NAN appears when doing time series problems.（做时间序列问题时候出现NAN）

[glotm]

Dataset Create

I test the effect of KAN through artificial data, and the process guarantees that NAN will not appear.

Training

when I first train,this is look fine

Problem

Then I use the "prune" function,and train again,the loss start begin NAN

I adjusted the learning rate of all the fit functions above to 0.001, but the training after pruning is still nan.

How can I solve this problem

[glotm]

I saw that other people's methods include adjusting "updata_grid=False". My training after pruning was successful, but when I used the "refine" function to continue training, it was still NAN.

[glotm]

PS.3
I adjusted my network structure

That is, to train on a smaller-scale network. At this time, all the training is normal, and I don't need to use "updata_grid=False" anymore.
This question makes me a little puzzled.

[yymydyy]

same question when I using deeper KAN