nengo / nengo-loihi

Run Nengo models on Intel's Loihi chip
https://www.nengo.ai/nengo-loihi/
Other
35 stars 12 forks source link

Installation documentation out of sync #244

Closed arvoelke closed 4 years ago

arvoelke commented 4 years ago

After #240 the installation documentation should mention both scipy and numpy.

In addition, the instructions refer to nxsdk==0.8.1 in a number of places, when that version is no longer supported (currently only 0.8.5 is supported)?

drasmuss commented 4 years ago

Where are you thinking about referring to scipy? We refer to numpy because it should be slightly faster when installed through conda, but installing scipy through pip or conda shouldn't make a difference. We don't explicitly list out all the requirements in that document (as they should all just be installed through pip install nengo-loihi) E.g. we also require jinja, but I don't think that needs to be explicitly mentioned either.

But we should definitely update the nxsdk version.

arvoelke commented 4 years ago

Didn't look too thoroughly but comments like:

If anything goes wrong during this process, it is likely related to installing NumPy.

SciPy seems as likely of a culprit but I could be wrong?

But moreso:

Nengo and Nengo Loihi’s emulator backend are pure Python packages that use NumPy to simulate neural models quickly. On your local machine, you only need to install Nengo Loihi and its dependencies, which include Nengo and NumPy. See Installation for details.

from the overview.

Technically this is all accurate the way it's written, and admittedly I made the mistake of thinking the link to "NumPy install instructions" was staying within the nengo-loihi docs (and not crosslinking to the nengo core docs). Mainly I just had the expectation that I would see scipy mentioned somewhere.

hunse commented 4 years ago

In my experience, Numpy's always the one that's been hard to install, particularly if you want it fast. It's actually much better than it used to be, but it's possible on some systems it may still be a bit tricky to hook up to whatever BLAS library you want to use.

I've never had a problem installing Scipy once Numpy is installed. The main thing I could see going wrong there is it does have a bunch of C and Fortran code, so if for some reason you had to install from source you'd need to have working C and Fortran compilers installed. But I've never had this happen to me, since so many systems some precompiled version available now.

drasmuss commented 4 years ago

We should also add a comment here https://github.com/nengo/nengo-loihi/blob/master/nengo_loihi/hardware/interface.py#L93 to update the installation docs when we change the nxsdk version.

arvoelke commented 4 years ago

Relatedly, python 3.5.2 is no longer in the standard conda channel, and so this is a reminder to update the installation documentation (to suggest adding the conda-forge channel).

hunse commented 4 years ago

Note: The command to add the conda-forge channel with python 3.5.2 is conda config --add channels confa-forge

arvoelke commented 4 years ago

This looks resolved now in the 0.10.0 release. :)