Added empircal_sinkhorn_divergence function (#13)

Co-authored-by: David Widmann <[email protected]>
Co-authored-by: David Widmann <[email protected]>

Author: davibarreira

Project.toml

@@ -1,7 +1,7 @@
 name = "PythonOT"
 uuid = "3c485715-4278-42b2-9b5f-8f00e43c12ef"
 authors = ["David Widmann"]
-version = "0.1.3"
+version = "0.1.4"
 
 [deps]
 PyCall = "438e738f-606a-5dbb-bf0a-cddfbfd45ab0"

docs/src/api.md

@@ -14,6 +14,7 @@ emd2_1d
 ```@docs
 sinkhorn
 sinkhorn2
+empirical_sinkhorn_divergence
 barycenter
 ```
 

src/PythonOT.jl

@@ -11,7 +11,8 @@ export emd,
     barycenter,
     barycenter_unbalanced,
     sinkhorn_unbalanced,
-    sinkhorn_unbalanced2
+    sinkhorn_unbalanced2,
+    empirical_sinkhorn_divergence
 
 const pot = PyCall.PyNULL()
 

src/lib.jl

@@ -244,6 +244,43 @@ function sinkhorn2(μ, ν, C, ε; kwargs...)
     return pot.sinkhorn2(μ, ν, PyCall.PyReverseDims(permutedims(C)), ε; kwargs...)
 end
 
+"""
+    empirical_sinkhorn_divergence(xsource, xtarget, ε; kwargs...)
+
+Compute the Sinkhorn divergence from empirical data, where `xsource` and `xtarget` are
+arrays representing samples in the source domain and target domain, respectively, and  `ε`
+is the regularization term.
+
+This function is a wrapper of the function
+[`ot.bregman.empirical_sinkhorn_divergence`](https://pythonot.github.io/gen_modules/ot.bregman.html#ot.bregman.empirical_sinkhorn_divergence)
+in the Python Optimal Transport package. Keyword arguments are listed in the documentation of the Python function.
+
+# Examples
+
+```jldoctest
+julia> xsource = [1];
+
+julia> xtarget = [2, 3];
+
+julia> ε = 0.01;
+
+julia> empirical_sinkhorn_divergence(xsource, xtarget, ε) ≈
+       sinkhorn2([1], [0.5, 0.5], [1 4], ε) -
+       (
+           sinkhorn2([1], [1], zeros(1, 1), ε) +
+           sinkhorn2([0.5, 0.5], [0.5, 0.5], [0 1; 1 0], ε)
+       ) / 2
+true
+```
+
+See also: [`sinkhorn2`](@ref)
+"""
+function empirical_sinkhorn_divergence(xsource, xtarget, ε; kwargs...)
+    return pot.bregman.empirical_sinkhorn_divergence(
+        reshape(xsource, Val(2)), reshape(xtarget, Val(2)), ε; kwargs...
+    )
+end
+
 """
     sinkhorn_unbalanced(μ, ν, C, ε, λ; kwargs...)