training slightly modified yolov4-tiny with imagenet

[huseyinuri]

I am planning to use yolov4-tiny for an embedded device (rockchip rk3399pro) but groups and group_id are not supported parameters for model conversion toolkit. If i understood correctly, route layer slices the input feature map into 2 parts (groups = 2) and forwards the second part (group_id = 1) to the next convolutional layer. It seems that replacing route layers (not concatenating ones) with 1x1 convolution layers are supposed to work. However, training fails and loss becomes nan after a while with modified cfg with yolov4-tiny.conv.29.

1. Original code piece
   

   Route layer replaced with 1x1 convolution
   

My questions

1. Is it ok to replace some route layers with 1x1 convolutions ?
2. Should i retrain yolov4-tiny backbone with imagenet from scratch first ? If yes, how ?
3. What are the alternative backbones for yolov4-tiny without having groups params ?

[huseyinuri]

Shufflenet backbone uses channel_slice for splitting feature map. Is it ok to replace some route layers with channel_slice? : #3750

[asmaamirkhan]

Is it solved? 👀
I have the same problem 🙄

[kadirbeytorun]

I asked a similar question and did not receive a proper answer.
@AlexeyAB @WongKinYiu

[diPDew]

Same question here. Any update so far?