Building a Large Language Model (GPT v2) from scratch

I built a Large Language Model (GPT v2) from scratch. In this project, I first went through the transformer architecture. Understand and describe the core building blocks of transformer model, and also descibe how it works, to give predictions. After that, I go through the GPT v2 model, how it follows the transformer architecture in a way, and how it has included some detailed implementations in it. I built everything in the project from every block wise module, from the scratch. The code for the project can be found here: GitHub.

Sentiment Analysis of google app reviews with BERT.

I built a sentiment analysis system with the BERT large language model from transformer library of the Pytorch framework. In this work, I described the basic building blocks of the BERT model, how it works, how it's embedding systems, and other criterias are set up, how it works for the classification tasks, etc. For training the model, we use google app reviews as our text data, and the review scores as the outputs. I address some imbalance issues, and the whole training process here too. The code for the project can be found here: GitHub.

Human Action Recognition from Videos using Resnet-50 and Bi-directional LSTM

I built a Human Action Recognition from Videos using the Resnet-50 model, for the feature extraction process, and the Multi Layered Bi-Directional LSTM model, for the prediction of the human activity from the videos. In my work, I describe, why the Resnet-50 is suitable for the feature extraction task, and why the sequence pattern of the video data is suitable with the LSTM model. We conduct the experiment of our work on the University of Central Florida's Youtube Activity Recognition Dataset. The code: GitHub.

Tennis video analysis system based on Resnet-50 and YOLOv8

I built a tennis analysis system that can detect the participating players in a tennis match, detect the tennis ball, which is being used to play the game, calculate the shot speed and the player speed of each player in the video. For the detection phase, I used Resnet-50 for detecting the keypoints in a tennis court. I used YOLOv5 model to train and predict the ball positions in each frmae of the video. This project also involved some complex calculations and manipulations based on logics, to solve some challenges. The code for the project can be found here: GitHub.

Building an Artificial Neural Network from Scratch (With Numpy) and training, predicting on a dataset

I built a Artificial Neural Network from Scratch with the use of only the Numpy library of python. I dive deep in the implementation of Neural Networks in this work. I discuss the process from the absolute mathematical ground up. After building the network, I trained it on the Digit Recognizer Dataset to get the optimal weights for my network, that is suitable to this dataset. The accuracy rate after training was 83% and I also gave some predictions by this model. The code: GitHub.

Anomaly Detection in Times Series Data with Robust Z score method, Isolation forest, Local Outlier Factor

In this work, I have built an anomaly detection system based on three different methods. They are, Robust Z score method, Isolation Forest method, and the Local Outlier Factor method. I implemented this work on the ec2 CPU Utilization dataset. I find the anomaly in the CPU usage, in different timestamps. Besides the detections, I also showed the results, compared with the ground level truths. The code: GitHub.

Human Object Interaction Detection with Positional Relation Features and Adjacency based Feature Refinements

This was my undergraduate thesis work. It focused on the HOI Detection topic. The HOI detection is a well reknowned area in the Computer Vision domain. It is the task of detecting and recognizing interactions between humans and objects in an image. The problem is a bit tricky because it includes multiple scenarios such as, varying number of human-object interactions in an image, a single human interacting with multiple objects, multiple humans interacting with one object etc. I conducted my experiment on the popular V-COCO (Microsoft Verbs in COCO dataset) dataset. The code for this research project can be found here: GitHub.

Building a Maio Like Platformer Pygame using the Python Pygame module

I built a platformer pygame with python's pygame module. The work is a massive python Object Oriented based project. It involved relations between various agents, such as Player, Enemy, Obstacle_Particles, Environment e.t.c. It required a lot of correlations with differnt objects from different classes, to get it finished. It helped me to understand the OOP concepts of python, on a very large scale project. The code: GitHub.

My works on Data Structures and Algorithmic concepts, and Programming Problems solving in this particular area.

I have studied and implemented the data structures and algorithmic concepts with the Python programming language. My studies include, Big(O) functions, Time, Space Complexity, Recursions, and different functionalities, and implementation of differnt operations of different Data Structures and Algorithmic concepts such as, Linked List Data Structures, Stack, Queue, Singuly Linked Lists, Doubly Lists, Binary Tree, Binary Search Tree, AVL Tree, Binary Heap, Trie, Sorting Algorithms (Quick Sort, Merge Sort, Selection Sort etc), Searching Algorithms (Binary Search, Linear Search), Graph Algorithms (BFS, DFS, Topological Sort, Single Source Shortest Path, All Pairs Shortest path), also some problem solving strategies such as, Dive and Conquer, BackTracking, Dynamic Programming - Top down and Bottom up approach, Greedy approach etc. I have also solved more than 150 programming problems in Leetcode. My Leetcode account link can be found in the 1st Slide of this. My fundamental DSA work code repo: GitHub.