Merge branch 'public-main' into node-features-v2

ultrack/core/interactive.py

@@ -264,7 +264,7 @@ def add_new_node(
     node = Node.from_mask(
         time=time,
         mask=mask,
-        bbox=bbox,
+        bbox=np.asarray(bbox),
     )
     if node.area == 0:
         raise ValueError("Node area is zero. Something went wrong.")

ultrack/core/linking/__init__.py

@@ -0,0 +1 @@
+from ultrack.core.linking.processing import add_links

ultrack/core/linking/processing.py

@@ -48,6 +48,61 @@ def _compute_features(
     ]
 
 
+def color_filtering_mask(
+    time: int,
+    current_nodes: List[Node],
+    next_nodes: List[Node],
+    images: Sequence[ArrayLike],
+    neighbors: ArrayLike,
+    z_score_threshold: float,
+) -> ArrayLike:
+    """
+    Filtering by color z-score.
+
+    Parameters
+    ----------
+    time : int
+        Current time.
+    current_nodes : List[Node]
+        List of source nodes.
+    next_nodes : List[Node]
+        List of target nodes.
+    images : Sequence[ArrayLike]
+        Sequence of images to extract color features for filtering.
+    neighbors : ArrayLike
+        Neighbors indices (current/source) for each target (next) node.
+    z_score_threshold : float
+        Z-score threshold for color filtering.
+
+    Returns
+    -------
+    ArrayLike
+        Boolean mask of neighboring nodes within color z-score threshold.
+
+    """
+    LOG.info(f"computing filtering by color z-score from t={time}")
+    (current_features,) = _compute_features(
+        time, current_nodes, images, [Node.intensity_mean]
+    )
+    # inserting dummy value for missing neighbors
+    current_features = np.append(
+        current_features,
+        np.zeros((1, current_features.shape[1])),
+        axis=0,
+    )
+    next_features, next_features_std = _compute_features(
+        time + 1, next_nodes, images, [Node.intensity_mean, Node.intensity_std]
+    )
+    LOG.info(
+        f"Features Std. Dev. range {next_features_std.min()} {next_features_std.max()}"
+    )
+    next_features_std[next_features_std <= 1e-6] = 1.0
+    difference = next_features[:, None, ...] - current_features[neighbors]
+    difference /= next_features_std[:, None, ...]
+    filtered_by_color = np.abs(difference).max(axis=-1) <= z_score_threshold
+    return filtered_by_color
+
+
 @curry
 def _process(
     time: int,
@@ -91,71 +146,86 @@ def _process(
         next_nodes = [row[0] for row in query]
         next_shift = np.asarray([row[1:] for row in query])
 
-    current_pos = np.asarray([n.centroid for n in current_nodes])
-    next_pos = np.asarray([n.centroid for n in next_nodes], dtype=np.float32)
+    compute_spatial_neighbors(
+        time,
+        config,
+        current_nodes,
+        next_nodes,
+        next_shift,
+        scale=scale,
+        table_name=LinkDB.__tablename__,
+        db_path=db_path,
+        images=images,
+        write_lock=write_lock,
+    )
 
-    n_dim = next_pos.shape[1]
-    next_shift = next_shift[:, -n_dim:]  # matching positions dimensions
-    next_pos += next_shift
+
+def compute_spatial_neighbors(
+    time: int,
+    config: LinkingConfig,
+    source_nodes: List[Node],
+    target_nodes: List[Node],
+    target_shift: ArrayLike,
+    scale: Optional[Sequence[float]],
+    table_name: str,
+    db_path: str,
+    images: Sequence[ArrayLike],
+    write_lock: Optional[fasteners.InterProcessLock] = None,
+) -> pd.DataFrame:
+
+    source_pos = np.asarray([n.centroid for n in source_nodes])
+    target_pos = np.asarray([n.centroid for n in target_nodes], dtype=np.float32)
+
+    n_dim = target_pos.shape[1]
+    target_shift = target_shift[:, -n_dim:]  # matching positions dimensions
+    target_pos += target_shift
 
     if scale is not None:
         min_n_dim = min(n_dim, len(scale))
         scale = scale[-min_n_dim:]
-        current_pos = current_pos[..., -min_n_dim:] * scale
-        next_pos = next_pos[..., -min_n_dim:] * scale
+        source_pos = source_pos[..., -min_n_dim:] * scale
+        target_pos = target_pos[..., -min_n_dim:] * scale
 
     # finds neighbors nodes within the radius
     # and connect the pairs with highest edge weight
-    current_kdtree = KDTree(current_pos)
+    current_kdtree = KDTree(source_pos)
 
     distances, neighbors = current_kdtree.query(
-        next_pos,
+        target_pos,
         # twice as expected because we select the nearest with highest edge weight
         k=2 * config.max_neighbors,
         distance_upper_bound=config.max_distance,
     )
 
     if len(images) > 0:
-        LOG.info(f"computing filtering by color z-score from t={time}")
-        (current_features,) = _compute_features(
-            time, current_nodes, images, [Node.intensity_mean]
-        )
-        # inserting dummy value for missing neighbors
-        current_features = np.append(
-            current_features,
-            np.zeros((1, current_features.shape[1])),
-            axis=0,
-        )
-        next_features, next_features_std = _compute_features(
-            time + 1, next_nodes, images, [Node.intensity_mean, Node.intensity_std]
+        filtered_by_color = color_filtering_mask(
+            time,
+            source_nodes,
+            target_nodes,
+            images,
+            neighbors,
+            config.z_score_threshold,
         )
-        LOG.info(
-            f"Features Std. Dev. range {next_features_std.min()} {next_features_std.max()}"
-        )
-        next_features_std[next_features_std <= 1e-6] = 1.0
-        difference = next_features[:, None, ...] - current_features[neighbors]
-        difference /= next_features_std[:, None, ...]
-        filtered_by_color = np.abs(difference).max(axis=-1) <= config.z_score_threshold
     else:
         filtered_by_color = np.ones_like(neighbors, dtype=bool)
 
-    int_next_shift = np.round(next_shift).astype(int)
+    int_next_shift = np.round(target_shift).astype(int)
     # NOTE: moving bbox with shift, MUST be after `feature computation`
-    for node, shift in zip(next_nodes, int_next_shift):
+    for node, shift in zip(target_nodes, int_next_shift):
         node.bbox[:n_dim] += shift
         node.bbox[-n_dim:] += shift
 
     distance_w = config.distance_weight
     links = []
 
-    for i, node in enumerate(next_nodes):
+    for i, node in enumerate(target_nodes):
         valid = (~np.isinf(distances[i])) & filtered_by_color[i]
         valid_neighbors = neighbors[i, valid]
         neigh_distances = distances[i, valid]
 
         neighborhood = []
         for neigh_idx, neigh_dist in zip(valid_neighbors, neigh_distances):
-            neigh = current_nodes[neigh_idx]
+            neigh = source_nodes[neigh_idx]
             edge_weight = node.IoU(neigh) - distance_w * neigh_dist
             # using dist as a tie-breaker
             neighborhood.append(
@@ -176,13 +246,14 @@ def _process(
 
     with write_lock if write_lock is not None else nullcontext():
         LOG.info(f"Pushing links from time {time} to {db_path}")
+        connect_args = {"timeout": 45} if write_lock is not None else {}
         engine = sqla.create_engine(
             db_path, hide_parameters=True, connect_args=connect_args
         )
         with engine.begin() as conn:
-            df.to_sql(
-                name=LinkDB.__tablename__, con=conn, if_exists="append", index=False
-            )
+            df.to_sql(name=table_name, con=conn, if_exists="append", index=False)
+
+    return df
 
 
 def link(
@@ -230,3 +301,40 @@ def link(
         multiprocessing_apply(
             process, time_points, config.linking_config.n_workers, desc="Linking nodes."
         )
+
+
+def add_links(
+    config: MainConfig,
+    sources: ArrayLike,
+    targets: ArrayLike,
+    weights: ArrayLike,
+) -> None:
+    """
+    Adds user-defined links to the database.
+
+    Parameters
+    ----------
+    config : MainConfig
+        Configuration parameters.
+    sources : ArrayLike
+        Sources (t) node id.
+    targets : ArrayLike
+        Targets (t + 1) node id.
+    weights : ArrayLike
+        Link weights, the higher the weight the more likely the link.
+    """
+    df = pd.DataFrame(
+        {
+            "source_id": np.asarray(sources, dtype=int),
+            "target_id": np.asarray(targets, dtype=int),
+            "weight": weights,
+        }
+    )
+
+    engine = sqla.create_engine(
+        config.data_config.database_path,
+        hide_parameters=True,
+    )
+
+    with engine.begin() as conn:
+        df.to_sql(name=LinkDB.__tablename__, con=conn, if_exists="append", index=False)

ultrack/core/tracker.py

@@ -16,7 +16,7 @@
     tracks_layer_to_trackmate,
     tracks_to_zarr,
 )
-from ultrack.core.linking.processing import link
+from ultrack.core.linking.processing import add_links, link
 from ultrack.core.main import track
 from ultrack.core.segmentation.processing import get_nodes_features, segment
 from ultrack.core.solve.processing import solve
@@ -165,6 +165,12 @@ def get_nodes_features(self, **kwargs) -> pd.DataFrame:
         nodes_features_df = get_nodes_features(self.config, **kwargs)
         return nodes_features_df
 
+    @functools.wraps(add_links)
+    def add_links(self, **kwargs) -> None:
+        self._assert_segmented("add_links")
+        add_links(config=self.config, **kwargs)
+        self.status |= TrackerStatus.LINKED
+
     @functools.wraps(add_nodes_prob)
     def add_nodes_prob(
         self,