The difference between original model and converted model

[xwhkkk]

Thank you for sharing your great work!
I noticed we can generate our own edge map with both the original model and the converted model. So which one is a better choice?
Besides, I try to generate a map with both table7_pidinet.pth(13-epoch) and table5_pidinet.pth (19-epoch) through pidinet-converted, but the result is different. What types of images are suitable for each of these two types of models?

The edge map of the same image is listed as follows. (I want to extract the edge of the steps)

table5_pidinet

table7_pidinet

[minh-nguyenhoang]

I guess the converted model are already loaded with the fused weight while the original one requires you to re-calculate the weight every forward pass so there will be a little overhead. The converted layer with fused weight will act as a normal convolution layer so you can export this to other format as well.