Inconsistent interface for weight-clipping

[kevinchern]

There's an inconsistent interface for methods that interact with weights.
For example, in sampling and in computing the quasi-objective function, weight clipping is applied when linear/quadratic range is supplied.
However, when computing energies in the forward method, no such option is available.
While computing energies without clipped weights does not impact gradient calculations, it can affect analysis.
Have I described the problem accurately @jquetzalcoatl ?

[kevinchern]

Exposing such an interface is undesirable because the design was motivated by explicitly separating the model from samplers. The model is always as is, ie, beta=1. It is the user's responsibility to ensure the sampler samples at beta=1. if it doesnt, one can rely on prefactor to correct for it. By doing so, we limit the prefactor's presence to sampling and only sampling.

Documentation needs significant improvement to make this clear.

[VolodyaCO]

should I own this @kevinchern ?

[kevinchern]

should I own this @kevinchern ?

@VolodyaCO To improve documentation? That'd be great! We should also check with @jquetzalcoatl

[VolodyaCO]

Ok, waiting for @jquetzalcoatl