By
@GamerBolt03
Trying to create a new architecture for object detection to perform much more efficiently and also run with extremely high FPS on low-end devices to random CPU's
16h 45m 41s logged
FunctionAI Devlog #4
Currently still working on improving the architecture and making it fast enough to learn object detection in at least under 30 mins.
Also thinking of working on the UI in the mean time since the project is almost complete
13h 20m 46s logged
FunctionAI Devlog #3
After completely re-coding the entire project from scratch the new model is able to reach an accuracy of 30% as of now with 15mins of training on 1080p data and run at a framerate of 1000fps, which is not bad for having re-coded the entire project but I need to make it much better.
Also, I created a simple logo, although I couldn’t really get a good idea on what I should make
13h 26m 56s logged
Well after a lot of tests and coding (approx. 13 hours) I decided to switch to a neural based AI instead of function based using my custom CTF architecture. Ive been training the model on the COCO dataset with 80 classes for object detection. Currently in under 5 mins of training on an RTX5090 the model reaches 71% accuracy with estimated 1000 FPS on the Core Ultra 9 255k CPU which is insanely fast compared to mobile net and other datasets. I believe with an increase time limit the model can reach around 95% accuracy however I would like to keep the training time as is at 5 mins just to prove the architecture works insanely well too. Although might have to change the name due to switching to CTF.
(I really like sushi, dogs and video games)
5h 4m 27s logged
This is a project I started creating to support AEVA which is another project based around a new AI architecture however I decided that this deserved its own project since after all it may have no use in AEVA
Currently I attempted to only use function approximation along with grounding using linear regression however it only managed to achieve an accuracy of 10% which has led me to think of new ideas using a transformer replacement I developed for AEVA known as CTF as the star choice. However, I would still like to work further on the neuron less architecture a little further.
