@Hitesh

Building the model turned out to be a lot tougher than I thought. I spent so many hours, sometimes whole days, planning, experimenting, and figuring out how everything would work. My main goal was to make the model bug free and working completely fine. There were times when I wasn’t sure if my code or project would actually work. Some of my friends had made fun of me for working so hard or thought the project was stupid, but I kept going anyway. I want to show them that hard work on even “stupid” projects can turn out good.

The hardest parts weren’t technical. I struggled with wanting to give up, worrying I wouldn’t reach my goal, and hearing discouraging words from people close to me. Even so, I kept going. Visiting Hugging Face for inspiration and learning more about how the code worked helped me learn more about OCR and Image Classification models. I’ve learned a lot along the way ngl.

After fixing some final bugs, I feel like I’m almost finished with my model. I just need to do a bit more research to make sure I’m headed in the right direction, and of course, I’ll keep posting these devlogs to share my progress/journey!

I actually timed myself here💀. Ok so there I was, trying to find data/images to train my image classification model, and I ran into a lot of problems😭. The biggest problem was the lack of suitable space images and astronomical text data. I tried Kaggle first, but that failed because most datasets weren’t specific to astronomy. Next I used Hugging Face for the images, and that also failed because most of them were either JSON files or plain text instead of actual space imagery. I finally settled on NASA’s Open Data Archive to get actual space images and train the model, and I even used an image scraper extension to scrape additional images to supplement my training data. I woke up at 4:30 in the morning because of my jetlag and spent those early hours trying to find good datasets, figure out which languages to use, and figure out how to actually get the datasets into my pipeline. After looking different things, I settled on Keras and TensorFlow to create the model since they are pretty good for deep learning and image classification and integrates with the language I’m good at, Python .For the demo and web interface, I used TensorFlow.js so I could deploy it directly in the browser. To make things a little bit simpler and faster, I used Teachable Machine’s scripts to get the model’s TensorFlow.js script by training the model again there, which saved me a lot of time on the export and conversion process. But I was extremely happy with the end result and was amazed at how good my model really was . Despite all the challenges from the jetlag affected morning coding sessions to the frustrating dataset search, the final result performed surprisingly well on astronomical images and I’m excited to keep improving it for my space and AI projects.