I'm a Java Architect turned AI MSc student. I found myself using `torch.nn.Linear` and `torch.nn.Conv2d` without fully grasping the tensor operations underneath.
So I started this project to rebuild these layers from first principles (using only Tensor operations and Autograd, no `nn.Module`).
Currently implemented:
- Linear Regression & Logistic Regression (Numpy manual grad vs Autograd)
- Multi-Layer Perceptrons (MLP)
- Convolutional Neural Networks (CNN) with pooling
Coming next:
- RNNs and Attention mechanisms (WIP)
The goal is education: stripping away the abstraction layers to see the math. Feedback on my CNN implementation is very welcome!
Show HN: Reimplementing PyTorch from scratch (MLP, CNN) to learn the internals | HN Companion
I'm a Java Architect turned AI MSc student. I found myself using `torch.nn.Linear` and `torch.nn.Conv2d` without fully grasping the tensor operations underneath.
So I started this project to rebuild these layers from first principles (using only Tensor operations and Autograd, no `nn.Module`).
Currently implemented: - Linear Regression & Logistic Regression (Numpy manual grad vs Autograd) - Multi-Layer Perceptrons (MLP) - Convolutional Neural Networks (CNN) with pooling
Coming next: - RNNs and Attention mechanisms (WIP)
The goal is education: stripping away the abstraction layers to see the math. Feedback on my CNN implementation is very welcome!