Experimental TensorFlow GPU acceleration repository

lmclouth said:
It appears someone else had a similar problem that I had with the repository not installing all of the files.
B

Thread 'Cuda Core script not working after 1.9.0 update'

I have been using Pixlnsight for a while and always used the Cuda Core script to have my GPU (rtx3060) boost the performance of the RC Astro add-ons (blur, noise and starX) but since i have updated to 1.9.0, my GPU is running at 0% and the RC tools are seriously slow.

I have tried looking for updates as usual but none found. Anyone have any updates on this? Or a updated script?

Win11
Rtx3060 gpu
Ryzen 5 5600G cpu
128Gb 3600mhz RAM

I replied with what I think is the best solution of putting everything into another directory (including the tensorflow.dll) and adding the path to it. I've discussed this with Russ and he's considering making a script or similar to automate this process.

Lamar
Click to expand...
Russ also has the directions on his site for simply doing the CUDA toolkit installation, which is very easy and straightforward and creates a proper CUDA installation on your system, at which point all you have to concern yourself with in the future is keeping a copy of the CUDA enabled tensorflow DLL on hand to put back into the PI bin directory after updates. Personally I think this is the best approach to use.
 
Thank you Russ Croman for setting this up for us. Big fan of your tools. Really why I started using Pixinsight was to get access to BlurXterminator. Just updated from this repository and it worked perfectly the first time out. Tested StarXterminator before and after on a wide field, dense star, 320MB image. Prior to installing a GPU, StarXterminator took 8 minutes 32 seconds to process. After installing a GPU and updating from this repository, StarXterminator ran the same image in 29.8 seconds! Sweet!

System details below if it will help anyone:

Motherboard: TUF Z390
Processor: Intel i5 9600K 3.7GHz
Windows 10 PRO x64
GPU: MSI Geforce RTX 3060, 12GB
Pixinsight version: 1.9.1 Lockhart x64 build 1629

Darrell DuBose
 
Hi,

I installed 1.9.2 and since then I have a problem with CUDA.

When I launch BlurX* for example, Pixinsight closes with the message that it has stopped working.

I replaced my old tenserflow.dll thinking that it was perhaps no longer up to date. Now it does not crash anymore but it does not use the GPU at all...

Any idea what I can do to make it work again?

I uninstalled PI, the updates, reinstalled everything. I updated my GPU, I restarted.

Now I'm stuck
😅


Thanks for your help.

PS: It works perfectly fine with Gra*pert or Starne*

Motherboard: Asus Prime X570
Processor: AMD Ryzen 9 3900X 12-Core 3.80GHz
RAM: 32go
Windows 10 Pro x64
GPU: nvidia Geforce RTX 2080 SUPER 8go
Pixinsight 1.9.2
 
Last edited:
Aya19th said:
Hi,

I installed 1.9.2 and since then I have a problem with CUDA.

When I launch BlurX* for example, Pixinsight closes with the message that it has stopped working.

I replaced my old tenserflow.dll thinking that it was perhaps no longer up to date. Now it does not crash anymore but it does not use the GPU at all...

Any idea what I can do to make it work again?

I uninstalled PI, the updates, reinstalled everything. I updated my GPU, I restarted.

Now I'm stuck
😅


Thanks for your help.

PS: It works perfectly fine with Gra*pert or Starne*

Motherboard: Asus Prime X570
Processor: AMD Ryzen 9 3900X 12-Core 3.80GHz
RAM: 32go
Windows 10 Pro x64
GPU: nvidia Geforce RTX 2080 SUPER 8go
Pixinsight 1.9.2
Click to expand...
How did you replace the tensorflow DLL? The one that PI installs is CPU only, and you need to use the correct GPU version in order to work properly with the RC tools. Which you can get here: https://www.tensorflow.org/install/lang_c
 
cloudbait said:
How did you replace the tensorflow DLL? The one that PI installs is CPU only, and you need to use the correct GPU version in order to work properly with the RC tools. Which you can get here: https://www.tensorflow.org/install/lang_c
Click to expand...

Before updating PI, I copied my tenserflow.dll elsewhere (which was working fine) and pasted it again in the bin folder after the update.
And now neither the old way or the experimental way works. But only with the RC.

So, I thought my tensorflow was outdated and I follow the link on RC and yours too, to download a recent one. But it didn't change anything. So basically, now, nothing works with cuda regarding the RC tools :cry:

I thought of downloading the same tenserflow I had when I installed cuda years ago. It was 2.3 I think with cuda 10.1. But I can't find it and I don't know if it will change something.
 
Aya19th said:
Before updating PI, I copied my tenserflow.dll elsewhere (which was working fine) and pasted it again in the bin folder after the update.
And now neither the old way or the experimental way works. But only with the RC.

So, I thought my tensorflow was outdated and I follow the link on RC and yours too, to download a recent one. But it didn't change anything. So basically, now, nothing works with cuda regarding the RC tools :cry:

I thought of downloading the same tenserflow I had when I installed cuda years ago. It was 2.3 I think with cuda 10.1. But I can't find it and I don't know if it will change something.
Click to expand...
I'm not sure what happens when you mix-and-match a proper CUDA installation and Russ's tool. They put files in different places, and might interact badly. I'd probably clean things out, make sure there are no CUDA files installed in PI folders (which you could do with a re-install of PI), then confirm (and repair as necessary) the regular CUDA installation, as described on the RC site.
 
cloudbait said:
I'm not sure what happens when you mix-and-match a proper CUDA installation and Russ's tool. They put files in different places, and might interact badly. I'd probably clean things out, make sure there are no CUDA files installed in PI folders (which you could do with a re-install of PI), then confirm (and repair as necessary) the regular CUDA installation, as described on the RC site.
Click to expand...
Okay, it works. I deleted everything about CUDA and made a new install with 11.8.
I don't know what was the problem, if it is a compatibilyty issue with PI 1.9 or if the files interacted in a bad way (it was not the case before the update). But it works!

Just one question: I thought the experimental method was to not have to replace the tenserflow when major update. So it means you already have cuda on, no?
 
Aya19th said:
Okay, it works. I deleted everything about CUDA and made a new install with 11.8.
I don't know what was the problem, if it is a compatibilyty issue with PI 1.9 or if the files interacted in a bad way (it was not the case before the update). But it works!

Just one question: I thought the experimental method was to not have to replace the tenserflow when major update. So it means you already have cuda on, no?
Click to expand...
I gather that Russ's utility doesn't do a normal CUDA installation, but populates the PI bin folder with the necessary CUDA files that would normally live under Program Files. And that does appear to work... but I can see how it might be an issue that also has a normal CUDA toolkit installation.
 
Straightforward & successful implementation:
System: Win 10 CPU Ryzen 9 3900x 12-Core Ram 64GB Nvidia Gforce GTX 1060 6gb
Previous manual installation of TensorFlow.
Before upgrade: PixInsight Core 1.8.9-3 File : 281,440kb Blurxterminator 43.596 s
Upgrade to 1.9.2 Lockhart & updated repositories no CUDA update Blurxterminator 03min :02.72 !
Add https://www.rc-astro.com/TensorFlow/PixInsight/GPU and update 1.484 gb
Blurxterminator 41.477 s a massive improvement of nearly 2 secs
No issues all went like a charm
 
Thank you for doing this, Russ. It worked perfectly for me!

Motherboard: MSI MAG B650 Tomahawk
Processor: AMD Ryzen 9 7900X 12-Core 4.70GHz
RAM: 64gb
Windows 11 Home x64
GPU: nvidia Geforce RTX 3050 6gb
Pixinsight 1.9.2
 
cloudbait said:
I gather that Russ's utility doesn't do a normal CUDA installation, but populates the PI bin folder with the necessary CUDA files that would normally live under Program Files. And that does appear to work... but I can see how it might be an issue that also has a normal CUDA toolkit installation.
Click to expand...
I have installed CUDA on a kubuntu system some time ago, updated to luvuntu 24.04 and currently use RC Astro’s tools on PI Lockhart so it seems compatible with a Linux based CUDA toolkit installation.
 
rcroman said:
I finally got around to preparing a repository containing the CUDA/cuDNN software libraries needed to enable GPU acceleration of AI-based tools. This didn't used to be possible due to onerous license restrictions, but those have since been relaxed a bit.

This is for Windows only at the moment: if this goes well I'll work on the Linux version. This isn't needed for MacOS users – the "CoreML" library provided by Apple is used, so Mac users with capable hardware get GPU acceleration of RC Astro tools out of the box.

THIS IS EXPERIMENTAL – please READ this post completely, and proceed at your own risk. I can only test on a small number of hardware configs. I recommend backing up your PixInsight installation so you can roll back if it goes sideways. At the very least make a copy of the existing tensorflow.dll file in PI's bin directory – putting that back and restarting PI should effectively revert.

There is a second repository to revert to CPU-only operation if something goes wrong. In the worst case, you may have to re-install PixInsight. Please roll with this if it happens, and provide kind feedback so I can try to fix it.

You need a capable NVIDIA GPU. This means one with ≥ 2GB of RAM and compute capability ≥ 3.5. Check this NVIDIA page if in doubt about your GPU's capabilities. You'll need at least several GB of free disk space.

The GPU repository installs version 2.10 of the "GPU-only" version of tensorflow.dll. On the TensorFlow project page, it clearly says "GPU only," but in my testing it ran in CPU-only mode just fine if there was no GPU installed. I have no idea if this will be the case on every machine. Along with the TensorFlow library, the necessary CUDA/cuDNN libraries (version 11.8) and the zlib compression library (version 1.2.3.0) are installed.

The CPU repository installs version 2.10 of the CPU-only build of tensorflow.dll.

These libraries are all the property of the respective organizations that created them. They are distributed via these repositories for use with PixInsight/RC Astro tools only in accordance with relevant license agreements. You can find these agreements in PI's etc/legal/licenses directory after installation.

To proceed, add the following to your PI repository list using Resources -> Updates -> Manage Repositories:


Then run Resources -> Updates -> Check for Updates, proceed with the download (be patient – it's a 1.5GB package), and then quit PI to complete the update. You should now enjoy GPU acceleration of RC Astro tools. This should "just work." No environment variables to set unless you want to constrain TensorFlow to only use as much GPU memory as needed (see the notes about TF_FORCE_GPU_ALLOW_GROWTH here). I've had limited success getting that setting to actually work – TensorFlow is very GPU-memory-grabby.

If it doesn't work – running BXT/NXT/SXT causes a crash or doesn't use the GPU – let me know your hardware details, PI version, OS version, etc., and the full text of any error messages. If you manage to get things working where others have not, please share.

You can revert to the CPU-only tensorflow.dll by changing the repository entry to


and going through the update process again. If you want to revert yet again to the GPU version, you'll need to change the repo entry again AND do a Resources -> Updates -> Reset Updates.

Thanks for testing... let me know how it goes!
Click to expand...
I added your GPU repository and everything seems to work fine including BTX and NTX.
My system is:
Windows 11 Pro
Nvidia GeForce GTX 1660 Super graphics card.
Intel i9-10900K CPU
ASUS ROG Maximus Xiii Hero motherboard
 
I am new to PI with a week into the trial licence and have been trying to follow Pete Zelinka "PI for Dummies". I am solely Ubuntu and when using BlurXT get a message about MLLoad() and Graphdef binaries not being consistent!! I'm using a Dell 7620 with an RTX-3060 GPU and after a week of failed TensorFlow installing problems I'm wondering if I have trashed the OS now. Firstly, where/how do I check the Graphdef binaries and are the stored order of repositories in PI a contributing factor?
 
Hi Cropduster
I am in the same boat, having spent hours on getting the GPU to run. In the end, I think it would be great if they. came out with an automated repository install for this type of thing, like they do for Windows/RC Astro plugins. The time would have been better spent learning PI for me.

Clear Skies,
Paul
 
Last edited:
promero said:
Hi Cropduster
I am in the same boat, having spent hours on getting the GPU to run. In the end, I think it would be great if they. came out with an automated repository install for this type of thing, like they do for Windows/RC Astro plugins. The time would have been better spent learning PU for me.

Clear Skies,
Paul
Click to expand...
Who is "they"? PixInsight doesn't use GPU processing, so it doesn't attempt to install anything. And shouldn't. And the RC tools don't depend upon it, so it's always been the responsibility of users to install the CUDA tools if they wanted the speed advantage. Really, a minor "hack". Russ figured out a way to automate this for his third party tools, and only with Windows so far.
 
Cloudbait, I dont use the RC Astro tools, and I am not familiar with them. My apologies for my ignorance on the subject. I assumed the tools were using GPU acceleration. I meant RC Astro by they.

Hope that helps,

Paul
 
promero said:
Cloudbait, I dont use the RC Astro tools, and I am not familiar with them. My apologies for my ignorance on the subject. I assumed the tools were using GPU acceleration. I meant RC Astro by they.

Hope that helps,

Paul
Click to expand...
PI does not use GPU acceleration at all. There is no reason to install it (which requires hacking the PI installation) unless you intend to run the RC tools.
 
Ok, so if the repository https://www.rc-astro.com/TensorFlow/PixInsight/CPU/ (or /GPU) uses Tensorflow then why do I get this error msg:

*** Error: ERROR: MLLoad() could not load tensorflow graph. Error 3: Converting GraphDef to Graph has failed. The binary trying to import the GraphDef was built when GraphDef version was 440. The GraphDef was produced by a binary built when GraphDef version was 1645. The difference between these versions is larger than TensorFlow's forward compatibility guarantee. The following error might be due to the binary trying to import the GraphDef being too old: NodeDef mentions attr 'explicit_paddings' not in Op output:T; attr=T:type,default=DT_FLOAT,allowed=[DT_HALF, DT_BFLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT32, DT_INT64, DT_UINT8, DT_INT16, DT_INT8, DT_UINT16, DT_QINT8]; attr=ksize:list(int),min=4; attr=strides:list(int),min=4; attr=padding:string,allowed=["SAME", "VALID"]; attr=data_format:string,default="NHWC",allowed=["NHWC", "NCHW", "NCHW_VECT_C"]>; NodeDef: {{node BXT/max_pooling2d_14/MaxPool}}. (Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary.).

Reading swap files...
3193.959 MiB/s
<* failed *>
 
You must log in or register to reply here.
Back
Top