Introduction

This directory contains PyTorch YOLOv3 software developed by Ultralytics LLC, and is freely available for redistribution under the GPL-3.0 license. For more information please visit https://www.ultralytics.com.

Description

The https://github.com/ultralytics/yolov3 repo contains inference and training code for YOLOv3 in PyTorch. The code works on Linux, MacOS and Windows. Training is done on the COCO dataset by default: https://cocodataset.org/#home. Credit to Joseph Redmon for YOLO: https://pjreddie.com/darknet/yolo/.

Requirements

Python 3.7 or later with all of the pip install -U -r requirements.txt packages including:

• torch >= 1.3
• opencv-python
• Pillow

All dependencies are included in the associated docker images. Docker requirements are:

• nvidia-docker
• Nvidia Driver Version >= 440.44

Tutorials

• GCP Quickstart
• Transfer Learning
• Train Single Image
• Train Single Class
• Train Custom Data

Jupyter Notebook

Our Jupyter notebook provides quick training, inference and testing examples.

Training

Start Training: python3 train.py to begin training after downloading COCO data with data/get_coco_dataset.sh. Each epoch trains on 117,263 images from the train and validate COCO sets, and tests on 5000 images from the COCO validate set.

Resume Training: python3 train.py --resume to resume training from weights/last.pt.

Plot Training: from utils import utils; utils.plot_results() plots training results from coco_16img.data, coco_64img.data, 2 example datasets available in the data/ folder, which train and test on the first 16 and 64 images of the COCO2014-trainval dataset.

Image Augmentation

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.

Augmentation	Description
Translation	+/- 10% (vertical and horizontal)
Rotation	+/- 5 degrees
Shear	+/- 2 degrees (vertical and horizontal)
Scale	+/- 10%
Reflection	50% probability (horizontal-only)
HSV Saturation	+/- 50%
HSV Intensity	+/- 50%

Speed

https://cloud.google.com/deep-learning-vm/
Machine type: preemptible n1-standard-16 (16 vCPUs, 60 GB memory)
CPU platform: Intel Skylake
GPUs: K80 ($0.20/hr), T4 ($0.35/hr), V100 ($0.83/hr) CUDA with Nvidia Apex FP16/32
HDD: 1 TB SSD
Dataset: COCO train 2014 (117,263 images)
Model: yolov3-spp.cfg
Command: python3 train.py --img 416 --batch 32 --accum 2

GPU	n	--batch --accum	img/s	epoch time	epoch cost
K80	1	32 x 2	11	175 min	$0.58
T4	1 2	32 x 2 64 x 1	41 61	48 min 32 min	$0.28 $0.36
V100	1 2	32 x 2 64 x 1	122 178	16 min 11 min	$0.23 $0.31
2080Ti	1 2	32 x 2 64 x 1	81 140	24 min 14 min	- -

Inference

detect.py runs inference on any sources:

python3 detect.py --source ...

• Image: --source file.jpg
• Video: --source file.mp4
• Directory: --source dir/
• Webcam: --source 0
• RTSP stream: --source rtsp://170.93.143.139/rtplive/470011e600ef003a004ee33696235daa
• HTTP stream: --source http://wmccpinetop.axiscam.net/mjpg/video.mjpg

To run a specific models:

YOLOv3: python3 detect.py --cfg cfg/yolov3.cfg --weights yolov3.weights

YOLOv3-tiny: python3 detect.py --cfg cfg/yolov3-tiny.cfg --weights yolov3-tiny.weights

YOLOv3-SPP: python3 detect.py --cfg cfg/yolov3-spp.cfg --weights yolov3-spp.weights

Pretrained Weights

Download from: https://drive.google.com/open?id=1LezFG5g3BCW6iYaV89B2i64cqEUZD7e0

Darknet Conversion

$ git clone https://github.com/ultralytics/yolov3 && cd yolov3

# convert darknet cfg/weights to pytorch model
$ python3  -c "from models import *; convert('cfg/yolov3-spp.cfg', 'weights/yolov3-spp.weights')"
Success: converted 'weights/yolov3-spp.weights' to 'converted.pt'

# convert cfg/pytorch model to darknet weights
$ python3  -c "from models import *; convert('cfg/yolov3-spp.cfg', 'weights/yolov3-spp.pt')"
Success: converted 'weights/yolov3-spp.pt' to 'converted.weights'

mAP

python3 test.py --weights ... --cfg ...

• mAP@0.5 run at --iou-thr 0.5, mAP@0.5...0.95 run at --iou-thr 0.7
• YOLOv3-SPP ultralytics is ultralytics68.pt with yolov3-spp.cfg
• Darknet results: https://arxiv.org/abs/1804.02767

	Size	COCO mAP @0.5...0.95	COCO mAP @0.5
YOLOv3-tiny YOLOv3 YOLOv3-SPP YOLOv3-SPP ultralytics	320	14.0 28.7 30.5 35.5	29.1 51.8 52.3 55.4
YOLOv3-tiny YOLOv3 YOLOv3-SPP YOLOv3-SPP ultralytics	416	16.0 31.2 33.9 39.2	33.0 55.4 56.9 59.9
YOLOv3-tiny YOLOv3 YOLOv3-SPP YOLOv3-SPP ultralytics	512	16.6 32.7 35.6 40.5	34.9 57.7 59.5 61.4
YOLOv3-tiny YOLOv3 YOLOv3-SPP YOLOv3-SPP ultralytics	608	16.6 33.1 37.0 41.1	35.4 58.2 60.7 61.5

$ python3 test.py --img-size 608 --iou-thr 0.6 --weights ultralytics68.pt --cfg yolov3-spp.cfg

Namespace(batch_size=32, cfg='yolov3-spp.cfg', conf_thres=0.001, data='data/coco2014.data', device='', img_size=608, iou_thres=0.6, save_json=True, task='test', weights='ultralytics68.pt')
Using CUDA device0 _CudaDeviceProperties(name='Tesla V100-SXM2-16GB', total_memory=16130MB)
               Class    Images   Targets         P         R   mAP@0.5        F1: 100% 157/157 [03:30<00:00,  1.16it/s]
                 all     5e+03  3.51e+04    0.0353     0.891     0.606    0.0673
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.409
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.615
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.437
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.242
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.448
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.519
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.337
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.557
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.612
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.438
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.658
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.746

Reproduce Our Results

This command trains yolov3-spp.cfg from scratch to our mAP above. Training takes about one week on a 2080Ti.

$ python3 train.py --weights '' --cfg yolov3-spp.cfg --epochs 273 --batch 16 --accum 4 --multi --pre

Reproduce Our Environment

To access an up-to-date working environment (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled), consider a:

• GCP Deep Learning VM with $300 free credit offer: See our GCP Quickstart Guide
• Google Colab Notebook with 12 hours of free GPU time: Google Colab Notebook
• Docker Image from https://hub.docker.com/r/ultralytics/yolov3. See Docker Quickstart Guide

Citation

Contact

Issues should be raised directly in the repository. For additional questions or comments please email Glenn Jocher at glenn.jocher@ultralytics.com or visit us at https://contact.ultralytics.com.