Live- Exploring Nvidia RAPIDS- Open GPU Data Science

Key Takeaways

and that only right so uh welcome everyone to this amazing session of rapids open open GPU data science I've explored before some amount of information in this and just imagine it says that by using this kind of um rapid APIs uh you can rapid AIS you can actually run your machine learning code in GPU so till now we were uh actually worried of running the machine learning libraries, machine learning algorithms in in u multi-course processors like in Intel you say that you have nine cores, eight cores, right? 10 cores. So you are actually trying to run in that. But let's see how it goes on and uh again uh this was been referred by many of the people also. So let's go ahead and try to understand what is the rapid open u GPU data science. So to first to begin with guys I really want to understand with respect to the about section see whenever you are learning something and I'm I'm also learning along with you okay I'm also learning along with you I'm learning with you itself and Nvidia if you consider GPUs if you consider deep learning CUDA libraries plays a very important role and just by changing this CUDA libraries they are able to execute some amazing things uh by using this uh rapids So uh just to understand the description the rapid AI the the rapid suit of opensource uh software libraries and APIs gives you the ability to execute end toend data science and analytics pipeline entirely on GPU each and every operation entirely on GPU guys. Okay. So it is definitely amazing because just imagine how fast the process will become, how fast the pipelines will start making u start running in short. Okay. Apart from that you can see that uh it is being incubated by Nvidia based on extensive hardware and data science experience. It utilizes Nvidia CUDA primitive for lower level compute optimization. exposes GPU parallelism and high bandwidth memory speed through user friendly Python interfaces. So uh GPU parallelism is required because we'll just see guys over here you'll be able to see that you also have examples of notebooks you know Google collab notebooks and we will try to execute this we'll try to see how it gets executed and all uh what all and there are a lot of dependencies that will probably be required uh so this is the code we need to actually focus on these two things that is cudf and cdl examples this is actually making the machine learning algorithms run in GPU or the whole data science project run on the GPU and uh we'll we'll discuss about these things u first let us just understand the about section u it also focuses on common data preparation task tasks like for analytics and data science this includes familiar data frame APIs that integrates with a variety of machine learning algorithms from end toend acceleration without paying typical serialization cost. It also includes supports for multi-node, multiGPU deployment, enabling vastly. So these are there a lot of things are there and probably you'll not be interested because no one is interested in reading so many things. So let's jump directly to the Google Collab. Now uh in Google Collab here there are some examples. Uh the STIPS CSV is a data set and probably if you really want to do some kind of operations, you know how it'll do in the GPUs and all that we'll try to see. Okay. Now before going guys if you really want to check which Nvidia it is basically using or what what is the CUDA version that is have present in this Google collab. So if you just execute this so you'll be able to see that as soon as I execute this it'll show you show you the CUDA version that it actually has and how to go over here guys after clicking on uh get started you can either go and click on this or you can also launch in collab okay any of this will be properly fine okay it is up to you uh how you want to go to this okay and there are some clone statements we'll discuss about this let's see now here it says that the CUDA version is 10.1. Okay. And probably it has Tesla K80. Um check out the make sure that you have allocated you have been allocated a Tesla T4, P4, P100. So here you can see that Tesla K80 has been allocated. Um so I think it'll work fine. Okay. And uh again I can do also this local uh in my local machine but I really don't want to do in my local machine. It'll be very very difficult to you. Okay. So I hope everybody has got till here. Can you just give me a go ahead scan clustering amazing things are there. So can I get a quick go ahead like something everyone do you do you like this kind of sessions where we study together so that you'll understand how is my study patterns and all. Okay perfect so today no dislikes. Okay. Okay perfect. Uh let's go ahead um with respect to the setup uh the setup script installs install most recent miniconda release compatible with Google collab python installed 3.6.7. So if you have in your local if you have a good GPU you can do this also or whatever the Anokonda environment you have you basically can uh use it. Apart from that uh you install you remove the incompatible libraries then you install rapid libraries okay set necessary environment variable copy rapid. So all these things are there this is with respect to all the setup that you actually want. Okay. So coming to the installing rapids. Okay, installing rapids and probably in the next session I'll also try to explore this in my local and I'll try to do it but right now let's do it in Google collab so that we can quickly learn and understand most of the things okay so all the steps that has been mentioned has been done over here so first of all we have cloned uh rapidsai CSP utils.git If I really go over here, these are the basic libraries I think that is required in order to support Rapids AI. Okay. So here it is. You can see this is the collab and this is all the script I probably think it'll run. Okay. So main bash scripts to properly install rapids onto a Google GPU instance. Okay. So for this why we are using this it is simple. Just go in the about section. Okay. It is clearly written over here that it is an open source libraries and APIs gives you the ability to execute end toend data science analytics pipeline entirely on GPU right now. Um if you do some of the operations with respect to data frame and all it won't run into the GPUs, right? It'll it'll run into the CPU. It'll take the cores of the CPU. But what if if you really want to execute everything in the GPU, right? Why not? This is an amazing opportunity, right? Since if you're keeping your GPU idle you know suppose if you have in your premise a very big GPU and if you're keeping it idle what is the use of it and uh if you have solved any machine learning problem statements all the pre-processing techniques usually happens in the CPU right what if all those operations is moved to GPU right it'll it'll be pretty much amazing yes so uh let's go so after git clone we are doing this bash uh rapid AIC TP utils.sh.stable. So this is probably a bash command and why they have used this this is just for the installation of the rapids AI. Okay. And then they are actually setting the path. It was mentioned is over here. Okay. This actually probably removes the compatible files. Then installs the rapid libraries. Okay. This is the rapid libraries. Set necessary environment variables is basically through this step. Okay. Then copy rapid. So file into the current working directory. So this is the step that is being done over here and then if running version 0.0 or higher updates PI okay pi arrow library is also required I think I I saw over here say rapid uh has its start from Apache arrow and go AI project based on the column inmemory data structures that deliver efficient and fast uh data interchange with flexibility to support complex data models. So this is actually there and that is why arrow is actually getting used. Okay. So if I execute this, this will probably take much time to do it. Till then we will be exploring things. Let's see. Okay. Unfortunately, collab did not give you a T4, P4 or P. But uh they said that Okay, this is this is really really frustrating. Now I have to find out uh unless until Collab does not provide me Tesla T4, I have to keep on executing this. This this is something not good. Why? Okay, this is getting executed. But I think it has got started executing, right? Make sure that you have configured collab request to let me see what is the run type. Run type is GPU only. But why it is not given me this one. Make sure that you configure to GPU instance. If you get a different error, I reset all the times. Let me see. Uh do we reset or just we reload it? I'll try to reload it guys because it will not work. Um so here it is. I'll execute this. Nvidia semi cell is already running where? Okay, this will not stop actually. Okay. So, first thing first, we'll execute this. This time I have again got Tesla K80. Why I am not getting Tesla other GPUs? Check out the I need to make sure that uh change the runtime run time guys just try to in your side whether you are able to get some other GPUs just see that or whether you are just getting Tesla just just let me know guys whether you are able to get other GPUs or you are getting Kat Tesla K80 I'm getting Tesla K80 only. Everybody's getting Tesla K80. How to allocate collab with Tesla T4? Globe collab but only Tesla KD. Yes, one issue has been raised in rapid AI. So it says this has been a growing issue over the past month. So and a bunch of people have been talking about it in the Slack to the best of my knowledge a few day lag. Nobody knows how why and how to fix it. Perfect. So the first issue is that we will not be able to run collab now has P instance and we have updated our script. We see you get a keep. Okay, let's see. Let's see in this. Okay, you you're getting Tarun Chob is getting Tesla P4. What did you do Tarun? What did you do separately? Anything separate things that you did? Rajat, how did you get Tesla T4? Did you just execute this or did you do something else? because my runtime code can uh can Rajat or Chaitan or Yoges tell how did you get Tesla T4 because it is coming as Tesla K80 only. Okay, everybody's trying but why I'm not getting bad luck. Huh? Uh, reset runtime. Let me restart the runtime. Just execute uh executive notepad codes. Everybody's getting I am only not getting. What the hell? No. Why? Let me open a new Google Collab. Tesla K. So we have to pay it. We need to have collab pro which is not available in India. Getting T4 in normal condition is very browser browser. Let me see. Nvidia cloud factory reset runtime. Okay. GPU save and execute. Factory reset runtime. Let's see. I'll execute from here. No. Tesla K80. Tesla K80. Sir, it's random. Oh K you have to try I think for many number of times okay J runtime confirm the things [Music] fresh too many terminate. Okay, fine. I will manage manage session. Oh, oh, okay. Okay, fine. Let me see. Got it. T4. But I guess how do we study? Oh, finally Tesla P. I got amazing. See this Tesla P. I got see over here I got Tesla P. What you do is that whatever notebook you have opened just make a copy of it. in India probably I think you'll be able to get this. So I've done this. Uh let me just execute this now. P00 amazing. So the installation has started. T4 is fine guys. T4 is fine. Vasanta just make a copy of it. See, I made a copy of it, right? When you launch in the collab, you'll get like this, right? Just make a copy of it. Copy to drive and then try to run. Okay. So, this is getting executed nicely. You can see the output. The output is coming up. Sir, you terminated. It worked. Yes. Yes, hit like for rapid AI finally. No, no, hit like for uh this GPU. GPU is that which we had actually worked so hard to get it right. The GPU name is Tesla P00 by dislike. Okay. So this will probably take some time. Please everybody execute till here. All the dependencies are getting executed. Then we'll go to the next step to understand what is CUDF and CUML. can't run. Hooray. Got Tund got using VPN. Okay, perfect. So, this will take some time guys. Till then, let's talk about something else. Uh, okay. So, this is basically the architecture. This looks like for Ap logos, themes, branding, and other guys. Take a look at our branding and guide page. Ease of use. You can see that GPU architecture on top of that you have CUDA on top of that you have CC++ and uh there you have Python but as you go down the performance will be the performance is increasing at this side as you go down with CUDA you'll be able to get performance uh better than CC++ programming language ease of use you can see that Python is probably the most easiest one okay so uh you'll be able to see this over here Right. And um visualization one our focus on Python allows rapid to play well with most of the data science visualization libraries for great for even greater performance. We are working towards deep integration with this libraries since a native GPU in memory data formats provides high performance high FBS values. So you can see over here 20 CPU nodes 30 CPU nodes how it is shorter is better. So this basically shows the time right with respect to data load plus feature engineering is basically indicated by this uh color. You can see that uh the last two nodes are basically with the GPU and this is basically with the CPU. You can see with respect to the CPU how much seconds it is actually taking. Okay. If I consider this whole pipeline, it is somewhere around 8,000 seconds. Okay. uh if I talk talk about 30 CPU nodes, it is taking somewhere around 6,000 seconds and if I talk about 50 CPU nodes, it is taking somewhere around 4,000. But if I take about GPUs, right, it is hardly taking you probably more less than 100 seconds, right? So this is good. Oh, it also has integration with deep learning libraries, libraries and APIs overflow. So probably we will start looking at CUDF and CUML. Okay. So let's see what is the status of this and whether it is working fine or not. It'll take time every everything this will take time. Let's go with respect to CUDF. Now what is CDF? CUDF. CUDF is uh called as GPU data frames. Okay. Uh CUDF is called as GPU data frames. uh you can see that it is built based on Apache arrow column memory format. CUDF is a GPU data frame library for loading, joining, aggregating, filtering and otherwise manipulating data. So whatever functionalities that we have actually done in pandas all those things all those things can be done with the help of cudf right. So whatever things you do with pandas then you know that pandas usually take time and it runs on CPU right. So if you really want it to work with GPU you have to use this library which is called a cudf. Okay. Yeah. Similarly like pandas. So it is saying that it is a GPU data frame. Pandas has a data frame right? GPU data frame library for loading, joining, aggregating and all. So these are some of the functionalities. As you can see that by using this they have actually calculated some percentage and we'll try to understand how much time it will take. Uh let this execution happen. So has the execution happened? Uh okay it is taking still time guys. Uh still for you also it is taking time right or is it only for me? Just tell me this answer. So this okay toolkit is getting downloaded right now. I have a very good internet speed since I'm getting time. Probably for you also it'll be taking time. Okay. Sure. Suman Paul Chri I'll be talking about open pose. Is it faster than CPU? Yes, obviously it'll be faster than CPU. But make sure that you have a GPU then only it'll execute all those codes in uh GPU. Right. We'll wait for some time. Okay. Uh finally verifying transaction. It's it is almost it is almost there. Yes, almost there. Then after this we will be running the next part that is cudf and cml guys. Uh if it is completed just give me a thumbs up so that I come to know that CUDA toolkit download or okay what we do here sir where ain we are actually exploring Nvidia you don't have pi arrow unloaded oh why your Google collab instance has rapids installed okay perfect let's see whether we'll be able to execute this code. I really want to see whether this will execute this code or not. If it gets executed perfectly, then I think this should not be a problem. You do not have pi arrow and probably. Okay, perfect. You can see over here it's running absolutely fine. back to internet problem. So don't worry, don't worry. Be happy. Yes, installation will take too much time, guys. That is obvious right okay let's see uh I will just try to probably I can use this right time how much time it will take to execute this oh whether I have not used correctly or what okay how much time it is taking so it is taking around 263 milliseconds okay to do this calculation If you really want to print uh what is let's print tf tips_df dot shape so by that you'll be able to understand like how many how many records are actually there so they are on 244 records and just to do this operation it is taking somewhere around uh 94.4 4 milliseconds. Right? So this is it. Loads a data set into the GPU memory resident uh data frame and performs a basic calculation. Everything from CSV parsing to calculation tip percentage is grouped. Average is done on the GPU. You must import NV strings and NV category before CUDF. Else you'll get an error. So here it is with respect to this particular execution. and see over here um browse our complete a API documentation. So the API good documentation is also given over here. If I talk about data frames so here it is you can see data frame and uh you can also do this kind of operations right coodf operations any kind of operations you can actually do like how they have done with respect to the data frame similar operations you you can see over here. So in short what happens is that when you're executing this code this whole code when you're executing it or let me just show you how much time it will take for this three operations. Okay, I think uh is it better to calculate or compute the time from here for doing the request because request is done from the uh Okay, let me do one thing. I'll just import time. Okay. And I will say start time is equal to time dot time. Okay. And uh I will say after this operation or after the group by operation I'll say uh end time end time is equal to time dot time. Okay. And then I will say print because I really I don't want to include this request.get because this depends on the uh internet right it depends on the internet speed. So I don't really want to use this. So what I'll do is that instead of using like this uh I'm going to use time from this operation still the this operation because I want to understand over here. So the total time uh is nothing but this much dot format and the format I'm going to use start time start time minus end time. Okay. So if I execute this so the total time why it is coming as minus okay sorry end time minus start time so very very less guys very very less so what I'll do is that I will just do like this and probably the total time is just 020 seconds yeah Tesla T4 is also okay the overall time for this is 95.5 milliseconds Right? If I talk about total time for executing this operation operation from here to here because this is all the things we'll be loading into the GPU and we'll be doing this right and for that it is taking somewhere around 02 that is uh around 20 milliseconds. So 70.5 milliseconds has been taken just to load this just to load this just imagine over here. Okay. Uh your screen looks blur guys just uh is everyone facing that issue? Just reload the page guys. Just keep your settings. If your screen is looking blood that basically means the internet there's the internet problem on your side. Okay. Increase the resolution because your internet connection is slow not mine. Yeah. Just reload the page. you will be able to see that. Okay, let me let me see with respect to this also how much time it will take. Um probably I will just use start time here too. Okay, and uh end time here too. Uh start time time and end time end time here too. after this and I will be preventing here as request request uh requests request time I'll say request time and probably because of that only you'll be able to see that [Music] what is request okay request time is taking somewhere around 24 milliseconds okay this is high right it's taking huge amount of time Right? 264 millconds this time. Right? So definitely uh our whole pre-processing of this right what kind of operations is there it is basically taking very very less time that is the total time is somewhere around 022 seconds. Okay so 22 millconds isn't it amazing everyone isn't it amazing right? uh more things you can refer from this I think that is the main reason built based on Apache arrow blah blah blah cudf provides a panda pandas like API that will be familiar to data engineer data scientists so that they can use it easily to accelerate their workflows without going into the details of CUDA programming which is pretty much amazing right and same example is given over here so after this session kindly have another session regarding resumeum building. Sure, I will do that. Okay, I'll do that too. But uh have you got the idea of uh CUDF guys? Can I can you give me a go ahead so that I will discuss about the CML and any kind of operations that you do like in data prep-processing and all you can do it. Okay. Yeah, perfect. Uh, next Sage Maker video I showed you, right? I just have to record it. I was just practicing it today. Understand, guys, I also work somewhere. So for me it'll become difficult to do all the things you have to give you have to be patience. Now the next thing that I'm going to discuss about is CML. Okay. So CML what it is? It is a suit of libraries that implement machine learning algorithm and mathematical primitive functions that share compatible APIs with other rapids project. CML enables data scientists, researchers and software engineers to run tabular ML task on GPU without going into the details of CUDA programming. One very very important thing that you have to note guys uh CUDA programming is very very difficult. It I'll not say u it it is up to you guys but there are lot of configurations that are involved and if they're providing a wrapper on top of it like this like cuml you'll be able to run your whole machine learning uh uh basically if you have a tabular data set you can run that you can run the ML task on the GPU without going into the detail of CUDA programming. So in most cases CML Python APIs matches the uh API from skyit learn. See it is also compatible with skyit learn and uh over here amazing benchmark results has been shown. Okay let me just show you some benchmark results again again all the credit goes to Nvidia. Um otherwise just imagine how training used to happen and all definitely not possible right deep learning would not be that fun. Let's see. I have to open this. How do I open this guys? Is this page opening for you all? Do I have to download it? Fine. I can also download it. Let me see tools. Okay. Benchmarks guys. Is it upload uh loading? Please do let me know. It's okay. T4 is also working fine. Let me just download this guys. CL. Okay. Because we really need to see the benchmarking. Okay. Okay. Fine. First of all, let let us explore. Okay. this thing for larger data set GPU based implementation on complex 10 to 15x faster than the CPU equivalents see according to the benchmarking it is basically saying that the speed usually becomes more than 50% speed right more than 50% speed uh 50 times you know of what it was trying to uh do it right so because of that you'll be able to see that it is faster than CPU definitely we'll have a look now this is an example of Python snippets load inputs and compute DB scan clusters on all GPUs. So let me see over here uh let's go to the thing. So over here uh the simple uh okay this snippet loads as above all the calculations performed on GP GPU. So you have a data frame. Okay, this is your probably your first uh probably your first column, second column and the third column. And you are actually performing a uh DB scan with epsilon value as 1.0. And CML has all the libraries I guess guys. Let's see. Insert code cell. Oops. No, I don't want text cell. Come on. I want a code cell. insert code cell. Okay. So I will just import CU CUML. Okay. And I will be inserting one more row code set. If you write cuml dot let's see whether any any kind of see there are so many things accuracy score rand score DB scan decomposition K means label encoder lasso linear regression logistic regression MSD MSD classifier MSGd regressor nearest neighbor PCA random forest classifier all the algorithms are there so it was said clearly over here what it was said it is in the API matches the API from SkyitLarn. Okay. So if you really want to tomorrow if you're planning to run your um uh machine learning code and if you really want to learn run everything quickly, you can actually use Google Collab and start executing on the GPUs. Right now it is just not applicable only to CPU cores but instead you can also execute it in GPUs. Right? So uh did you get it guys? Isn't it amazing? So let me just uh do it for you all and try to execute this and let's see what all labels we will be getting. So this is the labels you can see over here when we are actually trying to see three groups has been divided uh 0 1 2. So with respect to that you'll be able to see this right. This is to cuml there are more notebooks that has been given by rapid AIS right some more examples if you really want to see see for CUDA 10.0 0 for CUDA 10.1 how you have to install it is also given just write install and this it'll get automatically installed cudf also you can do it in your local machines very nice I think al together let's see I'll do it in my local I I'll see what all issues I may face right can't we do this in Jupyter notebook we are doing in Jupyter notebook also right when you do that whole installation in your local machine do it in Jupyter notebook. Google Collab is just like a Jupyter notebook only, right? Right. Everyone Okay. So guys, how was this session everyone? Was it pretty much amazing? Did you learn something? Everyone, did you learn something or not? Just tell me. Sir, right? Open source. open source then I will go with it otherwise I'll not go with it right if it is open source similarly a AWS SageMaker IBM Watson is there there are many things new new things that are coming right and uh you can actually explore that now this is with respect to just one operation now you try to do it with your own you know different different things yes the recording will be always available in the live project playlist so this will also be put up in the live project playlist. Okay. Now [Music] one thing that uh probably you have noted okay so this is basically the data frames of cudf right fine cdf readad csv this will automatically if you if you see uh the types of uh tips_df let me print uh tips type of tips DF. Let's see. So it basically says cuda cudf.core.datafataframera.dataf frame. See this in pandas it says something else right? Let's see what what does it says in pandas. Uh let me just create a code cell. Okay. Um the the the content okay I am going to use this. So I'm basically going to write import pandas pd right mpd dot read csv and I'm just going to copy this uh what is this io right? And let me just write it down in df. Uh I'm going to print type of df probably it says pandas.code.frame.dataf frame. Right. Right. So here it says you coodf.core.dat dataf frame and everything is remaining all are same guys. Whatever operation you do, whatever operation you want to do, you do it. Uh let me just show you that also all the options that are present in CUDF. Okay, again just explore guys. This is amazing, right? If you know pandas, why not you explore this cudf dot let's see uh you have category data type, you have flow divide from pandas. Oh, you can also convert from pandas. Okay, let's see from pandas. I will try to uh use df. Okay, let me go and write type of this. Let me print it. Here you can see right you have actually converted from data frame to coodf uh data frames. Isn't it good? Right. Amazing. Nice. coodf dot let's see let's let's explore some more option um oh they are merge merge sorted read csv read_c read hdf redundo park oh oh my god everything whatever is available in pandas uh I did not see read excel is it there okay so the best thing is that oh get dummies is also there. Okay, let that basically means I can even handle category features. I can completely do all the things, right? Why not execute some projects? Why not execute some machine learning projects? It's good. I will try to do some machine learning projects in the future classes. Okay, it's amazing. Uh let me just see some more things from pandas. Get dummies. Okay. Oh, okay. Oh, arc sign arc tan concat cuda dataf frame exponential flow divide from DLP pack DLP pack from pandas get underscore dummies uh grouper index io logical and logical not or melt merge merge sorted readers aro Oh, readers feather readers hdf read_json readers_rc readers park remainder rmm series is also there amazing right so if I if I write cu 2DF head. Okay. Oh, sorry. Not it should not be CUDF. Probably I'll save this in some data frame. C UF_DF is equal to this uh DF dot head. So if you do do head, what is the error? Okay, syntax error will be there always. So here you can see that it is showing you in the form of data frame itself. Right? So how to use C means clustering on more than 20 inteious features of data just apply. Okay. With respect to visualization if you want to visualize you just have to perform PCA and then uh visualization you you can just perform PCA into two dimension and then you can visualize. Okay. How can you make the front end more presentable? You have to focus on the front end. Use some Angular framework, ReactJS and all right. Okay. Uh so visualization visualization was also there right let's see uh visualization install oh Q graph let's see this there's something called as this graphs okay a Q graph is GPU graphic analytics the rapid G is a collection of GPU accelerated graph algorithm that process data found in GPU data frame the vision of CU graph is to make graphic analysis ubiquitous to the points that user just thinks in terms of analysis and not technology or framework. Let's see some example. Where should I bring this graph data? Is it somewhere there or not? to realize that vision CPU graphs operate at the Python layer on GPU data and they allowing for seamless passing of data between ETL task in CDF and machine learning task in CML. Okay. However, there is a less type and structure checking at C++. For more details, check this. What is this graph data? Where do I get this data set? I think it will be present in this um graph data data sets data sets data sets. Okay, I'll just go into Python. I think it is not present here. Let's see some examples. Uh probably there will be some examples then then then uh rapids rapid cu I have I'll search for it rapids craft rapids which graph was that craft okay we are seeing this. Okay, there are some documentation with respect to this also. Okay, we can use this all things. I really wanted to see some coding example. Okay, it is given at M of0 uh source destination the computed degree of corresponding vertex. Okay, let me just explore this. I think this requires more exploration when compared to other things. This will be requiring more exploration. Yes. uh some more examples that I really want to see. Okay, there are some blogs also that are written over here in medium, right? So in medium you can see on Microsoft Azure using rapids NLP and text processing by Vibhu Java. Uh amazing. Let me see. I don't know whether this will open or not. And again the credit goes to Vibhu Java. uh and this issue is always there. So whenever you find this kind of issue open in cognitive mode right so dogfish okay you can see over here do series codes some example has been given with respect to this right explore this guys I think uh you'll be able to see this let me see see some projects that I really want to create for u rapids AI and it'll be pretty much amazing right what is present in this notebook advanced notebooks or tutorial rapid customized kernels okay right now this is not opening I don't know for some reason getting started notebook books, intro tutorials. Oh, here you have. So all the things here you have. Oh, it is also using Dusk. Okay, let me just clone this repository guys so that it'll be handy much right uh getting started notebooks. Okay, intros and probably let me just change my internet it is given over here over here right notebooks contrib today files are not opening in GitHub why I think I have to do a clone and probably do you want one more session probably in the future we can have this session uh This will be amazing, right? Oh, Daskk is also used. Daskk is for multiparalism, you know, it'll help you to do your data loading faster also. Parallel processing also. No, not parallel processing. Sorry. If your data set is huge, you can actually take part by part. Okay, let me just explore this. I'm just going to keep uh into my favorite section so that I clone this, I start exploring this. Okay. Uh and what are things there? Um Okay. What about installation? Okay. Installation is also given over here. Build. So, you just have to do get clone, make, build, make run, and just open it in the local host. This will probably work in u Linux I guess or let's see make for make build we require Visual Studio. Okay. So let's see I'll just explore this and we'll come back to the next session. Okay. Uh so guys uh so data structures mentors come on Saturday Saturday and Sunday we will be discussing about the mentors and all. So let's uh go with the next session and uh but right now you have got some ideas right and probably we will be creating more kind of tutorials where we we do some in-depth analysis. So Chris if you run on your workstation RTX what will be there will there be a huge difference? uh my if I consider what was the uh just a second. Okay. So you can see over here uh P Tesla P right. So what is the cost of Tesla P? Let me just see the course right probably it is very very faster. Tesla P cost. So price analysis is done over here. Okay, so this is pretty much fast. Pesla PCIe 12GB. What is there in mine? PCIe. I don't know how much GB it is there. But I think I can uh take out this information. Can I? I have to install PyTorch and then find the memory. How much memory is there? Let me just execute this NVDI once again because it will tell us how much memory has been used. Okay, it is 16280 that is uh somewhere around 16 GB. 16 GB is costly. So probably this will work. So I'm using this one, right? It'll be pretty much faster. Yeah. So it'll be pretty much faster. uh pretty much amazing when compared to my Titan RTX. So better it is that sir I think from Nvidia as a gift you need to get a 100 instead of Titan RTX. No no no these are costly guys 4,7 374 uh with 16 GB. Uh I'm so lucky that I got in this Google collab this one right it is very very hard to get from that much time we were trying right is there a option okay there is no such option as such oh you have that emin test data sets and all. Okay, let me just explore this in something else. Probably I will be exploring this this whole uh where is that this whole notebooks and uh probably we'll keep one more session and uh then we'll try to explore it. Okay. So yes guys uh I hope you like it. Please do hit the like button. Uh it's almost 1 hour and uh we will again be coming and make sure guys uh just join my night session where I'll be playing games. The link of my YouTube channel, my gaming YouTube channel or vlogging YouTube channel. The name is Krishna Vlogs. Uh there you can exp go and subscribe press the bell notification icon. So that I'm coming tonight live. Uh I'm playing some other games and probably I'll be also taking your Q&A right. So, thank you all. Uh, bye-bye. Any questions that you have, I'll give you another five minutes, you can ask me. So, post a community post for your vlog channel. Sure, Krishna, I'll do that. Any any more questions, guys? I hope uh it is fine. Needs hand-on session like this twice a month. Sure, I'll do that. by Kashor. TensorFlow versus PyTorch. Really a great session. Thank you. Thank you everyone. G clone. See Git clone is pretty much simple guys. First of all, you need to install g. So you just write g install. I have Windows Windows 10, right? Just go over here, download the .exe file in Windows. Just download it. Install it. Okay. Then go to the GitHub file. Sorry, go to the git. Just go over and click over here. Copy this. Go into your working location, right? Any working location. Just write g clone uh get clone and just give the URL like this. and automatically the clone will happen. Confused what to choose DL or ML? You have to choose both. Good night. So, see you all at 10:30 p.m. guys. Uh make sure that you subscribe my other channel that is Krishnag Walls. The link is given in the description. Thank you. Bye-bye. Good night everyone. Yes, you can also save the Google uh collab file like uh you can see over here um if you if you go and see this it'll get saved. You can also download this. I'll try to open this file only in the next time. Okay. Okay. So, thank you all. Uh, bye-bye. Have a great day and yes, keep subscribe the channel, hit the like, make it 200 before going and thank you all. I'll see you all in the next video that is tonight. Bye-bye. Bye-bye everyone.