Add sparse Chebyshev approximation (#502)

* wip bcoo geodesic

* add sparse laplacian

* add test & pass test

* typo & fix type

* appease code formatter

* fmt

* fix laplacian type

* norm lap with elementwise multiplication

* unify tests

* fix sparse scan & sinkhorn test

* fmt

* rm mv to sparse since `@jesp.sparsify`

* rm sparsify wrapper and fix type

* fix type & mv fn & test memory

Author: guillaumehu

src/ott/geometry/geodesic.py

@@ -11,7 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Any, Dict, Optional, Sequence, Tuple
+from typing import Any, Dict, Optional, Sequence, Tuple, Union
 
 import jax
 import jax.experimental.sparse as jesp
@@ -22,10 +22,11 @@
 from ott import utils
 from ott.geometry import geometry
 from ott.math import utils as mu
-from ott.types import Array_g
 
 __all__ = ["Geodesic"]
 
+Array_g = Union[jnp.ndarray, jesp.BCOO]
+
 
 @jax.tree_util.register_pytree_node_class
 class Geodesic(geometry.Geometry):
@@ -106,13 +107,10 @@ def from_graph(
     if t is None:
       t = (jnp.sum(G) / jnp.sum(G > 0.0)) ** 2
 
-    degree = jnp.sum(G, axis=1)
-    laplacian = jnp.diag(degree) - G
-    if normalize:
-      inv_sqrt_deg = jnp.diag(
-          jnp.where(degree > 0.0, 1.0 / jnp.sqrt(degree), 0.0)
-      )
-      laplacian = inv_sqrt_deg @ laplacian @ inv_sqrt_deg
+    if isinstance(G, jesp.BCOO):
+      laplacian = compute_sparse_laplacian(G, normalize)
+    else:
+      laplacian = compute_dense_laplacian(G, normalize)
 
     if eigval is None:
       eigval = compute_largest_eigenvalue(laplacian, rng)
@@ -220,6 +218,33 @@ def tree_unflatten(  # noqa: D102
     return cls(*children, **aux_data)
 
 
+def normalize_laplacian(laplacian: Array_g, degree: jnp.ndarray) -> Array_g:
+  inv_sqrt_deg = jnp.where(degree > 0.0, 1.0 / jnp.sqrt(degree), 0.0)
+  return inv_sqrt_deg[:, None] * laplacian * inv_sqrt_deg[None, :]
+
+
+def compute_dense_laplacian(
+    G: jnp.ndarray, normalize: bool = False
+) -> jnp.ndarray:
+  degree = jnp.sum(G, axis=1)
+  laplacian = jnp.diag(degree) - G
+  if normalize:
+    laplacian = normalize_laplacian(laplacian, degree)
+  return laplacian
+
+
+def compute_sparse_laplacian(
+    G: jesp.BCOO, normalize: bool = False
+) -> jesp.BCOO:
+  n, _ = G.shape
+  data_degree, ixs = G.sum(1).todense(), jnp.arange(n)
+  degree = jesp.BCOO((data_degree, jnp.c_[ixs, ixs]), shape=(n, n))
+  laplacian = degree - G
+  if normalize:
+    laplacian = normalize_laplacian(laplacian, data_degree)
+  return laplacian
+
+
 def compute_largest_eigenvalue(
     laplacian_matrix: jnp.ndarray,
     rng: jax.Array,
@@ -242,7 +267,7 @@ def compute_largest_eigenvalue(
 
 
 def expm_multiply(
-    L: jnp.ndarray, X: jnp.ndarray, coeff: jnp.ndarray, eigval: float
+    L: Array_g, X: jnp.ndarray, coeff: jnp.ndarray, eigval: float
 ) -> jnp.ndarray:
 
   def body(carry, c):

src/ott/types.py

@@ -11,18 +11,14 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import Protocol, Union
+from typing import Protocol
 
-import jax.experimental.sparse as jesp
 import jax.numpy as jnp
 
-__all__ = ["Transport", "Array_g"]
+__all__ = ["Transport"]
 
 # TODO(michalk8): introduce additional types here
 
-# Either a dense or sparse array.
-Array_g = Union[jnp.ndarray, jesp.BCOO]
-
 
 class Transport(Protocol):
   """Interface for the solution of a transport problem.

tests/geometry/geodesic_test.py

@@ -14,6 +14,7 @@
 from typing import Optional, Union
 
 import jax
+import jax.experimental.sparse as jesp
 import jax.numpy as jnp
 import networkx as nx
 import numpy as np
@@ -29,6 +30,7 @@ def random_graph(
     n: int,
     p: float = 0.3,
     seed: Optional[int] = 0,
+    is_sparse: bool = False,
     *,
     return_laplacian: bool = False,
     directed: bool = False,
@@ -45,6 +47,8 @@ def random_graph(
       G
   ) if return_laplacian else nx.linalg.adjacency_matrix(G)
 
+  if is_sparse:
+    return jesp.BCOO.from_scipy_sparse(G)
   return jnp.asarray(G.toarray())
 
 
@@ -196,7 +200,8 @@ def laplacian(G: jnp.ndarray) -> jnp.ndarray:
     np.testing.assert_allclose(actual, expected, rtol=1e-5, atol=1e-5)
 
   @pytest.mark.fast.with_args(jit=[False, True], only_fast=0)
-  def test_geo_sinkhorn(self, rng: jax.Array, jit: bool):
+  @pytest.mark.parametrize("is_sparse", [True, False])
+  def test_geo_sinkhorn(self, rng: jax.Array, jit: bool, is_sparse: bool):
 
     def callback(geom: geometry.Geometry) -> sinkhorn.SinkhornOutput:
       solver = sinkhorn.Sinkhorn(lse_mode=False)
@@ -208,6 +213,8 @@ def callback(geom: geometry.Geometry) -> sinkhorn.SinkhornOutput:
     x = jax.random.normal(rng, (n,))
 
     gt_geom = gt_geometry(G, epsilon=eps)
+    if is_sparse:
+      G = jesp.BCOO.fromdense(G)
     graph_geom = geodesic.Geodesic.from_graph(G, t=eps / 4.0)
 
     fn = jax.jit(callback) if jit else callback
@@ -257,11 +264,29 @@ def callback(geom: geodesic.Geodesic) -> float:
   @pytest.mark.parametrize("normalize", [False, True])
   @pytest.mark.parametrize("t", [5, 10, 50])
   @pytest.mark.parametrize("order", [20, 30, 40])
-  def test_heat_approx(self, normalize: bool, t: float, order: int):
+  @pytest.mark.parametrize("is_sparse", [True, False])
+  def test_heat_approx(
+      self, normalize: bool, t: float, order: int, is_sparse: bool
+  ):
     G = random_graph(20, p=0.5)
     exact = exact_heat_kernel(G, normalize=normalize, t=t)
+    if is_sparse:
+      G = jesp.BCOO.fromdense(G)
     geom = geodesic.Geodesic.from_graph(
         G, t=t, order=order, normalize=normalize
     )
     approx = geom.apply_kernel(jnp.eye(G.shape[0]))
+
     np.testing.assert_allclose(exact, approx, rtol=1e-1, atol=1e-1)
+
+  @pytest.mark.limit_memory("150 MB")
+  def test_sparse_geo_memory(self, rng: jax.Array):
+    n = 10_000
+    G = random_graph(n, p=0.001, is_sparse=True)
+    x = jax.random.normal(rng, (n,))
+
+    graph_geom = geodesic.Geodesic.from_graph(G, t=1.0, order=10)
+
+    out = jax.jit(graph_geom.apply_kernel)(x)
+
+    np.testing.assert_array_equal(jnp.isfinite(out), True)