.Net Foundation .Net Community Standups, 9 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Machine Learning Community Standup - Sept 9th 2020 - Data Science and Machine Learning with SciSharp

Description

Learn from the SciSharp team how you can leverage libraries like TensorFlow .NET for your data science and machine learning applications.

Community Links: https://www.theurlist.com/mlnet-standup-2020-09-08

Featuring: Haiping Chen (@Haiping2015)

A

A

Hi everyone and welcome to the second machine learning.net community stand up. uh My name is bree, I'm the pm for ml.net on the.net team um and I'll. Let the rest of the team introduce themselves jake.

B

I'm jake, I'm the dev lead on the ml.net dueling side.

C

Hey I'm uh luis. I work on microsoft, docs.

A

Right and we're here with hyping who actually works on a psi sharp an open source project. uh It looks like uh we have another person that just joined hyphen. Do you want to do a quick intro and then we'll get you on later to talk about psyshart.

D

Yeah yeah sure so hello, everyone. So thank you for inviting us to attend this stand up, so I'm hyping, I'm I'm! Basically in my spare time I work for scishow libraries uh yeah. That's that's about me. I can talk about. I can talk more later.

A

Yeah awesome thanks, I think uh yeah. So the first thing we want to do is um go over some community links. So uh louise you want to share your screen and we'll we'll uh show off some things that the community has been doing with ml.net and machine learning.

C

Yeah, absolutely let me just go ahead and share that.

C

And cool, just let me know when you folks can see my screen yeah. You are all good all right awesome.

C

So let me pull this up here uh and uh you folks should be able to see that as well, um but essentially we have a lot of really really fun stuff uh since last time, uh one of the things we have here is um this ml.net uh crash course right and it was put together by by john uh who has the youtube channel. We talked about him uh on on the stream last time.

C

uh So essentially, you know if you're looking to learn about ml.net, and you know just sort of get up to speed with um how you can get started with it. uh This is a really great introduction to emerald.net. We also have for folks, it's really nice to see uh our international um sort of users right. uh We have this guide of. How can you can use devil.net with jupiter notebooks, um and this is in portuguese, so for our portuguese speaking uh audience right?

C

They can leverage this uh guide here and it kind of guides them through how you can essentially slow down and interact with which, uh if your posts aren't familiar, it's the among many things. It provides a kernel to run jupiter, notebooks uh and use either c sharp, f, sharp or powershell.

C

um We also have uh well shameless plug here. I put together a guide very similar thing with uh using jupyter, notebooks uh and running.net code. The difference is that I showed how you can use azure machine learning, compute instances which are essentially this computing environment. That is that has everything set up for you? So if you're looking to do machine learning, you set up a compute instance with azure machine learning, and you can use that as your development environment.

C

So one of the one of the benefits of doing that is that it brings jupiter notebooks installed out of the box. It has um all a lot of very popular data science, libraries and packages that you can just literally get up and running with your workflows as easy as possible.

C

So considering the fact that it already has super notebooks installed by default uh installing the net interactive so that you can basically write net code inside of jupiter notebooks in your compute instance is fairly trivial, and this blog post that I put together essentially just shows you that um we also have some really cool posts from another fellow community member bruno who one of them.

C

He talks about automl for ranking scenarios, and he basically just showed you how to set up a ranking experiment using using auto automl or automated machine learning for ml.net, and he has another one which he was very eager to go ahead and try. The uh the gpu uh support that was recently announced for model builder recently came out, so it's really nice to see that the community is going ahead and trying out these things right and then.

C

Finally, if you are interested, if you can't get enough ml.net, um you know after today's show you are more than welcome. I believe this is the global ai tour. That's putting this event together, um alexander, who was on the on the stream last time uh is actually giving a presentation at 5 00 p.m. uh Eastern time, I believe and he's doing it on make that a little bit bigger, so it's probably easier to see on emma lobs.

C

uh You know, basically, I meant and to end up racial operationalization of modernity, models uh within lobs.net and that's kind of the tool that he talked about last time when he was on stream. So, if you're looking to get a little bit more uh in-depth knowledge on that particular tool, I suggest you definitely go ahead and check out this uh check out. This talk.

A

Awesome uh so another link that we can add to the list there actually and I'll share my screen. So I can't remember if we talked about this last time, but louise and I and alexander and john who we've talked about previously.

A

um We all worked on a workshop together, which we presented first at the virtual, the first virtual community conference for ml.net and also at ndc melbourne. So essentially, if you want a really in-depth way to get started with ml.net from the very beginning stages to you know using deep learning and auto ml and mlaps, this is a great tool right here. uh You can it's a step-by-step. It can be a self-guided workshop that you can do or if you kind of already know these things or after you learn.

A

You can also take this and and use it to share as well. So if you wanted to teach your own workshop on, ml.net you're totally welcome to take um this workshop and as it is now or modify it to your needs and we'll we'll keep it updated as there's new updates with ml.net as well. So this is a great one to get to get started and actually, if you all know brady gaster he's the one who came up with this tool um for putting together the workshop.

A

It's really really awesome, step by step here um so yeah. This is a really awesome one, and so I also want to show off so last time we showed off ml.net model builder, which is one of our one of the easiest ways to get started, but I want to show the cli as well, which is our cross-platform offering, and let me pull that up here. So the ml.net ml.net cli is just a net global tool and I'll actually show this one first.

A

So if you do mlnet- and let me zoom in a bit so that you can see you can see here, these are the available commands. So right now the ml.net cli supports classification regression recommendation and a new one that's coming up. This is an internal build, is image classification, so this is really awesome. This is the same process as model builder, essentially except there's, no ui and again this is cross platform. So I'll show you an example here, let me pull up the data set that I want to use.

A

So let's say I want to do a regression task where I want to predict uh the price of taxi fare based on you know certain um certain factors or features so I'll show you what the data set looks like that, I'm going to train with- and here you can see, you've got a variety of columns, so you've got a vendor id a rate code. How many passengers were in the taxi? How long did it take? What was the distance if you're paying with card or cash and then the fair amount?

A

So we want to predict this fair amount. We want to train a model with this data. This is how you do it with the um with the cli, so you put in mlnet regression. uh All you have to do is input in your data set, how long you want it to train for and then you choose what the label column is or the what you want to predict in this case the fair amount and then uh once you hit that command with those options.

A

It'll start that training and again this is using automated machine learning or automl to iterate through different algorithms, uh to choose the best model for your data. So you can see it's going through a few different models here in the time that I gave it to train, uh looks like it went through four and the r squared, which is the metric used for regression um at the top.

A

One looks like it's: gonna be actually the first one that it trained and once it's done, training it'll actually generate that c sharp code so that you can easily consume it. So, if you go to uh here, you can see it created this. This sample regression folder with a console app and the dot and the class library- and this is the same- that was generated with model builder and, what's really cool, is if you open this up.

A

Actually, let me just open this up in vs and show you that it has a sample console app that you can actually quickly consume the model. It already has all that code for you so I'll. Let this load up- and this is this- uh is available already it's just like. I said a dot net global tool and regression is one of the ones that's already included with it.

A

And you can see here that console application and the this is actually class library uh project. So this has the trained model, the model input model output and I am enjoying visual studio, but I won't leave feedback right now, uh so program.cs for this console application that was generated.

A

uh So you can see it says you know it's auto generated file and in this case it's actually creating a new model input based on you know it takes the first one of my data set, but I can change this to whatever numbers. I want to start testing the model um and then you'll use it'll use the model to make a single prediction based on the inputs that you give it. So if we run this.

A

And we'll give it a second to run and this process is going to be the same or very similar for the recommendation task, which is like product recommendation uh as well as classification such as sentiment analysis spam versus not spam, github labeling, you know endless number of scenarios, uh so you can see here using the model to make a single prediction um you've got. This is the inputs we gave it and then it predicts the fair amount to be uh this amount.

A

So this has been out for a little while. I also want to show uh the newest scenario that we'll be adding, which is image classification, and so let me pull that up, and I have these images here which is weather. So it's broken up into cloudy rainy and sunny, and you can see each one has around 30 images of either cloudy um or rainy images, and then I want to train a model with that.

A

So uh for that one it will be image classification, and in this case all we have to do is indicate the data set uh and some image classification takes a bit longer, especially with the more images you add and more classes. You have so I've already actually trained that and I'll just show you um what that looks like, so it generated again sample image, classification and I'll open up open this one up in vs to show you how it's similar.

A

So give it a second to load, and then I will show um it still will generate that console application with the sample where you can test out test images, it will also have the trading code and uh and the trained model the model, input, model, output and so on all right. So this is the console up that it generated. It looks very similar to the regression task, but in this case the only input is the image source and I'll actually show you what this image looks like.

A

So this image is not part of the data set that was used to train. I already set it as something else, and it's going to be this one right here.

A

Which uh you can see kind of looks like I was going to say seattle in the winter, but it doesn't quite rain that hard here, maybe more like florida. So then I've, that's the one that I'm going to.

A

Test again, image classification is about to be released, is not quite released yet, um but it will be added to the cli, because we want to make sure that any scenarios that we have as part of model builder. We also have those scenarios in the cli. We want to keep that scenario. Parity for you know our cross-platform users or people who don't use visual studio.

A

And actually something that's coming on the road map. We actually plan on open sourcing, our tooling, so that's something to look forward to uh so you can see it started.

A

It always has this kind of generic text here and all right so took that image source and it predicted it as rainy, and this is kind of the probability or the score of how rainy or or so in this case you could see the middle one indicates the rainy one, um this one's probably sunny and cloudy, and it just returns this array of scores so yeah. So that's uh how the cli works um and that's the so regression's already out there image classification will be soon um but yeah.

A

So now that we've shown off, you know our own tooling. I want to give lots of time for the psyshart folks to uh to talk about. You know what they've been working on, so I'm going to add on hyping.

A

And also, you know all right.

D

A

Welcome to the stream, uh so uh do you all again hyping for people who might have not been on earlier? Do you both want to introduce yourselves.

D

Okay, okay, I can do first.

D

First of all, so thank you for inviting us to attend. This stand up yeah. We are appreciating that and I'm the key contributor of scishow stack community. So this community is the open source community focused on donet, based like machine learning tools so yeah before before we doing work at sunshine.

D

I've been working on like aiml for several years and still in the process of transiting from traditional software development, to like aiml engineer, updating my skill set and I'm being involved into some of the ai projects in the past years, such as like ai chatbot platform, like recommendation systems in like graph, embedding and computer vision in deep learning and other stuff related to machine learning in python and the donut programming language.

D

So that's basically yeah about myself so I'll head off to like.

E

Hannah hello, my name is manuel freitas, um I'm known as hinon on github, and I I am a c-sharp fan and uh I was looking for for ai libraries and and stuff and and and that's how I learned get to got to know hyping.

E

um Then we uh we worked together on on tensorflow.net for some time um and um and I then started to work on uh numpy, net and and and python included. Those are my main projects, um so hyping is is, is investing a lot of work in in the native port of tensorflow um he's translating all the python code into c, sharp and- and that takes really a huge amount of effort to do that, and I took a quite a different road.

E

I didn't have so much time to to really go into deep and and translate everything. So I tried to to go a different road. I I found that awesome project, that's called python net and I used that to to wrap the complete numpy api uh for c-sharp, so that you can put that and pull that in as a nougat and then start start using numpy. And for that I also introduced a project that is called python included, which is essentially a a deployment tool, an installer that will take embedded python installation.

E

D

E

Embed it in in a net assembly, or you can download it from the internet um and use that to to install whatever python version you need on your computer without having to rely on a pre-existing python environment.

E

So what people are doing with python I mean with python included- is, for instance, they they use that to to create azure services that depend on some python library or on numpy or whatever they embed all the python tools that they need and load that up onto an asia server and and consume it from c-sharp. So that's a very interesting thing and I've got lots of a positive response to it um and uh for uh with respect to numpy net.

E

um I wrapped all the numpy uh ap calls which are about 500 um and did this in only two weeks. So it's a it's a very a very clever approach that that allows us to to supply a whole library for c-sharp that wouldn't be uh there in in that completeness, without the the idea of generating all the c-sharp functions that call into python net.

E

But there is, of course, there are drawbacks, so python that is generally very slow and if you, if you are trying to do things like going over the data, do something in numpy copy the thing back into c sharp doing something there and copy it back to python and and looping.

D

E

Data all the time, then it will be much too slow to to do that with numpy net.

E

So um but of course, numpy is a very nice library where you can put a huge data array in you copy it, in with marshall copy, it's very fast actually to copy a huge bunch of data to python, then uh leave it there and and do the whatever tasks you want to do in python in in the native part of numpy, and and for that it's really useful, and I think that maybe that approach would also work well with uh tensorflow.

E

So I'm thinking of doing something like that for tensorflow too, um because uh the knife thing the nice thing about tensorflow is you you download the data into it and and then you have that computation graph and it runs in tensorflow, and you don't have to copy the data back and forth. So.

E

It wouldn't work well for eager execution, but hyping is working now on, which is very fast in c sharp, it's even faster than python, um but maybe the graph execution that maybe the the title net approach that I'm doing with numpy could be quite a nice thing to do, um but I'm not there yet.

E

So if, if, if people would like to see this, then I really uh they should come up to me and I would start working on this because I already have the code generator and I uh just need to write the parser for the documentation of tensorflow and then use my code generator to to generate all the ap cores.

E

Okay, that's in a nutshell, what I was doing.

A

Awesome yeah thanks for that, um so I want to well so a lot of people on the stream. Probably that are watching, probably don't have data science backgrounds or you know, know a ton about machine learning. Quite yet so um do you want to maybe, as you know, as you're explaining things, you can explain what is tensorflow um and so on. Just to make sure that people who might not know can we'll know what that is, I'm also hyping. I think you were gonna share your screen right and show it show us some things.

D

A

D

I can show- maybe some people don't know sunshine yet, so I can give give people like some quick induction induction about sunshine.

D

D

So first thing I want to talk about is like why sign shop exists.

A

Think are you sharing your screen yet.

D

Not yet maybe later so, so I need to talk about like why we need sideshow, and what does sancho do so? So we are trying to create a pure donut-based machine learning tool set. So this so, though, those tools are like consistent with the python ecological experience like it will make model migration easier and have the lower learning curve, because the syntax is the same as the existing documentation in the python world.

D

So why right do I have this kind of idea so from myself. So as a donut software developer, it's not easy to enter the field of like machine learning. World software development and machine learning are like two very different patterns, especially for like machine learning model development, which is completely different from like web development.

D

So web development is all about like concurrency security database, synchronization and like design patterns and but for like ai world model. Development is kind of more like switching data extract the data like evaluate the result, accuracy and the algorithm, and all about statistics, math and math.

D

So it's it so it. It sounds like a very totally different thing for donald developer.

D

So what we are trying to solve is how to like have the net developers to make their own ml model in their own familiar way.

D

So, obviously, we need to develop a series of like basic tools to enable developers to do like machine learning tasks, but at this stage, how do we achieve this goal?

D

So in the python world it's been involving like decades.

D

There are a lot tons of like learning, materials and documentations and papers on the internet, so but most of the code is written in python and people have to use, get have been used to the the python based machine learning ecology, especially in terms of like grammar and the thinking pattern.

D

If so, if we can like make tools that trying to like mimic the syntax of python and add the complex, the compilation error detection in for the donut community, I think that would be like benefit the donut developers think about it. People can use almost like 90 exactly the same syntax, to develop ml models without like adopting new syntax.

D

That means people can like lower the learning curve curve and invest less effort to use machine learning in their projects.

D

So, for the first principle in our libraries, it's like a keep same syntax and per even for the parameter definition same as the python functions as much as possible.

D

Even for the like naming conversion. I know some people like hate these makes making them like uncomfortable. Some people uh say I it's it. It's it's like a normal. It's it it's different with the like the mainstream donet projects, so people leave it.

D

But it's a hard decision in terms of like compromising compromisation.

D

So, but there are still many people like it because they don't have to learn new functions, don't have to learn new apis and the most advantage is they can like copy paste. The existing python code, the model the model algorithm, just add some like, uh like var, uh keyword before the the variables so everything it will work, everything's working.

D

So it's yeah, it's the valuable thing I I will demo.

D

E

We have uh we have some code snippets on on the front pages of our github projects, where you can see the python code and the sharp code side by side, and you see how how uh how similar it looks is just just the language uh differences are there.

E

The rest is all the same and yeah that's a huge advantage. Even though everything is a lowercase and some.

D

People like it, some people hate these so.

B

One one thing I wanted to call out too, is being really cool about your your projects is just that um it's not just like a syntax thing and familiarity, but also like how you publish your applications, how you get them out to users. All of that can stay the same with psi sharp, so you don't have to you know, learn python, to go, use those libraries and then figure out how to package them and how to get them on people's machines or on you know their their other devices.

B

So I think it's a really really cool project to bring it down to the dotnet.

D

So, okay, that's all the introduction about the scishow community and I because they are we, we still have like half hour right. Maybe we have 10 or 15 minutes to demo how how the sunshot library works. Okay, I will share my.

D

Screen you should be second to.

D

Okay, okay, can you.

A

You should be able to see now.

D

Okay, so, first of all, I want to show you the links sunshine repo links is here. The links is here, so we have several uh libraries uh one of the most one of the most uh popular library repo is tensorflow.net, is a tensorflow binding and the another one is blackboard. Shop workshop is a a chatbot platform and the first, the other, is the nine shop. My shop is totally written in in in internet standard 2.0, it's not a binding. It's totally totally delay between c-sharp code.

D

So this is, the numpy.net is, is maintained by hanan and the carrier's done is determined by d-cap.

D

Sharp service is a library for that. It's also uh the binding that the open save is abandoned. Okay, I have a question.

B

With the with the keras.net, we've actually been getting a lot of um community members asking when, when dotnet will get a keras like api, um can you tell us a little bit about the state of this and whether or not it's okay ready so.

D

This this this class.net is, is rely on python environment, so it is, it doesn't. So it's an is another approach, as hanan said, is a like shortcut to build up the ecosystem because for c-sharp developer they can we can. They can use the c-sharp syntax to use the existing uh keras you we can see from this.

D

So but this library is relying on python environment, so another thing is, I want to show. You is like I am working on. The um like native keras version is here it's a very cool thing see from these examples.

D

By the way this example, the repo you can find.

D

From here yeah, this is the main example example repo.

D

So there are like several examples to help donate developer to get started with the machine learning.

D

And so we are working on the master branch master branch is is built based on tensorflow 2.0.

D

So it's not it's not stable, yet we are in development intensively.

D

If you want to use it in your production environment, you should go to like tensorflow. Like one point, this branch, this branch is, it is production ready. So let me give you a like a quick demo for how does the tensorflow.net.

D

D

When you click the start, so you will see all the examples we are working on, because this is the master branch is not stable, so some of the example doesn't work, so you you have to wait. So we will fix that.

D

I will run a very simple one, like basic operations.

D

So I I set a breakpoint here, so I want to show you the excited features for these for the master branch and it's it's a binding for tensorflow 2.0. So the most exciting thing is you: it can runs in the eagle mode.

D

So what's the eager mode, your mode is the. If you have experience in tensorflow 1.0, so you definitely like it because you can see the result immediately. Even in the debugging model uh I should show. I can go through like one step by step, step by step.

E

Let me let me just just to say that tensorflow has two modes: one is the graph mode and one is eager mode and in graph mode you you have to design your computation graph up front and then it will calculate everything inside and just give you a result, and it is hard to to look into the calculation and see any uh intermediate results and eager mode is completely different.

E

It uh you can debug the complete calculation, because you uh you do every step in in your code by yourself, and so it's it's a lot easier to to learn with that and to decode that if you are a beginner and if you don't know exactly what the things are doing.

D

So, thank you yeah, thank you and then for explanation of that. So here in in tensorflow 2.0, you will see the result immediately in the debugging like a window see. This is like t, tensorflow, add three and a two add three that equals five. You can immediately see the result, so you don't have to define your like placeholder and the variables in advance.

D

So this is the extreme difference between tensorflow 2.0 and 1.0 tensorflow 102. That's the huge difference.

D

So here see all the results comes up immediately.

D

See this is the four leg of practice. Then you can see the result here. So that means the this scientist can use doughnut to develop their like a models and debug the models very convenient.

D

So this is a very basic uh operations demo. We can run a more complicated example.

D

We can show this linear regression.

D

D

This example is still running in the eager model, so you can see every step how the model running I will show you. Let me show you.

D

Well, this is the basic linear model formula, so you will see the result immediately from the debugging model and the gradient present and apply the gradient descent immediately see the result.

D

D

See once once the apox runs like 100, you can print out the um the result of the model. So this is the loss. You will see the value, the loss value in this window and you can print it as well.

D

This is the stochastic gradient descent optimizer. You can define before the.

D

D

Okay, this is the.

D

Linear regression example. So another example is we can do like keras, so I will show you how the class works.

D

It should be eight.

D

This is that means user means. The date is that.

A

Do you want to talk a little bit more about what that is, for people who might not know.

D

uh Mist data is like publicly set for, like uh handwriting handwritten the digitals.

D

uh I think we can get a lot of information from means.

D

Okay, I can show you the how the carriers works for keras. We have two different way to create the keras model. One way is: is customize your model in a programming way like here.

D

This is to here so yeah. This is the this round approach to create your carriers model. Is you can inherit from the base class model? Then add your layers here.

D

The most important thing is, you have to override this function,.

D

So this is so this. This class is very, like a very uh look, looks very similar with the python.

C

Version so so uh sorry to interrupt here, uh but um so you're showing off keras like functionality here, uh and this is inside of sensorflow.net correct. So uh why would I want to first of all what is keras and why would I want to use that to build my neural networks.

D

Oh, if keras is a high level api to like to design for deep learning when you use keras build your neural network, it will be very, very simple. You just add layers.

D

It if so the keras is like uh keras is like uh building blocks, so you can like add whatever layers you can add more specifically like more uh neurals. Whatever you want, you can add layers whatever number you want to add, and it supports many it. It has been in many layers. If you take a look at here, uh you will see many building layers.

D

You can do do it immediately like dense convolutional, like lstm and pulling layer it's already built in so you just need to like come combine them into a new neural network, so you can develop your neural network very, like very easily looks like a building a lego.

D

So another another approach to build your keras is like use the building models like sequential. I have another example here here, so this is a. This is like a more like easy way to build your neural network. You just add different layer in the network, so I think people most of people like this way to build his own network.

D

D

Another way, I just said, is it's more like you want to control more in your network. You should use.

D

You should use this way. You should use this way yeah if I customize your model, this way is have more power control to your model,.

D

This way is more simple but less control, so I think that time I spent too much time to talk about this bryce, so any plan. I think yeah.

A

No, that was awesome you're, showing that off.

E

You should you should quickly add that uh one thing that blew me from from my seat is when he compared the eager mode in c-sharp against eager.

D

E

In in python, it was faster in c-sharp.

D

E

I haven't prepared that maybe.

D

D

Yeah I have, I haven't.

A

Seen a demo and.

D

I'm not prepared.

A

Oh, that's! Okay. um There is one thing that I wanted to mention, which we haven't talked about is so uh ml.net actually offers an uh high level image classification, api for training, image, classification models, using a method called transfer learning and I'll actually share my screen.

A

This was a blog post about it that cesar wrote a while back and it actually runs on top of tensorflow.net, which is really really awesome. So um you can kind of see here uh what uh actually.

E

A

Looks like um so you've got your tensorflows.net c sharp bindings and then on top of that, we've got ml.net and it looks, I think, there's some sample code here yeah. So you can see what that looks like if you've seen ml.net code before you know that you first have your ml context in this case you're, adding a trainer for image classification- and this is the where that image classification api comes in, and this is built on top of tensorflow.net. So all pipings work um but yeah.

A

uh So we've got about 10 minutes left for questions um jake or louise. Did you have any questions uh top of mind.

C

um I do, but uh I could probably hold them off until you know. Perhaps the some of the questions from the chat are answered.

A

um So I will try and go through the chat and see.

B

Too, that we're we're we're, then we're being model builder in the clr, then written on top of this image classification api, which is using the tensorflow.net. So we're we're a huge um huge fan of tensorflow.net and trying to make it really easy to consume with model builder and the cli.

A

Comments there's a lot. Was there anything.

C

So there's there's one that kind of um you know. Part of it was answered, uh but maybe uh high ping or add-on could go ahead and maybe provide additional context to this, which is, uh is, has citeshop rewritten these libraries to run on.net or have has it just adapted? The native wants to support that net.

D

Hannah, can you explain this.

E

um Okay, so there is uh the tensorflow.net is, uh is based on the c plus plus core of uh of tensorflow, which is, uh as everyone knows, from google, but it also has a python layer on top and the python layer is always thin. It is very complex and what we did is we completely removed the python layer and directly call into the c plus plus uh core, which is the the the number crunching engine uh from uh from google and.

E

The the same goes for some other libraries like num sharp. They are completely designed or written in c sharp and also the sharp cv.

E

But there are some libraries where we do just call into python, like numpy net like kerasnet and uh and there's also a torch net, which is not complete, but there is those two approaches, because it is enormously much a huge amount of work to to to convert all the python code, and uh I don't know how many developer developers are working at google uh on tensorflow, but they are chucking out new versions faster than we can uh convert the old versions to to c-sharp. So there is kind of those two approaches.

E

Hyping is is really awesome, all the work he puts into that, but he's always trailing behind so um and and- um and uh that is why I was focusing more on the on the generative approach and and on to calling into python so that um some libraries are can be, can be supplied to to see sharp users and consumers uh quite fast and and in a very uh complete way.

D

Yeah, thank you.

A

Nice uh louise: what do we got next question wise.

C

Yeah so uh you mentioned uh well. First of all, we can maybe ask this one here: uh are there rapids ai in this or, I guess, rapids, ai support.

E

uh I don't know.

D

I don't know what's rapidly.

C

Could you uh could you maybe clarify that question janisco and, in the meantime, we'll perhaps move to another one? um We have this other one. Also that asks you mentioned that it was faster and there was a question for clarification in terms of what exactly was faster when you ran. uh You know some of these algorithms between some of these models with.

D

C-Sharp, I can answer this question, so the faster means- I'm saying you know in for tensorflow.

D

The many people know tensorflow is all about python part. They many people doesn't know the c plus path c api underlying the python. So python is a binding uh above the api.

D

For for the tensorflow one point, one version one many work, many work. For example, when you define uh when you define a tensor, you have to use like a python interface, uh so you.

D

Like you have to like write an array like one, two, three so first of all, the python environment uh allocated the memory for this array and it copied into in the past into the c api in this.

D

In the api you copy the memory to to like copy a memory and it converts to like c convert to the c c. Tensor is programming c, so there are many. You know the python is a weak language, so you have to convert, uh convert the data type, keep can like keep converting between uh between the python and the say, so there will cause many like cost between in the communicate communication.

D

D

I'm not sure if I answer the question.

B

Yeah, I think that helps so it's mainly in like the type conversions between the the native and python, where we think a lot of the performance improvement came from.

E

Yeah and then c sharp is faster than python and and so, um especially with the eager mode, there is a huge performance uh increase. If you use c sharp.

B

And I have a follow-up question on that, so eager mode just means like you're feeding all the data to it, versus it being grabbed like one by one or what's what's how do you? How would you describe the difference between the two modes.

D

The I think the most different different uk for for for any user is in graph mode. You cannot get the result immediately. You have to run the whole graph from beginning to end. Then you can print out. You can get the result, but for eager mode you can print every result in every step, one by one step.

D

That's the big difference: cool yeah.

A

All right we've got, I think, a few more questions here. uh Let's see, we've got how's the relationship between cysharp and.net, interactive efforts. Are there collaborations or overlaps.

D

um I have to say when we start start the jupiter notebook, the microsoft hasn't released a like stable version for jupiter notebook kernel, so we have to write, write it in our own. So we uh we have our own, like ic sharp core for jupiter kernel.

D

But after like half a year, so microsoft launched his like official version for the jupiter kernel and then like child.net.

D

Yeah they, I think, there's no collaboration between this library and are not interactive, but absolutely they have overlap.

A

Any questions that I'm missing please see.

C

Yeah, there's there's there's a few, but I noticed that we're also running running low on time. um So in the interest of time. What is the best way to reach you folks, and how can you know if uh folks, on the call want to contact you and ask you more questions? What's the best way to get a contact.

D

I mean for sancho yep, uh we you can uh all the work we we have done uh in progress or is all in the public is in github. You can you. If you have any question, you can submit issues on the responding projects and you can also email to us to the official email address and at the same time, at the same meantime, every pr and the comments are all welcome.

E

We have a github channel, that's all yeah.

D

E

D

E

Yeah we have that beautiful website and you find everything you need to know about scishow on this website. It's https psi sharp dot, github dot, io.

D

Yeah thanks for like hannah and create this beautiful website.

A

Love the purple.

A

Thanks a ton for for coming on today and talking about scishart um so yeah anyone else who might have questions for them, um you can reach them on their github or on their on this site right here they have a contact us that you can reach out so yeah and.

B

We will have you on the community stand up again thanks for thanks for joining us.

E

Yeah thanks for having us.

D

Yeah thanks yeah appreciate it.

A

Yeah thanks for everyone tuning in um and you know you can always reach out to any of us with questions. If you got them so.

D

A

All right thanks.

A