Hello. I have a question about the iTransformer in neuralforecast. When I use the transformer with the same data, I always encounter the following error at the same point. Upon checking, it turns out that the values in a 32×90×32 tensor are all NaN. What could be the possible reasons for this occurring? Especially, I want to know what kind of things in _base_multivariate.py.

It probably means your loss is NaN. From the posted code, I'd guess that most likely culprit is using

identity

as scaler_type. I'd suggest to use

robust

or

standard

and try again.

Yes, my loss is nan! Ok, I try it on your suggestion. Thanks !!

First, the loss function should inherit from losses.pytorch.BasePointLoss. See also how MAE or RMSE has been defined in losses.pytorch. Second, this loss function doesn't really make sense imho. As you sum all values, there's no real 'incentive' for the network to perform well on individual timestamps. If you want to weight timestamps, I'd use horizon_weight parameter in RMSE, which makes more sense, and then set the weights comparable to your alpha/betta/gamma parameters.

Thank you! I'll make the adjustments as per your advice and give it a try. I have one more question. The tensor sizes are as follows: •

insample_y

→

torch.Size([32, 16, 32])

•

outsample_y

→

torch.Size([32, 90, 32])

•

output

→

torch.Size([32, 90, 32])

In the

output

tensor, the dimensions are

32 × 90 × 32

. I understand that: •

32

= batch size •

90

= predict length (h) •

32

= What is this?

You gave n_series=32 as input

Got it! Thanks for everything, Olivier.