error (please note I installed pytorch CPU version as described by David Malan)

(base) D:\Users\abhishekt>conda install pytorch cuda90 -c pytorch
Solving environment: failed

UnsatisfiableError: The following specifications were found to be in conflict:

  • backports.functools_lru_cache
  • pytorch
    Use "conda info " to see the dependencies for each package.

pytorch supports working with CPU as well, and I do not see why the installation should not work. Unfortunately I have no idea concerning the particular error you get.
However I question if it makes sense to use a CPU - you need the efficient parallel matrix / tensor operations and the deep learning algorithms of NVIDIA to get the performance you need. A CPU has a handful of cores whereas a GPU has thousands. Instead I would use the virtual environments suggested in the course: Paperspace, AWS, Crestle… good luck!

@jeremy
For step8:
For Windows 10 in developer mode and Git for Windows (I’m running 2.17.1(2)) I believe symlinks are properly created.
del fastai
will delete all the files within the directory instead of deleting the shortcut. I was then confused as to why the fastai imports didn’t work

Perhaps a warning note for step 8 is required for Windows 10?

My situation is this:

  1. I’d like to take the fast.ai course using a paperspace GPU.
  2. I’ve never used linux/ubuntu, but always only windows.

I see two options:

  1. I could buy acces to a paperspace machine with the fast.ai (linux based) public template. Downside is, that I’d probably have all kinds of problems using ubuntu, which would slow down my progress in deep learning.
  2. I could buy acces to a paperspace machine with a windows 10 template. Downside is, that I’d not have easy acces to the data sets (and probably also other stuff) used in the fast.ai course.

Am I understanding my options correctly? Is there a third alternative? Have any of you been in the same situation as me, and if so, what are your experiences?

Best,
Kristian


hey, after executing “conda env update” it show me this error, when i retry the command it process the last three “pytorch-0.3.1-” packages then show me the same error, i tried to install pytorch using CMD by “pip3 install http://download.pytorch.org/whl/cpu/torch-0.4.0-cp36-cp36m-win_amd64.whl” but it doesn’t change any thing here?

i have the same GPU, i’m trying to start but i’m stuck here, can you help me?

They are just connection errors, either by your end or by the remote server.

i tried so many times, couldn’t solve it, i’m looking for other ways of installing it

I was thinking you could download the repository, and then modify the installation script to read the repository locally. Try and send a PM to Jeremy about how to do it exactly.

Tensorflow has started supporting these versions.

@ayush1397,
thanks.
I have switched to ubuntu 16.04 with cuda 9.0 and cudnn 7.0
so i am good.

1 Like

Hi everybody
I install fastai as @Jeremy described, but I have error when I run the first cell
“Matplotlib requires six>=1.10; you have %s” % six.version)

What is wrong?
with pip show Matplotlib I have 1.11

I am using Windows 10 with an MSI-GeForce GTX 1070 8GB Video Card. I updated the drivers and began to the following steps. I downloaded Anaconda, Git, and proceeded to the following steps. The only part that seemed a bit off was when I did conda env update it gave me these recommendations:

Not sure what to do there but continuing. Once I got to step 8 on the del fastai part I get this:

fastai1

Now lastly when I go into lesson1.ipynb, and I try to proceed but I get this error over and over again:

As an aside, I am a Mathematician by education and I am getting very interested in ML/DL so I just want to get my system set up to take my journey learning about ML/DL to the next level. If anyone can help I would really appreciate it.

to me, it looks like your fastai symlink is not created. you deleted the “bogus” one that works out of the box for unix, but do you see the error message when you try to create the link? You need to run that command as an administrator.

Somewhere in this forum is a link where symlinks can be created from the GUI using the RMB pop-up menu. That is how I set mine up.

Actually, I was able to get it to work by doing the steps above as an administrator. Although, when I closed out and tried again making sure to activate fastai I launched jupyter and still got the same error that no module named ‘fastai’. I am not sure what the problem is now when I just got it to work.

the installation process of the windows is very easy. you get the proper ideas from hp support number. you will get the download link from this site which is very much useful for all the user.

Quick note: I have noticed that learning rate finder version 2 doesn’t work with the old pytorch (0.3.xxx) as it crashed on x.data (Tensor autograde issue).

After installing the latest version of PyTorch (0.4) via “conda install pytorch torchvision -c pytorch” I ran into an import error regarding some DLL. The following fixed it for me: “conda install mkl mkl_fft intel-openmp numpy”. Running fine now, with LRF2 :smiley:

You don’t have permissions, try using Administrator access and if you are on Win 10x64, you can use WSL to create the symlink for you…(just a simple command which can be easily Googled)

The easiest way to use fastai is to just clone it in a particular directory, and just point your jupyter notebooks to the parent directory of which fastai is a part of…
That way I can create all my notebooks at one place(the con side) but I can surely do git pull to use the bleeding edge version…

1 Like