Darknet: Darknet53 cfg file

[fishman2008]

Hi Joseph, could you also upload the Darknet53 cfg file? The model weight file is available, but not the cfg. Thanks.

[TheMikeyR]

@fishman2008 https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg

[rky0930]

Is there any cfg file for Darknet53 not yolo v3.
I need the information after conv 74 which are Avgpool, Connected, Softmax.

[cebolan]

@fishman2008 @rky0930 I attempted to recreate the darknet53 based on the yolov3 tech report and the cfgs for yolov3 and darknet19: https://gist.github.com/cebolan/f4e7c9b1e81d74e8097d9c59c8de7d05

[lupotto]

Hi guys,

Maybe is a very basic question but I don't understand what is the main difference between darknet53.conv.74 and yolov3.weights.

According to what I understand, yolov3.weights is trained in COCO dataset but which layers are loaded, all of them? And darknet53.conv.74 just the convolutionals and trained in Imagenet?

If I have a model which I train with yolov3.weights initialization weights and I want to train it with darknet53.conv.74 how would you solve it?

Thanks in advance

[cebolan]

@lupotto you've got the first part right: yolo3.weights includes weights for all layers which were trained in COCO while darknet53.conv.74 includes weights for the convolutional layers trained in Imagenet.

Regarding your last question, I don't know what you would want to accomplish with that. The training that resulted in yolov3.weights, if I'm not mistaken, started from darknet53.conv.74. Loading its weights again would probably only mess up the training done in the upper layers.

[lupotto]

Absolutely @cebolan . Thanks for the explanation

[rdfong]

Anyone know why the routed layers are at layers 61 and 36 instead of 62 and 37? It seems like it would make sense to use the very last feature map of a certain resolution instead of the one before the resnet block routing layer. Maybe it doesn't make a huge difference either way though.

[rdfong]

In fact for the coarsest output it directly uses the layer AFTER the routing layer of the previous res block, so I feel like the same should be the case for the rest of the outputs.