Details of Updating (fit!
)
Core Principles
- Stats are subtypes of
OnlineStat{T}
whereT
is the type of a single observation.- E.g.
Mean <: OnlineStat{Number}
- E.g.
fit!(o::OnlineStat{T}, data::T)
- Update
o
with the single observationdata
.
- Update
fit!(o::OnlineStat{T}, data::S)
- Iterate through
data
andfit!
each item.
- Iterate through
Why is Fitting Based on Iteration?
Reason 1: OnlineStats doesn't make assumptions on the shape of your data
Consider CovMatrix
, for which a single observation is an AbstractVector
, Tuple
, or NamedTuple
. If I try to fit!
it with a Matrix
, it's ambiguous whether I want rows or columns of the matrix to be treated as individual observations.
See eachrow
and eachcol
(after typing using LinearAlgebra
).
x = randn(1000, 2)
fit!(CovMatrix(), eachrow(x))
fit!(CovMatrix(), eachcol(x'))
CovMatrix: n=1_000 | value=[0.989115 -0.0513987; -0.0513987 1.08399]
Reason 2: OnlineStats works out-of-the-box with many data structures
Tabular data structures such as those in JuliaDB iterate over named tuples of rows, so things like this just work:
using JuliaDB
t = table(randn(100), randn(100))
fit!(2Mean(), t)
A Common Error
Consider the following example:
julia> fit!(Mean(), "asdf")
ERROR: The input for Mean is Number. Found Char.
This causes an error because:
"asdf"
is not aNumber
, so OnlineStats attempts to iterate through it- Iterating through
"asdf"
begins with the character'a'