PyTorch V1.0 Stable Release for Linux & Windows

I noticed this morning on PyTorch.org the availability of PyTorch V1.0 Stable Release for Linux & Windows, to include Cuda 10.0.

conda install pytorch cuda100 -c pytorch
pip3 install torchvision

Any issues upgrading from PyTorch 0.3.1 which I’ve used from the start?

You’ll have to upgrade to fastai v1. Note that support for windows in fastai v1 is experiments for now, and we’ll fix the related issues starting next week.

Thanks for the fastai status update. I dual-boot Win10 & Linux Mint, depending on my applications work. I’ll stay with the old version in Win10 and the latest Linux version for now. Will monitor the Win10 fastai upgrade. Looking forward to making both environments V1.

I’ve just pushed support for Windows that handles everything except for 3 transforms with num_workers>0:

  • pad
  • crop
  • crop_pad

(note that default get_transforms uses these, so you’ll need a custom transforms list.)

You’ll need github master for this BTW.

I installed both pytorch (conda install pytorch torchvision -c pytorch) and fastai earlier today on Win7 (clean anaconda install with Python 3.7). It was dead slow with cpu only (quad core i7) especially with text. I tried https://github.com/fastai/fastai/blob/master/examples/text.ipynb and did not complete the first 4 epochs :slight_smile:

1 Like

Thanks to @313V this limitation is now removed in master :slight_smile:

I think the issues discussed there still persist ? Recursion error; fastai v1.0.27, Windows 10

Does anyone have a pytorch 1.0 Windows 10 fast ai installation tutorial/post somewhere? All the Windows 10 stuff I can find is a bit outdated.

Thanks.

There is no tutorial, just follow the instructions of the home page.