Week 7: Training Phase (Halfway Checkpoint and Model Training)

This week, Mr. Roshan conducted a halfway checkpoint session for all students, allowing everyone to present their progress and gain insights into their work. It was interesting to see how others were approaching their projects and to observe the different methodologies being used. The session was not just about showcasing our advancements but also about receiving constructive feedback, both from Mr. Roshan and our peers. Hearing about the challenges others were facing and the solutions they were implementing gave me a fresh perspective on my own approach. It was reassuring to know that many of us were encountering similar obstacles, and discussing potential solutions as a group was incredibly helpful. Seeing other projects at different stages of development also gave me motivation to keep refining my own work and to push forward with the training process.

A significant part of this week involved properly structuring my dataset to ensure efficient training for the YOLOv8 model. To keep everything organized, I split the dataset into two main directories: one for images and another for labels. Within these, I created three separate subsets: training, validation, and testing. The images folder contained a diverse collection of pictures capturing the ball from various angles, lighting conditions, and positions. This variety is essential for helping the model learn how to recognize the ball in different environments and scenarios, improving its overall accuracy. Meanwhile, the labels were automatically generated and organized by Roboflow during the annotation process. Each annotation was stored as a text file containing bounding box coordinates for its corresponding image. Roboflow handled this formatting seamlessly, ensuring the data was clean and ready for model training. The platform also helped structure the dataset into separate folders for training, validation, and testing—an essential step that allows the model to learn effectively, validate its accuracy during the process, and ultimately be evaluated fairly using unseen data.

With the dataset fully prepared, I began the initial stages of training the YOLOv8 model. I started by keeping things simple, focusing on training the model to recognize a ball before expanding its capabilities. The early stages were slow, requiring multiple adjustments to get the training process running smoothly. Since deep learning models require careful tuning, I had to make several modifications to the Python scripts handling the training pipeline. This included tweaking hyperparameters, adjusting the dataset configuration, and integrating additional PyTorch files that were essential for optimizing the model’s learning process. At times, it was challenging to get everything to function correctly, as minor issues in the code or dataset formatting could affect training stability. However, with each adjustment, I could see gradual improvements in how well the model was recognizing the ball.

As training progressed, I began experimenting with different variations of the dataset and made changes to the model settings to fine-tune its accuracy. The goal was to create a detection system that could reliably track a ball in real time, so I had to ensure that the training data was well-balanced and diverse enough for the model to generalize effectively. Although I encountered a few roadblocks along the way, such as unexpected errors in the dataset formatting and inconsistencies in detection accuracy, I continued refining the process by testing different configurations and analyzing the results. Each iteration brought new insights, helping me understand what needed to be adjusted for better performance.

Moving forward, I will be focusing on optimizing the training phase, experimenting with more complex detection tasks, and refining the model's ability to handle real-world scenarios. The progress so far has been steady, and with further improvements, I aim to develop a highly efficient system for ball tracking that can be integrated into practical applications.