Build a Machine Learning App From Scratch with Flask & Docker

hi everyone in this video you will learn how to dockerize a machine learning application and this tutorial consists of two parts in the first part we build our flask app and apply the machine learning model and in the second part we dockerize everything and for this i actually got help from francesco who is a docker captain and a real expert in this field hello everyone and thanks for the introduction i'm francesco docker captain and i like to help developers understand continuation concepts you can find me on twitter and youtube in the links below in the second part we will take care to containerize the application that patrick is about to create let's start so the second part is available on his channel and i will put the link in the description below and now let's get started alright so let's start from stretch by developing our machine learning model and our web app and for this i have one empty directory that i called app and then i have another one where that i called mldf so this is where we do our machine learning development and yeah here we don't start completely from scratch so i already have the data here and a notebook where we train and come up with our model so this is not the focus of this video but you will learn how we transfer our notebook to a production ready web app so let's do this and yeah let's go into the ml def folder and if you are working with machine learning stuff then the chances are pretty high that you use a conda environment so i want to use the same in here so we start from scratch by setting up a conda environment so i say connor create minus n and then i give it a name and i also like to specify the python version i won so here i want python 3.9 so hit enter and you don't need to use a conda environment but i recommend that you use at least one kind of virtual and management so you could also just use the build in virtual and this works just fine but in this case i want to demonstrate how to do this with a conda environment so now we activate it and the great thing later will be that when we have our docker container then we no longer have to worry about this local setup here so then everything should work in our docker container but for now we have to set up this locally so yeah now we have this activated and now we can install all the packages we want so the first thing i install is two pointer and ipi kernel for the two fighter notebook support then i also want to install the a kernel for this environment so i say ipython kernel install minus minus user minus minus name and for the kernel name i give it the name as the environment so also nlp and now i need to install all the packages i need to develop the model so in this case i'm going to use scikit-learn but the same approach works also with tensorflow models or pytorch models so in this example i want to keep it simple but again the approach is the very same for any kind of machine learning or deep learning app so now we have this and in this example i need one more package and this is the nltk package to work with text and now we should have everything we need so now we can start our notebook by saying two fighter notebook and this fires up the server and here we see the data and this is the notebook i prepared so this is the notebook and i will upload this to github and put the link in the description and now we go over this only very briefly so what we're doing here is we work on a project to classify twitter tweets and say if they are positive or negative so the first thing i want to do is change the kernel and make sure that we use the kernel from the environment that we just created so we select this nlp and then let's run the first cell and see if all the imports work and here we get a error module not found so i forgot to install this in the environment so let's go back to a kernel to the terminal and then activate this in a new window conda activate nlp and then we say conda install pandas and hit enter alright so now we have panda so now let's go back and run this again and now the imports work so yeah and then now let's go over this briefly so here we load the data and again i will put the link for the data set in the description below so you can download this as well and then we do some text pre-processing so we here we define emojis and stop words and then i created one function pre-processed that gets the raw text data and then here for example i remove your l patterns the user patterns and then i remove stop words and i apply the slammatizer and the lemmatizer is a popular technique in nlp that groups together inflected forms for example better and good so we instantiate this here and this is available in the nltk package and when you run this the first time you might get a error and a note that you need to execute this command so we have to say import nltk and then download this package and this package so this is important because we have to remember this for later because we have to do the same commands in our docker container but yeah this is the pre-processing so we define this function then we call this with the pr and get the pre-processed text and here i do a train test splits and then i set up all the different models or pipeline steps i need so the first thing here is to use a vectorizer and so i set this up and call vectorizer fit and then vectorize a transform and then i try out different classification models as last step so here i have this helper function to evaluate the model and yeah here i create different models for example this bernoulli naive bayes and then call model fit and model evaluate and yeah once we've done this and decided for the final best model now the next thing to do is to save all the different steps and all the train models and then load this in our application but in this case i want to show you one best practice that you can use with the sk learn package and for this we use the sklearn pipeline and the way it works is that we define a pipeline object where we put in all the different steps so the first step is the vectorizer and the second step is the final classification model and then we have everything together so now i have to only call one fit command and then one predict command and don't have to call the vectorizer and the bernoulli model separately so yeah i recommend that you set up a pipeline like this and then here i evaluate it again to see if this still gets the same results and now we have to only save one trained pipeline so after having set pipeline dot fit we now save this pipeline and for this we use the pickle module and then we load this again and here i call model evaluate to check if this is still the same and yeah this approach is the same for every machine learning pipeline so if you use tensorflow or pytorch then also at some point you want to save either all the steps separately or maybe you could save the whole pipeline in one step and then later in your application we call a model load again so yeah this is how we save and load the model and then here at the very end i have one function to predict the text so here we call this pre-process function so this step is still separately because this is not a model that we can put in our pipeline but then here we only call model predict so this will apply the vectorizer and the final classification model and then it returns the label so here i test this with some example tweets so now let's run all the cells and hope that this works all right so the code run through and now we see the final output so here we have the classification i hit twitter this is class zero which means negative may the force be with you this is one which means positive and then here again we have one negative tweet so yeah our classification works and now we have all the code we need in the notebook and now the next step is to transfer the code that we need from here to our application so let's go back to our folder and if we go to the ml dev folder then here we should see our pickled file so this is the pipeline so the first thing i want to do is i want to copy this over to the app and now let's go back to the terminal and here we can quit the server so we no longer need our notebook and now let's go back and into the app folder and here i fire up my code editor so in here the first thing i want to do is i create another folder and i call this api and then in here again a new folder and i want to call this models and then i move the saved model into this folder so now we have this and then in the api folder i want to create two files so i want to have one app dot pi and and another one and we call this utilities dot pi so yeah so now we have the after pi the utility stop pi and the models with the saved model and now the first thing to do is to bring over the code from the notebook to the utilities file so let's put this next to each other and now let's bring over all the code that we need so we need a few imports so we need import ray and pickle then we need this lemmatizer so let's copy and paste this and then if we go down so we don't need this data set but we need the emojis and the stop words for the preprocessing so again let's copy this and i will make this smaller for now so yeah here we copied this then we need this pre-processing and the lemmatizer so let's copy and paste this whole function and then again here we don't need this um train test split stuff and we also don't need the stuff where we set up the different models and fit them and evaluate them so the only thing we need at the end we also don't need this pipeline stuff we only need this where we load the pickled file again so let's copy and paste this as well and here we have to be careful so this is now in the models folder so we say models slash pipeline dot pickle file and then let's go down further so yeah we could also use this predict function and maybe this with if name equals equals main to test the code so now we should have all the codes that we need and now let me reformat this a little bit so i want to put the lemmatizer up here so that we see it right away then i also want to have this where i load the model at the top so that i see this right away as well and then let's create one final function and i will call this define predict pipe line and this only gets the final text and then here we can say return predict because this function still gets the model so down here it gets the model and the text so here we say predict and then we put in the load at pi and the text and then this is basically the only function we have to call and then down here here let's say predict pipe line and then it only gets the text and then i want to i can close the notebook and make this larger again and now let's open the terminal in here so in here i want to run this and make sure that this works so now let's make this larger again and the first thing in the terminal i want to do is make sure that i'm in the correct environment so like you can see here this is a different one that was activated by default so i deactivate this and say conda activate nlp and now we first let's go into the api folder and then we can run python utilities dot pi and yeah this works so here we should now see the same output that we've seen in our notebook so this works as well now we have all the code for our model classification that we need here and the last thing to do is now to implement the flask app so for this we use flask and of course we have to install this and here i'm simply going to install this in our conda environment but i install it with pips so i can say pip install flask and this works just fine all right so now we have flask so now let's import the stuff we need so we say from flask we import flask then i also want to sonify and i also want request and then we say from [Music] utilities utilities we want to import the predict pipeline function this is all we need from there then we set up our app so we say app is our flask app with the underscore name convention and then we need to create one function and we call this let's call this predict function and for now we say pass and we have to decorate this of course so here we say at and then app dots and now this is new since since flask 2.0 we can say app.post so for a post request and then we define the url endpoint so we say this should be at slash predict and here what we want to do is we want to basically send a text to this endpoint as json data and here we get the text and then call the predict pipeline so the first thing we do is we say data equals request dot json and then since we want to deploy this to production i also want to do a couple of error checking here so i wrap this in a try except block and then say the sample equals the data with the key let's call this text so now the user should put in a json file with the key text and otherwise we get a key error so we say accept and then we catch a key error and here we can say we want to return jsonify and then this should be a dictionary where we put in the key error and then as a message we could maybe say no text was sent so yeah otherwise it works and then we have the sample and this is a raw string right now so we want to put this in a list so we need a list with at least one element so we say sample equals a list with this item and then we call the predictions pipeline so we say predictions equals predict pipeline and this only gets the one sample now and then again we could and do some more error checking with this but yeah i leave this up to you but now for the very end i want to do one more error checking so here i want to say result equals and then again i want to chase sonify this and then i put in the first predictions so this should again have only one element so i can say predictions and then zero and if this does not work then this will or this could maybe throw a type error and i show you why in a second so we catch a type error as e and then here again we can return a error as json object and here i want to make the exception message as a string and otherwise always say this is also the result and then at the very end we return the result so yeah this is basically all we need for our predict endpoint and now again i want to say if underscore name equals equals underscore main then i want to test my flask app and i want to say app dot run and here i give it um host the host object equals as a string 0.0 and while i'm debugging this and trying this i can simply still say debug equals true this uses the internal flask development server so this is fine because later we use a production ready server and not this one so now let's test this again by going to our terminal and then this case we run python app.pi and this should start the flask server so yeah this is up and running and now we can go to a separate terminal and then use a curl command to send a post request to this endpoint so let me copy and paste this in here so we say curl and then a post request and we need to send this json data with it so here we have the key text and then we use this tweet and then here we have this endpoint at part 5000 and then slash predict so let's run this and see what happens and we get a response back so the endpoint work but we get an error object of type in 64 is not chase and serializable and so what happened here is that in our application if we go back and have a look at here so here we catch this type error and this is exactly what happened because now if we go back to the predict pipeline and in the predict function so here this prediction is a numpy data type and not a python int so we have to convert this to a integer and in fact i want to do this a little bit different so instead of putting a tuple in here i want to create a dictionary so this will look nice in json format so let's create a dictionary from this so this needs a key so here we put in the key text and then the text then we put in the key let's call this spread for prediction and then here we want to put in the label and yeah this is the label to prediction or other way right the other way around prediction to label so yeah so now we have this so let's save this and run this and this is still in debug mode so if we go back then this should be reloaded so now let's do the curl command again and now it says connection reviews so i think i accidentally quit it this because here i have a syntax error so if we go back then of course here this should be before the colon so now it should work so now let's run our app again and see if this is up and running so yeah now our app is running again so now let's send a call command again and yeah now it works so now we get the correct prediction back where we see the label the prediction and the original text so now let's also test this with the negative one so here we send the text i hate twitter and then we get the negative labels so yeah this worked so now our app is running and now we want to dockerize this so before we do this i want to do one more thing so in our api folder i create one file requirements dot txt and this is where i put in all the packages that i need so um what we can do here is if we go back to the terminal and quit our application so since we used a conda environment we can say conda list and this will list all the packages that we have here but here we basically only need to grab the ones that i installed in the beginning so conda installs a bunch more so we need scikit-learn and we also see the version here and i highly recommend that you put in the version in the requirements txt file as well and pin this because if we have a difference between the versions where we trained our model and then later when we load this and do it in our app if there is a difference then our app or our model could break or produce different results so we want to have the same version for scikit-learn and i recommend the same if you use tensorflow or pytorch so we pin the version here and then we also need the nltk so let's search for this so here we have nltk and then let's also copy and paste this and pin the version to 3.6.6 and then we also need flask so let's search for flask and copy and paste this and yeah for flask it doesn't really matter but i recommend that this should be at least 2.0 because then we can use this app.post command so let's pin this to at least 2.0 so we can say greater equals 2.0.0 and then this is all that we need and i also want to write a comment here so i copy and paste this so this is a command that we need to import nltk and download the packages for nltk so i put this here as a comment as well and then yeah francesco can see this and now we have everything we need for our web app so we have this one directory with the api and here we have our flask up and running and yeah as a next step we want to dockerize this and as i said in the beginning you can find this on francesco's channel so i hope you enjoyed this tutorial and then i hope to see you next time bye

Channel: Patrick Loeber

Views: 53,225

Keywords: Python

Id: S--SD4QbGps

Length: 25min 39sec (1539 seconds)

Published: Sat May 14 2022