This directory contains software developed by Ultralytics LLC, and **is freely available for redistribution under the GPL-3.0 license**. For more information on Ultralytics projects please visit:
The https://github.com/ultralytics/yolov3 repo contains inference and training code for YOLOv3 in PyTorch. Training is done on the COCO dataset by default: https://cocodataset.org/#home. **Credit to Joseph Redmon for YOLO** (https://pjreddie.com/darknet/yolo/) and to **Erik Lindernoren for the PyTorch implementation** this work is based on (https://github.com/eriklindernoren/PyTorch-YOLOv3).
**Start Training:** Run `train.py` to begin training after downloading COCO data with `data/get_coco_dataset.sh`. Training runs about 1 hour per COCO epoch on a 1080 Ti.
Each epoch trains on 120,000 images from the train and validate COCO sets, and tests on 5000 images from the COCO validate set. An Nvidia GTX 1080 Ti will process about 10-20 epochs/day depending on image size and augmentation. Loss plots are shown here using default training settings.
`datasets.py` applies random OpenCV-powered (https://opencv.org/) augmentation to the input images in accordance with the following specifications. Augmentation is applied **only** during training, not during inference. Bounding boxes are automatically tracked and updated with the images. 416 x 416 examples pictured below.
Run `test.py` to validate the official YOLOv3 weights `weights/yolov3.weights` against the 5000 validation images. You should obtain a .584 mAP at `--img-size 416`, or .586 at `--img-size 608` using this repo, compared to .579 at 608 x 608 reported in darknet (https://arxiv.org/abs/1804.02767).
Run `test.py --weights weights/latest.pt` to validate against the latest training results. Default training settings produce a 0.522 mAP at epoch 62. We are currently exploring how to improve this.
