What is the real architecture(layers) of YOLOv3?

user_13123102 0
Viewed 310

Recently i'm reading yolov3's paper and code and i found a question.

In yolov3 it is darknet-53, which means it has 53 convolutional layers, but when i see this picture and count, i only get 52 convolutional layers. this picture is from it's paper

And i also see the yolov3.cfg, totally it has 107 layers(from 0 to 106), and my result is:

75 convolutional layers + 23 shortcut layers + 3 yolo layers + 4 route layers + 2 upsample layers = 107 layers

I want to know did i misunderstand something? or where is the 53th convolutional layer in the picture above?

Thanks in advance.

darknet yolo conv-neural-network
1 Answers

Like most of the neural network models, in darknet-53, 53 refers to the total trainable layers (like 19 in VGG-19). So 52 CNN layers that you counted plus the last connected layer gives darknet-53. Note, the connected layer can also be created using CNN layer, that's why yolo is being called fully convolutional neural network.

For your second query, I personally think you are right. There is a total of 107 layers in yolov3.cfg file.

52 layers are taken from darknet-53 (of course excluding connected layer), 27 other convolutional layers are added including 3 YOLO layers. This gives a total of 79 trainable convolutional layers in total.

user_97285730
Related