about the yolov3-tiny-prn.cfg

[weixiaolian21]

hi, thanks for your work. can you tell me where is the PRN module in this cfg? I just find some shotcut layer and the shortcut layer seems to add two tensor with same dims. So how to implete the partial add in this cfg. thank you.

[WongKinYiu]

https://github.com/WongKinYiu/PartialResidualNetworks/blob/master/cfg/yolov3-tiny-prn.cfg#L182

[weixiaolian21]

https://github.com/WongKinYiu/PartialResidualNetworks/blob/master/cfg/yolov3-tiny-prn.cfg#L182

thanks for your reply. that's just make me confused shotcut layer in yolov3 can add two feature map, but it need two same dimension feature map. how to implete partial add add two different dimension feature map.

[WongKinYiu]

Hello, shortcut layer of darknet has no limit to add same or different dimension feature maps. The only issue you need to notice is that the size of output feature map is same as the size of the layer just before shortcut layer.

[weixiaolian21]

thanks a lot， I have another question. for example, In your paper, you add half of the feature maps to the next layer.

The PRN we propose is a stack of partial residual connection blocks, and the structure of partial residual connection is shown in Figure 1.

In the cfg, it seems that they all have 128 feature maps.

[WongKinYiu]

figure 1 is for darknet53-prn. figure 4 is for tiny-prn head.

and in https://github.com/WongKinYiu/PartialResidualNetworks/blob/master/cfg/yolov3-tiny-prn.cfg#L182 channel of 8th layer is 256.

[weixiaolian21]

In your new paper，CSPNET， you use a parameter γ to control the number of channels add to next layer，both the layer didn't use all the channels， how to change the γ， does it need to change the source code？

[WongKinYiu]

it need not to change the source code. just adjust the number of filters in cfg file.

however, shufflenet-v2 (eccv18) and HarDNet (iccv19) shows that parameter γ = 0.5 is the best trade-off. so i think we can just follow this optimal value.