Potential bug: Loss = Nan when likelihood = "bernoulli-gamma"

alan-turing-institute / deepsensor

A Python package for tackling diverse environmental prediction tasks with NPs.

https://alan-turing-institute.github.io/deepsensor/

MIT License

94 stars 16 forks source link

Potential bug: Loss = Nan when likelihood = "bernoulli-gamma" #138

Open Opio-Cornelius opened 1 week ago

Opio-Cornelius commented 1 week ago

Description

The loss calculation returns NaN values when the likelihood parameter is set to "bernoulli-gamma" in the ConvNP model

Reproduction steps

1. Imported an xarray zarr file of precipitation that has gaps in it
2. Called the "gap-fill" option in the task loader
3. Used a likelihood = "bernoulli-gamma" in the ConvNP model
4. Used loss as RMSE
5. Got NaN as both the training and validation loss

Version

0.4.2

Screenshots

![Image](Screenshot 2024-11-08 at 15.21.32.png)

OS

Linux

Opio-Cornelius commented 1 week ago

Screenshot 2024-11-08 at 15 21 32