[MRG] Sliced and 1D Wasserstein distances : backend versions (#256)

* add numpy and torch backends

* stat sets on functions

* proper import

* install recent torch on windows

* install recent torch on windows

* now testing all functions in backedn

* add jax backedn

* clenaup windowds

* proper convert for jax backedn

* pep8

* try again windows tests

* test jax conversion

* try proper widows tests

* emd fuction ses backedn

* better test partial OT

* proper tests to_numpy and teplate Backend

* pep8

* pep8 x2

* feaking sinkhorn works with torch

* sinkhorn2 compatible

* working ot.emd2

* important detach

* it should work

* jax autodiff emd

* pep8

* no tast same for jax

* new independat tests per backedn

* freaking pep8

* add tests for gradients

* deprecate ot.gpu

* worging dist function

* working dist

* dist done in backedn

* not in

* remove indexing

* change accuacy for jax

* first pull backend

* projection simplex

* projection simplex

* projection simplex

* projection simplex no ci

* projection simplex no ci

* projection simplex no ci

* pep8

* add backedn discusion to quickstart guide

* projection simplex no ci

* projection simplex no ci

* projection simplex no ci

* pep8 + better doc

* proper links

* corect doctest

* big debug documentation

* doctest again

* doctest again bis

* doctest again ter (last one or i kill myself)

* backend test + doc proj simplex

* correction test_utils

* correction test_utils

* correction cumsum

* correction flip

* correction flip v2

* more debug

* more debug

* more debug + pep8

* pep8

* argh

* proj_simplex

* backedn works for sort

* proj simplex

* jax sucks

* update doc

* Update test/test_utils.py

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update docs/source/quickstart.rst

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update docs/source/quickstart.rst

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update docs/source/quickstart.rst

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update docs/source/readme.rst

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update test/test_utils.py

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update ot/utils.py

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update docs/source/readme.rst

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* Update ot/lp/__init__.py

Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

* begin comment alex

* comment alex part 2

* optimize test gromov

* proj_simplex on vectors

* add awesome gradient decsnt example on the weights

* pep98 of course

* proof read example by alex

* pep8 again

* encoding oos in translation

* correct legend

* new backend functions for sliced

* small indent pb

* Optimized backendversion of sliced W

* error in sliced W

* after master merge

* error sliced

* error sliced

* pep8

* test_sliced pep8

* doctest + precision for sliced

* doctest

* type win test_backend gather

* type win test_backend gather

* Update sliced.py

change argument of padding pad_width

* Update backend.py

update redefinition

* Update backend.py

pep8

* Update backend.py

pep 8 again....

* pep8

* build docs

* emd2_1D example

* refectoring emd_1d and variants

* remove unused previous wasserstein_1d

* pep8

* upate example

* move stuff

* tesys should work + implemù random backend

* test random generayor functions

* correction

* better random generation

* update sliced

* update sliced

* proper tests sliced

* max sliced

* chae file nam

* add stuff

* example sliced flow and barycenter

* correct typo + update readme

* exemple sliced flow done

* pep8

* solver1d works

* pep8

Co-authored-by: Rémi Flamary <remi.flamary@gmail.com>
Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>

1 files changed, 85 insertions, 0 deletions

diff --git a/test/test_1d_solver.py b/test/test_1d_solver.py
new file mode 100644
index 0000000..2c470c2
--- /dev/null
+++ b/test/test_1d_solver.py
@@ -0,0 +1,85 @@
+"""Tests for module 1d Wasserstein solver"""
+
+# Author: Adrien Corenflos <adrien.corenflos@aalto.fi>
+#         Nicolas Courty <ncourty@irisa.fr>
+#
+# License: MIT License
+
+import numpy as np
+import pytest
+
+import ot
+from ot.lp import wasserstein_1d
+
+from ot.backend import get_backend_list
+from scipy.stats import wasserstein_distance
+
+backend_list = get_backend_list()
+
+
+def test_emd_1d_emd2_1d_with_weights():
+    # test emd1d gives similar results as emd
+    n = 20
+    m = 30
+    rng = np.random.RandomState(0)
+    u = rng.randn(n, 1)
+    v = rng.randn(m, 1)
+
+    w_u = rng.uniform(0., 1., n)
+    w_u = w_u / w_u.sum()
+
+    w_v = rng.uniform(0., 1., m)
+    w_v = w_v / w_v.sum()
+
+    M = ot.dist(u, v, metric='sqeuclidean')
+
+    G, log = ot.emd(w_u, w_v, M, log=True)
+    wass = log["cost"]
+    G_1d, log = ot.emd_1d(u, v, w_u, w_v, metric='sqeuclidean', log=True)
+    wass1d = log["cost"]
+    wass1d_emd2 = ot.emd2_1d(u, v, w_u, w_v, metric='sqeuclidean', log=False)
+    wass1d_euc = ot.emd2_1d(u, v, w_u, w_v, metric='euclidean', log=False)
+
+    # check loss is similar
+    np.testing.assert_allclose(wass, wass1d)
+    np.testing.assert_allclose(wass, wass1d_emd2)
+
+    # check loss is similar to scipy's implementation for Euclidean metric
+    wass_sp = wasserstein_distance(u.reshape((-1,)), v.reshape((-1,)), w_u, w_v)
+    np.testing.assert_allclose(wass_sp, wass1d_euc)
+
+    # check constraints
+    np.testing.assert_allclose(w_u, G.sum(1))
+    np.testing.assert_allclose(w_v, G.sum(0))
+
+
+@pytest.mark.parametrize('nx', backend_list)
+def test_wasserstein_1d(nx):
+    from scipy.stats import wasserstein_distance
+
+    rng = np.random.RandomState(0)
+
+    n = 100
+    x = np.linspace(0, 5, n)
+    rho_u = np.abs(rng.randn(n))
+    rho_u /= rho_u.sum()
+    rho_v = np.abs(rng.randn(n))
+    rho_v /= rho_v.sum()
+
+    xb = nx.from_numpy(x)
+    rho_ub = nx.from_numpy(rho_u)
+    rho_vb = nx.from_numpy(rho_v)
+
+    # test 1 : wasserstein_1d should be close to scipy W_1 implementation
+    np.testing.assert_almost_equal(wasserstein_1d(xb, xb, rho_ub, rho_vb, p=1),
+                                   wasserstein_distance(x, x, rho_u, rho_v))
+
+    # test 2 : wasserstein_1d should be close to one when only translating the support
+    np.testing.assert_almost_equal(wasserstein_1d(xb, xb + 1, p=2),
+                                   1.)
+
+    # test 3 : arrays test
+    X = np.stack((np.linspace(0, 5, n), np.linspace(0, 5, n) * 10), -1)
+    Xb = nx.from_numpy(X)
+    res = wasserstein_1d(Xb, Xb, rho_ub, rho_vb, p=2)
+    np.testing.assert_almost_equal(100 * res[0], res[1], decimal=4)