Jump to content

  •  

CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.

Photo

StarNet v2 help requests

  • Please log in to reply
225 replies to this topic

#26 rdr

rdr

    Lift Off

  • -----
  • Posts: 19
  • Joined: 02 Sep 2007

Posted 28 March 2022 - 06:09 PM

By the way, I have rebooted PixInsight and the MAC itself, but that does not help



#27 johnpane

johnpane

    Ranger 4

  • *****
  • Posts: 351
  • Joined: 13 Dec 2014
  • Loc: Wexford, PA, USA

Posted 28 March 2022 - 06:24 PM

I am running PixInsight on MAC OS BigSur (11.2.3) on a MacBook Pro with M1 chip. My PixInsight is 1.8.9 Ripley build 1544 released 2022-03-13.  I installed the PixInsight today.

 

I downloaded Starnet and have placed the contents of the folders in the PixInsight/bin folder.

 

When I go to Modules/Install Modules.  the FIRST TIME PixInsight informs me that StarNet2-pxm.dylib cannot be opened because it is from an unrecognized vendor.  So I go the the MAC OS System Preferences and authorize access to StarNet2-pxm.dylib.   (this is routine for vendors not recognized by apple).

 

But subsequent attempts to install through the Modules/Install process fail to find Starnet.  Search returns no results.  Browse does "see" the StarNet2-pxm.dylib, but it is greyed out.

 

I don't seem to be able to get thru MAC securty.  How do I do this?

I was able to get it installed by using the script in this post. In addition to the appropriate block of commands for the M1 Mac, you will also have to run the line of code mentioned at the end.


Edited by johnpane, 28 March 2022 - 06:25 PM.


#28 Palmito

Palmito

    Vostok 1

  • -----
  • Vendors
  • Posts: 181
  • Joined: 16 Nov 2018
  • Loc: Lausanne, Switzerland

Posted 29 March 2022 - 11:58 AM

Good question! The only way to be sure is to try and run it) However, tensorflow should be able to run on multiple GPUs even without SLI, so there are no fundamental problems here!

So I just ran a test using a GTX1080 and a GTX680.

Unfortunately tensorflow runs only on the 1080 (confirmed through GPU RAM usage as well as measuring average duration on baseline tests).

 

It seems to me, reading tensorflow api documentation, that there are specific parallelism strategies to setup, tf.distribute.Strategy (https://www.tensorfl...g_multiple_gpus).

Else it will only use the GPU with the lowest id.

 

Is it something you might implement in the near future? Can I help?

It seems fairly straight forward and convenient to develop as virtual GPUs can be configured for development and testing purposes with one physical gpu.

 

I am asking because I am currently looking into buying a new GPU, or two lower-end GPUs if applicable :-D

 

--edit: I just read the GPUs should be of the same series for this to work, so my tests might be invalid...


Edited by Palmito, 29 March 2022 - 04:32 PM.


#29 codydorris1

codydorris1

    Lift Off

  • -----
  • Posts: 8
  • Joined: 04 Mar 2022

Posted 10 April 2022 - 03:57 PM

Help! I have the newest version of Pixinsight (1.8.9) on windows. I download starnet v2 and it worked once, now the icon is completely gone. The only thing there is the v1. I don't get it. It shows the Pixinsight module is downloaded. The same thing happened a few weeks ago with the standalone application. Worked once then never again.



#30 Xylophilist

Xylophilist

    Lift Off

  • -----
  • Posts: 1
  • Joined: 16 Apr 2022

Posted 16 April 2022 - 06:45 PM

hi, nope; I've havent installed v1 on the linux version of Pix. I have only v1 and v2 installed on windows.

 

Here is a screenshot of my Pix folder

attachicon.gif2022-03-17_20-26.png

Hiya,

 

IME the StarNet++ v2 PI plugin worked out of the box on Windows, but the Linux build is clearly broken.

In /opt/PixInsight/bin/lib/ both the libtensorflow and libtensorflow_framework libraries have come out as text files with their own filenames in; they should be symlinks to the real libraries.

 

Consider:

cd /opt/PixInsight/bin/lib/
ln -sf libtensorflow.so.2.6.0 libtensorflow.so
ln -sf libtensorflow.so.2.6.0 libtensorflow.so.2
ln -sf libtensorflow_framework.so.2.6 libtensorflow_framework.so
ln -sf libtensorflow_framework.so.2.6 libtensorflow_framework.so.2

HTH


  • dswtan likes this

#31 alvise

alvise

    Lift Off

  • -----
  • Posts: 5
  • Joined: 26 Mar 2021

Posted 25 May 2022 - 11:10 AM

Hello, everyone !
Please redirect me if this is the wrong thread in which post support/bugfix requests...

I've been using StarNet2 since the release of V2 and it just works well! Thanks Nikita !
 

There is only one thing I would improve (or maybe is a small bug)... I am pasting my message from the PixInsight forum (in which they suggested to come here and ask the same thing here):
-----------

Quoting:
[...]
One of them is how to save Process Icons permanently to the workspace, then quickly drag and drop the Icons themselves onto my images (speeding up the GUI interaction process or even calling them by JavaScripts).

Normally if I want a customization of the parameters of a Process Icon, I first set the values (parameters that require a value number or boxes that need to be checked) and then drag the little triangle onto the workspace. Next time, that Process Icon will run with the parameters I had set.
 

This works for several processes (LRGBCombination, ChannelExtraction, etc.), except for StarNet2. Specifically, if I check the "Linear" box, this setting remains throughout the life of the PI program in the Icon created by dragging the usual small triangle on the workspace, but at the next restart of PixInsight, if I double click my custom StarNet2 Icon, the box "Linear" will not be checked anymore (so StarNet2 will not work well on a linear image by just dragging the Icon onto the image, unless I double click that Icon and check the box, which is what I am trying to automatize, especially if I have to do on many images or if I want to do from a script). I checked the XML file for my workspace Icon sets and in fact there is no trace of some "linear" parameter saved persistently.
 

I don't understand where I'm going wrong or if there's something I'm missing in this Process Icons saving procedure. But my usual procedure works for other Processes...

So it sounds like a small bug maybe easy to fix that would improve quite a lot my automatization of workflows.

 

END QUOTING.

-----------

 

Thank you in advance,

Alvise


  • nekitmm likes this

#32 nekitmm

nekitmm

    Vendor - StarNet Software

  • *****
  • Vendors
  • topic starter
  • Posts: 227
  • Joined: 25 Apr 2018

Posted 27 May 2022 - 02:46 PM

Hello, everyone !
Please redirect me if this is the wrong thread in which post support/bugfix requests...

I've been using StarNet2 since the release of V2 and it just works well! Thanks Nikita !
 

There is only one thing I would improve (or maybe is a small bug)... I am pasting my message from the PixInsight forum (in which they suggested to come here and ask the same thing here):
-----------

Quoting:
[...]
One of them is how to save Process Icons permanently to the workspace, then quickly drag and drop the Icons themselves onto my images (speeding up the GUI interaction process or even calling them by JavaScripts).

Normally if I want a customization of the parameters of a Process Icon, I first set the values (parameters that require a value number or boxes that need to be checked) and then drag the little triangle onto the workspace. Next time, that Process Icon will run with the parameters I had set.
 

This works for several processes (LRGBCombination, ChannelExtraction, etc.), except for StarNet2. Specifically, if I check the "Linear" box, this setting remains throughout the life of the PI program in the Icon created by dragging the usual small triangle on the workspace, but at the next restart of PixInsight, if I double click my custom StarNet2 Icon, the box "Linear" will not be checked anymore (so StarNet2 will not work well on a linear image by just dragging the Icon onto the image, unless I double click that Icon and check the box, which is what I am trying to automatize, especially if I have to do on many images or if I want to do from a script). I checked the XML file for my workspace Icon sets and in fact there is no trace of some "linear" parameter saved persistently.
 

I don't understand where I'm going wrong or if there's something I'm missing in this Process Icons saving procedure. But my usual procedure works for other Processes...

So it sounds like a small bug maybe easy to fix that would improve quite a lot my automatization of workflows.

 

END QUOTING.

-----------

 

Thank you in advance,

Alvise

Got it, I will look into this. I might add this functionality, but to be honest I am leaning towards just having star mask and linear checkboxes checked by default and maybe removing those checkboxes from the GUI! I don't see any reason to keep any one of those un-checked...


  • dswtan and alvise like this

#33 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 27 May 2022 - 03:18 PM

there are still cases where you might want to run starnet on an already processed image... please don't remove the linear control checkbox. i don't know how i feel about the star mask control.

 

sounds like there is just some code missing from the module to set those variables and UI state based on the invocation from the process icon?

 

rob



#34 kvastronomer

kvastronomer

    Mariner 2

  • -----
  • Posts: 246
  • Joined: 11 Nov 2020

Posted 27 May 2022 - 03:52 PM

Got it, I will look into this. I might add this functionality, but to be honest I am leaning towards just having star mask and linear checkboxes checked by default and maybe removing those checkboxes from the GUI! I don't see any reason to keep any one of those un-checked...

If you default both parameters as stated by you above we would be able to use V2 within EZ Processing Suite which is supposed to be used on linear image.



#35 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 27 May 2022 - 07:46 PM

If you default both parameters as stated by you above we would be able to use V2 within EZ Processing Suite which is supposed to be used on linear image.

i dont know if that is correct because since starnetV1 could not operate on linear images, the EZ script must do a reversable stretch and then call starnet, then reverse the stretch.

 

you'd have to remove the code in EZ that stretches and unstretches the image if you called starnetV2 in linear mode.

 

rob



#36 alvise

alvise

    Lift Off

  • -----
  • Posts: 5
  • Joined: 26 Mar 2021

Posted 28 May 2022 - 05:17 AM

Got it, I will look into this. I might add this functionality, but to be honest I am leaning towards just having star mask and linear checkboxes checked by default and maybe removing those checkboxes from the GUI! I don't see any reason to keep any one of those un-checked...

Thank you nekitmm for looking into it ! Really appreciated.

 

As someone said, I would also not remove that option, but just make its default to "true"(checked) would be great (at least in my case I remove stars in linear stage to avoid blowing the stars 90% of the times, and sometimes I remove in non-linear from tutorial images already partly processed).

Thank you very much.

 


Edited by alvise, 28 May 2022 - 05:26 AM.

  • dswtan likes this

#37 slabkoff

slabkoff

    Sputnik

  • -----
  • Posts: 27
  • Joined: 22 Mar 2019

Posted 05 June 2022 - 11:44 AM

Hello -

I've recently gotten a new MacBook Pro M1 14" with an M1 Max chip an 64 GB ram and 2TB of SSD space - plenty of power/space to do pretty much anything I think.

PixInsight Version 1.8.8.9-1

 

 

I reinstalled PixInsight today, but for some reason StarNet (the built in version) did not reinstall properly.  I then went to download the StarNet2 update and put the files into the BIN directory as instructed.

When I went to update/install modules, StarNet2 was flagged as not being signed - I then allowed it to install w/ a bypass of security to allow it to install, but when I went back to PI, it did not see it anymore... and it does not install.  And I can't get the original StarNet nor the new StarNet2 to install and be seen. As this was a fresh install, I'm frankly a little confused as to what happened here... I followed the instructions and watched a video on how to install this properly... still zippo...  It's not even being seen - this is not about the weights library being in the wrong place...

 

Has anyone seen this behavior, does anyone have any suggestions about how to deal w/ this?

Thanks in advance.

 

Steve



#38 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 05 June 2022 - 10:42 PM

what's happened is that PI is an x86-64 application and so needs to be translated to ARM by the Rosetta2 software layer. this layer does not implement a bunch of x86-64 vector instructions, which the tensorflow libraries that ship with PI use (AVX2/FMA). therefore rosetta just throws an illegal instruction trap if starnet loads, and PI crashes. rather than solve this problem, juan just made the OSX installer delete the StarNet module if the architecture is detected to be ARM.

 

the solution to the problem is to find a libtensorflow*dylib that was compiled without those vector instructions. as it happens, the tensorflow library that shipped with the original command-line macosx version of starnetv1 fits the bill. however, starnetV1 is not available for download anymore. so you need to scrounge up the library from somewhere else.

 

the tensorflow library that ships with StarNetV2 does not have the vector instructions, so V2 should be compatible with M1 macs. however, for reasons which i have not been able to discover, PI is not able to find the tensorflow libraries unless they are copied to /Applications/PixInsight/PixInsight.app/Contents/Frameworks/ - this is a hidden folder and needs to be accessed from the command line, or you need to set up the finder to show hidden folders. additionally the starnetV2 tensorflow libraries are newer than the ones that StarNetV1 was compiled against, so i think V1 will not work right with the tensorflow libraries that come with V2.

 

this is all a huge shame because it doesn't seem like the AVX2/FMA instructions confer any real speed increase, so it would have been better if the libraries that shipped with PI's StarNetV1 had those instructions turned off.

 

rob



#39 Jabl

Jabl

    Lift Off

  • -----
  • Posts: 2
  • Joined: 02 Apr 2021
  • Loc: Florida

Posted 12 June 2022 - 10:43 AM

I was able to get Mac to accept the files by going to finder, right clicking and selecting open. That seemed to work..... but, when running Starnet v2 in PI I get the following error:

 

Error: Could not find checkpoint file in  '//Applications/Pixinsight/bin/'! 

 

I have triple checked and all files are in the correct folders. Any idea what is wrong?.



#40 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 12 June 2022 - 02:41 PM

what do you get if you open terminal and do "ls -lt /Applications/PixInsight/bin" (without the quotes). is there a file ending with .pb there?

 

rob



#41 Jerome Ni

Jerome Ni

    Explorer 1

  • *****
  • Posts: 98
  • Joined: 21 Jul 2010

Posted 22 June 2022 - 10:31 AM

I'm running into an issue on an Ubuntu 18.04 machine with a first-gen AMD threadripper. I followed the following steps below from https://www.cloudyni...rnet-v2/page-12

 

"

1.   Close PixInsight

2.   Download from https://www.starnetastro.com/download/

 

3.   Extract the compressed file

 

4.   Go to extracted folder

      > cd ~/Downloads/StarNetv2PI_linux

5.   Change file permissions

     > sudo chown root:root StarNet2-pxm.so
     > sudo chmod 775 StarNet2-pxm.so

    > sudo chown root:root StarNet2_weights.pb
    > sudo chmod 775 StarNet2_weights.pb

6.  Move these to the PixInsight library folder

     > sudo mv StarNet2-pxm.so /opt/PixInsight/bin
     > sudo mv StarNet2_weights.pb /opt/PixInsight/bin

 

7.   Start PixInsight

8.   Process menu -> Modules -> Install Modules

9.   Click the "Search" button and Starnet 2 should show up

10.   Click the "Install" button

"

 

except that, I could not find the StarNetv2 module in step 9, and it says

"

0 additional PixInsight module(s) were found on directory:
/opt/PixInsight/bin
(Nonrecursive search)

"

 

I verified that StarNetv1 that shipped with PixInsight works as expected. Has anyone encountered this issue?

 

Thanks!


  • dswtan likes this

#42 Linwood

Linwood

    Gemini

  • *****
  • Posts: 3,414
  • Joined: 09 Jul 2020
  • Loc: Ft Myers, Florida, USA

Posted 30 June 2022 - 12:42 PM

Sorry, initially posted in the announcement thread, will copy here and try to delete there. .. 

 

I am using Pixinsight 1.8.9-1 on windows 10x64 (current) with Starnet2 version 2.0.0-34 (I don't know if that is current as the download page does not list version numbers).  The PC is a 3970X 32 core with 128gb regular memory, with a pretty beefy GPU.

 

It runs fine and works well except I have a GPU memory leak.  You can see it in the below.  GPU memory goes dramatically up as it starts running, then stays there. The time of this extends beyond the finish of the process (which was successful and fast).  It will stay in use until I close Pixinsight.  StarnetV2 is all I ran during this PI session, but I have found that running Starnetv2 a couple times, while also doing other things, will substantially slow down PI itself (as well as other system functions).

 

It would appear that StarnetV2 is not giving back GPU memory it uses in the plugin.

 

gpu.jpg

 

Postscript:   I realized I was following the original (Dark's?) instructions for tensorflow, building with the GPU version from NVIDIA (or some such, I forget the details).  I tried instead with the version included in the PI process version at the starnetv2 download site.  It doesn't use the GPU at all.  I started it running on my sample image and waited seemingly forever it was at 4%.

 

I put back my Tensorflow (which is about 3 times the size by the way) and it now flies along as before. And as before uses GPU memory and doesn't give it back (so now I am sure I am running the latest plugin).

 

So at least for me, the downloaded tensorflow doesn't use the GPU at all.  Which to be fair may or may not be addressed in the last 4 pages of notes, I'll try to review in more detail.

 

But there still seems to be a GPU memory leak.  Workaround: Close PI and re-open.



#43 nekitmm

nekitmm

    Vendor - StarNet Software

  • *****
  • Vendors
  • topic starter
  • Posts: 227
  • Joined: 25 Apr 2018

Posted 30 June 2022 - 08:36 PM

Weird, no one reported this kind of problem before. Do you observe this memory leak on CPU?

 

I started it running on my sample image and waited seemingly forever it was at 4%.

You say you got "pretty beefy CPU"? I myself only run starnet on CPU on my laptop and can not complain about performance so this statement makes me raise my eyebrow tbh.

 

It could be some incompatibility between your drivers/cuda/cudnn or something...


Edited by nekitmm, 30 June 2022 - 10:10 PM.


#44 Linwood

Linwood

    Gemini

  • *****
  • Posts: 3,414
  • Joined: 09 Jul 2020
  • Loc: Ft Myers, Florida, USA

Posted 30 June 2022 - 08:40 PM

Weird, no one reported this kind of problem before. Do you observe this memory leak on CPU?

 

You say you got "pretty beefy CPU"? I myself only run starnet on CPU on my laptop and can not complain about performance so this statement makes makes me raise my eyebrow tbh.

 

It could be some incompatibility between your drivers/cuda/cudnn or something...

No memory leak observed in the regular memory though to be fair I have so much memory it would take a lot more of a leak (12gb vs 128gb) to show up. 

 

Yes, it certainly could be the tensorflow I am using.  But it is working properly other than the leak, it runs very fast, uses the GPU.  The tensorflow included with the kit just ignored the GPU. 

 

I've got an easy workaround if it is just me, was just reporting it in case it helped.



#45 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 30 June 2022 - 09:59 PM

i'm running macosx so its a different situation, but i did find that tensorflow doesn't really clean up after itself. meaning all of the threads spawned by tensorflow seem to hang around even when starnet finishes and no tensorflow job is running. i suppose if it does the same thing with the GPU this could explain things. on the other hand you'd think that everyone running GPU based StarNet on PI would have reported this since it seems to have an effect on PI itself.

 

based on how most tensorflow programs seem to work (run, produce output, quit) it's possible that this is a longstanding issue that google has never realized is a problem.

 

i think both nikita's and linwood's experience with CPU-based runtime could be simultaneously true if nikita is running on 4:3 size sensor data and linwood is using his full-frame camera.

 

rob



#46 Linwood

Linwood

    Gemini

  • *****
  • Posts: 3,414
  • Joined: 09 Jul 2020
  • Loc: Ft Myers, Florida, USA

Posted 30 June 2022 - 10:05 PM

And I am (full frame, drizzled, slightly cropped), so a lot of pixels. 

 

But I'm wondering if the bigger clue is not in the fact the downloaded tensorflow dll does not work for me at all.

 

But again, if I'm the only one having an issue, I have an easy workaround.



#47 nekitmm

nekitmm

    Vendor - StarNet Software

  • *****
  • Vendors
  • topic starter
  • Posts: 227
  • Joined: 25 Apr 2018

Posted 30 June 2022 - 10:30 PM

I just double-checked and I think all the necessary tensorflow-related objects are deleted at the end of the session. I tend to keep PI open all the time, never restart it and have not noticed any potential leaks. Just made a run and memory usage was back to pre-starnet level.

 

This said, tensorflow is very bad at using GPU memory. Unlike CPU runs, it always allocates all the available GPU memory even if actual need is a tiny fraction of it. It could be that for some reason it fails to clean afterwards, but you would think that by now more people would notice this... 

 

 

i'm running macosx so its a different situation, but i did find that tensorflow doesn't really clean up after itself. meaning all of the threads spawned by tensorflow seem to hang around even when starnet finishes and no tensorflow job is running.

 

How bad is it? What do you mean by threads spawned by tensorflow hanging around?


Edited by nekitmm, 30 June 2022 - 10:30 PM.


#48 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 30 June 2022 - 11:43 PM

I just double-checked and I think all the necessary tensorflow-related objects are deleted at the end of the session. I tend to keep PI open all the time, never restart it and have not noticed any potential leaks. Just made a run and memory usage was back to pre-starnet level.

 

This said, tensorflow is very bad at using GPU memory. Unlike CPU runs, it always allocates all the available GPU memory even if actual need is a tiny fraction of it. It could be that for some reason it fails to clean afterwards, but you would think that by now more people would notice this... 

 

 

How bad is it? What do you mean by threads spawned by tensorflow hanging around?

i was having some strange performance problems with SNPro - i had a 16C/32T machine and i noticed that SN would never even come close to saturating the CPU (vs on my prior 6C/12T machine where i would see 100% cpu utilization). so i profiled the application and i could see all the tensorflow threads going but never really gained an understanding of why there wasn't enough work to keep all the threads busy. i hacked the code to respect PI's thread limit so that i could lower the number of threads and see if there was just "too much" CPU available, or had hit some kind of wall with all those threads wasting time communicating with each other. i never really figured anything out, but i was surprised to see all the tensorflow thread names still active in the trace after starnet had finished execution and returned control to PI. maybe it doesn't mean anything, or maybe it is a bug in apple's profiler, but i thought it was strange.

 

i did look at the teardown code and it looks correct, which is what led me to say above that if this is a real problem, maybe google's engineers never noticed it as i assume most people's TF programs just quit back to the OS after they have finished running.

 

 

And I am (full frame, drizzled, slightly cropped), so a lot of pixels. 

 

But I'm wondering if the bigger clue is not in the fact the downloaded tensorflow dll does not work for me at all.

 

But again, if I'm the only one having an issue, I have an easy workaround.

the DLL that comes with SN doesn't run on the CPU? it just does nothing? that is weird.

 

rob



#49 Linwood

Linwood

    Gemini

  • *****
  • Posts: 3,414
  • Joined: 09 Jul 2020
  • Loc: Ft Myers, Florida, USA

Posted 01 July 2022 - 09:19 AM

the DLL that comes with SN doesn't run on the CPU? it just does nothing? that is weird.

It doesn't seem to be used.  I just did this again to confirm: 

 

- PI not running

- Copy tensorflow.dll from zip file downloaded to C:\Program Files\PixInsight\bin  (this one is 145,206kb dated 11/27/2022 inside zip)

- Processed a linear image to create a starmask also (13000 x 8940 image, greyscale) 

- Observed no change in GPU memory or GPU utilization, CPU utilization went to 65% +/-

- At 3 minutes I was at 10%

 

- Stopped starnet and closed PI

- Copy tensorflow.dll from my saved copy (this one is 466,371kb dated 9/8/2021) 

- Processed same image, it was complete in 1 minute 2 seconds, consumed all GPU memory but very little GPU other resources. 

 

Projecting the first linearly it would have taken about 30 minutes, or 25-30 times as long, so the GPU was a huge help.

 

I got the tensorflow I am using long ago from following these instructions: 

 

https://darkskies.sp...t-starnet-cuda/

 

though subsequently I know I did a few upgrades of various types and not at all sure that's the same tensorflow.  There are a bunch here maybe i got one o fthose.  I just don't recall, and there's no version number shown in the properties/details of the dll. 

 

https://www.tensorfl...tall/source#gpu

 

I should have kept better notes.  I did also do the whole Nvidia/developer stuff, install cuDNN, change environmental variables, etc. 

 

But definitely -- when i run the one in the install kit, it appears to ignore the GPU, but it flies with the older one.  But doesn't release memory. 

 

So it certainly may be because of using the "wrong" tensorflow.  But the one I use just flies.  smile.gif



#50 pfile

pfile

    Fly Me to the Moon

  • -----
  • Posts: 6,402
  • Joined: 14 Jun 2009

Posted 01 July 2022 - 10:23 AM

what you are seeing with the stock tensorflow is normal. the DLL distributed with SNv2 runs tensorflow on the CPU. it's going to be extremely slow on a 2x drizzled image from a 6200. the 65% utilization is what i was talking about above; on smaller thread count machines i saw 100% CPU usage but adding more threads/cores does not speed things up past some point. that's not necessarily unusual for parallel codes.

 

rob




CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.


Recent Topics






Cloudy Nights LLC
Cloudy Nights Sponsor: Astronomics