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 the following pip3 install -U -r requirements.txt packages:

• numpy
• torch >= 1.1.0
• opencv-python
• tqdm

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: n1-standard-8 (8 vCPUs, 30 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: 100 GB SSD
Dataset: COCO train 2014 (117,263 images)

GPUs	batch_size	images/sec	epoch time	epoch cost
K80	64 (32x2)	11	175 min	$0.58
T4	64 (32x2)	40	49 min	$0.29
T4 x2	64 (64x1)	61	32 min	$0.36
V100	64 (32x2)	115	17 min	$0.24
V100 x2	64 (64x1)	150	13 min	$0.36
2080Ti	64 (32x2)	81	24 min	-
2080Ti x2	64 (64x1)	140	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 weights/yolov3.weights

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

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

Pretrained Weights

• Darknet *.weights format: https://pjreddie.com/media/files/yolov3.weights
• PyTorch *.pt format: https://drive.google.com/drive/folders/1uxgUBemJVw9wZsdpboYbzUN4bcRhsuAI

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

• test.py --weights weights/yolov3.weights tests official YOLOv3 weights.
• test.py --weights weights/last.pt tests most recent checkpoint.
• test.py --weights weights/best.pt tests best checkpoint.
• Compare to darknet published results https://arxiv.org/abs/1804.02767.

ultralytics/yolov3 mAP@0.5 (darknet-reported mAP@0.5)

	320	416	608
YOLOv3	51.8 (51.5)	55.4 (55.3)	58.2 (57.9)
YOLOv3-SPP	53.7	57.7	60.7 (60.6)
YOLOv3-tiny	29.0	32.9 (33.1)	35.5

$ python3 test.py --save-json --img-size 608
Namespace(batch_size=16, cfg='cfg/yolov3-spp.cfg', conf_thres=0.001, data='data/coco.data', img_size=608, iou_thres=0.5, nms_thres=0.5, save_json=True, weights='weights/yolov3-spp.weights')
Using CUDA device0 _CudaDeviceProperties(name='Tesla T4', total_memory=15079MB)
                Class    Images   Targets         P         R       mAP        F1: 100% 313/313 [07:40<00:00,  2.34s/it]
                  all     5e+03  3.58e+04     0.119     0.788     0.594     0.201
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.367 <---
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.607 <---
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.387
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.208
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.392
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.487
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.297
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.465
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.495
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.332
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.518
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.621

$ python3 test.py --save-json --img-size 416
Namespace(batch_size=16, cfg='cfg/yolov3-spp.cfg', conf_thres=0.001, data='data/coco.data', img_size=416, iou_thres=0.5, nms_thres=0.5, save_json=True, weights='weights/yolov3s-ultralytics.pt')
Using CUDA device0 _CudaDeviceProperties(name='Tesla T4', total_memory=15079MB)
                Class    Images   Targets         P         R       mAP        F1: 100% 313/313 [07:01<00:00,  1.41s/it]
                  all     5e+03  3.58e+04      0.11     0.739     0.569     0.185
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.373
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.577
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.392
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.175
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.403
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.537
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.313
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.482
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.501
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.266
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.541
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.693

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.