add maximum mean discrepancy metric #56
Conversation
| @@ -0,0 +1,89 @@ | |||
| import numpy | |||
There was a problem hiding this comment.
probably better to use torch instead of numpy for consistency unless numpy is needed for a particular reason
There was a problem hiding this comment.
would be nice if this worked with arrays of strings as well since that's a common data structure. We allowed numpy arrays here for the same reason: https://github.com/Genentech/beignet/blob/main/src/beignet/_farthest_first_traversal.py
There was a problem hiding this comment.
that being said I could see wanting this both differentiable and GPU-enabled... any thoughts on that? implement two versions?
There was a problem hiding this comment.
casting to tensor opens the tokenization and padding can of worms and usually you just want something simple
| Y, | ||
| distance_fn=None, | ||
| kernel_width: float | None = None, | ||
| eps: float = 1e-16, |
There was a problem hiding this comment.
What's the motivation for lower bounding the MMD at sqrt(eps), rather than 0?
There was a problem hiding this comment.
I suppose I was trying to forestall division by 0 errors but upon reflection that's a decision downstream users should make, will revert lower bound to 0
|
ok I have attempted to make this compliant with the python array standard API and NEP 56 to support both numpy and pytorch arrays |
| @@ -9,6 +9,7 @@ requires = [ | |||
| [project] | |||
| authors = [{ email = "[email protected]", name = "Allen Goodman" }] | |||
| dependencies = [ | |||
| "numpy>=2.0.0", | |||
There was a problem hiding this comment.
Let torch manage the numpy dependency
There was a problem hiding this comment.
I think torch doesn't depend on numpy now
There was a problem hiding this comment.
yeah that what I was seeing
|
@kleinhenz finally fixed the broken test, are we good to merge? |
|
I do think there is a valid question around whether we want to force beignet users to upgrade to NumPy 2.0 (to support the array API standard). @0x00b1 thoughts? |
estimates maximum mean discrepancy from samples