StarNet++ module in PI 1.8.8-6 on Mac OS Mojave?

myxlflik

myxlflik

Member
I have seen conflicting reports about whether or not the StarNet++ module works on Mac OS Mojave. I have found that when I first run PI 1.8.8-6 after installing it, it crashes immediately unless I remove the StarNet-pxm.dylib file form the /Applications/PixInsight/bin directory. If I do that PI runs and can update. If I put the file back in the directory, it does not show up when I launch PI, but it doesn't crash PI either. If I go to Modules > Install Modules and select recursive and search it crashes immediately. If I remove the StarNet-pxm.dylib file from the directory then the search runs and yields no results. Am I limited to running an older version of PI if I want StarNet++ to work? Is this something that can be resolved so that I can run it on my Mojave machine?
 

Attachments

  • Mac Pro (No Serial).jpg
    Mac Pro (No Serial).jpg
    68.5 KB · Views: 53
  • Crash.jpg
    Crash.jpg
    500.5 KB · Views: 58
it works on mojave.

the problem is your CPU does not support AVX2/FMA instructions - it's too old. tensorflow and libRaw depend on these instructions.

rob
 
That's what I was afraid of. I don't suppose tensorflow and libraw could be compiled without needing the AVX2/FMA support could it?
 
in theory you can compile the source at tensorflow.org and target your machine, which would prevent clang from emitting those vector instructions. this might be easier said than done depending on your skill level with Xcode/software development in general.

another problem is that the PI core application has the "hardening" feature turned on under OSX which requires all linked libraries to be signed with a paid apple developer certificate, and perhaps the certificate has to be the same as juan's. this might prevent you from substituting your own tensorflow library into the PI application bundle. having said that, there are ways around the signing stuff which of course are not recommended.
 
I am guessing that the only viable workaround would be to get the package built for machines without that support from the original developer. Not likely that will happen, but never know. I have another Mac and a Linux box that are capable of running StarNet++ so I can do that part of my workflow on those machines if needed. It's a shame my beefy Mac Pro just isn't new enough in that regard. It is awesome for everything else.
 
well the original developer of tensorflow is google and they will not help beyond whatever info webpages are on tensorflow.org.

the command line versions of starnet might work on your CPU as the tensorflow library distributed with it might be compatible with your CPU. that's on sourceforge still (there is a python version and a c version).
 
I tried installing the dylib's included in the StarNet_PI_OSX.zip file from sourceforge and that did not work either. It's not the end of the world as I have other machines I can run it on. :)
 
StarNet is fun, but I find the novelty soon wears off. If your objective is to create "arty" starless images, then the ghostly residual halos of bright stars are an unwelcome artefact; if your objective is the temporary suppression of stars in order to do fine "non-stellar background" image processing, followed by adding the stars back again, then I think more traditional methods using the careful construction of multiple star masks are really needed to produce good results.
 
myxlflik said:
I tried installing the dylib's included in the StarNet_PI_OSX.zip file from sourceforge and that did not work either. It's not the end of the world as I have other machines I can run it on. :)
Click to expand...

yes those may also have the AVX2/FMA instructions; i can't remember anymore. i don't think i have any machines with xeon CPUs so all my machines had those instructions.

fredvanner said:
StarNet is fun, but I find the novelty soon wears off. If your objective is to create "arty" starless images, then the ghostly residual halos of bright stars are an unwelcome artefact;
Click to expand...

really all you need to do is stretch your image gently and then run starnet, optionally undoing the stretch to get back to linear. besides that the starmasks produced by starnet are wonderful since they capture all scales at once. it's certainly possible to touch up the starmasks if there are any artifacts.

rob
 
Thanks tdgm. Well, I put the files where the "read me" file said to but all I can find in Pixinsight is a "Starnet" process but not a "Starnet++" process. And, the Starnet process has a couple of red "x" circles on a couple of fields and fails. Intriguing process but the instillation might be above my pay grade.
 
Starnet++ is now integrated into PI as the Starnet process. You need to run this process and use the "settings" (wrench icon) to point to the two weights files (usually in the ...Pixinsight/library folder). Then Starnet runs just like Starnet++.
 
I am running PixInsight 1.8.8-9 on MACBook Pro with the new M1 chip. StarNet does not even show up as a process with PixInsight, so there appears to be no mechanism to install it. Any thoughts? There is a lot of discussion of StarNet on the MAC, and the general consensus seems to be a "no go"?
Thanx.
 
starnet as shipped with PI works fine on intel macs. however the tensorflow libraries juan bundled with PI contain AVX2 and FMA instructions which the Rosetta2 x86-64 emulator on M1 does not support. so the PI installer just deletes the Starnet-PXM.dylib file, which contains the starnet module.

there are 2 fixes. one is to use the zipfile package that airscottdenning put together that contains tensorflow libraries which don't have those vector instructions. the package also contains the starnet module (Starnet-PXM.dylib) so you don't have to go dig it out of the PI installer (which can be done with Pacifist)

the other fix is to install StarNetV2. nikita shipped this with similar tensorflow libraries (no vector instructions) so it will run out of the box on M1. however, you need to be sure to delete all the old tensorflow library files from your PI installation as the old libraries will cause PI to crash on the M1, and wouldn't work on an intel mac either because the new StarNet V2 expects a newer version of the tensorflow libraries.

rob
 
kfinley1 said:
Does anybody know if Starnet++ will run on a Mac running Big Sur V 11.1?
Click to expand...
I have a problem with my iMac 17.1 of 2015 & Catalina + Older version Pixi 1.8.8.8 (last version for Catalina) and Starnet++ not possible to use. But after uptdate Catalina to BigSur 11.6.1, and new reinstall Pixi to last version 1.8.8.11 the Starnet++ its all Ok.
 
Last edited:
You must log in or register to reply here.
Back
Top