[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in `ot.bregman` #437

6Ulm · 2023-02-22T11:17:38Z

Types of changes

Add argument of warmstart/initialization of dual vectors to Sinkhorn-based methods.

Motivation and context / Related issue

The current Sinkhorn solvers sinkhorn, empirical_sinkhorn, empirical_sinkhorn_divergence use default intialization of dual vectors, and does not allow to take other initializations as input. This can be unnecessarily inefficient, in terms of time and computational cost, if meaningful initializations exist.

How has this been tested (if it applies)

These changes have been tested on toy examples. The script of the tests is added to test_bregman.

PR checklist

[x ] I have read the CONTRIBUTING document.
[ x] The documentation is up-to-date with the changes I made (check build artifacts).
[ x] All tests passed, and additional code has been covered with new tests.
[ x] I have added the PR and Issue fix to the RELEASES.md file.

…s in

codecov · 2023-02-22T11:27:43Z

Codecov Report

Merging #437 (2571802) into master (80e3c23) will increase coverage by 0.01%.
The diff coverage is 80.95%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #437      +/-   ##
==========================================
+ Coverage   94.69%   94.70%   +0.01%     
==========================================
  Files          24       24              
  Lines        6593     6608      +15     
==========================================
+ Hits         6243     6258      +15     
  Misses        350      350

This reverts commit f3d36b2.

rflamary

Small change in position of parameter wramstart and few doc changes.

Thanks for the PR @6Ulm

RELEASES.md

rflamary · 2023-02-23T09:51:21Z

ot/bregman.py

@@ -93,6 +93,10 @@ def sinkhorn(a, b, M, reg, method='sinkhorn', numItermax=1000,
        those function for specific parameters
    numItermax : int, optional
        Max number of iterations
+    warmstart: tuple of arrays, shape (dim_a, dim_b), optional
+        Initialization of dual vectors. If provided,
+        the dual vectors must be already taken the logarithm,


Suggested change

the dual vectors must be already taken the logarithm,

the dual potentails should be given (that is the logarithm of the u,v sinkhorn scaling vectors)

also remove the following line

rflamary · 2023-02-23T09:52:34Z

ot/bregman.py

@@ -24,7 +24,7 @@
 from .backend import get_backend


-def sinkhorn(a, b, M, reg, method='sinkhorn', numItermax=1000,
+def sinkhorn(a, b, M, reg, method='sinkhorn', numItermax=1000, warmstart=None,


move the warmstart after the stpthr

rflamary · 2023-02-23T09:52:42Z

ot/bregman.py

@@ -154,35 +158,35 @@ def sinkhorn(a, b, M, reg, method='sinkhorn', numItermax=1000,
    """

    if method.lower() == 'sinkhorn':
-        return sinkhorn_knopp(a, b, M, reg, numItermax=numItermax,
+        return sinkhorn_knopp(a, b, M, reg, numItermax=numItermax, warmstart=warmstart,


rflamary · 2023-02-23T09:52:49Z

ot/bregman.py

                                        stopThr=stopThr, verbose=verbose,
                                        log=log, warn=warn,
                                        **kwargs)
    else:
        raise ValueError("Unknown method '%s'." % method)


-def sinkhorn2(a, b, M, reg, method='sinkhorn', numItermax=1000,
+def sinkhorn2(a, b, M, reg, method='sinkhorn', numItermax=1000, warmstart=None,


rflamary · 2023-02-23T09:53:17Z

ot/bregman.py

@@ -407,6 +415,10 @@ def sinkhorn_knopp(a, b, M, reg, numItermax=1000, stopThr=1e-9,
        Regularization term >0
    numItermax : int, optional
        Max number of iterations
+    warmstart: tuple of arrays, shape (dim_a, dim_b), optional


rflamary · 2023-02-23T09:53:29Z

ot/bregman.py

@@ -546,7 +561,7 @@ def sinkhorn_knopp(a, b, M, reg, numItermax=1000, stopThr=1e-9,
            return u.reshape((-1, 1)) * K * v.reshape((1, -1))


-def sinkhorn_log(a, b, M, reg, numItermax=1000, stopThr=1e-9, verbose=False,
+def sinkhorn_log(a, b, M, reg, numItermax=1000, warmstart=None, stopThr=1e-9, verbose=False,


same here move warmstart please

rflamary · 2023-02-23T09:54:03Z

ot/bregman.py

@@ -746,7 +771,7 @@ def get_logT(u, v):
            return nx.exp(get_logT(u, v))


-def greenkhorn(a, b, M, reg, numItermax=10000, stopThr=1e-9, verbose=False,
+def greenkhorn(a, b, M, reg, numItermax=10000, warmstart=None, stopThr=1e-9, verbose=False,


rflamary · 2023-02-23T09:54:31Z

ot/bregman.py

@@ -789,6 +814,10 @@ def greenkhorn(a, b, M, reg, numItermax=10000, stopThr=1e-9, verbose=False,
        Regularization term >0
    numItermax : int, optional
        Max number of iterations
+    warmstart: tuple of arrays, shape (dim_a, dim_b), optional
+        Initialization of dual vectors. If provided,


rflamary · 2023-02-23T09:55:01Z

ot/bregman.py

@@ -2857,7 +2899,7 @@ def jcpot_barycenter(Xs, Ys, Xt, reg, metric='sqeuclidean', numItermax=100,


 def empirical_sinkhorn(X_s, X_t, reg, a=None, b=None, metric='sqeuclidean',
-                       numIterMax=10000, stopThr=1e-9, isLazy=False, batchSize=100, verbose=False,
+                       numIterMax=10000, warmstart=None, stopThr=1e-9, isLazy=False, batchSize=100, verbose=False,


6Ulm added 2 commits February 20, 2023 18:04

Allow warmstart in sinkhorn and sinkhorn_log

7ba5f03

Added argument for warmstart of dual vectors in Sinkhorn-based method…

eabeabe

…s in

6Ulm changed the title ~~Huy tran~~ Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman Feb 22, 2023

6Ulm added 2 commits February 22, 2023 12:29

Add the number of the PR

cdd9373

[WIP] CO-Optimal Transport

f3d36b2

rflamary changed the title ~~Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman~~ [WIP] Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman Feb 22, 2023

6Ulm and others added 4 commits February 22, 2023 14:39

Revert "[WIP] CO-Optimal Transport"

53e50e5

This reverts commit f3d36b2.

reformat with PEP8

96ea795

Fix W291 trailing whitespace error in pep8 test

96449a0

Merge branch 'master' into huy-tran

49b4975

rflamary reviewed Feb 23, 2023

View reviewed changes

Rearange position of warmstart argument and edit its description

2571802

rflamary changed the title ~~[WIP] Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman~~ [MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman Feb 23, 2023

rflamary merged commit a313e21 into PythonOT:master Feb 23, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in `ot.bregman` #437

[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in `ot.bregman` #437

6Ulm commented Feb 22, 2023

Uh oh!

codecov bot commented Feb 22, 2023 •

edited

Loading

Uh oh!

rflamary left a comment

Uh oh!

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

rflamary Feb 23, 2023

Uh oh!

Uh oh!

	the dual vectors must be already taken the logarithm,
	the dual potentails should be given (that is the logarithm of the u,v sinkhorn scaling vectors)

[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman #437

[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in ot.bregman #437

Conversation

6Ulm commented Feb 22, 2023

Types of changes

Motivation and context / Related issue

How has this been tested (if it applies)

PR checklist

Uh oh!

codecov bot commented Feb 22, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

rflamary left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in `ot.bregman` #437

[MRG] Added argument for warmstart of dual vectors in Sinkhorn-based methods in `ot.bregman` #437

codecov bot commented Feb 22, 2023 •

edited

Loading