search

timaeus-research / devinterp

Tools for studying developmental interpretability in neural networks.
73 stars 14 forks source link

Need to pass temperature in multiple places #88

Closed jqhoogland closed 1 month ago

jqhoogland commented 3 months ago

You need to pass a temperature in both the callbacks and the optim_kwargs. We should give a warning if the user only provides one or the other.

jqhoogland commented 3 months ago

One of the examples may currently leave out the temperature, thus default to a value and confuse users

svwingerden commented 3 months ago

I think this fallback behavior is well-documented in both https://timaeus-research.github.io/devinterp/devinterp.optim.html#module-devinterp.optim.sgld and https://timaeus-research.github.io/devinterp/devinterp.slt.html#module-devinterp.slt.llc, so I think a warning is overkill. I can add in a print statement, though.

wz-ml commented 2 months ago

Looks like nbeta is now a required parameter in the callbacks that take them: LLCEstimator and MalaAcceptanceRate. I've added print notices in sample and estimate_learning_coeff_with_summary for both directions, just in case this changes. See PR #93 .