Discover more from Matt Rickard
Defensible Machine Learning Model Naming
There's a curious case of a neural network for object recognition called YOLO – You Only Look Once. While many object detection models were two-pass (one for identifying bounding boxes, the other for classifying), YOLO was single-pass. This makes YOLO fast and small.
I used a modified version of YOLO for my model ScapeNet: Real-time Object Detection in Runescape. Except, my YOLO wasn't really the same YOLO. There are almost a dozen different models, each claiming to be YOLO, written by other authors.
Which is the real YOLO? Does it matter? What makes a model "win"?
YOLOv3 was forked by researchers at Baidu in a model called PP-YOLO.
YOLOv4 (2020) was released by a different author, Alexey Bochkovskiy. This repo is a fork of Redmon's original repository and is closest in architecture.
YOLOv5 was written by Glenn Jocher, and implemented in PyTorch.
Meituan released a model MT-YOLOv6, which is also called YOLOv6.
Bochkovskiy (author of v4) also released a new model called YOLOv7.
Machine learning model naming is tough. Most users won't dive into the architecture of how it works. Some versions might differ on "non-research" elements: better developer experience, different implementation or framework, or different end-user API.
The threat of the hard fork might be even greater with open-source model architecture.