Training nano-plus-m with backbone shufflenetv2_1.0x from scratch get significantly lower APs

[NrealWJX]

Hi, I train nano-plus-m (image size is 384*384) with default settings and get 0.262 AP at 300 epoch and 0.251 AP at 90 epoch.

However, when I train from scratch (set pretrain=False in yaml), I get 0.213 AP at 90 epoch.

Have you try to train those models from scratch? How are those results?

• with pretrained weights

• without pretrained weights

[RangiLyu]

For tiny backbones, pre-training is very important.

[NrealWJX]

For tiny backbones, pre-training is very important.

@RangiLyu Thx for your reply. I will pre-train the backbone before using it on target datasets.

I also read the tips from yolov5: Tips-for-Best-Training-Results. However it recommends to pre-train or not according to dataset scale instead of model size. How do you like this?