Implementation of "Scaled-YOLOv4: Scaling Cross Stage Partial Network" using PyTorch framwork.

Last update: Jan 02, 2023

Overview

YOLOv4-large

This is the implementation of "Scaled-YOLOv4: Scaling Cross Stage Partial Network" using PyTorch framwork.

Model	Test Size	AP^test	AP₅₀^test	AP₇₅^test	AP_S^test	AP_M^test	AP_L^test	batch1 throughput
YOLOv4-P5	896	51.4%	69.9%	56.3%	33.1%	55.4%	62.4%	41 fps
YOLOv4-P5	TTA	52.5%	70.3%	58.0%	36.0%	52.4%	62.3%	-

YOLOv4-P6	1280	54.3%	72.3%	59.5%	36.6%	58.2%	65.5%	30 fps
YOLOv4-P6	TTA	54.9%	72.6%	60.2%	37.4%	58.8%	66.7%	-

YOLOv4-P7	1536	55.4%	73.3%	60.7%	38.1%	59.5%	67.4%	15 fps
YOLOv4-P7	TTA	55.8%	73.2%	61.2%	38.8%	60.1%	68.2%	-

Model	Test Size	AP^val	AP₅₀^val	AP₇₅^val	AP_S^val	AP_M^val	AP_L^val	weights
YOLOv4-P5	896	51.2%	69.8%	56.2%	35.0%	56.2%	64.0%	`yolov4-p5.pt`
YOLOv4-P5	TTA	52.5%	70.2%	57.8%	38.5%	57.2%	64.0%	-
YOLOv4-P5 (+BoF)	896	51.7%	70.3%	56.7%	35.9%	56.7%	64.3%	`yolov4-p5_.pt`
YOLOv4-P5 (+BoF)	TTA	52.8%	70.6%	58.3%	38.8%	57.4%	64.4%	-

YOLOv4-P6	1280	53.9%	72.0%	59.0%	39.3%	58.3%	66.6%	`yolov4-p6.pt`
YOLOv4-P6	TTA	54.4%	72.3%	59.6%	39.8%	58.9%	67.6%	-
YOLOv4-P6 (+BoF)	1280	54.4%	72.7%	59.5%	39.5%	58.9%	67.3%	`yolov4-p6_.pt`
YOLOv4-P6 (+BoF)	TTA	54.8%	72.6%	60.0%	40.6%	59.1%	68.2%	-
YOLOv4-P6 (+BoF*)	1280	54.7%	72.9%	60.0%	39.4%	59.2%	68.3%
YOLOv4-P6 (+BoF*)	TTA	55.3%	73.2%	60.8%	40.5%	59.9%	69.4%	-

YOLOv4-P7	1536	55.0%	72.9%	60.2%	39.8%	59.9%	68.4%	`yolov4-p7.pt`
YOLOv4-P7	TTA	55.5%	72.9%	60.8%	41.1%	60.3%	68.9%	-

Model	Test Size	AP^val	AP₅₀^val	AP₇₅^val	AP_S^val	AP_M^val	AP_L^val
YOLOv4-P6-attention	1280	54.3%	72.3%	59.6%	38.7%	58.9%	66.6%

Installation

# create the docker container, you can change the share memory size if you have more.
nvidia-docker run --name yolov4_csp -it -v your_coco_path/:/coco/ -v your_code_path/:/yolo --shm-size=64g nvcr.io/nvidia/pytorch:20.06-py3

# install mish-cuda, if you use different pytorch version, you could try https://github.com/thomasbrandon/mish-cuda
cd /
git clone https://github.com/JunnYu/mish-cuda
cd mish-cuda
python setup.py build install

# go to code folder
cd /yolo

Testing

# download {yolov4-p5.pt, yolov4-p6.pt, yolov4-p7.pt} and put them in /yolo/weights/ folder.
python test.py --img 896 --conf 0.001 --batch 8 --device 0 --data coco.yaml --weights weights/yolov4-p5.pt
python test.py --img 1280 --conf 0.001 --batch 8 --device 0 --data coco.yaml --weights weights/yolov4-p6.pt
python test.py --img 1536 --conf 0.001 --batch 8 --device 0 --data coco.yaml --weights weights/yolov4-p7.pt

You will get following results:

# yolov4-p5
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.51244
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.69771
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.56180
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.35021
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.56247
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.63983
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.38530
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.64048
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.69801
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.55487
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.74368
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.82826

# yolov4-p6
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.53857
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.72015
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.59025
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.39285
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.58283
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.66580
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.39552
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.66504
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.72141
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.59193
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.75844
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.83981

# yolov4-p7
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.55046
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.72925
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.60224
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.39836
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.59854
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.68405
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.40256
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.66929
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.72943
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.59943
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.76873
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.84460

Training

We use multiple GPUs for training. {YOLOv4-P5, YOLOv4-P6, YOLOv4-P7} use input resolution {896, 1280, 1536} for training respectively.

# yolov4-p5
python -m torch.distributed.launch --nproc_per_node 4 train.py --batch-size 64 --img 896 896 --data coco.yaml --cfg yolov4-p5.yaml --weights '' --sync-bn --device 0,1,2,3 --name yolov4-p5
python -m torch.distributed.launch --nproc_per_node 4 train.py --batch-size 64 --img 896 896 --data coco.yaml --cfg yolov4-p5.yaml --weights 'runs/exp0_yolov4-p5/weights/last_298.pt' --sync-bn --device 0,1,2,3 --name yolov4-p5-tune --hyp 'data/hyp.finetune.yaml' --epochs 450 --resume

If your training process stucks, it due to bugs of the python. Just Ctrl+C to stop training and resume training by:

# yolov4-p5
python -m torch.distributed.launch --nproc_per_node 4 train.py --batch-size 64 --img 896 896 --data coco.yaml --cfg yolov4-p5.yaml --weights 'runs/exp0_yolov4-p5/weights/last.pt' --sync-bn --device 0,1,2,3 --name yolov4-p5 --resume

Citation

@InProceedings{Wang_2021_CVPR,
    author    = {Wang, Chien-Yao and Bochkovskiy, Alexey and Liao, Hong-Yuan Mark},
    title     = {{Scaled-YOLOv4}: Scaling Cross Stage Partial Network},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {13029-13038}
}

Acknowledgements

Expand

Implementation of "Scaled-YOLOv4: Scaling Cross Stage Partial Network" using PyTorch framwork.

Related tags

Overview

YOLOv4-large

Installation

Testing

Training

Citation

Acknowledgements

Owner

Kin-Yiu, Wong

Geometric Sensitivity Decomposition

Neural Scene Graphs for Dynamic Scene (CVPR 2021)

FedTorch is an open-source Python package for distributed and federated training of machine learning models using PyTorch distributed API

Direct Multi-view Multi-person 3D Human Pose Estimation

Machine learning evaluation metrics, implemented in Python, R, Haskell, and MATLAB / Octave

The audio-video synchronization of MKV Container Format is exploited to achieve data hiding

PyTorch implementation of DeepDream algorithm

Pose estimation for iOS and android using TensorFlow 2.0

Unofficial implementation of Proxy Anchor Loss for Deep Metric Learning

The official codes for the ICCV2021 Oral presentation "Rethinking Counting and Localization in Crowds: A Purely Point-Based Framework"

Expand human face editing via Global Direction of StyleCLIP, especially to maintain similarity during editing.

Code for project: "Learning to Minimize Remainder in Supervised Learning".

A Runtime method overload decorator which should behave like a compiled language

A CNN implementation using only numpy. Supports multidimensional images, stride, etc.

PyTorch implementation of SmoothGrad: removing noise by adding noise.

Using machine learning to predict and analyze high and low reader engagement for New York Times articles posted to Facebook.

This repository contains pre-trained models and some evaluation code for our paper Towards Unsupervised Dense Information Retrieval with Contrastive Learning

Object tracking and object detection is applied to track golf puts in real time and display stats/games.

Improving Machine Translation Systems via Isotopic Replacement

This is a work in progress reimplementation of Instant Neural Graphics Primitives