pytorch-yolo3

Introduction

Pytorch implementation of YOLOv3. Tensorflow2.0 version can be found here

Quick Start

1. Download yolov3.weights and darknet53.conv.74 from YOLO website.
2. Download COCO dataset
3. Modify the dataset root and weights root in the config file

python main_coco.py --resume load_yolov3 --do_test --net_size 608 --batch_size 8

Training

1. run the following command to start training

python main_voc.py/main_coco.py --resume load_darknet --net_size 480 --batch_size 12

Visualization

The Tensorboard is origanized like TF-ObjectDection-API

Performance on VOC2007 Test(mAP)

Initial backbone weights	train scales	baseline	data augmentation	+multi test	+flip
darknet53	480	0.532	0.738	0.753	0.769
darknet53	448,480,512	-	0.727	0.737	0.754
coco pretrained	448,480,512	-	0.817	0.834	0.845

Note: all experiments trained for 100 epochs with learning rate dropped 10 times at the 70 and 90 epoch.

Supported Attributes

• Data agumentation
• Multi-scale Training
• Multi-scale Testing(including flip)
• Focal loss
• ....

TODO

• Update VOC performance
• Update COCO performance
• Support distribute training
• Support Custom dataset

Reference

gluon-cv

tf-eager-yolo3

keras-yolo3

stronger-yolo

yolo3-pytorch