A CUDA based modular neural network library
Likely my final Neural Network repository, I've been working on this neural netork projects for a couple months now. But I've limited by the fact I had a AMD card, so I was stuck to slow CPU based training.
Now that I have a much faster Nividia card I've modified my previous NumPy and SciPy based library to work with tensorflow and cupy. And it's over 300x faster
I stated before that my goal was to create a object detection model based off of a YOLO type architecture. So far I've been able to replicate Yolo V1 to Yolo V3 with somewhat accurate results
Heres a couple showcases:
My previous model was purely sequential, meaning the output from the previous layer would only go straight into the next layer. I've now added both a ResidualBlock and a ConcatBlock layer which can skip connections. This is very important for feature pyramid architectures.
I've also fixed a couple errors and misconceptions in my code.
I've also added new activation functions, loss functions and schedulers.
The code is also much more optimized, apart from being GPU based.
I plan to add proper documentation soon, but for now just look at the examples
- https://www.youtube.com/watch?v=Lakz2MoHy6o&t=337s
- https://www.youtube.com/watch?v=pj9-rr1wDhM
- https://optimization.cbe.cornell.edu/index.php?title=Adam
- https://paperswithcode.com/method/he-initialization#:~:text=Kaiming%20Initialization%2C%20or%20He%20Initialization,magnitudes%20of%20input%20signals%20exponentially.
- https://builtin.com/machine-learning/adam-optimization
- https://www.youtube.com/watch?v=Lakz2MoHy6o&t=337s
- https://towardsdatascience.com/yolo-v3-explained-ff5b850390f