yolov4-tensorflow2

This is a re-implementation of yolov4, using tensorflow2.2. ⭐

Various tricks in the paper are implemented. You can train a custom data set by specifying the classes file. The perfomance can reach 19~22fps in single thread, while the hardware configuration is Intel(R) Core(TM) i5-10300H CPU @ 2.50GHz 2.50 GHz, GeForce GTX 1660Ti, RAM 16G.:collision:

I also made some notes, which is some experience in model training, click here to view.:notebook_with_decorative_cover:

Contact the author: qq: 419289453

Directory:bookmark_tabs:

1. Demo

2. Requirements

3. Quick-Start

4. Work

5. Note

6. Train

7. Test

Demo

Requirements

Python 3.7.7+

tensorflow-gpu 2.2.0+

Pillow 7.1.2+

python install -r requirements.txt

Quick-Start

1. Clone this repository ☀️

git clone https://github.com/NiceRingNode/yolov4-tensorflow2

2. Download pretrained weights here, extract code: a768. This directory contains two .h5 files,

pretrained weights name	dataset based	classes number
yolov4_voc_weights.h5	VOC2007+2012+COCO2014	20
yolov4_coco_weights.h5	COCO2017	80

3. Put the weights files in folder model_data and put image for inference test in folder img
4. Change working directory in cmd,

cd/d yolov4-tensorflow2

​ and run predict.py

python predict.py --img_root <your path of the image>

​ or just type

python predict.py

Work

• CSPDarkNet53 🌊

• PAN, SPP ⚡

• Mish Activation Layer 🎨

• CIOU 📷

• Label Smoothing 📜

• Mosaic Data Augmentation 🐑

• Cosine Decay 🌾

• LetterBox Image 🐡

Note

The pretrained weights yolov4_coco_weights.h5 are trained based on images that have size 608×608，but it was modified to 416×416 in the code while the anchors are calculated in the size of 608×608. You can change it to 608×608 if needed.

Train

1. Here we use the VOC2007 or VOC2012 dataset to train the model, so download the VOC dataset and put it under the root directory.
2. Here we use the pretrained model to finetune, so we should download the pretrained weights, and put them in the folder model_data.

├─demo
├─font
├─img
├─logs
├─model_data
│  ├─coco_classes.txt
│  ├─voc_classes.txt
│  ├─yolov4_coco_weights.h5
│  ├─yolov4_voc_weights.h5
│  └─yolo_anchors.txt
├─video
└─VOCdevkit
    └─VOC2007

2. Change working directory in cmd, and run voc2yolov4.py

python voc2yolov4.py

3. Then run voc_annotations_convert.py, now the data format is converted to <image path> xmin ymin xmax ymax class_id

python voc_annotations_convert.py

4. As data prepared, we can now begin training. 🐚

python train.py --classes_path <the path of classes file> --pretrained_weights_path <the path of pretrained weights> --logs_path <the path of storing log files> --batch_size 2 --epoch 100

If you want to train the model on custom dataset, you should change the root of dataset, and specify the classes file when training by modifying the --classes_path argument.

or just

python train.py

Test

This step is similar to Quick Start, but you can use the pretrained weights trained by yourself instead of the pretrained weights downloaded above. Also you can run video.py to test the real-time performance.

python video.py