edward

edward copied to clipboard

This is a request for a feature that allows to provide a global flag to edward which essentially sets validate_args to True in debug mode. Currently, when a model is ill-specified e.g. wrong parameter values are provided for the distributions, or data values dont match the output distribution, the program silently fails and produces nans in the output. The user has no way to debug this easily because validate_args is False by default, and manually adding validate_args=True while initializing each distribution can be tedious for the user.

I would be happy to contribute this feature to edward if this seems reasonable. Do you have any suggestions on how to go forward?

That's an interesting suggestion. I can see that as a potentially useful utility. Can you raise this in TensorFlow? It seems like a feature request that should be considered upstream in TF Distributions.

Hi @dustinvtran , I tried suggesting this in tensorflow but they didnt accept my proposal. But it was accepted in pytorch

Do you think we can add this in edward? I will be happy to contribute this feature.