This message was deleted.

Hi! Sure. Let us take a look and come back to you.

I think separate classes for

MQ_*

are no longer the approach, dropped in favour of the MQ loss function. Try on the import line:

from neuralforecast.models import NHITS

from neuralforecast.losses.pytorch import MQLoss

and then further down pass the class the loss function:

levels = [5, 50, 95]
model = NHITS(..., loss=MQLoss(level=levels))

Thanks! @Kin Gtz. Olivares, any more pointers?

if there's interest I can look at putting up an MR that brings https://github.com/Nixtla/transfer-learning-time-series/blob/main/nbs/Transfer_Learning.ipynb up to date with

v1.0.0

Indeed @Valeriy and @Chris Gervais We deprecated the creation of MQmodels in favor of the modularity of the loss input parameter. Cristian and I are working on replicating the transfer learning experiments with the refactored neuralforecast code In case you want to continue running the previous experiments use !pip install neuralforecast==0.1.0 We will let you know when the new version is ready!

@Chris Gervais, thanks for offering help. We will definitely come back to you. ;)

Thanks @Max (Nixtla) @Kin Gtz. Olivares and @Chris Gervais

Guys, we just merged the notebook.

Hi @Chris Gervais and @Valeriy Here is a working example of Transfer-learning where we pretrain on M4 and test on Box-Cox AirPassengers improving StatsForecast AutoARIMA: https://nixtla.github.io/neuralforecast/examples/transfer_learning.html

amazing, will try to check it out this week and circle back if we have questions / feedback

Great stuff