Scaled-YOLOv4-tensorflow2

A Tensorflow2.x implementation of Scaled-YOLOv4 as described in Scaled-YOLOv4: Scaling Cross Stage Partial Network

Update Log

[2021-07-02]:

• Add support for: Exponential moving average decay for variables. Improve mAP from 0.985 to 0.990 on Chess Pieces dataset.

[2021-06-29]:

Major Features and Improvements:

• Add support for: Sharpness-Aware Minimization(SAM_sgd,SAM_adam).

Bug Fixes and Changes:

• Fix the nan loss error when using adam optimizer
• Set default optimizer as SAM_adam
• Change default running mode from 'fit' to 'eager mode'

[2021-06-27] Add support for: resuming training from checkpoints.

[2021-02-21] Add support for: model.fit(dramatic improvement in GPU utilization); online coco evaluation callback; change default optimizer from sgd to adam

[2021-02-11] Add support for: one-click deployment using tensorflow Serving(very fast)

[2021-01-29] Add support for: mosaic,ssd_random_crop

[2021-01-25] Add support for: ciou loss,hard-nms,DIoU-nms,label_smooth,transfer learning,tensorboard

[2021-01-23] Add support for: scales_x_y/eliminate grid sensitivity,accumulate gradients for using big batch size,focal loss,diou loss

[2021-01-16] Add support for: warmup,Cosine annealing scheduler,Eager mode training with tf.GradientTape,support voc/coco dataset format

[2021-01-10] Add support for: yolov4-tiny,yolov4-large p5/p6/p7,online coco evaluation,multi scale training

Demo

ScaledYOLOv4_p5_detection_result:

ScaledYOLOv4_tiny_detection_result:

Installation

1. Clone project

git clone https://github.com/wangermeng2021/Scaled-YOLOv4-tensorflow2.git
cd Scaled-YOLOv4-tensorflow2

2. Install environment

• install tesnorflow ( skip this step if it's already installed,test environment:tensorflow 2.4.0)

• pip install -r requirements.txt
  

Note:

I strongly recommend using voc dataset type(default dataset type), because my GPU is old, so coco dataset type is not fully tested.

Training:

• Download Pre-trained p5 coco pretrain models and place it under directory 'pretrained/ScaledYOLOV4_p5_coco_pretrain' :
  https://drive.google.com/file/d/1glOCE3Y5Q5enW3rpVq3SmKDXzaKIw4YL/view?usp=sharing
  

• Download Pre-trained p6 coco pretrain models and place it under directory 'pretrained/ScaledYOLOV4_p6_coco_pretrain' :
  https://drive.google.com/file/d/1EymbpgiO6VkCCFdB0zSTv0B9yB6T9Fw1/view?usp=sharing
  

• Download Pre-trained tiny coco pretrain models and place it under directory 'pretrained/ScaledYOLOV4_tiny_coco_pretrain' :
  https://drive.google.com/file/d/1x15FN7jCAFwsntaMwmSkkgIzvHXUa7xT/view?usp=sharing
  

• For training on Pothole dataset(No need to download dataset,it's already included in project):
  p5(single scale):

  python train.py --use-pretrain True --model-type p5 --dataset-type voc --dataset dataset/pothole_voc --num-classes 1 --class-names pothole.names  --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 200 --batch-size 4 --multi-scale 416 --augment ssd_random_crop 
  

  p5(multi scale):

  python train.py --use-pretrain True --model-type p5 --dataset-type voc --dataset dataset/pothole_voc --num-classes 1 --class-names pothole.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 200 --batch-size 4 --multi-scale 320,352,384,416,448,480,512 --augment ssd_random_crop 
  

• For training on Chess Pieces dataset(No need to download dataset,it's already included in project):
  tiny(single scale):

  python train.py --use-pretrain True --model-type tiny --dataset-type voc --dataset dataset/chess_voc --num-classes 12 --class-names chess.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 400 --batch-size 32 --multi-scale 416 --augment ssd_random_crop 
  

  tiny(multi scale):

  python train.py --use-pretrain True --model-type tiny --dataset-type voc --dataset dataset/chess_voc --num-classes 12 --class-names chess.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 400 --batch-size 32 --multi-scale 320,352,384,416,448,480,512 --augment ssd_random_crop
  
  

• For training with SAM_sgd on Chess Pieces dataset:
  

  python train.py --optimizer SAM_sgd --use-pretrain True --model-type tiny --dataset-type voc --dataset dataset/chess_voc --num-classes 12 --class-names chess.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 400 --batch-size 32 --multi-scale 416 --augment ssd_random_crop 
  

• For training with ema(Exponential Moving Average) on Chess Pieces dataset:
  

  python train.py --ema True --use-pretrain True --model-type tiny --dataset-type voc --dataset dataset/chess_voc --num-classes 12 --class-names chess.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 400 --batch-size 32 --multi-scale 416 --augment ssd_random_crop 
  

Tensorboard visualization:

• Navigate to http://0.0.0.0:6006

Evaluation results(GTX2080,mAP@0.5):

model	Chess Pieces	pothole	VOC	COCO
Scaled-YoloV4-tiny(416)	0.985
Scaled-YoloV4-tiny(416)+ema	0.990
AlexeyAB's YoloV4(416)		0.814
Scaled-YoloV4-p5(416)		0.826

• Evaluation on Pothole dataset:
• Evaluation on chess dataset:

Detection

• For detection on Chess Pieces dataset:

  python3 detect.py --pic-dir images/chess_pictures --model-path output_model/best_model_tiny_0.985/1 --class-names dataset/chess.names --nms-score-threshold 0.1
  

  detection result:

  

• For detection on Pothole dataset:

  python3 detect.py --pic-dir images/pothole_pictures --model-path output_model/best_model_p5_0.827/1 --class-names dataset/pothole.names --nms-score-threshold 0.1
  

  detection result:

  

Customzied training

• Convert your dataset to Pascal VOC format(you can use labelImg to generate VOC format dataset)
• Generate class names file(such as xxx.names)

• python train.py --use-pretrain True --model-type p5 --dataset-type voc --dataset your_dataset_root_dir --num-classes num_of_classes --class-names path_of_xxx.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 200 --batch-size 8 --multi-scale 416  --augment ssd_random_crop 
  

Deployment

TensorFlow Serving is a flexible, high-performance serving system for machine learning models, designed for production environments.it include two parts:clients and server, we can run them on one machine.

• Navigate to deployment directory:

  cd  deployment/tfserving

• Generate a docker image which contains your trained model (it will take minutes，you only have to run it one time):

  ./gen_image --model-dir ScaledYOLOv4-tensorflow2/output_model/pothole/best_model_p5_0.811

• Deploy model:
  

  • Server side( docker and nvidia-docker installed ):

    ./run_image

  • Client side(no need to install tensorflow):
    

    1. install client package

       pip install tfservingclient-1.0.0-cp37-cp37m-manylinux1_x86_64.whl

    2. predict images

       python demo.py --pic-dir xxxx --class-names xxx.names

References

• https://github.com/WongKinYiu/ScaledYOLOv4
• https://github.com/ultralytics/yolov5
• https://github.com/dmlc/gluon-cv