search

ciemss / pyciemss

Causal and probabilistic reasoning with continuous time dynamical systems
Other
12 stars 4 forks source link

Calibration resulted in tensor errors #526

Closed mwdchang closed 4 months ago

mwdchang commented 4 months ago

We are seeing this error when running calibration, it seems to happen with specific models.

ValueError: Expected parameter scale (Tensor of shape (1000,)) of distribution Normal(loc: torch.Size([1000]), scale: torch.Size([1000])) to satisfy the constraint GreaterThan(lower_bound=0.0), but found invalid values:

Attached are the inputs for a problematic case we encountered:

dataset.csv model.json request.json trace.log

SamWitty commented 4 months ago

@mwdchang , this model does not have a distribution over any of the parameters. In other words, the model is claiming that there is no uncertainty, and thus no parameters can be updated when presented with data. We can add a more informative error message, but the fact that this does not calibrate is expected.