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+】
git clone https://github.com/robbebluecp/tf2-yolov4.git
train :-> 😅
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
just run: python3 predict.py -m xxx.weights(or xxx.h5) -i xxx.jpg
If you want to use pre-trained weights, just deliver your weights path into YOLO class in train.py
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.
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