there appears to be an inconsistency in the sigma passed to the regression model at training and inference time. Training has random sigma. Inference has sigma = 1. Not sure if it would have any impact. I imagine that the regression model quickly learns to ignore the randomized sigma during training.
We should probably still fix it. Training should have a constant sigma.
nbren12
commented
2 months ago
Version
main
On which installation method(s) does this occur?
No response
Describe the issue
there appears to be an inconsistency in the sigma passed to the regression model at training and inference time. Training has random sigma. Inference has sigma = 1. Not sure if it would have any impact. I imagine that the regression model quickly learns to ignore the randomized sigma during training.
We should probably still fix it. Training should have a constant sigma.
Minimum reproducible example
No response
Relevant log output
No response
Environment details
No response