Elisha Odemakinde Hosts Roboflow ML Engineer, Jacob Solawetz

Key Takeaways

all right so i'm going to show you a new link where we are going live is that okay yeah yeah sounds good five nice i can share it too okay so that's the new link let me just tweet it let me just switch the new link so that people can view us live from the day yeah yeah perfect perfect yeah it's coming through on mine so i'll also put it here my i just want to do five minutes broadcast we're going to be starting in five minutes so i just want to do the five minutes broadcast which is the new link okay sounds good sounds good so okay so i just want to do a quick um sweets [Music] of the new link okay let's do it all right so um we have six people watching us right now all right um hello guys uh we are so glad and excited to um have you all um please uh we're gonna be talking this um in a couple of seconds from now um we would love to have you all would love to have you all um share this new um new streaming link on your social media your platforms and your friends just to tell them that um we are live now we are live now we are just um this there was a bit of amnesia so um please give me a thumbs up give me a thumbs up if you can hear me you can hear me live please give me a thumbs up alright awesome all right so let's let's just do this let's share the link in five minutes we begin um this awesome session on roboflow yeah have everyone and uh yeah as alicia said um feel free to share the link with everyone in your network and we'll uh we'll have a great session here really looking forward to it awesome awesome i can see very awesome comments here and there oh cool cool all right so uh in three minutes we should go like we should begin the session [Music] so all right guys um i'm so glad to um welcome you all and i'm so excited okay um my name is adam mcindel just a bit about myself um i'm a researcher at data science in nigeria and um [Music] of course a lot of you guys know me to be community you know community at this science nigeria we handle communities coming today all right so uh just uh just a quick one i'm so excited to um post everyone here and um i'm so excited to also have um jacob here and i will just be sharing a little insight into the reason why we are actually having this session please um as we move on with this session i would love to have that scene from the um what's up from the youtube chat session i would love also talk about this robo flow of the thing as jacob is dishing out those you know punch lines i would love us to just be responding with our comments because i'm i'm watching you guys also i'm also on youtube saying everything i'm going to be seeing a lot of your um interactions and discussions your questions will be duly noted we are going to be answering all those questions so i just want you guys to be um calm just uh try your best ask the questions that um try ask your questions of all those questions all right so just a quick one about myself um of course i work with data science nigeria as a researcher and more importantly as a community manager so i have jacob with me here jacob is actually a machine learning engineer at roboflow.ai what actually um caused us having this this meeting is because of the fact that um there was a time i was just trying to uh think of um i was thinking of doing something more than image image classification uh like i just wanted to just work something related to um i wanted to do something in real time so and i was thinking okay let me get this done this way let me get this done this way then i was uh thinking okay let me just do and then i came across the roboflow.ai platform and um i discovered it was very very fascinating and awesome in the way you it was um very very it was very very direct in putting you out to computer vision and that's the reason why i'm committed to you know sharing this great opportunity with everyone that is watching us live because it's it's it's going to make a lot of sense if i love you guys i must have heard of roboflow that hey i would i'm quite a lot of people from community they do not um know so that's reason why i'm like bringing this this up to you guys that no we get we we have to just have something awesome we have to have uh we have to in action and that's the reason why i had to you know jacob of course i i i was in contact with you called that slack channel and we began talking about some things relating to object detection and um computer vision and because of that i was just so excited to just say okay jacob let's let's just do this let's just do this you know a lot of people want to see this great idea they want to see these awesome intentions they want to see this um awesome work that roberto does ai is doing to ease computer vision end to end and you know that was what you know pushed me to [Music] writing about uh roboflu the ai uh actually writing about the articles that you guys have been reading on medium so i wrote an article on a hint of hair detection with um with um n2w detection with efficient plates which was um an objection architecture i'm sure a lot of you guys must have read it you know you know when i released out the video a lot of you guys were like awesome school you know so i was like no for people to be so excited to see something like this then let us let us let us introduce them introduce to them the end-to-end platform that gets this point for them in fusion that's the reason why we're having this session so i'm more than excited to actually invite jacob so exactly you know brad and then joseph so [Music] to accept my um my willingness to have jacob on this platform so jacob i would just love you to just say a few things about yourself before we actually going to do the deal of the day sure um so first of all this is uh this is an immense honor to uh to be here and uh to be able to speak with you all this this really is kind of uh the embodiment of the uh mission of roboflow which is to uh make computer vision easier for any developer to use anywhere in the world um so the fact that you know we are seeing traction you know with people starting to deploy these technologies um anywhere in the world including nigeria exciting fact for me and so very excited to be talking about this and um got to give you guys a little bit more info about rebel flow today a little bit about my i um i've gotten interested in ai over the years and i was working actually in the originally in the in the market trying to predict stock prices were moving and things like this and then i started working building nlp chat bots and virtual assistants um came across this opportunity to join um some long time friends of mine um on the roboflow project um so they've been working on rebel flow for over a year now um i've been here for for about five months and um we're really exciting about the growth that we've seen and the the product that we've been able to build um i primarily work on the machine learning side which is um kind of and working on the technologies underlying uh training uh computer vision models and also working on our product to kind of try to make it even easier to get to that moment where you went from just having raw images to that moment where you see a model doing inference which is um sort of similar to the experience that alicia is talking about with using efficient debt to to detect objects and object detection so that's that's a little bit about me and uh really looking forward to diving in deeper oh awesome awesome you know that's actually the spirit and that's actually you know one of the most important reasons why i'm like no we just have to bring this down here you know i i was like you know when i was thinking about it's a lot of things done and i came across this your platform um and it it made a lot of things very easy for me you know to do a lot of computer vision tasks and projects that i've been engaging myself in and that's the reason why we are having this session all right so guys i know we are more than excited to actually have gekko to start a stage so you guys have heard literally about him but before i go ahead i just want to just uh share my screen and then just point out few things that i want you guys to do i want us to tweet and um i want also use the following [Music] hashtags on social medias so let me just share my screen in a couple of seconds all right so all right guys so i want to believe you guys can actually see this so i want us to to follow roboflow.ai on twitter yeah use this um i want us to follow roberto.real on twitter tweet your experience about the ai what you've been learning what you're going to what you'll be learning from jacob the live station and the live demo and everything um i wanted to use that attribute ai astra computer vision hashtag ai and ask roboflow a high all right so we'll now be handing over to um jacob i'll be handing over to jacob to start a session so jacob over to you thank you all right guys now your screen right now to make your presentation and then start your life then thank you sure yeah so i'm gonna share my screen and let's see if we can get both uh both of us up here see both of us here in the corner um and uh so yeah um everybody can see everything okay okay for you alicia yeah it's okay okay great um okay great um okay so awesome so here i'm going to kind of dive in and give you a little bit of a uh just a pro of what the mission of roboflow is and what what we're um sort of working on uh doing day to day um so basically we're starting with the mission of taking raw images and transforming them into a trained computer vision model um so here uh this is an example of uh some boats and houses that we took pictures of um via a drone and then we took we were able to pass them through a computer vision model that could now understand how to label things like boats houses and docks even without um ever having seen uh the test images before so that's a little bit um on roboflow so it rebelflow builds uh developer tools that enable any developer to use computer vision so it's a really good way to just kind of get started if you're just starting to dabble in computer vision and you want to try to get from images to getting your model off the ground very quickly by loading in your data and doing some conversions on your data some augmentations on your data and then exporting your data over to a training environment and then then you're able to kind of also share your data and your models uh via the roboflow via the rebelflow platform and we're going to dive in a little bit deeper on all that in in a demo later but this is just kind of the broad idea of what uh roboflow is doing with with uh conversion and analysis um okay so this is uh sort of a exciting time to be getting into computer vision the the technologies are just starting to get good enough like with the release of papers like efficient debt which allows basically anyone to kind of pick up a model architecture and then apply it to their problem and already see really great results um and because this this this is a unique time and kind of human history with with computer vision becoming more applicable to all sorts of different industries we're basically seeing like a cambrian explosion of computer vision apps uh coming out all over the world for an example um here we've put some some fields that we've seen coming through the roboflow um through our notifications of people what what so we've seen people uh using uh rebel flow to detect weeds in fields so if you're flying a drone over and taking pictures of the field or you have equipment um that you you want to spray weeds with people are actually kind of identifying what's a normal plant and what's a weed there are people who are looking at pipes where they see a pipe and they there's gas leaking out of the pipe and the computer vision model is able to kind of sit over top of the pipe and see where these leaks are and then kind of automatically uh alert those who are running the pipeline that that they might uh into this area which is very important for things like environmental regulation and um you know kind of making sure that that systems are safe um we have uh measuring fish actually measuring the number of uh up a river um and kind of counting salmon migrations as they're up through the river um one other interesting area is a hard hat detection so this is where people are kind of trying to determine um you know if someone has a hard hat in or not and this is really good for compliance and and keeping people safe um and so that that's just a few examples but i think you kind of get the idea basically that anything you have in your head that you think is sort of a replicatable image that uh sort of tell the semantic difference between objects quite quite easily and you think you can construct that into a data set and then you create a computer vision model that that idea that you're having is probably um actually pretty well founded and getting started with a tool like rebel flow or anywhere else with object with with computer vision technologies is a great way um to get started and just kind of see if if the thing that you are is actually feasible um and sort of for a uh for a fun example here we actually have uh one person building an application to identify different kinds of sushi so as the sushi is coming through on a little um conveyor belt in their restaurant they're identifying what what type it is and then they're able to uh eat uh eat different kinds of uh or they're able to identify different kinds of sushi for the customer um so that's just a few examples that uh thousands of developers are coming through rebel flow and uh using using computer vision for different technologies but certainly this is a movement that's uh much broader than us but i think uh i guess here we just have some proof that um all these and if you have an idea of using something then then i certainly recommend uh getting started in and reaching out and kind of getting some traction on it so now i'm going to talk a little bit about since you know this is a data science meetup we're going to kind of get a little bit more uh in depth on exactly um what the process looks like to um go from your images to to the model so basically this is kind of the way that we see things which is there's a few main steps um so you'll have your images um and you'll need to gather them but then once you have a data set of images you you want to label these so uh this is the labeling step where um basically you're providing annotations to the computer able to understand um what exactly you're trying to teach it to do so um the most uh the thing you usually start off with is classification um which is just kind of like um making a distinction between if this is an image of some of one thing or another thing so you could say um i i recently trained a classification model to uh tell whether a flower is a daisy or a dandelion um and uh that's that's sort of like uh the basic level then there's other types of labeling you can do where you can actually draw a box around the um around the object that you want to detect and that's called object detection um that is a little bit more popular and a lot of people are using that for a few reasons because you can count actually the number of things in the image and then you can also localize objects in the image um and then that that's kind of a little bit more exactly like the way that humans see and then going even further is there's a modeling technique called semantic segmentation we're actually drawing lines and outlines around objects so that's kind of more even getting closer to the way that we actually see things which is uh via outlines but in any event you'll you'll make labels uh your data set and then this these will be your annotations which are sort of informing the computer as to uh what what things that should be paying attention to and what things it should be learning to model um and there are a few solutions there um there's all sorts of open source free solutions which are are things that i recommend like um cvat um or vot or um or uh yeah cvat and vot and label image um these are all free things that you can just um you can uh start labeling your images for free um i'll provide some links there for for those as well and we have a lot of content on the roboflow blog of how to label um but then there's other solutions like if you want automatic labeling you can use amazon mechanical turk or for really big projects you can use scale ai anyways there's all these labeling services and all of them whatever the format data format is it can be exported which is something that we're very big about about supporting and we make sure that you can ingest your data in any format came from and better yet rebelflow actually has a lot of public data sets that are already labeled so you can just kind of go ahead and lift off the data set that you found and sort of get started right away with something that's already been constructed especially if you just want to learn about the technology i highly recommend just kind of taking a public data set labeled and sort of using it to just get a feel for for the technology then uh the next step in the computer line is to organize your data so you can imagine the first time you go through it's pretty easy to just kind of have one data set and you'll just be kind of pushing that data set through the process and no no problems but at the end of the day running a ton of these different experiments you're going to be wanting to just hear rotate images here or you're going to want to actually not try to model this class and going to be all kinds of different uh branches that your your data can take and it's important to kind of keep that uh process in an organized fashion another thing that roboflow helps you do um is you can kind of version your data and you can be organizing it and kind of all in a in a um clean fashion which experiments you're going to next of course all that done outside wrote before via folder structures and such but um this is this is just kind of uh one one tool where we we are trying to add some value in the in the pipeline um then after you kind of have your data or organized uh you can process it so uh process in incorporates uh two two distinct things here uh the first one is uh the pre-processing step so pre-processing is uh some sort of image transformation that you're gonna make that's going to span uh all your whole data sets so across your training sets this is something that you're doing to kind of standardize your data to basically remove some elements uh that you want to kind of simplify your model or simplify the um one example here would be um the probably the most common example here is uh the simple resize step so you might want to make a smaller model by resizing your image down um and if you're doing an easy task then resizing down is going to be a really good way to kind of make a smaller model and just send in a smaller image but that's an example of a pre-processing step another example of a pre-processing step might be to use a static crop so you're um then for a static crop you would be actually cutting out just a small box um and because you know that the object that you want to be focusing on is always in the same spot um so that's going to apply to all images and it's just something that you kind of take for granted as a standardization step across the whole data set the other kind of processing that we allow you to do in roboflow is data set augmentation and augmentation is something that only applies to your training set so only to the images you're going to be showing your model and it allows you to generate more training images from the base images so this is where you're kind of trying to think i have this data set and i want to try to think ahead of problems that my model might be having in advance and i'm going to kind of sniff these out by thinking about ways that i can transform the base images that i have into a way that will sort of model the future and sort of anticipate the way that uh the things that the model is going to be struggling with in the future so for example um if you're uh making application to uh model receipts so you want to be able to cut out little boxes out of a receipt you might be zooming in or zooming out with the picture you're taking so it would make sense then to zoom your images in and out and this would be we could do this via a random crop so the random crop augmentation kind of can kind of crop your images randomly or bind them out which is simulating the the zoom effect a lot of more advanced augmentations like mosaic data augmentation where uh you can do clever things um that it's sort of like kind of me it's more so um sort of addressing problems in the way that that the model uh is memorizing things like so mosaic augmentation takes an image and it tiles it into four different tiles and then um stitches all those tiles together and then kind of makes a new image out of out of four different images and what that does is it sort of teaches your model to localize in different spots because your objects might kind of all be uh i'll all be localized in a single spot and be memorizing that location so that that's that's an example another advanced example of of augmentation these are all things you can do in rebel flow and you can do with python libraries um to sort of make your data set even bigger so now we've gotten through the first three steps um and uh after after that you you want to be starting to move into the modeling process so this is actually where the machine learning comes in and i'm sure a lot of people here are really experienced with getting hands-on with machine learning models and uh we'll be really excited about this step but um sort of the hard part is done you've gotten your data set um into the spot you want and now uh you want to get it through into into a model so um that's where you can do this there's uh makeml there's pie torch there's tensorflow um and just like the way that rebel flow will ingest any um labeling tool that you have or wherever your data set comes from it will also export to any training environment that you might have and this removes a very tedious step in which you often have data in one spot but you want to get it to another and so you have to do all these complex data transformations even though at the end of the day the underlying semantics of a bounding box or a classification label is the same but you want to make sure that the data is transformed into the right format so then you can send it through into whatever uh the training pipeline that you want to start training so during training uh the model will be the feedback that you're giving it via the images and it will be using gradient descent to back propagate through its parameters to start learning um kind of the structure of of the thing you're trying to teach it so that's uh that's the training step and i'll show a little bit more on that on the uh via the rebelflow model library and uh the rebel flow train features here in a minute um then after you've trained your model this is where things can uh often get a little bit more complicated too is is deploying it so um the uh blog post that alicia was mentioning was where he trained an efficient debt model here in the trade step and then he went over to deploy it onto a raspberry pi so this is where you're taking a model that you you've built and you're actually bringing it into real life by putting it onto a device and and that is where a lot of people actually spend their time where they're actually just working on just taking a train model and putting on deploying seeing how fast they can get it to run um and and big thing here big concept is real time uh so the way the human eye sees is is that um so many frames a second and people are trying to get that to the exact level that it seems like the model is actually detecting things in in real time um but deploy is a very important step and it's kind of what ends up bringing your model to life and then once you're there um you actually display the results via however you want maybe you're you're just tracking things in a database or you're displaying them on an app um the next step is that's uh that's that's on to you so this this this is where kind of rebelflow sits and helping the developer uh push their way through the computer vision pipeline um but uh at the end of the day the developer will be kind of seeing the whole landscape here um and making their own application and coming up with their own their own usage for for the technology so that's a little bit on the computer vision pipeline i know i covered a a ton of ground uh throughout that whole uh that whole process there um so feel free to uh drop some questions in the chat and and we can uh kind of zone in on your own uh individual interests on that on that step down the road so i know i keep talking about examples of applications but this is a little of uh what the rebel flow founders were able to do with computer vision um so they're originally interested in um uh augmented reality so here's an example of magic sudoku where that kind of can scan over uh sudoku board automatically identify what numbers are in it and then send that into uh the revolver so if you've ever seen sudoku this is kind of a cool augmented real example of uh how to how to do it um uh i was going to talk a little bit about computer vision modeling the train step here because this is kind of my personal interest so as as you're doing all that data management part that's all that designed here to uh create the right thing to pass through your network so this is this is an example of a network where the images are being passed through convolutional layers and they're forming features based on based on the underlying image and then there's a whole kind of downstream modeling process where all these are being mixed together and then the prediction is coming out at the end so it's important here to think like this this is a pretty complex process and you really want to make sure you get this original input step right uh before moving forward and that's that's uh going to really really help you out um so a little bit of a pause see uh alicia if you have any thoughts and um then we'll a rebel flow live demo cool cool cool so um let's just wait for uh you want to take some questions is that what you want jacob yeah sure this could be a good time to uh to take some questions okay okay so that was just happens from you guys let's have some few questions from you guys before we go live with the uh demo all right guys so questions questions questions okay so if there's no if there is no question then um we might just have to i think it's because the thing is a bit lagging behind that's the reason why we are we haven't seen we haven't had our responses oh yeah yeah that makes sense yeah i know there's a couple minute delay um yeah so i i did see actually one i did see one interesting question come through a little bit earlier uh so i see someone asked um oh that i started as a quant and why did i switch um so this is a good this is good um i i think um at the end of the day i was very much interested in in quant uh finance because uh because it was a really complex problem and and i got to use a lot of interesting technologies and it was was a good way to make money in in a in for a little bit um but uh at the end of the day i wasn't super fulfilled with the work and i wasn't really like felt like i was building anything for the world it was just kind of like you make an application you spin it up it's up there and then you don't really see any of the impact so the fact that you know rebel flow is is kind of already reaching uh corners of places like nigeria and stuff is like something that really gets me motivated and like that we're actually you know computer vision is something that's actually going to transform the world and i think working on it is is a really good thing for for uh humanity to just kind of move forward with technology and build new things and make the world a better place so um ultimately that's kind of why i jumped ship and honestly i think on the tech side a lot of the problems you'll encounter are just as interesting um and um so that's that's my two cents on that topic okay awesome uh jacob so we are having some interesting questions opening up from my screen over here um so somebody is asking uh okay okay okay for you that you have staff today i hope there will be a video for us to watch later yes it's always unbelievable on youtube on this channel um okay so somebody asked can you use roboflow to detect moving objects uh yeah so um rebel flow allows you to so you can uh kind of detect each object in each frame um but that is actually a good point of uh sort of getting a sense for the same object being localized and i will say that that's an area that i would say we probably are able to help in the data pre-processing and augmentation aspect but being able to localize the same object um involves models that i personally haven't um started to get uh get acquainted to yet but um please be curious to hear more about about the project and um and uh what what you end up finding with it so also so i hope that's actually a such question roboflu connection is actually used for you can use it for moving objects all right so another person asks i'm wondering how transfer learning fits into this i have a small little data i want to find you on okay um i think that's a complete question if i could rephrase it um he's trying to hacks how you can use transform learning using openflow for um is small data yeah definitely all right so would you want to answer that question i don't want to answer any question for you i'm going to ask any questions for you so okay okay so what's going to be the answer regarding that yeah so um on the transfer learning front um rebelflow uh does certainly um uh help you uh use transfer learning so as you're making your um original data set you are kind of tuning into a certain area that you want to be modeling and then a lot of our tutorials on modeling will show you how when you take your images you can send them through to train based on a pre-trained network so most of the network networks you'll leverage especially in object detection will be um they'll be pre-trained on the coco data set which has over 200 000 images with 80 classes so this is kind of designed to teach the computer how to just identify sort of any small feature in an image and then uh after uh after the computer sort of learned to identify uh those basic features and then you apply it to your own data set uh you'll see good results um so um so yeah and then and then furthermore um if you want to be transfer learning um from uh you can actually kind of if you start to have a ton of experiments then you can transfer learn from uh checkpoints like maybe you already have a aerial satellite imagery checkpoint um where you're kind of like you already have started a lot of different experiments there and you can you can start um from that spot so definitely good to look ahead on on transfer learning uh it's definitely an advanced advanced technique but i think it's super important a quick one we're going to just answer uh two more questions that we can go into the live you know demo session because that's actually the most important parts so i'm just going to take the next two questions somebody said um how you switch ripple flow in finance how'd you switch your book to in finance oh yeah um so i haven't i haven't done anything on that um we have some people who are like uh processing receipts and stuff um so they're doing kind of like the back end uh back ends uh like sort of uh processing of documents and and financial documents and such by identifying portions of text and then sending that into a ocr type app um but other than that um i don't know i guess the application is sort of up to you i know there are some people who use computer vision to like estimate how how uh heavy or light an oil tanker is as it's uh crossing the ocean so it's like you're actually kind of getting an idea for um flows of oil and then you can make trades based on that but um i don't know it's up to you yeah yeah actually actually it's it's it's engineering question actually it depends on the aspect of what exactly in finance how you it depends exactly on what exactly you can actually do that for i will basically see more oh see how ocr actually but um let's just quickly take this last question all right he said as for a non-python developer it's a p-corporate flow to use for predicting production level in a manufacturing company that's a very very important question and i really want you to answer that question yeah so honestly should i repeat it oh yeah i heard it um uh the the non-python applications okay awesome yeah um yeah so uh that that is a question actually that's um sort of very top of mind for us these days which is um as as i go through a live demo here you'll see we have this data management platform which is all no code it's kind of you're up on the um you're up on the application and you're able to click through things and then we would traditionally kind of funnel the data into a python type environment where you could start modeling things on your own so a new feature that we worked on is sort of automatic training so you can take your data and use rebelflow to just one click automatically train your model and then you have it up on an api endpoint where you can pass images through via command line or no javascript or whatever you're developing in um and uh so that's uh that's a feature we're really excited about um rolling out here in the next couple weeks um but but there is early access so um you can always uh reach out to me and we can always try to get get started on um on the app but yeah we we do have a few manufacturing examples where people are just hitting the api and and um they uh they didn't they didn't write any python code to train it they they're just just kind of leveraging a trained model exactly exactly for train modules that's that's almost the easiest way to do in fact the one i did the one i did the other detection efficient dead on you know temperatures and uh it was it was to treat buddhists it's not going to be very easy you know training from scratch and all these things all right so um jacob i will i want us to go into the live demo right now so all questions can all other questions i'm seeing a lot of exciting questions coming up on my screen over here there are things that we can uh answer after the uh after the live session of course they would see the live session and you they would they would have answers but yeah this is the live station all right so can we begin the live station sure sure here i will do that yep okay let's see sorry one second okay um you guys can see here yeah okay great um so yeah so if you want to get started i mean if you're at your desktop here you can actually go ahead and go to rebelflow.com and we can we can sort of do a little bit of an example here but yeah so this is uh rovo.com we say transform raw images into a trained computer vision model in minutes uh i think i think minutes is uh is a little bit uh fast but so that's the idea it's supposed to be supposed to be quick to uh to get there but um so i'll go ahead and just sign in um and uh when you log in you'll you'll see a screen here um and the screen um basically is sort of like your your command center for your data so you have your data sets um and they're all here um there's different versions of things so here we have tires and chest pieces and phones and such those are the things that i've been thinking away on the last couple of days um but uh you you get in and so what you can do is to get started you can just hit create a data set so here i'll create a chess sample for youtube with annotation class pieces and then the other thing here is to decide what kind of data set it is so object detection is the most common which is what mostly is coming through which is just kind of the bounding boxes um you can choose to just classify your images or segment them or you know there's uh there's the types we're we're talking about but we'll do object detection here um and then you go ahead and hit create data set uh and then this is the upload flow so this is uh uh the area here where roboflow developers have made it uh super easy to just kind of bring in whatever data set type you have so here um for example i might have this chest sample and it's uh here there's example images so there's there's an image of a chess chess set and then here's the labels so these could be in whatever format kind of doesn't doesn't exactly matter um but but you can you can bring your your data set in in any format so you can just go ahead and drag and drop your data set in there this is all in the uh onboarding flow too so um you can kind of get started we like to use chess as an example because there's all the uh the chess ai uh links so here um we'll we'll kind of uh use the chess set to uh uh sort of get a feel for for the way that things are are working um so here's a a new preview we just put in these new previews we're really excited about um where you can see uh your different classes um so and then after your data is loaded in you can go ahead and hit finish upload and then you have an opportunity here to trip split between train valid and test so the train set's going to be what we show to our model to learn from the validation set's going to be what we check we check our model with each iteration to see kind of how well it's doing and then the test set is thing we hold out for the very very end uh to evaluate on and and uh really make sure that uh the model did a good job of modeling on our test images um so there we go then that go ahead and that that uploads it up into the removal flow cloud so one one thing i should mention here is roboflow is free for the first um 1000 images and so you can kind of be getting a feel for for using it for projects and and oftentimes you know that's going to be that's going to be plenty of time to get started with your um or that's going to be pl plenty of images to get started um with uh with whatever your project is because the models are getting better and better now so they're they're able to uh they're able to um learn very quickly so once your data set is loaded in um you'll see it here um and it will be kind of in this uh data set version page um and so this is again where you can kind of just get a feel for like oh did i get my images in um in the correct fashion you know are they here are the class labels um all here and you can sort of do some data set introspection there the other thing i really like to use is this data set health check um so this uh basically lets you see how many images um what are the classes so um makes sense for chess right that the pawns are the most common and then going down the line with different pieces and then the other thing you can see is the annotation heat map which shows you the localizations of your objects so i found this to be useful sometimes if the model is like unable to identify an object in a given spot then that shows me that i should gather some data that's a little bit more um varied or in different places um and uh and yeah and so that's that's kind of uh it for the dataset health check um and then back into the modify stages this is what i was talking about in the um processing step uh for computer vision where you're trying to take your images and you're you're processing them ahead of time so uh pre-processing is uh sort of the first thing um where you can do those things like i was talking about like static crop um or you can do like a um tiling this is something uh that can also be useful or uh you can uh modify your classes so maybe i wanna rename or only export certain classes uh you can do that here um and then uh the other side of the processing is uh augmentation step so this is uh the part where you're actually making transformations to your images so you can flip you can flip things like this or you can um you can rotate them like so um you know you can be making kind of like i said trying to see ahead to problems that your model is going to be have having as it's modeling things and making those augmentations to make your training set even bigger um so this is where you kind of decide how many multiplications you want so you can go up to like 50 generated images per um per normal image so then once you're ready to go to make your new version you uh just hit generate images and that um that basically goes through and this is going to just send that processing job to our back end and all these cloud servers will be spinning up and doing all kinds of different things especially if you have a big data set um and uh you you just don't even have to think about all that stuff it just kind of shoots out um and uh yeah so then um then the big thing here is the exportation exporting so um this is what i said is really useful is like some people actually just use riboflow just as a conversion tool so they'll come in with one dataset um annotation type and then they'll come out uh via a different one so these are all the ones we can export to i'll go ahead and click yellow darknet here and so this this gives you a zip or download code of where to get your data um so this here is kind of a example of behind the scenes of like kind of where you would go if you wanted to actually start modeling and start doing experimentations yourself you could take this link and go forward as we will here as we continue to go with the live demo um but um but otherwise you know the thing i was talking about is this roboflow train which is something i've been working on pretty heavily lately which is to just kind of spin up a gpu on the cloud load in your data set do all the training and post the endpoint um without uh without having to go through all all of that yourself but certainly there's there's different there's different cases here where like if you want to be using rebelflow more hands-on for all your different experimentations and such to to get a really good feel for computer vision uh modeling on your own then then um you want to go through the export route um but yeah so here we can see we have some different versions going on with our data and um now we have a data set that's been transformed in these ways um yeah so so now you're uh you're ready to move out of the rebelflow platform and and uh into modeling um so uh going i guess before we uh before we go into that one one last thing on the dataset side which i want to show you guys is this public datasets link so this is something you can get started with without even gathering your own data so this is the blood cell detection data set which i use in a lot of different tutorials and i think it's a really good way to kind of get started with things so public data sets um you can go ahead and just grab one of these and uh just sort of get going on this next step without even having to having to gather your own images um yeah so another thing i want to show you guys is this rebel flow model zoo um so this is uh our model library which is uh i think a really cool place to get started i know alicia and i are working on adding a better version of this because i wrote this one but this one's not so good but here you have all kinds of different models where you can deploy things so the classification if you're doing classification right now i think this resnet 34 is a really good one to use um otherwise this yellow v5 is a really good one to use and uh this is object detection and so we'll we'll have a few things we have a video uh where you can see us kind of going through it in video as we're going to be sort of showing here um there's uh also the colab notebook um which is a really good place to get started so you can just kind of jump over to the code um so yeah so this is uh this is an example notebook here um where we're in colab which is uh google's providing us uh free free gpus which is really nice and and you can kind of just log in there and um so you know this whole process like i said just getting started uh doesn't cost you anything it's just just a resource out there to to um start familiarizing with the technology and get started um but yeah so um i'm not actually going to run through this uh run the code but i'll just kind of go quickly through the steps here so we can kind of go from the theory i was talking about earlier into how it actually uh manifests itself in in um in a real training job so here uh we'll just take a peek at the gpu so um this is uh what google collab provides you for free um which is pretty exciting so uh that's that's the gpu coming in um and then here we're using darknet which is uh where um the yolo model came out of so this is a very popular small object detection model that runs really fast and so any any notebook you're going to have is going to start with a step like this where you're like kind of installing dependencies and you're building things and hopefully rebelflow's gotten the code right so this doesn't break and so you can just kind of move move forward without having to redo it um but we also have our our support slack if you if you have any questions on that and then here's the step where roblox comes in so you can download your data set in um and uh it's coming down and here are all the images this was from the public blood cell uh set that i was talking about so now your images are coming in they're loaded into your computer um and you want to tell the computer where the files are points the files to the images and then this this part you're like making the network so the stuff is usually you don't have to edit it but um you can later if you want to get get into the weeds um but uh yeah so you build the network you load in the images and then um here's here's the network actually printed out and then you get down to this step and you're like okay so i got these images i got this network defined i'm going to run the training job and i'm going to start passing those images through the network and start training it and uh you watch here as the loss goes down and your model's starting to learn and you're getting excited and you can bring in more images later um so that's the training step and then once once it's done training um then you can actually use it so you can be in the notebook here doing a little research and you're like okay so now i'm gonna uh pass an image through for inference um so this is an example of an image that the model has never seen but it was trained to understand these objects so it knows what a red blood cell is it knows what a platelet is and it knows what a white blood cell is now um just by being shown training images you know we don't have to install open cve and do a bunch of complicated color matchings to to get these things you know it's uh it's just learned from the data so um that's a good way to get started i think and you can export there you can run it on video you can rent it on webcam and then kind of build uh build applications around it so um so yeah that's an example of uh getting started and uh the roboflow uh suite of technologies and and things that we have available for you guys so [Music] we can keep the discussion going [Music] um [Music] okay okay all right but guys before we go ahead to you know answer a lot of questions from you guys um i said something from the start of this program it's very very important let me go ahead and share my screen it's very very important guys i want you guys to um i want you guys to i want you guys to tweet at roboflow.ai want you guys to use the hashtag hashtag computer vision hashtag ai hashtag ai you know i really want us to you know see uh you guys do this do that let's do that let's do that all right so um yeah let's go ahead to have questions so jacob we have some questions here and i'm sure a lot of them should have answers by now um [Music] okay the first answer says can roboflu okay you've answered the foster cake now the one i want you to understand before the first one i want to join is can rubber flow and you predict streams directly uh yes yeah um so that's um sort of in the so after your model's been trained on the images um then you can run the model through a video stream and you'll see um the predictions coming out um of it so but but obviously that's um like once once you get to that step and you have the model ready to go then that's kind of where uh the work comes in into being able to run it in real time and such but um that's uh yeah super important step okay okay um i think the question is more than just like asking if you know who has this um um everything you need into developing [Music] a computer vision model whether you have um you you you person you you can do your people you can export your data in terms of the example of files the object detection aspect you can do classification semantic segmentation and a lot of things like that so the person is more or less like more specific in the sense that is there you know after doing all of all these things such that you've been able to export the model you know train um is that more or less like you're just passing the real time the image and then you get your results in real time yeah um so instead of going uh so instead of actually going through the routes i went in my um in my um article that i wrote last week so you know i had to pass do all the pre-processing in roboto export the data then um load them into google coll