Log cosh #421
Log cosh #421
Conversation
While waiting forhttps://github.com/JuliaStats/Distributions.jl/pull/1157
For a 0.14.8 release
For a 0.14.9 release
For a 0.15 release
For 0.15.1 release
This reverts commit 9939b92.
Codecov Report
@@ Coverage Diff @@
## dev JuliaAI/MLJBase.jl#421 +/- ##
=======================================
Coverage 81.83% 81.83%
=======================================
Files 38 38
Lines 2703 2703
=======================================
Hits 2212 2212
Misses 491 491
Continue to review full report at Codecov.
|
| function (log_cosh::LogCosh)(ŷ::Vec{<:Real}, y::Vec{<:Real}) | ||
| check_dimensions(ŷ, y) | ||
| return sum(log.(cosh.(ŷ-y))) / length(y) | ||
| end |
There was a problem hiding this comment.
Suggestion (see C below): return a vector of per-observation losses: return log.(cosh.(ŷ-y))) / length(y)
oops I mean't to drop the /length(y) as well.
There was a problem hiding this comment.
The sum is missing in this definition
There was a problem hiding this comment.
@ven-k Thanks for this contribution and for taking the time to wrap your head around the API!
C
As the loss is just the mean of a per-observation loss, can I suggest the changes above? By default this loss will be aggregated in resampling using the mean (aggregation(...) = Mean()). In this way the per-observation losses will be available, for example, to hyper-parameter optimisation strategies that use this information (some Bayesian methods).
So then the code will look basically the same as l1 which has this behaviour:
julia> l1([1,2,3], [2,3,4])
3-element Array{Int64,1}:
1
1
1
src/measures/continuous.jl
Outdated
|
|
||
| Log-Cosh loss: | ||
|
|
||
| ``\\text{Log-Cosh} = m^{-1}∑ᵢ log(cosh(ŷᵢ-yᵢ))`` |
There was a problem hiding this comment.
One more thing update the docstring to reflect new behavior
There was a problem hiding this comment.
@OkonSamuel Yeah it is a bit confusing. Some measures that could be implemented as per-observation losses are not, eg, mae (which I guess essentially duplicates l1, which is per-observation loss).
@ven-k For the record, measures is overdue for a thorough review and migration out of MLJ https://github.com/alan-turing-institute/MLJBase.jl/issues/299 .
|
@ven-k Now live (see above) |
|
Thanks! @ablaom and @OkonSamuel I am following JuliaAI/StatisticalMeasures.jl#17 and #416. I am interested in StatisticalMeasures and possible visualisation package. |
log_cosh = sum (log (cosh (yhat - y))) / n