The wonderful thing about PyTorch is that it's high-level and low-level at the same time.
It abstracts away backprop via its shadow computation graph, which is nicely high-level. And at the same time its forward-feed layers are little more than wrappers over the basic matrix multiplications.
The combination of these attributes makes implementing architectures from papers fast and easy.
As part of my graduate program in computer science from UT Austin, I took Deep Learning in 2023, studying the theory and mathematics and the implementation of a variety of architectures. This repo holds the convolutional networks.
Each folder houses a different model and a readme with any relevant notes I felt like adding.
I recommend torchinfo over torchsummary. Usage is exactly the same, but torchinfo catches shape errors much better, and groups output by class, which is much easier to read.
Like all my libraries, this repo was created and is managed with Terraform, and was based on a template.