Unofficial Setup thread (Local, AWS)

Part 1 (2019)
162

I found the following conda formula worked better for setting up on a personal machine:
conda install -c pytorch -c fastai fastai pytorch torchvision cudaxxx
Where cudaxxx matches the version of CUDA you have installed.

4 Likes
163

Thank you @neuradai, indeed your instructions work for me, too. They provide fastai=1.0.42 and pytorch=1.0.0. (I hadnā€™t realized that Pytorch 1.0 had been released)

The instructions at the top of this thread and in the AWS instructions are out of date, it would seem. Iā€™ve only just restarted, so Iā€™m not confident enough to edit the instructions above. Perhaps someone with more experience would look over the bits where they install pytorch-nightly & update per the prior post instead?

1 Like
164

@rogerallen I just submitted a PR for the conda formula change to the AWS instructions via GitHub. If they agree, maybe the fastai maintainers will propagate it to other instructions, as wellā€¦

1 Like
165

I ended up creating a GCP instance for the V3 version only. I guess there are posts on how others have used AWS to create both versions together.

166

Hi all,
Apologies for not keeping this post updated. I had created this however, I didnā€™t update it with the latest instructions over time.

Iā€™ll update it soon this week and re-comment. Sorry for the issues that everyone had to face because of this.

1 Like
168

This formula worked great. Thanks for posting.

1 Like
169

I found that if I wanted to update fastai with conda it would include pillow as an update which would overwrite my compiled pillow-simd

Here is what you can do to avoid this. Write a file call ā€˜pinnedā€™ in your conda-meta directory. You will find this in your
~/anaconda3/envs/fastaiv3/conda-meta
The file ā€˜pinnedā€™ I create as such
vim ~/anaconda3/envs/fastaiv3/conda-meta/pinned
and placed the text as follows
pillow-simd ==5.0.3.post0

There maybe other files from the build of pillow-simd I need to add to this.
Search the conda online docs for ā€˜pinnedā€™
Please comment

170

I have built my own box and these instructions helped immensely, thanks. I have followed them as written, cuda is successfully installed etc. I get as far as
conda install -c fastai torchvision-nightly

and get the following error output

Collecting package metadata: done
Solving environment: failed

PackagesNotFoundError: The following packages are not available from current channels:

  - torchvision-nightly

Current channels:

  - https://conda.anaconda.org/fastai/linux-64
  - https://conda.anaconda.org/fastai/noarch
  - https://repo.anaconda.com/pkgs/main/linux-64
  - https://repo.anaconda.com/pkgs/main/noarch
  - https://repo.anaconda.com/pkgs/free/linux-64
  - https://repo.anaconda.com/pkgs/free/noarch
  - https://repo.anaconda.com/pkgs/r/linux-64
  - https://repo.anaconda.com/pkgs/r/noarch

To search for alternate channels that may provide the conda package you're
looking for, navigate to

    https://anaconda.org

and use the search bar at the top of the page.

how should I resolve this?

thank you

171

If I instead try conda install -c fastai torchvision it wants to downgrade pytorch from pytorch-nightly-1.1.0ā€¦ to pytorch-0.4.1

172

*

prob
prob.png3414Ɨ1674 178 KB
*

I am working on mx150 gpu , which although have some cuda cores in it, torch is giving false when i am checking for cuda cores.

I am very new to this part of deeplearning , been using keras so far

173

Please follow the instructions in the official repo.
fastai now has a single command install for setting up cuda, torch, fastai instead of multiple pip installs.

Iā€™d also suggest using conda over pip.