Swift - Linux jupyter install

[codeck (Kousik)]

I am trying to install swift in linux and its been a pain, do you guys know an easy workaround? If not I’ll post on how i managed to solve the issues that i have/will come across. Also trying to get it work on jupyter-notebook without tensorflow.

Lemme know!

PS: Ubuntu 16.04, just wanna get familiar with swift!

[RogerS49]

Hi
i am trying to build swift for tensorflow on ubuntu 16.04 using virtual box this is a complex build process. Ready builds are for 18.04 10.3.* OS X. If Jeremy is using this for 2 lessons or perhaps it’s 1 but spreads over 2.
I would imagine there will be something in place to enable use to work on the notebooks or what ever.
If I build this image perhaps it could get uploaded somewhere but there would be no guarantees with it. Cheers

EDIT:
I found the link to the linux 16.04 swift-tensorflow download

Ubuntu 16.04 swift-tensorflow build Jan 04 2019

The link was hidden in plain sight but YAY! I don’t have to build it. Followed the ‘Older build link’
Have no idea if it’s cpu or gpu capable.

It looks like the ‘harebrain’ category is where we need to be for Swift stuff

[marcrasi (Marc Rasi (S4TF Team))]

I found the link to the linux 16.04 swift-tensorflow download

I would not recommend using the older builds because Swift for TensorFlow is moving very fast, and builds from Jan are already completely outdated.

[sayko]

how I installed on my 18.04:

1. installed clean nvidia drivers (this was pain to do)
2. installed CUDA 10.0, cudnn 7 for CUDA 10.0 (requires to make account on nvidia) - !first check what cuda and cudnn you need!
3. installed swift (followed instructions from here: https://github.com/tensorflow/swift/blob/master/Installation.md)
4. installed jupyter swift ( followed instructions from https://github.com/google/swift-jupyter )

And it works. Just stick to official notebooks

[a_yasyrev (Andrei Yasyrev)]

I used Dockerfile from swift-jupyter.
All you need - nvidia drivers installed, then start docker image and you has jupyter with swift.
Without nvidia gpu - just swift without gpu acceleration.

[j.laute (Johannes Laute)]

I tried building s4tf on manjaro for a couple hours but gave up. seems like for now docker and swift-jupyter are the best solution for non ubuntu linux.

[eunosM3]

Leaving this for future reference…

You can use Swift for Tensorflow in Google’s Colab, as mentioned here. Google has provided a blank Colab notebook with a S4TF kernel. It is convenient, and if I recall, Colab just upgraded the GPUs available.

[ste (Stefano Giomo)]

This is another option for linux users. I’m using it on ubuntu 16.04 and it works with no problems (there is only a little problem related to jupyter, but the workaround is on the post).

Swift for tensorflow using GPU with docker

[image] INSTALL Install nvidia docker from nvidia website: https://github.com/NVIDIA/nvidia-docker. NOTE: be sure to see the result of nvidia smi on test container after install. Clone https://github.com/google/swift-jupyter and build nvidia docker image git clone https://github.com/google/swift-jupyter.git cd swift-jupyter # from inside the directory of this repository docker build -f docker/Dockerfile -t swift-jupyter . Launch docker daemon: sudo systemctl start docker RUN To…

[RogerS49]

I entirely agree and as noted

I would not recommend using the older builds because Swift for TensorFlow is moving very fast, and builds from Jan are already completely outdated

I moved from 16.04 to 18.04. I have found this a much better platform to work on. I am able to take advantage of most of the latest Nvidia products and have the drivers etc updated by the inbuilt upgrade system.