Is it recommended to use the development snapshot nightly versions over the release Swift for Tensorflow versions available here, or does it not matter whichever is used?
It depends on your goal. The course used the old 0.4 toolchain. But the master branch of swiftai is on par with the nightlies.
After following the steps mentioned, I installed swift env in GCP and the kernel reflected with Swift. But when I open a swift notebook from tutorials I get the following error message.
Just reporting that I managed to do this, too! I’ve basically no experience with Linux let alone installing stuff on it, so it took a while. For other newbies also running into problems, I had to make these changes to Jeremy’s full gist, but that’s probably because some errors on my part:
- Used
pip3instead ofpipeverywhere, as the latter wasn’t installed - Added
~/anaconda3/binto PATH (and~/.bashrc) manually to be able to run conda - Created these directories manually before line 67:
/usr/local/cuda/includeand/usr/local/cuda/lib64 - On line 93, I couldn’t get
conda install -c pytorch -c fastai fastai pytorchto work because of an UnsatisifableError, so I ranconda install pytorch torchvision cudatoolkit=10.0 -c pytorchinstead, but dunno if that was very clever
- Copied the swift-jupyter kernelspec json output by register.py at the very end manually to
~/.local/share/jupyter/kernels/myswift/kernel.json, as the kernel wasn’t listed withjupyter kernelspec list
Kalle