No luck with StarNet++ on RTX 3070.

mat_blue said:
I got it working on Windows 10 for a 3060 with tensorflow for GPU version 2.9.0, CUDA 11.2 and CUDNN 8.1.1.33. You can get tensorflow from https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-gpu-windows-x86_64-2.9.0.zip
The CUDA and CUDNN files from the nvidia developers site. You need version 460.89 or later of the nvidia device driver. Follow the instructions In the Starnet2 install to set the environment variables.
This setup should work on a 3080.
Click to expand...
So this worked, kind of. It is using the GPU and is like 1000% faster, but only hitting 30% on CUDA in task manager. Any ideas on how to increase that?
 
I only get 30% CUDA usage with my 3060 with a stride of 256 but it only takes 12 seconds. With a stride of 128, it uses 60% of CUDA and takes 22 seconds. Somebody else said they were only getting 60% usage with a 3090 (which would be a lot more CUDA cores than 60% on my 3060).
 
Ar3s said:
Would you be able to fill me in on how you got it working? I've got a 3080 and have had 0 luck on getting this to work on 1.8.9. I thought the tensorflow files from darkskies were outdated.
Click to expand...

just followed the updated darkarchon instructions, being careful to download the latest versions, as indicated in this thread above
 
Looks like tensorflow-gpu 2.10.0 is the end of the line for now:

" Caution: TensorFlow 2.10 was the last TensorFlow release that supported GPU on native-Windows. Starting with TensorFlow 2.11, you will need to install TensorFlow in WSL2, or install tensorflow-cpu and, optionally, try the TensorFlow-DirectML-Plugin "

As a reminder, here's the link to the files I posted earlier in this thread.

No luck with StarNet++ on RTX 3070.

Thanks a lot guys. Changed from a 20 Series card to a 3070TI Worked with: libtensorflow-gpu 2.4.0 CUDA 11.0 cudnn v8.1.1.33 Fantastic! Starnet still works with CUDA, however, recently I encountered a bug. As soon as Starnet has done its thing and I maximize any open window, PI freezes and...
pixinsight.com
 
Hello,
for me works since yesterday with 3070 TI :
- tensorflow for GPU version 2.4.0 (higher Versions do not work)
- CUDA 11.0.2 and
- CUDNN 8.1.1.33

CS
Chris

Edit:
i played a little bit with the different versions;)
for me (RTX 3070 TI) works:
- CUDA 11.2
- CUDNN 8.1.1.33
than i tried dif. versions of tensorflow.dll
2.4 not works, 2.10 is the slowest (the amount of dedicaded RAM Usage was smaller, 5,4 GB vs 6,0 GB in the other Versions)
here are the times that i tested 5 times at the same Picture:
(for me v 2.8 was the fastest)


2.10 (5,4GB)
25
17
17
18
18,5

2.9 (5,9GB)
26
15
14
14
15

2.8 (5,9GB)
25
13,7
13,5
13,5
13,7

2.7 (5,9GB)
25
14
14
14
14

2.6 (5,9GB)
24
13
14
15
14

2.5 (5,9GB)
25
14
14
14
14
 
Last edited:
Inspired by the previous post, I did some tests with my RTX3080 and StarXTerminator 2.0.5 AI v11 large overlap. Running the recommended cuDNN 8.1 and CUDA 11.2. It's clear how much slower the 2.10.0 version is, even though it shows slightly higher average load. On my setup 2.9.0 was the fastest on all runs, but we're talking 0.3s max, so I rounded everything to full seconds. Results from multiple runs didn't show any major variation excluding the first one (caching). Only one image tested, though.

versiontimeload
2.10.059s92%
2.9.047s90%
2.8.047s90%
2.7.047s90%
2.6.047s90%
2.5.047s90%

I also ran the same test with the CPU version of 2.9.0; it took over 22 minutes on a 5950X. 😬
 
Last edited:
Tried some newer CUDA and cuDNN versions, and got a hefty speed boost; the same test from the post above completed in only 39s!

CUDA 11.8 https://developer.download.nvidia.c...rk_installers/cuda_11.8.0_windows_network.exe
cuDNN 8.7 https://developer.nvidia.com/downloads/c118-cudnn-windows-8664-87084cuda11-archivezip

With the newer versions you also need ZLIB!
http://www.winimage.com/zLibDll/zlib123dllx64.zip (right click -> save as)
From that archive extract zlibwapi.dll into ...Toolkit\CUDA\v11.8\bin.

Tensorflow 2.9.0
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-gpu-windows-x86_64-2.9.0.zip

Use darkarchon's instructions to install:
https://darkskies.space/pixinsight-starnet-cuda/
Just adjust for the new version numbers, and add the ZLIB!


This will also work on RTX 4000 series cards.
 
Last edited:
RT-- said:
Tried some newer CUDA and cuDNN versions, and got a hefty speed boost; the same test from the post above completed in only 39s!

CUDA 11.8 https://developer.download.nvidia.c...rk_installers/cuda_11.8.0_windows_network.exe
cuDNN 8.7 https://developer.nvidia.com/downloads/c118-cudnn-windows-8664-87084cuda11-archivezip

With the newer versions you also need ZLIB!
http://www.winimage.com/zLibDll/zlib123dllx64.zip
From that archive extract zlibwapi.dll into ...Toolkit\CUDA\v11.8\bin.

Tensorflow 2.9.0
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-gpu-windows-x86_64-2.9.0.zip

Use darkarchon's instructions to install:
https://darkskies.space/pixinsight-starnet-cuda/
Just adjust for the new version numbers.


This will also work on RTX 4000 series cards.
Click to expand...
I tried the above. GPU went up to 100% usage then abruptly crashed PI.
 
I had the same problem as Moorefam, it performed the first setup step, but when it went to actually run the AI (when it shows % processing) it crashed PI.

FIX: I then double checked and found the zlibwapi.dll copied into \CUDA\v11.8\bin didn't take (or I missed) once I copied that DLL over it worked.

for others, here's my performance comparison, both on v11.8
  • 48 seconds (22sec setup, 26sec AI processing) cudnn v8.2.4.15 & libtensorflow v2.7 - this was my current setup
  • 27 seconds (13sec setup, 14sec AI processing) cudnn v8.7.0.84 & libtensorflow v2.9 - tried this based on RT-- & Zorion's posts
So that's about double speed.

For reference these times are on a 26MP image (ASI2600MC) and I run a RTX4090.
 
Last edited:
kring said:
I had the same problem as Moorefam, it performed the first setup step, but when it went to actually run the AI (when it shows % processing) it crashed PI.

FIX: I then double checked and found the zlibwapi.dll copied into \CUDA\v11.8\bin didn't take (or I missed) once I copied that DLL over it worked.

for others, here's my performance comparison, both on v11.8
  • 48 seconds (22sec setup, 26sec AI processing) cudnn v8.2.4.15 & libtensorflow v2.7 - this was my current setup
  • 27 seconds (13sec setup, 14sec AI processing) cudnn v8.7.0.84 & libtensorflow v2.9 - tried this based on RT-- & Zorion's posts
So that's about double speed.

For reference these times are on a 26MP image (ASI2600MC) and I run a RTX4090.
Click to expand...
Thanks kring I had the same issue and your post saved me ! :)
 
mxcoppell said:
Folks,

Anybody got experience how to make RTX 3070 Cuda working with StarNet++?

I've followed darkarchon's tutorial: https://darkskies.space/pixinsight-starnet-cuda/. When I execute StarNet, it stuck at:

"Restoring neural network checkpoint: C:/Program Files/PixInsight/library/mono_starnet_weights.pb"

Tried to figure if the Cuda libraries and Cudnn libraries are too old for RTX 3070. Tried to deploy the following packages:
  1. cuda_11.1.1_win10_network
  2. cudnn-11.1-windows-x64-v8.0.5.39
  3. libtensorflow-gpu-windows-x86_64-2.4.0-rc2

Now the "Restoring neural network checkpoint:..." hanging is gone. But it doesn't look like StarNet++ is using Cuda. It's still running on the CPU.
** I've also tried libtensorflow-gpu-windows-x86_64-2.3.1 (built on 10/02/2020). Still only on CPU.

My graphics card is ASUS Dual NVIDIA GeForce RTX 3070 OC Edition Gaming Graphics Card.

Thanks in advance for any suggestions!

-Min
Click to expand...

GPU Acceleration for StarNet++ with CUDA | William Li Photos

A quick tutorial on how to speed up the StarNet neural net using CUDA GPU Acceleration. Compatible with StarNet V2 and PixInsight 1.8.8-12.
www.williamliphotos.com www.williamliphotos.com

Try this tutorial, it worked for me after a couple of tries, you must follow it exactly though, miss any steps and it will not work for you
 

GPU Acceleration for StarNet++ with CUDA | William Li Photos​

A quick tutorial on how to speed up the StarNet neural net using CUDA GPU Acceleration. Compatible with StarNet V2 and PixInsight 1.8.8-12.
www.williamliphotos.com
www.williamliphotos.com

Try this tutorial, it worked for me after a couple of tries, you must follow it exactly though, miss any steps and it will not work for you. You can use the latest 12.*v for the Cuda files etc. But make sure all the corresponding files are the same versions and you set them in the system variable's as well.
 
dpatterson said:

GPU Acceleration for StarNet++ with CUDA | William Li Photos

A quick tutorial on how to speed up the StarNet neural net using CUDA GPU Acceleration. Compatible with StarNet V2 and PixInsight 1.8.8-12.
www.williamliphotos.com www.williamliphotos.com

Try this tutorial, it worked for me after a couple of tries, you must follow it exactly though, miss any steps and it will not work for you
Click to expand...
dpatterson said:

GPU Acceleration for StarNet++ with CUDA | William Li Photos​

A quick tutorial on how to speed up the StarNet neural net using CUDA GPU Acceleration. Compatible with StarNet V2 and PixInsight 1.8.8-12.
www.williamliphotos.com
www.williamliphotos.com

Try this tutorial, it worked for me after a couple of tries, you must follow it exactly though, miss any steps and it will not work for you. You can use the latest 12.*v for the Cuda files etc. But make sure all the corresponding files are the same versions and you set them in the system variable's as well.
Click to expand...

This thread is over two years old, and the issue has been solved many times over; no need to reply to the original poster.
Post #49 contains the necessary links to get going.

That tutorial you linked, is overly complicated, and wants you to install unnecessary stuff. If it took you more than one try to make it work, is it even worth advertising? Also, you contradict yourself by saying "You can use the latest 12.*v for the Cuda files etc. But make sure all the corresponding files are the same versions." There is no cuDNN for CUDA 12 yet, so it is not advisable to use it. Many people in this thread have been blindly mixing and matching different versions of the required software, and causing themselves unnecessary issues.
 
RT-- said:
Did you copy the ZLIB dll?
GPU drivers up-to-date?
Click to expand...
Yes I did and it still crashed. The only thing that isn't quite clear to me at least is whether to copy the relevant contents of the bin and lib files from the cudnn-windows-X86.......cuda11-archive dir into the C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\11.8 bin and lib directories or copy the bin and lib directories across intact and if so where? When I start PI and BXT it states that v 1.1.1 is running, then says initializing, thinks about it for about 20 secs then PI crashes and vanishes. I have tried for a few hrs now and it's a shame as if I could speed up the 11 mins BXT takes to run it might save me thinking about an expensive and time consuming rebuild. GPU drivers are bang up to date.
 
Hi,sorry if this is out of sync but after a lot of problems I managed to download starnet++ v2 but both that and the original starnet do not take out all of the stars. I can’t afford to buy starxterminator at the moment so Illhave to put up with anything other than a starless image, unless someone could help.
 
RT-- said:
Tried some newer CUDA and cuDNN versions, and got a hefty speed boost; the same test from the post above completed in only 39s!

CUDA 11.8 https://developer.download.nvidia.c...rk_installers/cuda_11.8.0_windows_network.exe
cuDNN 8.7 https://developer.nvidia.com/downloads/c118-cudnn-windows-8664-87084cuda11-archivezip

With the newer versions you also need ZLIB!
http://www.winimage.com/zLibDll/zlib123dllx64.zip
From that archive extract zlibwapi.dll into ...Toolkit\CUDA\v11.8\bin.

Tensorflow 2.9.0
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-gpu-windows-x86_64-2.9.0.zip

Use darkarchon's instructions to install:
https://darkskies.space/pixinsight-starnet-cuda/
Just adjust for the new version numbers, and add the ZLIB!


This will also work on RTX 4000 series cards.
Click to expand...
Thanks for this! I followed the DarkArchon instructions and your file links and got it working on the first try. I used an RTX3050 and with my IMX455 subs got 58 seconds on BXT and 1:19 on SXT. That's between 5 and 10X faster. Worth the $339 GPU upgrade and the effort!
 
You must log in or register to reply here.
Back
Top