IanButterworth/YOLO.jl

Package wishlist

IanButterworth opened this issue · 3 comments

Models

  • YOLO v2
  • YOLO v2-tiny
  • YOLO v3
  • YOLO v3-tiny
  • custom model

Datasets

  • Optional downloading of datasets
  • VOC
  • COCO
  • ImageNet
  • Custom datasets & dataset locations (possible but untested)

Testing

  • Elegant minibatching
  • Move away from ImageView given Gtk issues on MacOS, to Makie

Training

  • Loss function etc.

Backends

  • Switchable backends
  • Knet
  • Flux

Looking forward to see this realized. What can I do to help?

@stemann I just merged a PR that makes this functional! Can you test, using the example code in the README? Much appreciated!!

Hi,
I managed to write the YOLOv2 chain and to load the pretrained weights (manually, no automatic generation). Forked repo here: https://github.com/iuliancioarca/YOLO.jl
I also updated the Readme with an example on how to load and use YOLOv2. Posprocessing and visualization were simply plug and play from V2_tiny.