nonuniform global

src/cryo_challenge/_map_to_map/gromov_wasserstein/gw_weighted_voxels.py

@@ -36,7 +36,7 @@ def make_sparse_cost(idx_above_thresh, dtype):
     coordinates = coordinates.reshape(-1, 3)
     sparse_coordiantes = coordinates[idx_above_thresh.flatten()]
     pairwise_distances = torch.cdist(sparse_coordiantes, sparse_coordiantes)
-    return pairwise_distances
+    return pairwise_distances, sparse_coordiantes
 
 
 def normalize_mass_to_one(p):
@@ -52,15 +52,15 @@ def prepare_volume_and_distance(
     volume = downsample_volume(volume, n_downsample_pix).numpy().astype(numpy_dtype)
     idx_above_thresh = return_top_k_voxel_idxs(volume, top_k)
     marginal = normalize_mass_to_one(volume[idx_above_thresh].flatten())
-    pairwise_distance = (
-        make_sparse_cost(idx_above_thresh, dtype=torch_dtype)
-        .numpy()
-        .astype(numpy_dtype)
-    )
+    pairwise_distance, sparse_coordiantes = [
+        x.numpy().astype(numpy_dtype)
+        for x in make_sparse_cost(idx_above_thresh, dtype=torch_dtype)
+    ]
+
     if normalize:
         pairwise_distance /= pairwise_distance.max()
     pairwise_distance = (cost_scale_factor * pairwise_distance) ** exponent
-    return marginal, pairwise_distance
+    return marginal, sparse_coordiantes, pairwise_distance
 
 
 def gw_distance_wrapper_element_wise(
@@ -371,35 +371,45 @@ def setup_volume_and_distance(
 
     marginals_i = np.empty((len(volumes_i), top_k))
     marginals_j = np.empty((len(volumes_j), top_k))
+    sparse_coordinates_sets_i = np.empty((len(volumes_j), top_k, 3))
+    sparse_coordinates_sets_j = np.empty((len(volumes_j), top_k, 3))
     pairwise_distances_i = np.empty((len(volumes_i), top_k, top_k))
     pairwise_distances_j = np.empty((len(volumes_j), top_k, top_k))
 
     for i in range(len(volumes_i)):
-        volume_i, pairwise_distance_i = prepare_volume_and_distance(
-            volumes_i[i],
-            top_k,
-            n_downsample_pix,
-            exponent,
-            cost_scale_factor,
-            normalize,
+        volume_i, sparse_coordinates_i, pairwise_distance_i = (
+            prepare_volume_and_distance(
+                volumes_i[i],
+                top_k,
+                n_downsample_pix,
+                exponent,
+                cost_scale_factor,
+                normalize,
+            )
         )
         marginals_i[i] = volume_i
+        sparse_coordinates_sets_i[i] = sparse_coordinates_i
         pairwise_distances_i[i] = pairwise_distance_i
     for j in range(len(volumes_j)):
-        volume_j, pairwise_distance_j = prepare_volume_and_distance(
-            volumes_j[j],
-            top_k,
-            n_downsample_pix,
-            exponent,
-            cost_scale_factor,
-            normalize,
+        volume_j, sparse_coordinates_j, pairwise_distance_j = (
+            prepare_volume_and_distance(
+                volumes_j[j],
+                top_k,
+                n_downsample_pix,
+                exponent,
+                cost_scale_factor,
+                normalize,
+            )
         )
         marginals_j[j] = volume_j
+        sparse_coordinates_sets_j[j] = sparse_coordinates_j
         pairwise_distances_j[j] = pairwise_distance_j
 
     return (
         marginals_i,
         marginals_j,
+        sparse_coordinates_sets_i,
+        sparse_coordinates_sets_j,
         pairwise_distances_i,
         pairwise_distances_j,
         volumes_i,
@@ -419,6 +429,8 @@ def main(args):
     (
         marginals_i,
         marginals_j,
+        _,
+        _,
         pairwise_distances_i,
         pairwise_distances_j,
         volumes_i,

src/cryo_challenge/_map_to_map/gromov_wasserstein/qp.py

@@ -141,6 +141,8 @@ def main(args):
         marginals_j,
         pairwise_distances_i,
         pairwise_distances_j,
+        _,
+        _,
         volumes_i,
         volumes_j,
     ) = setup_volume_and_distance(