Chess_champion

Hey, try changing the gradient clipping parameter as well as play with learning rate.

@dorienh can you elaborate more? If a features takes 170 different values you do not want one-hot encoding anyway, you want to do embedding. If you have 170 different features...

No, you just include them as the static categorical variables or time varying categorical variables, then the packages will perform the embedding automatically. No need for any embedding. Add static...

Can you change it to GroupNormalizer or something else to see if this is normalized?

It is strange, which suggests maybe the `x,y = next(iter(dataloader))` is not a right way to see if the features are normalized or not. I experiment with normalization through different...

I think tsd.data['target'] returns the raw values.

@jdb78 Can you please verify if normalization works as expected?

So can you check dataloader .dataset.data['reals'], this should be normalized.

What is the dimension of your sale_cols?

I think the reals should already contain the target as well, since you include it as unknown_reals.