Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
  • 15:51
    AvenSun opened #376
  • 13:09
    ahsanGoheer edited #459
  • 13:05
    ahsanGoheer edited #459
  • 13:05
    ahsanGoheer opened #459
  • 11:44
  • Dec 07 21:50
  • Dec 07 21:45
    slamj1 starred SciSharp/BotSharp
  • Dec 07 21:45
  • Dec 07 17:51
    Nucs labeled #372
  • Dec 07 17:51
    Nucs labeled #372
  • Dec 07 17:49
    Nucs labeled #373
  • Dec 07 17:49
    Nucs labeled #373
  • Dec 07 17:49
    Nucs labeled #374
  • Dec 07 17:49
    Nucs labeled #374
  • Dec 07 17:49
    Nucs commented #374
  • Dec 07 17:48
    Nucs labeled #375
  • Dec 07 17:48
    Nucs assigned #375
  • Dec 07 17:47
    Nucs commented #375
  • Dec 07 17:40
    solarflarefx opened #375
  • Dec 07 14:29
    xela-trawets starred SciSharp/NumSharp
Haiping Chen
@Oceania2018
@cveld Are you trying to find the open source solution or paid .NET solution?
Carl in 't Veld
@cveld
ideally no licensing costs
I was assuming SciSharp is completely royalty free?
Apache 2.0 license
Haiping Chen
@Oceania2018
Yes, you can use it for free. What's your specific requirements? ex: Emotion API Explorer.
Carl in 't Veld
@cveld
object recognition and image classification
Brendan Mulcahy
@BrendanMulcahy
@sir-deenicus Thanks for the pointer. I'm going to try tf.scan first, but I can use that as a backup plan
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 Is Func<ITensorOrOperation> how you normally translate python callables to C#? Trying to implement tf.scan :)
Haiping Chen
@Oceania2018
@BrendanMulcahy Yes, Func<> shoud be passed into tf.scan 's fn parameter.
Brendan Mulcahy
@BrendanMulcahy
Should it be Func<Tensor, Tensor> ?
I think that looks right to me
Haiping Chen
@Oceania2018
@BrendanMulcahy Right.
Haiping Chen
@Oceania2018
@BrendanMulcahy What version are you using? I'm using tensorflow 1.15.
Brendan Mulcahy
@BrendanMulcahy
I forked master so not sure
Matthew Moloney
@moloneymb
Has anyone looked at doing multi-gpu training?
Matthew Moloney
@moloneymb
@Oceania2018 Just had a chance to test GPU on linux, works great!
Иван Сердюк
@oceanfish81_twitter
can you form a benchmarking report?
I might share with some people, which care about the algorithms - maybe they would poke around optimisations
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 I have some half-baked code written on tf.scan: SciSharp/TensorFlow.NET#444 . Could use some help with the questions I listed there. Modifying control_flow_ops.while_loop looks like a bit of a rabbit hole but I think it needs to be done
Haiping Chen
@Oceania2018
Found an interesting example: Starcraft 2 in reinforcement learning. https://github.com/sammuttaja/Starcraft2_c-_Bot/tree/TensorflowNet
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 It looks like WhileContext.cs, line 239, there is some kind of hack to get packed_vars to work, but this is not flexible enough to make a working version of tf.scan, as its bodyItem has type of (Tensor, TensorArray[], TensorArray[]), so it's not able to cast TensorArray to Tensor.
Haiping Chen
@Oceania2018
@BrendanMulcahy You can have your own BodyItemInRnnWhileLoop. and make a switch wrap it. or you have a better idea?
John Ansa
@Blogrammer
@Oceania2018 I have checked for issues but none is available. What can I work on?
John Ansa
@Blogrammer
Also, how can I access the Articulate with Botsharp interface?
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 I'll see what I can do
Haiping Chen
@Oceania2018
@Blogrammer Can you raise your requirement with details?
Brendan Mulcahy
@BrendanMulcahy
I keep running into issues with while_loop. I got my own class working and I have an idea of how to fix the hacky part. I am using tensorflow.dll 1.14.1, which looks like what tensorflow.net is built with. Any thoughts?
System.EntryPointNotFoundException HResult=0x80131523 Message=Unable to find an entry point named 'AddControlInput' in DLL 'tensorflow'. Source=TensorFlow.NET StackTrace: at Tensorflow.c_api.AddControlInput(IntPtr graph, IntPtr op, IntPtr input) at Tensorflow.Operation._add_control_input(Operation op) in D:\Programming\GitHub\TensorFlow.NET\src\TensorFlowNET.Core\Operations\Operation.Control.cs:line 44 at Tensorflow.Operations.WhileContext._AddOpInternal(Operation op)
Brendan Mulcahy
@BrendanMulcahy
Oh nvm I got that working, I was using the wrong tensorflow.dll in my project
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 I got tf.scan implemented and I wrote some test cases based on the documentation. The output is the right shape but all the values are 0. Any idea how to debug this?
Haiping Chen
@Oceania2018
How can I check your code?
Brendan Mulcahy
@BrendanMulcahy
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 Let me know if you see something. I'm sort of stuck. I read through all the code another time and I didn't see anything obviously wrong
And compared it against the python version
Haiping Chen
@Oceania2018
@BrendanMulcahy I'll check later.
Matthew Moloney
@moloneymb
I don't quite understand Elis reponse to SciSharp/TensorFlow.NET#455
My expectation is that if I have to do a call back every time I want to add to the model that it will be rather onerous.
It would look rediculous in Jupyter
Eli Belash
@Nucs
I'm working right now on a static variable IsSingleThreaded when set to true, different threads will always point to the same session and graph stack.
So in Jupyter you'll only need to call ops.enforce_singlethreaded() once at the start of the notebook.
Matthew Moloney
@moloneymb
that would be perfect
uh, could we make it the default though? And set it to false for testing?
it would throw errors for people who don't know they need to first call that
Eli Belash
@Nucs
For testing it has to be true, most testing frameworks run the tests parallely which 99% of the runs cause memory corruption.
tensorflow's c_api is very sensitive in some parts to memory corruption caused by racing conditions on therefore we are limited by them.
Matthew Moloney
@moloneymb
Ah, I was under the impression that the multi-threading was put in for testing, not sure when you'd use it otherwise
Eli Belash
@Nucs
3 months ago before I added the the multithreading architecture we had randomally inconclusive failing tests.
John Ansa
@Blogrammer
@Oceania2018 I would like to see a working example of BotSharp in action or a guide on how to set it up.
Brendan Mulcahy
@BrendanMulcahy
@Oceania2018 Did you find anytime to look at the tf.scan implementation?
I've got time this weekend to work on it more