INTRODUCTION

This edition of yolov4 frame has been smoothly transferred from tf1+ and keras version. With tf2+ used extensively, many older functions have been optimized even won't be used any more. Keras has been embedded into tensorflow and won't be supported by official, except fixing bugs. For more readable, I've made a lot of changes while transferring, such as layer aggregation, image augment and some other changes from tf1 to tf2.

【Now, IT ONLY SUPPORTS TF2+】

INSTALLATION

git clone https://github.com/robbebluecp/tf2-yolov4.git

FILES

• data :-> to put some temporary files like jpgs or pngs and so on
• font_data :-> a font file for text on pic (will be removed)
• model_data :-> core config files mainly including class file and anchors file
• model_train :-> for saving models or weights files
• tools :-> helper functions
• config :-> configuration file
• convert :-> for converting weights file to h5 which trained by darknet using tf2+ (do not support tf1+)
• eval :-> a part of predicting
• generator :-> a generator of data by loading image files by batch
• loss :-> core loss function
• models :-> core yolo4 model
• predict :-> for predicting
• prepare :-> prepare config

• train :-> 😅

  HOW TO PREDICT

  If you use yolov4.weights file

• download yolov4.weights file from yolov4.weights
• put it into folder model_train/
• REMEMBER to change 【classes_path = 'model_data/voc_classes.txt'】into 【classes_path = 'model_data/coco_classes.txt'】 in config.py
• run: python3 convert.py -w model_train/yolov4.weights -m model_train/yolov4.h5
• then you'll see yolov4.h5 in your model_train folder
• run: python3 predict -m model_train/yolov4.h5 -i data/test.png

• then you'll see the following information
  56 chair 0.98 x:2 y:1037 x:448 y:1628
  0 person 0.83 x:1 y:553 x:335 y:1176
  0 person 0.95 x:257 y:463 x:440 y:943
  0 person 1.00 x:345 y:454 x:788 y:1195
  and if you have visual interface you can see a pic with rectangles and class name on it

  If you use your own train model

• just run: python3 predict.py -m xxx.weights(or xxx.h5) -i xxx.jpg

  HOW TO TRAIN

  Here is an example to show you how to train on voc2007. Images and labels are reorganized so that you can easily divide into your training

• download voc2007 data from voc2007.zip
• unzip it into /opt (of course some other paths are also ok..., if you choose other paths, you have to change 【label_path = '/opt/voc2007/labels.txt'】 into 【label_path = your_path/label.txt】 in config.py)
• make sure 【classes_path = 'model_data/voc_classes.txt'】 in config.py
• then, run python3 train.py

• If you want to use pre-trained weights, just deliver your weights path into YOLO class in train.py

  RESULT

  As you can see in loss.png and predict.png, I've trained voc2007 for about 160 epochs, loss line seems much healthier but val_loss looks like a little unstable. But I think with more epochs or loading model with pre-trained weights will be better and better. So, you can just try with pre-trained weights while loading model if you have a nice machine to develop. Whatever, there are a lot of difference between c release and python release, if you want to improve more, just step to
  http://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects for a more learning.
  
  

  RELATIONS

  http://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects
  http://github.com/Ma-Dan/keras-yolo4
  http://github.com/qqwweee/keras-yolo3
  https://arxiv.org/pdf/2004.10934.pdf