Numerial Stability Issue for MultiDecoder

[SwagJ]

Hi @stepankonev,

When I attempted to train the multi-decoder version with default config (final_RoP_Cov_A_fMCG.yaml), the traing process would be interrputed by infinity error at first epoch.
Setting trainable_cov to False will not help solve it.
If chaning running_mean_mode to "sliding", the training loss will be huge negative, and it becomes quite unstable

Did this happen when you train the network? And is this normal? Thank you in advance

[hushunda]

i have meet same problem. Have you deal with it? Can you share your resolution?
Thank you