cvplayground.ai is a place to play with computer vision models. It's currently a work in progress.
CVPlayground is designed to explore various methods of defining, training, and inferring computer vision models. The core motivation is to experiment with different frameworks and hardware implementations.
Check out the README in nn/ to learn more about the model being studied, BabyCNN, and the different ways it's implemented.
The project aims to implement BabyCNN using various methods:
- High-level frameworks like PyTorch, JAX, and TensorFlow, and MLX.
- Low-level implementations in C and CUDA (and numpy, for an intermediate abstraction level).
- Custom CUDA kernels.
- Advanced optimization techniques using Triton and CUTLASS.
In the future, there are plans to extend the project to work at the scale of DINOv2, but this is very non-trivial.