Experimental YOLO Nano Model

[AyushExel]

Experiment in brief

I'm trying to test an implementation of YOLO Nano model. Here is my approach:

1. First I'll try to experiment with a separate yolo_nano module.
2. If the results seem promising, I'll then work on yolo_nano.yaml file that'll be tightly integrated with the existing YOLOv5 model hub ecosystem.

Motivation for this experiment:

The numbers mentioned in the YOLO nano research paper are quite striking. A 4MB model with 69.1% mAP on VOC 2017 should be a major improvement in terms of robustness for edge devices and online computation.

@glenn-jocher Are there any steps that I should keep in mind when running this study? Like for example, to make this reproducible, is it recommended to set/modify some random seeds?

[glenn-jocher]

@AyushExel sounds like a good research direction!

VOC can return some pretty high mAP numbers in comparison to COCO, so it's hard to say if 69.1% is good or not. We have a public VOC project here showing about 86% and 88% [email protected] (VOC metric) with YOLOv5s and YOLOv5s6.
https://wandb.ai/glenn-jocher/VOC

Their 'computation cost (ops)' doesn't seem to align well with our own YOLOv3-tiny measurement of 13.2 GFLOPS at 640x640 either, not sure where the discrepancy lies, may be inference size asumptions https://github.com/ultralytics/yolov3#pretrained-checkpoints

Note for VOC our default train script is shown in the Google Colab notebook here:
https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb#scrollTo=BSgFCAcMbk1R&line=1&uniqifier=1

# VOC
for b, m in zip([64, 48, 32, 16], ['yolov5s', 'yolov5m', 'yolov5l', 'yolov5x']):  # zip(batch_size, model)
  !python train.py --batch {b} --weights {m}.pt --data voc.yaml --epochs 50 --cache --img 512 --nosave --hyp hyp.finetune.yaml --project VOC --name {m}