How do you do transfer learning based on the models from the model zoo?

[devedse]

I found the model zoo useful because I want to start training a model on my 3080. I decided that the YOLOv4-Mish-416 would probably fit fine.

For the normal yolov4.weights there's also a yolov4.conv.137 available. I can't find a similar file to the YOLOv4-Mish-416 file though. Is there a way to generate this file yourself or obtain it somewhere else?

Next to that, another related question: I saw Scaled-YoloV4 also being mentioned a lot of times. Are the models from the model zoo actually using YoloV4 scaled or not?

[AlexeyAB]

Look at these examples how to generate your own pre-traineed weights like yolov4.conv.137 from yolov4.weights https://github.com/AlexeyAB/darknet/blob/master/build/darknet/x64/partial.cmd

• Use https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.conv.137 to train YOLOv4, YOLOv4-Mish, YOLOv4-SAM-Mish https://github.com/AlexeyAB/darknet/wiki/YOLOv4-model-zoo

• Use https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-csp.conv.142 to train Scaled-YOLOv4 yolov4-csp.cfg

• Use https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4x-mish.conv.166 to train Scaled-YOLOv4 yolov4x-mish.cfg

• Use https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/cspx-p7-mish_hp.344.conv to train Scaled-YOLOv4 https://github.com/AlexeyAB/darknet/blob/master/cfg/cspx-p7-mish.cfg

[devedse]

So how do you know what that last number is? E.g. 137 for yolov4, but what is it for leaky or Mish?

[AlexeyAB]

Use the number of layer prior to the first [yolo] layer.

[devedse]

Hmm, I tried this on the yolov4.cfg. I'm not experienced with darknet as of yet so I assume that everything in between brackets ([...]) is a layer. I counted all instances of that and came to a total of 140.

I assume the first one [net] shouldn't be counted but that still results in a count of 139. Which is 2 more then expected.

I counted everything up until the red line (the first yolo layer)

[devedse]

From what I read regarding scaled YOLO it seems that the CSP model might actually be one of the better ones.

However I don't think the scaled models are described in the model zoo?: https://github.com/AlexeyAB/darknet/wiki/YOLOv4-model-zoo

Or am I misunderstanding some things?

[AlexeyAB]

Yes, YOLOv4-CSP it the best (but not for small objects): https://github.com/AlexeyAB/darknet/issues/7087#issuecomment-750924784

CSP model is described there: https://github.com/AlexeyAB/darknet/issues/7087

And there: https://alexeyab84.medium.com/scaled-yolo-v4-is-the-best-neural-network-for-object-detection-on-ms-coco-dataset-39dfa22fa982?source=friends_link&sk=c8553bfed861b1a7932f739d26f487c8

---

To get layer number just run detection on any image:

[devedse]

@AlexeyAB , i have been training for about 24 hours and this is the result. Is this to be expected?