[Loss] Tweedie DistributionLoss not working

[Antoine-Schwartz]

What happened + What you expected to happen

I get an error during the fit with the Tweedie DistributionLoss.
It seems that at the first step, 1) the loss is NaN, and 2) the output doesn't respect the correct typing.

Versions / Dependencies

Databricks Runtime : 14.3 LTS ML

• pip install neuralforecast==1.7.5

Reproduction script

from neuralforecast import NeuralForecast
from neuralforecast.models import NHITS
from neuralforecast.losses.pytorch import DistributionLoss
from neuralforecast.utils import AirPassengersPanel

Y_df = AirPassengersPanel

nf = NeuralForecast(
    models=[
        NHITS(
            h=12,
            input_size=48,
            max_steps=100,
            scaler_type=None,
            loss=DistributionLoss('Tweedie', rho=1.5, validate_args=False),
            enable_model_summary=False,
            enable_checkpointing=False,
            logger=False
        )
    ],
    freq="M"
).fit(Y_df)

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[elephaint]

Thanks for using NF!

I've PR'd a fix for this issue (#1164)