Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
GeorgeS2019
@GeorgeS2019
Is there interest here to bring some of the SciSharp Stack examples to web using blazor with charts (2D and 3D) or even with 3D models?
Haiping
@Oceania2018
@malisimo Thanks for your questions.
  • No plany yet, we're focused on TensorFlow and Keras for now.
  • Sure, we're interested in IPlot, although we have a similar library Plot.NET but there is nobody maintain it. Do you mind join SciSharp STACK community for that library ? You'll still be the owner.
GeorgeS2019
@GeorgeS2019
https://github.com/sps014/BlazorML5 (Blazor for Machine Learning)
Meinrad Recheis
@henon
wow @malismo your plot library produces awesome looking plots. you should really add more of those sexy pictures to the github readme
theolivenbaum
@theolivenbaum
@henon @malisimo if you guys are looking for plotting, we maintain an auto-generated plotly library for c# here: https://github.com/curiosity-ai/plotly-sharp - it supports almost every method from plotly - and we try to keep it in sync with plotly releases
And in case anyone is interested, another useful lib for embeddings search: https://github.com/curiosity-ai/hnsw-sharp
matt
@malisimo
Hi @Oceania2018 - yes would be very happy for IPlot to join SciSharp STACK community! Maybe when the lib is more stable would be good to do that, and I'm not particularly worried about ownership (but will continue maintaining).
Thanks @henon haha, have added a few pics but all the sexiness is from Plotly/HighCharts. Dark theme was definitely a decision though =)
I was not aware of plotly-sharp @theolivenbaum at all - thanks for that, its really impressive. Any plans to get this working in dotnet interactive?
Haiping
@Oceania2018
@malisimo One plan is that I'd like to visualize the training process in IPlot, I'm not sure how if it can achieve the goal.
In order to support Machine Learning Visualization, the library need to plot in high performance.
lqdev
@lqdev:lqdev.tech
[m]
Hmm... Kind of like Tensorboard?
Haiping
@Oceania2018
@Nucs Deployed Tensorflow v2.4.0
Eli Belash
@Nucs
@Oceania2018 awesome, I can get started with tf.experimental.numpy
Eli Belash
@Nucs
@malisimo pretty work, welcome :)
theolivenbaum
@theolivenbaum
@malisimo what is needed to get it working there? I remember playing with it during the alpha tests, and it worked fine!
I think all you needed was to return the plot.Render() as an HTML object - something like:
var testData = new[]
{
    new float[]{ 1, 20, 30 },
    new float[]{20,  1, 60 },
    new float[]{30, 60,  1 }
};

var chart = new Plot(Plot.traces(Traces.heatmap(Heatmap.z(testData))));

return HTML(chart.Render());
Haiping
@Oceania2018
@malisimo Is it possible to support Tensor in the TF.NET, for example:
var testData = tf.constatn(new[,]{ {1f, 20f}, {20f, 1f} })
var chart = new Plot(Plot.traces(Traces.heatmap(Heatmap.z(testData))));
return HTML(chart.Render());
matt
@malisimo
@theolivenbaum To get plotly-sharp working in dotnet interactive you would need to register your formatter using the Microsoft.DotNet.Interactive libs and spit out some binaries in specific folders in the nuget package - take a look at Philip's PR on XPlot: fslaborg/XPlot#153. From my tests it looks like you need to target .Net 5 rather than netstandard to support this, which is hopefully a transient situation!
matt
@malisimo
@Oceania2018 I agree the interop between .NET and JS via JSON does not particularly support massive data sets, although in terms of visualisation the plotly OpenGL trace types (e.g. HeatmapGL in your example) would support large numbers of polys as it uses hardware acceleration. This should work well for you, once the data is loaded in the browser. HighCharts deals with large data sets through the use of boost module (I haven't tested this with IPlot yet) though it seems you are not looking at scatter/time series. BTW I would love to support methods/functions accepting DataFrames and Tensors directly, something for later down the line.
theolivenbaum
@theolivenbaum
thanks @malisimo , will check it!
Aiden-Enztec
@Aiden-Enztec

Hi everyone. i am an student intern working on a game dev project. i have gotten tensorflow.net to work with unity and having it detect when they are doing certain things. but it seems to be taking 70-100ms to process a frame. i have narrowed it down to sess.run(). the model i am using has been made/tested in python and runs significantly faster than it does in unity. any tips on how to make it more optimised to run faster. tried googling the issue but can't find any solid solutions. here's a paste bin to the whole class for some context
https://pastebin.com/vpCJsdKu
the method is ```
public void AIProcessing(Mat image) {
Shape shape = new Shape(1, image.Width, image.Height, 1);
NDArray img = new NDArray(image.GetRawData(), shape);
using (s_SimulatePerfMarker.Auto()) {

        //Tensor tensor = new Tensor(image.GetRawData(),tf.float16);


        var input_operation = tf.get_default_graph().get_operation_by_name("import/x");
        var output_operation = tf.get_default_graph().get_operation_by_name("import/Identity");
        FeedItem item = new FeedItem(input_operation.outputs[0], img);
        s_SimulateAIMarker.Begin();
        var results = sess.run(output_operation.outputs[0], item);
        s_SimulateAIMarker.End();
        results = np.squeeze(results);

    } 

```
any help is appreciated

that code formating is weird
Aiden-Enztec
@Aiden-Enztec
hi everyone. turns out turning my resolution of the image when using sess.run() dramatically increases the performance (cutting the impact by a factor of 10 for 1/4 the resolution). if anyone knows how to increase the performance without having to reduce the resolution that would be great
Haiping
@Oceania2018
@Aiden-Enztec Did you only measure once ? it should be faster when it runs second time. For the first time, it needs to cold start that relatively slow.
Aiden-Enztec
@Aiden-Enztec
i run it every frame that the webcamtexture is updated. so its like every 20-30th frame it updates as i am running on a 144hz monitor.
minhduc66532
@minhduc66532
i'm trying to get my app to run on the nvidia jetson nano (ubuntu 18.04 ARM)
but when i tried to run it
image.png
image.png
as u can see
i have installed tensorflow 2.3.1 on the nano
but somehow my app still refuse to run
image.png
all of the packages in my app
Haiping
@Oceania2018
@minhduc66532 I don't have Jetson Nano, so I can't reproduce it. TensorFlow is able to run on Jetson Nano according to the doc https://docs.nvidia.com/deeplearning/frameworks/install-tf-jetson-platform/index.html
Dustin Dobransky
@Inventor22
Hi, does anyone know the reason the RetrainImageClassifier.cs example was removed from https://github.com/SciSharp/TensorFlow.NET? There seems to be a related closed issue regarding this too: dotnet/machinelearning-samples#585
Dustin Dobransky
@Inventor22
Thanks @Oceania2018!
Dustin Dobransky
@Inventor22
Hi @Oceania2018, are there any examples of training a custom multiple object detector using TensorFlow.NET?
Haiping
@Oceania2018
@Inventor22 Not yet.
GraphSam
@GraphSam
Hello... Thanks for the great work you are doing there, it's massive work in all its aspects.
GraphSam
@GraphSam
I have tried for the first time installing Numsharp v 0.20.5 using nuget in visual studio 2015 with .net v 4.7.1 framework but it failed to install it because of .net compatibility problems. Do you have any idea how to solve such problem?
Ed Cashin
@edtyping_twitter
Hi! I'm starting with NumSharp, using it in an F# program that is working with audio samples. I don't see np.quantile, though, the first thing I tried. It's in numpy 1.15.0 (https://numpy.org/doc/stable/reference/generated/numpy.quantile.html), but I can't find a way to tell what numpy functions are in NumSharp and what aren't. Is there a list? Is quantile available some other way? Thanks!
GraphSam
@GraphSam
Finally installation of Numsharp 0.20.5 is successful and looking next to try it with matplotlib.net though unsure if the latter works in c# environment. Do you have any idea how to use matplotlib.net?
@Oceania2018 .... Do you have any idea how to use matplotlib.net?
Иван Сердюк
@oceanfish81_twitter
Who is interested to give a remote talk?
Jean-Sylvain Boige
@jsboige

Here are the new bounties:
https://www.bountysource.com/issues/91873084-inputs-was-changed-by-accidently
https://www.bountysource.com/issues/91873276-loading-of-hdf5-files-from-keras-not-possible
https://www.bountysource.com/issues/91873312-build-graph-implement-switch-statement-tests
https://www.bountysource.com/issues/78645236-implement-tf-image
In order to collect the bounty, when the issue is closed, I believe you have to click on "Solve issue", and after a 2 weeks verification period you should be able to cash out with a 10% fee or redeploy at no fee.

Congrats to @Oceania2018 and carb0n (https://www.bountysource.com/people/80162-carb0n), who were able to claim Bounty#1 and #4.
2 more to go !

Congratulations again to @Oceania2018 for clearing up the bounties. I'm very glad those could be redeployed and you were able to claim them timely. Keep up the good work !

Haiping
@Oceania2018
@jsboige Thank you for orgainizing those informations.