Nishant Rao Guvvada

Hi. I'm Nishan Guwada. I'm currently working as a visualization, uh, and analytics specialist at Rackspace Technology. My text tag includes SQL, Python. I'm currently learning JavaScript, React, uh, Express, basically the Monstag. My professional journey began as a PA developer, building, uh, dashboards and reporting solutions and even, uh, automating sub processors using Python. Due to my inclination word, Python and automation, uh, I was given opportunities to, uh, get into the machine learning side of things. For the past 3 years I've been involved in multiple machine learning projects. Uh, during my tenure at better.com, which is an online mortgage provider, I, um, uh, created a a proof of concept of, uh, basically, a mortgage approval prediction system, um, and, uh, that was using a simple logistic regression. My tenure at Better, uh, was for around 2 years. Better is a startup when I had joined it. During my tenure, uh, here at Rackspace, I was again, uh, due to my inclination into machine learning and I think I was, uh, recommended for Google's advanced solution lab for machine learning which is a 2 months, uh, cohort, uh, where they brief you about the basics of machine learning, uh, as well as utilizing their own, that is Google's machine learning tools such as Vertex AI, AutoML, etcetera. My during my tenure, uh, at Better here, uh, I've been, uh, involved in multiple machine learning projects such as, I've built a churn prediction system using the land and forest. Uh, that was also a POC. Uh, one of the projects that, uh, that I, uh, moved to production was, uh, via data deliver, which is enterprise level chatbot. It is meant for providing insights into, uh, basically, if a particular information exists in a particular, uh, place, uh, in within the enterprise. That particular chatbot was backed by a rack implemented system. Uh, so, uh, such that, uh, the chatbot is only, uh, has only context with regard to the form and nothing else. And that is something that I don't, uh, tell production. In in my company, we have, uh, different teams for building the model and then deploying the model. We have different team and have been part of both the teams uh, for a certain capacity. And I'm looking forward to, um, to rules which have basically a full time machine learning rule because, uh, right now I'm a cross functional resource. Uh, I'm trying to enroll this again to look into the reporting solutions and automation. Um, but I'm as a cross functional resource also working, uh, in machine learning.

In order to deploy a Python based machine learning model to production environment, there are, uh, multiple stages and steps that need to be followed. Uh, the first step is building a base model and, uh, iterating iteratively building models and defining those models, uh, until you get a satisfactory model that gives you a level of precision or accuracy or whatever metric that you are following. Uh, until unless you achieve that metric, you keep on iterating, uh, and improving your model and, uh, even, uh, refining your data if required in order to get that level of satisfaction. Once that is done in order to deploy a model to production, uh, there are multiple ways. Uh, I have personally used, uh, hosting sites such as replicate, uh, hugging face, as well as, streamlet. Uh, I can just, uh, download the model using, uh, the pickle library. We can save the model, uh, even in TensorFlow. We can save the model and, uh, we can upload those models to a particular container and, uh, uh, we can upload those models to, uh, the hosting environment and, uh, we can run those models there. I have done that personally on Huddl phase and streamlet, um, as well as replicate. But, uh, if I talk about the enterprise level, uh, move to production, uh, I have personally done it using uh, Google's, uh, platform. Uh, the Google Vertex AI, um, is basically Google's AI platform where you can deploy your machine learning products, um, to an endpoint, basically. And, uh, using that endpoint where, uh, the particular, uh, where the particular, uh, model is hosted, you can utilize that, uh, to, uh, you know, make predictions or you can use that particular API and connect it to a front end built on whatever you like, like, Flutter or, uh, yeah.

There are multiple, uh, feature selection techniques that can be implemented, uh, uh, within a dataset. Um, some of the, uh, feature selection techniques are, uh, such as the recursive feature selection. You can, uh, train a particular model recursively, uh, giving the model incrementally the number of features, uh, increasing the number of features every time in every iteration you feed to the model. What that does is that basically trains the model with, uh, a minimum number of features and keeps on increasing that feature or keeps on decreasing the number of features. Uh, that depends on the type of, uh, technique that you apply. So, uh, the one I'm talking about is, uh, recursive feature elimination where you give a particular, uh, model, a set of features, and uh, you train the model on those set of features. And then in the next cycle of training or the next iteration, you reduce to the number of features and train it again. What happens is, uh, the best model performance is achieved, uh, when you give the correct number of, uh, features, and that is how you are able to select those, uh, features. There are other metrics that can be used such as, um, uh, correlation, uh, and covariance. Some other, uh, metrics are, uh, chi squared as well as, uh, ANOVA score. Uh, you can utilize these to, you know, select the features. These are some of the techniques that, uh, you can apply on the dataset to extract out, uh, uh, the best features. Uh, if I talk about, uh, a library, a Python library called sklearn or Skykit learn. Uh, that library provides, uh, some features selection techniques that I just stated. But along with that, uh, a random forest, uh, classifier also has a feature selection capability. And, uh, even the library, uh, provides that particular capability along with, uh, the random forest. And you can use that random forest feature selection, uh, to get the correct number of features that, uh, that is ideal for the, uh, model to perform the best.

Uh, there are multiple techniques, uh, that can be, uh, taken or, uh, methods that can be applied to mitigate overfitting. Firstly, overfitting is that, uh, when the model model is trained, the model, uh, understands or learns, uh, the data very well. At least that is what is, uh, presented or appears to be. But when you, uh, give the same model, uh, new data that it has not seen, uh, then you see that the accuracy of the prediction is not that good. This is because the model has kind of learned the data itself, and that is overfitting. The, uh, steps that you can take, uh, in cases of overfitting is basically, uh, if you can increase the, uh, dataset, the size of the data, uh, so that, uh, model has enough of the data to, uh, you know, to, uh, basically generalize. Uh, you can also utilize, uh, LASCO and Ridge, uh, regularization techniques, which are basically, uh, in mathematical terms, adding, uh, a particular term to the cost function so that now the model not just have to, uh, minimize the cost function, but, uh, another term that has been added, uh, to the cost function. Generally, that term for lasso regression is, uh, modulus of the weights or, um, in terms of uh, residual regression, it is basically squared, uh, of the weights. So it is basically, uh, Manhattan distance from Lasso and, uh, Euclidean distance for, uh, ridge, uh, ridge regularization technique. These are the two methods that you can use. You can also, uh, basically, um, you can increase the complexity of the model in order to mitigate overfitting. Um, and, uh, in that manner, you can, uh, achieve a better performing model. Uh, in supervised learning model, uh, because you have all the, uh, labels there, it is better generally to give, uh, the model more data, uh, so so that it does not, uh, memorizes the data training data and, um, overheads.

There are certain libraries that can be used, um, such as, uh, the SIS library, uh, PSUTIL library. Uh, we can utilize these libraries to, uh, perform multi threading. That means that the execution will happen, uh, parallelly. In the code itself, you can, uh, you know, flip some of the flags as true, uh, in order to execute, uh, the, uh, the you can, uh, in order to make the execution, uh, run parallel, uh, you can utilize that, uh, in order to get the, uh, Python script run parallelly, um, since you you can, uh, have it, uh, on a multi core machine. Currently, I I cannot, uh, actually name the exact library that can perform multi threading. Uh, I believe it is, uh, the SIS and the PSUTIL library because it can also uh, provide you, uh, information on how much memory your, uh, Python script is actually taking. In TensorFlow, uh, there's a certain flag that you can utilize. Uh, I believe the name of the flag is auto auto tune, um, which enables, uh, which enables, uh, uh, processing of of the task, you know, machine learning use case.

Uh, certainly, I can talk about the machine learning project, um, uh, where we had to basically optimize the Python code, uh, to meet the memory and computational budget. So in terms of memory, as I said, there are certain libraries such as, psu till, uh, the SIS library, the resource library, and the module uh, that can help us, uh, visualize or see the amount of memory that is being utilized by, uh, our Python code. And in order to, uh, optimize, uh, a Python code, uh, we basically, um, utilize generators instead of list, uh, which, uh, captures a lot more, uh, memory. Uh, we cleaned up the code where, uh, certain variables, uh, were not being used anywhere else. We, uh, kind of cleaned the code and, uh, optimized it in that manner. Used generators, obviously, instead of lists. Um, so, uh, so these, uh, these small changes basically help us in, you know, utilizing less memory for, uh, for a certain amount of Python code. And that helped us in, uh, you know, to to work within the constraints of the, uh, budget that we had. Also, uh, I worked on a project called BI Data Delivered. It's, uh, basically a chatbot that provides a code summarization or it helps you locate information within the data data warehouse that we have and that's basically Google's BigQuery. There, uh, because, uh, all the components, uh, of the model, uh, was posted and was trained on Google. Uh, for example, we used to, uh, Abel in order to build that model and work on it. So, uh, there we already have the the the metrics that we can, uh, see for, um, for how much are we spending in order to train the model, in order to run our query and stuff like that. So yep.

In the, uh, Python code snippet, um, the normalized data function, uh, is the receiving, uh, data, in this case, as a list, as input. We are utilizing NumPy to calculate the mean of that list of the elements in the list. Uh, we are also calculating standard deviation from the list. Um, normalized data is basically data, uh, is basically every element of the list, uh, subtracted by the mean and divided by standard deviation. Uh, I believe because we are dealing with list, subtracting mean from the list and dividing it, um, using standard deviation. The normalized data, uh, needs to be a list again. Um, I believe this should have been an iteration so that every element is subtracted by mean and divided by the standard deviation. And the normalized data is being returned, And I think this is something that we can, uh, do. There are other, uh, there are other methods to, uh, normalize as well such as, uh, min max scaling that is dividing each, uh, element of the data by, uh, the subtraction of maximum win.

Uh, so in this particular function, we are, uh, we have, uh, predictions and labels as the arguments. Um, the correct predictions are a sum of one every time the prediction is equal to the label. I believe this should be, uh, we can also add an else clause which says 0 if, uh, prediction is not equal to the label. Total predictions is equal to the length of the predictions is equal to the length of the predictions. Accuracy is given by correct predictions divided by total predictions, and we are returning the accuracy.

The main strategy of writing clean and maintainable code, uh, in machine learning projects is, uh, creating modules, uh, based on the functions. Never write the whole code, the machine learning code, uh, in in a single script. We the best strategy is to divide the code basis, the type of function, um, or, uh, the type of, uh, aim or the goal that is, uh, being trying to be achieved. Uh, as an example, uh, we can create a different module, uh, in order to, uh, I know, reprocess the data. That's a model for training the data, uh, the machine learning model and so on. In that manner, we have a structured and clean code. We can also write in order to write clean code, we can utilize data structures, uh, such as classes and functions, uh, in order to create and write daily clean code, um, so that every everything is wrapped in, uh, in a particular class that is that kind of signifies a certain, uh, function. So the first one was dividing the script into modules and other scripts. Uh, second one is, uh, utilizing, uh, the hoops concept, uh, that is classes, basically. The other thing is writing, uh, comments, uh, in order to have that, uh, code daily explainable, Especially in machine learning projects, uh, there are certain lines of codes that needs to be actually explained. So comments are something that would be really handy, um, for other developers who so I've been working.

I hope I've understood the question correctly. Uh, Python facilitates the comparison and decision making process by having because there are, uh, multiple libraries that we can utilize to, um, run our model, um, such as if I talk about SK learn library. We can utilize models from there. Uh, we can run multiple models or train multiple models on a particular set of data, um, and, uh, run them and understand their metrics, the accuracy that uh, they are achieving. And, uh, basis that basis those metrics, uh, if that's accuracy, then we can utilize, uh, this particular metric as, uh, as a decision, uh, uh, as as a threshold for making decisions whether we want a model, uh, that is providing this much accuracy versus the other model. And because Python is so widely used and we have multiple libraries that, uh, provides machine learning models that we can run-in a few line lines of codes. It is, uh, it becomes very easy for us to, um, basically, uh, compare and make make the decision.

I would say that some of the libraries that help us understand the memory requirement and write writing optimized codes such as PSUTIL, it says, um, as well as the resource module. Uh, they are, uh, modules that help us write optimized codes because that helps us give the memory that is being used uh, for a particular script or a function. Uh, because Python is is executed at run time, um, and, uh, and the execution happens line by line. So it has a better language for debugging.