refactor: extract fixed image patch in features

src/deep_neurographs/generate_proposals.py

@@ -109,8 +109,8 @@ def run(
     if trim_endpoints_bool:
         radius /= RADIUS_SCALING_FACTOR
         long_range, in_range = separate_proposals(neurograph, radius)
-        neurograph = run_trimming(neurograph, long_range, radius)
-        neurograph = run_trimming(neurograph, in_range, radius)
+        neurograph = run_trimming(neurograph, long_range, radius, progress_bar)
+        neurograph = run_trimming(neurograph, in_range, radius, progress_bar)
 
 
 def init_kdtree(neurograph, complex_bool):
@@ -297,7 +297,7 @@ def separate_proposals(neurograph, radius):
 
 
 # --- Trim Endpoints ---
-def run_trimming(neurograph, proposals, radius):
+def run_trimming(neurograph, proposals, radius, progress_bar):
     n_endpoints_trimmed = 0
     long_radius = radius * RADIUS_SCALING_FACTOR
     for proposal in deepcopy(proposals):
@@ -312,7 +312,8 @@ def run_trimming(neurograph, proposals, radius):
         elif neurograph.dist(i, j) > radius:
             neurograph.remove_proposal(proposal)
         n_endpoints_trimmed += 1 if trim_bool else 0
-    print("# Endpoints Trimmed:", n_endpoints_trimmed)
+    if progress_bar:
+        print("# Endpoints Trimmed:", n_endpoints_trimmed)
     return neurograph
 
 

src/deep_neurographs/groundtruth_generation.py

@@ -16,7 +16,6 @@
 
 from deep_neurographs import geometry
 from deep_neurographs.geometry import dist as get_dist
-from deep_neurographs.utils import graph_util as gutil
 from deep_neurographs.utils import util
 
 ALIGNED_THRESHOLD = 4
@@ -132,8 +131,8 @@ def is_component_aligned(target_graph, pred_graph, nodes, kdtree):
     Returns
     -------
     bool
-        Indication of whether connected component "nodes" is aligned to a connected
-        component in "target_graph".
+        Indication of whether connected component "nodes" is aligned to a
+        connected component in "target_graph".
 
     """
     # Compute distances

src/deep_neurographs/inference.py

@@ -152,7 +152,9 @@ def run(self, fragments_pointer):
         t, unit = util.time_writer(time() - t0)
         print(f"Total Runtime: {round(t, 4)} {unit}\n")
 
-    def run_schedule(self, fragments_pointer, search_radius_schedule):
+    def run_schedule(
+        self, fragments_pointer, search_radius_schedule, save_all_rounds=False
+    ):
         t0 = time()
         self.report_experiment()
         self.build_graph(fragments_pointer)
@@ -161,7 +163,12 @@ def run_schedule(self, fragments_pointer, search_radius_schedule):
             round_id += 1
             self.generate_proposals(search_radius)
             self.run_inference()
+            if save_all_rounds:
+                self.save_results(round_id=round_id)
+
+        if not save_all_rounds:
             self.save_results(round_id=round_id)
+
         t, unit = util.time_writer(time() - t0)
         print(f"Total Runtime: {round(t, 4)} {unit}\n")
 
@@ -263,7 +270,7 @@ def run_inference(self):
         )
         self.accepted_proposals.extend(accepts)
         print("# Accepted:", util.reformat_number(len(accepts)))
-        print("% Accepted:", len(accepts) / n_proposals)
+        print("% Accepted:", round(len(accepts) / n_proposals, 4))
 
         t, unit = util.time_writer(time() - t0)
         print(f"Module Runtime: {round(t, 4)} {unit}\n")

src/deep_neurographs/machine_learning/feature_generation.py

@@ -28,7 +28,7 @@
 )
 from deep_neurographs.utils import img_util, util
 
-CHUNK_SIZE = [64, 64, 64]
+CHUNK_SIZE = [48, 48, 48]
 N_BRANCH_PTS = 50
 N_PROFILE_PTS = 16  # 10
 N_SKEL_FEATURES = 22
@@ -192,7 +192,7 @@ def get_profile_specs(xyz_1, xyz_2, downsample_factor):
     voxel_2 = img_util.to_voxels(xyz_2, downsample_factor=downsample_factor)
 
     # Store local coordinates
-    bbox = img_util.get_minimal_bbox(np.vstack([voxel_1, voxel_2]), buffer=1)
+    bbox = img_util.get_fixed_bbox(np.vstack([voxel_1, voxel_2]), CHUNK_SIZE)
     start = [voxel_1[i] - bbox["min"][i] for i in range(3)]
     end = [voxel_2[i] - bbox["min"][i] for i in range(3)]
     specs = {

src/deep_neurographs/machine_learning/feature_generation_graphs.py

@@ -230,10 +230,10 @@ def proposal_skeletal(neurograph, proposals, radius):
                 neurograph.proposal_length(proposal) / radius,
                 neurograph.n_nearby_leafs(proposal, radius),
                 neurograph.proposal_radii(proposal),
-                neurograph.proposal_directionals(proposal, 8),
                 neurograph.proposal_directionals(proposal, 16),
                 neurograph.proposal_directionals(proposal, 32),
                 neurograph.proposal_directionals(proposal, 64),
+                neurograph.proposal_directionals(proposal, 128),
             ),
             axis=None,
         )

src/deep_neurographs/neurograph.py

@@ -660,16 +660,14 @@ def proposal_directionals(self, proposal, window):
 
     def merge_proposal(self, proposal):
         i, j = tuple(proposal)
-        somas_check = not (self.is_soma(i) and self.is_soma(j))        
+        somas_check = not (self.is_soma(i) and self.is_soma(j))
         if somas_check and self.check_proposal_degrees(i, j):
             # Dense attributes
             attrs = dict()
             self.nodes[i]["radius"] = 7.0
             self.nodes[j]["radius"] = 7.0
             for k in ["xyz", "radius"]:
                 combine = np.vstack if k == "xyz" else np.array
-                self.nodes[i][k][-1] = 8.0
-                self.nodes[j][k][0] = 8.0
                 attrs[k] = combine([self.nodes[i][k], self.nodes[j][k]])
 
             # Sparse attributes
@@ -701,7 +699,7 @@ def check_proposal_degrees(self, i, j):
         one_leaf = self.degree[i] == 1 or self.degree[j] == 1
         branching = self.degree[i] > 2 or self.degree[j] > 2
         return one_leaf and not branching
-        
+
     def upd_ids(self, swc_id, r):
         """
         Updates the swc_id of all nodes connected to "r".

src/deep_neurographs/utils/graph_util.py

@@ -462,7 +462,7 @@ def prune_branch(graph, leaf, prune_depth):
         # Check whether to stop
         if np.sum(node_dists) > prune_depth:
             break
-    return branch[0:min(4, len(branch))]
+    return branch[0:min(5, len(branch))]
 
 
 def smooth_branch(swc_dict, attrs, edges, nbs, root, j):

src/deep_neurographs/utils/img_util.py

@@ -142,8 +142,7 @@ def read_tensorstore_with_bbox(img, bbox):
     try:
         shape = [bbox["max"][i] - bbox["min"][i] for i in range(3)]
         return read_tensorstore(img, bbox["min"], shape, from_center=False)
-    except Exception as e:
-        print(type(e), e)
+    except Exception:
         return np.zeros(shape)
 
 
@@ -186,7 +185,7 @@ def read_intensities(img, voxels):
         Image intensities.
 
     """
-    return [img[tuple(voxel)] for voxel in voxels]
+    return [img[voxel] for voxel in map(tuple, voxels)]
 
 
 def get_start_end(voxel, shape, from_center=True):
@@ -405,7 +404,7 @@ def get_bbox(origin, shape):
         return None
 
 
-def get_minimal_bbox(voxels, buffer=0):
+def get_minimal_bbox(voxels):
     """
     Gets the min and max coordinates of a bounding box that contains "voxels".
 
@@ -425,7 +424,16 @@ def get_minimal_bbox(voxels, buffer=0):
     """
     bbox = {
         "min": np.floor(np.min(voxels, axis=0) - 1).astype(int),
-        "max": np.ceil(np.max(voxels, axis=0) + buffer + 1).astype(int),
+        "max": np.ceil(np.max(voxels, axis=0) + 1).astype(int),
+    }
+    return bbox
+
+
+def get_fixed_bbox(voxels, shape):
+    centroid = np.round(np.mean(voxels, axis=0)).astype(int)
+    bbox = {
+        "min": [centroid[i] - shape[i] // 2 for i in range(3)],
+        "max": [centroid[i] + shape[i] // 2 for i in range(3)],
     }
     return bbox