Merge pull request #86 from francois-drielsma/develop

Vectorized PPN loss cluster restriction routine, now much cheaper

Author: francois-drielsma

spine/model/layer/cnn/ppn.py

@@ -522,30 +522,24 @@ def get_ppn_positives(coords: torch.Tensor,
         closests : torch.Tensor
             (N) tensor of the closest label point index
         """
-        # Initialize the pixel assignment and the closest index tensors
-        device = coords.device
-        positives = torch.zeros(coords.shape[0], device=device, dtype=torch.bool)
-        closests = -torch.ones(coords.shape[0], device=device, dtype=torch.long)
-
-        # Loop over all unique particles with points within this entry
-        part_ids = ppn_labels[:, PPN_LPART_COL]
-        for part_id in torch.unique(part_ids):
-            # Restrict the set of labels points and voxels to one particle
-            point_index = torch.where(part_ids == part_id)[0]
-            index = torch.where(labels == part_id)[0]
-            points = ppn_labels[point_index][:, COORD_COLS]
-
-            # Compute the pairwise distance between the particle voxels and its
-            # label points
-            dist_mat = cdist_fast(coords[index], points)
-
-            # Generate a positive mask for all particle voxels within some
-            # distance of its label points
-            positives[index] = (dist_mat < resolution).any(dim=1)
-
-            # Assign the closest label point to each particle voxel
-            min_return = torch.min(dist_mat, dim=1)
-            closests[index] = point_index[min_return.indices] + offset
+
+        # Compute the distance from the PPN labels to all the image points
+        dist_mat = cdist_fast(ppn_labels[:, COORD_COLS], coords)
+
+        # Mask out particle voxels for which the particle ID disagrees
+        bad_mask = ppn_labels[:, [PPN_LPART_COL]] != labels
+        dist_mat[bad_mask] = torch.inf
+
+        # Generate a positive mask for all particle voxels within some
+        # distance of their label points
+        positives = (dist_mat < resolution).any(dim=0)
+
+        # Assign the closest label point to each postive particle voxel
+        pos_index = torch.where(positives)[0]
+        closests = torch.full(
+                (len(labels),), -1, dtype=torch.long, device=labels.device)
+        closests[pos_index] = offset + torch.argmin(
+                dist_mat[:, pos_index], dim=0)
 
         return positives, closests