Why is there a big difference in the number of parameters for the s0 model in train and deploy mode?

[eehoeskrap]

While testing the s0 model, I noticed that the number of train parameters is different from the number written in the paper.

So, I found this link, and found that the size of the s0 model differs greatly between train mode and deploy mode.

I wonder why there is such a large difference in the number of parameters.

[abc20002929]

s0 has 4 branch convs when training, s1~s4 only 1 branch