31 replies to this topic

[nekitmm]

Hi, everyone!

 

I was made aware that threads in Vendor Announcements, which I was using to get feedback and receive help requests, should only be used for announcements, so I've decided to create a couple of threads here, in Experienced Deep Sky Imaging. Not sure how far this will go, but oh well.

 

This thread is for experience share: will be glad to hear back from you about your experience with StarNet, good or bad, examples of processed images, etc.

 

The thread to request and get help is here: https://www.cloudyni...-help-requests/

 

I will start by sharing a couple of results I get on my end:

 

Hyades and Aldebaran using Zeiss 135 mm lens and QHY16200A:

 

Edited by nekitmm, 06 March 2022 - 02:28 PM.

[nekitmm]

VdB 152 region using FSQ106 and QSI683 (from DSW):

 

Edited by nekitmm, 06 March 2022 - 02:23 PM.

[nekitmm]

M78 using ASA500 (Chilescope)

 

[nekitmm]

Very dense region around Barnard's E (FSQ106 + QSI683, from DSW):

 

[astrokari]

Starnet v2 is excellent with PixInsight on my M1 Mac. This is the first version of the program I've used, and I'm delighted with the results. 

Here's NGC 6188 in narrowband from iTelescope.net's T30 (Planewave 20" CDK + FLI-PL6303E  CCD, imaged from Eastern Australia):
 

 

Edited by astrokari, 22 March 2022 - 10:10 PM.

[Woodbridge_Dave]

I'm enjoying Starnet 2, but I don't use it to produce starless images.  Starless images always look distorted, as if someone manually clone stamped the stars out.   My use of Starnet is to separate stars, stretch the nebula slightly, then rejoin the stars and continue processing.  The ability to do this while linear is fantastic.  Thanks for making this interesting software.  Any plans for Starnet 3.0 ?

[mikeyL]

I don't have examples to share, but can simply say that as nice as V1 was, V2 is 20 times better, and the ability to use it to create linear starless masters is wonderful, as I too use it to remove stars and process nebulosity separately before adding RGB stars back form other data collected at the same time. Just a wonderful job on this, really amazing work! 

 

ML

[WidgetsDad]

I just tried out the WIndows GUI.  Works great, but how do you get the star mask?  I was under the impression that you would get that image also.

[Holtdoa]

Anyone installed V2 to PI on Linux and gotten it to work? I follow the directions and PI never finds the process when I tell it to search. The libtensor* files in the lib directory didn't look like they had typical shared lib permissions, tried changing those, still no dice. The file mentioned in the troubleshooting section is in the specified spot. This looks like some requirement is not being met...

[idclimber]

Anyone installed V2 to PI on Linux and gotten it to work? I follow the directions and PI never finds the process when I tell it to search. The libtensor* files in the lib directory didn't look like they had typical shared lib permissions, tried changing those, still no dice. The file mentioned in the troubleshooting section is in the specified spot. This looks like some requirement is not being met...

I tried earlier today, but was not able to copy the files into the PI directory due to permission issues. I don't know why they don't use a repository link so we can actually use the normal installer for this stuff. 

Edited by idclimber, 05 May 2022 - 10:42 PM.

[Manitu]

I just tried out the WIndows GUI.  Works great, but how do you get the star mask?  I was under the impression that you would get that image also.

Same here. I also used the Windows GUI but wonder how I could get the star mask with stars only.

I see many examples which show the starless object and a star mask in the Tipps & Tricks section ont he starnet website.

But there is no option in the Windows GUI to save a star mask.

 

Anyone can help out here? 

Is that possible in PI only? And if yes how to get it there?

 

CS Tom

[galacticinsomnia]

In your favorite image editor.  Place the original photo, and the starnet output image next, then just change the layer property of the starless photo to difference, and instant star mask.

Clear Skies !!

[TheRealEleven]

I cannot for the life of my get this to work as a plugin on the Pixinsight linux distribution in EZ Processing Suite. I've followed all the instructions, but I'm starting to suspect that it's a file permission issue. 

 

/opt/Pixinsight/bin/ and it's child directories are only accessible with sudo. I think maybe having to get sudo access in order copy the Starnet2 files where they need to be might be somehow interfering with PI actually using it correctly. 

 

Specifically, it fails when trying to use EZ Processing Suite. EZ Decon uses Starnet2 to make a mask and it simply fails out.

 

Anyone had this issue? 

 

 

(EDIT)

 

Soooo....of course I figure it out right after posting. Every tutorial on this that I've found said to put the files from the Linux download into /opt/PixInsight/bin and /opt/PixInsight/bin/lib but never once mentioned having to set the Weights File in the original Starnet process. THAT is what Starnet2 looks for. Once I realized this, the process went a little farther before it would error out, but it still failed.

 

Then I looked for exactly where those weight files were in linux, and they were in /opt/PixInsight/library so I had the obvious idea to put the new StarNet2_weights file there. I also threw in the libtensorflow and libtensorflow_framework files that came with the StarNet++ linux download and lo and behold, EZ Processing Suite Deconvolution now works correctly. 

 

Yowza!

Edited by TheRealEleven, 11 September 2022 - 02:18 PM.

[lock042]

Hello nekitmm.

Just for your information.

In the next big release of Siril (1.2.0), it will be possible to use starnet.

 

Edited by lock042, 25 December 2022 - 09:32 AM.

[lock042]

How it looks like

Attached Thumbnails

• 

[Manitu]

I already used it in an unofficial version of siril and it works well. Very easy to use. Also with linear data. And pretty good results like the stand alone version.
With siril you need to install the command line version.

CS Manitu

[nekitmm]

Hello nekitmm.

Just for your information.

In the next big release of Siril (1.2.0), it will be possible to use starnet.

 

Capture d’écran du 2022-12-25 15-26-22.png

Nice, I love it!

 

Thanks for letting me know!

[ktran]

Just tested the new "experimental" ARM MacOS CLI via Siril. Yes, it's fast

[SilverLitz]

SN2 no longer appears in my PI Processes after I reinstalled the latest PI Win10 64bit.

 

I recently downloaded Bill Blanshan's ScreenStars script, and it worked nicely.  Later when I was trying BXT on a linear/mono image it hung up, and I terminated PI.  Afterward, ScreenStars script was no longer showed in PI Scripts.  I looked in the script folder and ScreenStars was still there, but would not show in PI.  I then tried uninstall/reinstall ScreenStars, and still no luck.  So, I reinstalled PI, and afterwards ScreenStars again was there, but SN2 process was gone.

 

Also, now BXT seems to run MUCH slower than before, 7.5 mins on QHY268 mono master, and does not use my RTX3070Ti (0% CUDA usage in Task Manager).

 

What is going on?

[pfile]

if you reinstall PI, all third party modules and scripts are deleted. this is because they reside in the pixinsight application folder, and that whole folder gets deleted when PI is reinstalled.

 

those 3rd party modules and scripts that have a repository will be replaced when PI checks for updates, and you apply those updates.

 

SN does not have a repository, so you need to install it by hand every time you re-install PI.

 

on windows it might be better to use "repair" if the installer gives you that option. to my knowlege that won't delete anything but cause built-in PI modules and scripts that already exists in the PI application folder to be overwritten with fresh copies.

[SilverLitz]

if you reinstall PI, all third party modules and scripts are deleted. this is because they reside in the pixinsight application folder, and that whole folder gets deleted when PI is reinstalled.

 

those 3rd party modules and scripts that have a repository will be replaced when PI checks for updates, and you apply those updates.

 

SN does not have a repository, so you need to install it by hand every time you re-install PI.

 

on windows it might be better to use "repair" if the installer gives you that option. to my knowlege that won't delete anything but cause built-in PI modules and scripts that already exists in the PI application folder to be overwritten with fresh copies.

Thanks!

 

All my 3rd party scripts were not effected by the reinstall, though it require the normal update after the initial opening of PI after reinstall.

 

Naively, I thought all Processes would come through with the PI install.  Apparently not all Processes come through natively.  BXT process came through without any thing extra, even though it is definitely an outside (extra paid for) process.

 

This version of PI 1.8.9.1 has been out since May of 2022, by far the longest time between updates I have seen, and I must have forgot about these things.

 

It seems that BXT needs SN2 (or SXT) for speed, or is it just my imagination that it is much slower without SN2 or SXT?

[pfile]

Thanks!

 

All my 3rd party scripts were not effected by the reinstall, though it require the normal update after the initial opening of PI after reinstall.

 

Naively, I thought all Processes would come through with the PI install.  Apparently not all Processes come through natively.  BXT process came through without any thing extra, even though it is definitely an outside (extra paid for) process.

 

This version of PI 1.8.9.1 has been out since May of 2022, by far the longest time between updates I have seen, and I must have forgot about these things.

 

It seems that BXT needs SN2 (or SXT) for speed, or is it just my imagination that it is much slower without SN2 or SXT?

well, this is what i was trying to say - BXT has a repository - if you put that in your list of repositories, then upon checking for updates, if there is an update, PI will download the update. the first thing PI does after a clean install is to automatically check for updates.  since no 3rd party modules/scripts are installed out-of-the-box, all of them get an "update" - and so that's how you got your scripts and the RC-Astro stuff back with no fuss.

 

anything else, notably SN and i guess the new denoise module, needs to be copied by hand into the pixinsight bin/ directory and then you need to manually ask PI to scan for new modules.

 

i think the RC-astro tools use TensorFlow on Windows. i'm not sure if the RC-astro tools also install CUDA support for tensorflow, but if they do, that will also dramatically speed up SN2 since it also uses tensorflow. i've actually not installed PI on a windows machine in several years so i don't know exactly how the RC-astro stuff works on windows. i do know that getting CUDA running on windows seems to be a bit of a hassle due to needing the exact right versions of stuff, so i am somewhat skeptical that the RC-astro tools can install all the right stuff from within PI's update framework.

 

rob

[SilverLitz]

I went to the website and got the Zip of the needed SN2 files and copied them to the PI \bin folder and installed the SN2 module.

 

Now, SN2 process shows up as expected, but upon running, it is no longer using my RTX3070Ti as there is no CUDA use and it took 7 min to run on a QHY268 color combined image.  Normally this would take ~50sec using the GPU.  I have not made any changes to my CUDA (believe it is 11.7) that I installed almost a year ago.

[pfile]

I went to the website and got the Zip of the needed SN2 files and copied them to the PI \bin folder and installed the SN2 module.

 

Now, SN2 process shows up as expected, but upon running, it is no longer using my RTX3070Ti as there is no CUDA use and it took 7 min to run on a QHY268 color combined image.  Normally this would take ~50sec using the GPU.  I have not made any changes to my CUDA (believe it is 11.7) that I installed almost a year ago.

 

you need a version of the tensorflow.dll that knows how to talk to CUDA. the one that ships with SN2 uses the CPU.

 

rob

Edited by pfile, 18 July 2023 - 04:58 PM.

[SilverLitz]

Rob, Thanks!

 

I previously had Cuda 11.7, and I followed a tutorial on Cuiv's YouTube, which referrences the following:

https://rikutalvio.b...sight-cuda.html

 

- Download the following 4 files:

NVIDIA CUDA Toolkit 11.8.0

NVIDIA cuDNN 8.7.0 for CUDA 11.x

TensorFlow 2.9.0

zLib DLL x64 1.2.3 (Right click > Save as...)

 

This tutorial was referencing CUDA 11.8, so that was what I used (though I 1st went to NVIDIA's website and downloaded 12.2).   

 

After this everything worked as expected, and seemed a little faster than 11.7, as a SN2 on QHY268M color combined image used only 39sec, using stride 256.  This sure beats 7-8 mins using my 24 Xeon Cores.

 

This only used 33% of my CUDA, while previously it seemed to use ~65%.  So it seems the 3070Ti has more than SN2 needs.

 

Since I also have CUDA 12.2 (the latest), would it be better to move CUDA 12.2 up higher in the Environmental variables?  (e.g. can CUDA 12.2 use TF 2.9.0?)

Edited by SilverLitz, 19 July 2023 - 12:58 PM.