Repository navigation
Document how initializers read custom parameter shapes - #678
Open
georgeguimaraes wants to merge 1 commit into
Open
georgeguimaraes wants to merge 1 commit into
georgeguimaraes wants to merge 1 commit into
Conversation
Axon.param's docs finish the sentence about shape functions and say how the variance scaling initializers compute fans on a parameter with a leading axis that is not a receptive field, and how to scale them in that case. Also drops a trailing space in loop_test.exs that fails mix format on main.
georgeguimaraes
force-pushed
the
param-initializer-docs
branch
from
September 17, 2026 14:55
153e5b6 to
24916c2
Compare
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two doc fixes on
Axon.param/3: the sentence about shape functions was cut off mid-way, and there was nothing saying that the variance scaling initializers treat every leading axis of a parameter as receptive field. A per-group kernel shaped{groups, in, out}therefore startssqrt(groups)too small unless you passglorot_uniform(scale: groups), which is what we had to work out by hand for soothsayer's per-series trend.Also removes a trailing space in
test/axon/loop_test.exsthat makesmix format --check-formattedfail on main right now.