Running GraXpert from within PixInsight: not sure exactly what it's doing?

[BrendanC]

Hi all,

My latest obsession - trying to figure out exactly what's going on when I use the new method of running GraXpert from within PixInsight.

Quick recap: GraXpert is a very nice gradient removal package that has, until recently, been a standalone product, not running within PI. So, to use it, you had to save your image, exit PI, open GraXpert, process it, save it, go back into PI, open it. Recently, Jürgen Terpe developed quite a neat way of running it within PI, as part of his Toolbox script set. It's not a native PI module or anything, just a script that calls the GraXpert executable and passes information to it and back into PI (I think).

If you've used it, you'll know that it does seem to work - load up your image, run the script, it hops off over to GraXpert, comes back with a gradient-free image. All good. I thought so too, but on looking more closely, I can't quite figure out exactly what it's doing, and this bothers me a little.

So, this is what I get when I do a very rough import of my Ghost Nebula linear image into GraXpert, by manually opening GraXpert, loading the linear image, processing it using the RBF method, and saving the stretched version (it seems to show what's going on better than stretching within PI):

Nice. This is what I get when I do the same thing, but using the newer AI method. As you would expect, it's slightly different (it's not immediately apparent - save them and flick between them using Photoshop or some such thing and you'll see the slight differences). Not necessarily better, in this case, although I've tried it on emission nebulae and it does seem to preserve more of the nebulosity:

Nice. Again. So, the new third option: running GraXpert from within PI. All was well and good until I noticed this in the process console:

2023-12-13 17:11:25,531 MainProcess absl INFO Fingerprint not found.
2023-12-13 17:11:25,713 MainProcess tensorflow WARNING No training configuration found in save file, so the model was *not* compiled.
statechange: 0
finished:0 / 0
Compile it manually.Saved model loading will continue.1.0.1. you can change this by providing the argument '-ai_version'

 

"Bit odd, that," I thought. I looked around for other people who have reviewed this, and noticed that Cuiv the Lazy Geek shows his process console while demoing it, and it appears on his console too (see https://youtu.be/KBR2xsZ-NmM, but blink and you'll miss it - screenshot below with relevant part of process console highlighted):

 

It looks to me like it's failing to find an AI model. So, is it defaulting to RBF? Or actually using AI? Let's take a look :  in this version, I've use the new script in PI to call GraXpert and remove the gradient, but then saved it as a linear image in PI, then opened it in GraXpert, but not processed it, just stretched and saved, which means I get the same stretch as the other two images:

It doesn't seem to match either the RBF or the AI versions. As with the RBF vs AI versions, the difference is subtle, and you have to flick between them to see this. You could argue it's so subtle as not to matter, but I do prefer to know what's going on with these things. I'd hate to find out, a year or so down the line, that I've been processing my images not to the best of my ability, simply because I didn't ask.

So, I'm not sure what it's doing, exactly. Is it running RBF, or AI, or something else? Does it default to the latest model used when run outside of PI? If so, why is the image generated by using the script to call GraXpert, different from both of the images created using GraXpert directly? Can I control which version it's running, because I'd like to force it to use the AI version via the '-ai version' argument, but I don't know where to specify that? Or, should I just get a life?

Any takers?

Thanks
Brendan

[Stuart1971]

You are correct, it should show the AI version it is using in the console window, which is 1.0.1 I think,  mine does, and I think this is an ongoing issue that some people are having, look on the Dark Matters discord group under GraXpert and then support for more info….

[BrendanC]

I'm correct? Wow. First time for everything. Thanks, I think I'll just use GraXpert manually for the time being. I've dipped my toe in Discord in the past and it's just noise.

Thanks anyway.

[Stuart1971]

1 hour ago, BrendanC said:

I'm correct? Wow. First time for everything. Thanks, I think I'll just use GraXpert manually for the time being. I've dipped my toe in Discord in the past and it's just noise.

Thanks anyway.

If you look at the top of the console in the image you posted, you will see it shows the AI version being used in Cuiv’s video, so he was using the AI version…

see below with yellow underline

 

[BrendanC]

Well so it does! Never spotted that. And I just took another look at mine, and mine says so too. In which case, why the messages about no training configuration found, or the model not being compiled, and why is the resulting output different when driven by the script? Curiouser and curiouser.

[Stuart1971]

6 minutes ago, BrendanC said:

Well so it does! Never spotted that. And I just took another look at mine, and mine says so too. In which case, why the messages about no training configuration found, or the model not being compiled, and why is the resulting output different when driven by the script? Curiouser and curiouser.

I have asked on the discord, will report back….👍🏻