Docs for onecold
#2134
Docs for onecold
#2134
Conversation
Describe the behaviour of `onecold` when given a probability distribution as input, since it's not what one would expect.
docs/src/data/onehot.md
Outdated
| The inverse is `onecold` (which can take a general probability distribution, as well as just booleans). | ||
| The inverse is `onecold`. This function can take a general probability distribution, as well as just booleans. However, when given a probability distribution it doesn't sample from it but just returns the element with the highest probability. |
There was a problem hiding this comment.
Maybe "probability distribution" is misleading, it doesn't care, takes any old array of numbers --- it's just argmax except that it translates an index to some given labels.
I think "inverse" is misleading too. It un-does onehot, but is normally used on arrays that look nothing like what onehot produces.
There was a problem hiding this comment.
I agree that "probability distribution" is misleading. (I called it that because that's what the current text calls it, but it probably shouldn't.)
I don't know how to update the pull request, but maybe something like this would be better:
There is also a onecold function, which is an inverse of onehot. It can also be given an array of numbers instead of booleans, in which case it performs an argmax-like operation, returning the label with the highest corresponding weight.
There was a problem hiding this comment.
That sounds great.
You should be able to edit your branch, here:
https://github.com/nathanielvirgo/Flux.jl/blob/patch-1/docs/src/data/onehot.md
Further commits will show up in this PR.
|
Thanks! |
Describe the behaviour of
onecoldwhen given a probability distribution as input, since it's not what one would expect.minor documentation change
PR Checklist