Chibueze Benneth

OKay so i haven’t updated my devlog in literally forever so it has accumulated several hours of work over a month. That being said, i ran into several issues and procastinated so much in this project until this point. I spent so long trying to set up the libraries to run the model; for some reason, tensorflow and torch were tedious to set up. Debugging was long and painful and the learning curve was a nightmare. I had several nights of zero or negative progress. When i finally set up the torch library, I had to learn that CNNs don’t accept pixel input directly but it has to be converted to a tensor and normalized. This converts pixels into numerical inputs that the model can interprete. After that, I was able to set up a simple model and I got it to run. But low and behold, it had a 100% accuracy. Cap. It overfit so bad I genuinely wanted to cry and laugh at the same time. I then spent a number of hours together with chatgpt troubleshooting the overfitting. Now, after applying variations to the model like adding a dropout rate, random transformations to the images before converting them to tensors, increasing the processing layers (I increased it to 3 layers of convoluting and pooling), and then adding weight decay to prevent memorization, I finally got it to act predictably. Now i have run into an issue; It only has a validation accuracy of 67% maximum which isn’t the best (since random is 33%) and a loss of 0.4 which could be better. Oh well, I will update you guys when I make progress with it

I just finished sampling and processing the image data. It was a bit of a hassle because i had to learn how to use the OpenCV module in Python to process images. Turns out my images were too HIGH RES, so I had to literally downscale them to about 1/10th of their original size and quality, while preserving the core info. lol

I’m working on my first project! This is so exciting. I can’t wait to share more updates as I build.