Do you know why just one small change (1 line) greatly improves accuracy +0.7 on Pytorch-yolo?

[AlexeyAB]

@WongKinYiu Hi,

Do you know why just one small change (1 line) greatly improves accuracy + 0.7 on Pytorch-yolo? https://github.com/WongKinYiu/CrossStagePartialNetworks/blob/pytorch/README.md

Model	Size	NMS	1080ti fps	BFLOPs	AP	AP50	AP75	cfg	weight
CSPResNeXt50-PANet-SPP	512×512	0.5	44	71.331	39.2	59.5	41.8	cfg	-
CSPResNeXt50c-PANet-SPP	512×512	0.5	-	71.734	39.9	60.1	42.6	cfg	-

Full diff:

[WongKinYiu]

@AlexeyAB

Hello, 273 epochs is not sufficient for training from scratch of panet. From the log file, I think it not yet converge.

In rethinking imagenet pre-training, they also shows that training form scratch need more epochs.

[AlexeyAB]

So +-0.7 AP can be just fluctuation on early stages of training. Or partial-residual connections require more iterations for training than common-residual connections?

[WongKinYiu]

I think both yes.