Merge pull request #80 from francois-drielsma/develop

Cluster label adaptation speed-up

Author: francois-drielsma

bin/larcv_check_valid.py

@@ -44,7 +44,13 @@ def main(source, source_list, output):
     keys_list, unique_counts = [], []
     for file_path in tqdm(source):
         # Count the number of entries in each tree
-        f = TFile(file_path)
+        try:
+            f = TFile(file_path)
+        except OSError:
+            keys_list.append([])
+            unique_counts.append([])
+            continue
+
         keys = [key.GetName() for key in f.GetListOfKeys()]
         trees = [f.Get(key) for key in keys]
         num_entries = [tree.GetEntries() for tree in trees]

spine/data/out/particle.py

@@ -485,7 +485,7 @@ class TruthParticle(Particle, ParticleBase, TruthBase):
     orig_interaction_id: int = -1
     orig_parent_id: int = -1
     orig_group_id: int = -1
-    orig_children_id: np.ndarray = -1
+    orig_children_id: np.ndarray = None
     children_counts: np.ndarray = None
     reco_length: float = -1.
     reco_start_dir: np.ndarray = None

spine/model/full_chain.py

@@ -23,8 +23,8 @@
 from spine.utils.calib import CalibrationManager
 from spine.utils.logger import logger
 from spine.utils.ppn import get_particle_points
-from spine.utils.ghost import (
-        compute_rescaled_charge_batch, adapt_labels_batch)
+from spine.utils.ghost import compute_rescaled_charge_batch
+from spine.utils.cluster.label import ClusterLabelAdapter
 from spine.utils.gnn.cluster import (
         form_clusters_batch, get_cluster_label_batch)
 from spine.utils.gnn.evaluation import primary_assignment_batch
@@ -173,8 +173,7 @@ def __init__(self, chain, uresnet_deghost=None, uresnet=None,
                 self.uresnet_ppn = UResNetPPN(**uresnet_ppn)
 
         # Initialize the relabeling process (adapt to the semantic predictions)
-        # TODO: make this a class which holds onto these parameters?
-        self.adapt_params = adapt_labels if adapt_labels is not None else {}
+        self.label_adapter = ClusterLabelAdapter(**(adapt_labels or {}))
 
         # Initialize the dense clustering model
         self.fragment_shapes = []
@@ -495,9 +494,8 @@ def run_segmentation_ppn(self, data, seg_label=None, clust_label=None):
             if seg_label is not None and clust_label is not None:
                 seg_pred = self.result['seg_pred']
                 ghost_pred = self.result.get('ghost_pred', None)
-                clust_label = adapt_labels_batch(
-                        clust_label, seg_label, seg_pred, ghost_pred,
-                        **self.adapt_params)
+                clust_label = self.label_adapter(
+                        clust_label, seg_label, seg_pred, ghost_pred)
 
                 self.result['clust_label_adapt'] = clust_label
 

spine/utils/cluster/label.py

@@ -0,0 +1,306 @@
+"""Class which adapts clustering labels given upstream semantic predictions."""
+
+import numpy as np
+import torch
+from torch_cluster import knn
+from scipy.spatial.distance import cdist
+
+from spine.data import TensorBatch
+
+from spine.utils.gnn.cluster import form_clusters, break_clusters
+from spine.utils.globals import (
+        COORD_COLS, VALUE_COL, CLUST_COL, SHAPE_COL, SHOWR_SHP, TRACK_SHP,
+        MICHL_SHP, DELTA_SHP, GHOST_SHP)
+
+__all__ = ['ClusterLabelAdapter']
+
+
+class ClusterLabelAdapter:
+    """Adapts the cluster labels to account for the predicted semantics.
+
+    Points wrongly predicted get the cluster label of the closest touching
+    cluster, if there is one. Points that are predicted as ghosts get invalid
+    (-1) cluster labels everywhere.
+
+    Instances that have been broken up by the deghosting or semantic
+    segmentation process get assigned distinct cluster labels for each
+    effective fragment, provided they appearing in the `break_classes` list.
+
+    Notes
+    -----
+    This class supports both Numpy arrays and Torch tensors.
+
+    It uses the GPU implementation from `torch_cluster.knn` to speed up the
+    label adaptation computation (instead of cdist).
+
+    """
+
+    def __init__(self, break_eps=1.1, break_metric='chebyshev',
+                 break_classes=[SHOWR_SHP, TRACK_SHP, MICHL_SHP, DELTA_SHP]):
+        """Initialize the adapter class.
+
+        Parameters
+        ----------
+        dtype : str, default 'torch'
+            Type of data to be processed through the label adapter
+        break_eps : float, default 1.1
+            Distance scale used in the break up procedure
+        break_metric : str, default 'chebyshev'
+            Distance metric used in the break up produce
+        break_classes : List[int], default
+                        [SHOWR_SHP, TRACK_SHP, MICHL_SHP, DELTA_SHP]
+            Classes to run DBSCAN on to break up
+        """
+        # Store relevant parameters
+        self.break_eps = break_eps
+        self.break_metric = break_metric
+        self.break_classes = break_classes
+
+        # Attributes used to fetch the correct functions
+        self.torch, self.dtype, self.device = None, None, None
+
+    def __call__(self, clust_label, seg_label, seg_pred, ghost_pred=None):
+        """Adapts the cluster labels for one entry or a batch of entries.
+
+        Parameters
+        ----------
+        clust_label : Union[TensorBatch, np.ndarray, torch.Tensor]
+            (N, N_l) Cluster label tensor
+        seg_label : Union[TensorBatch, np.ndarray, torch.Tensor]
+            (M, 5) Segmentation label tensor
+        seg_pred : Union[TensorBatch, np.ndarray, torch.Tensor]
+            (M/N_deghost) Segmentation predictions for each voxel
+        ghost_pred : Union[TensorBatch, np.ndarray, torch.Tensor], optional
+            (M) Ghost predictions for each voxel
+
+        Returns
+        -------
+        Union[TensorBatch, np.ndarray, torch.Tensor]
+            (N_deghost, N_l) Adapted cluster label tensor
+        """
+        # Set the data type/device based on the input
+        ref_tensor = clust_label
+        if isinstance(ref_tensor, TensorBatch):
+            ref_tensor = ref_tensor.tensor
+        self.torch = isinstance(ref_tensor, torch.Tensor)
+
+        self.dtype = clust_label.dtype
+        if self.torch:
+            self.device = clust_label.device
+
+        # Dispatch depending on the data type
+        if isinstance(clust_label, TensorBatch):
+            # If it is batch data, call the main process function of each entry
+            shape = (seg_pred.shape[0], clust_label.shape[1])
+            clust_label_adapted = torch.empty(
+                    shape, dtype=clust_label.dtype, device=clust_label.device)
+            for b in range(clust_label.batch_size):
+                lower, upper = seg_pred.edges[b], seg_pred.edges[b+1]
+                ghost_pred_b = ghost_pred[b] if ghost_pred is not None else None
+                clust_label_adapted[lower:upper] = self._process(
+                        clust_label[b], seg_label[b], seg_pred[b], ghost_pred_b)
+
+            return TensorBatch(clust_label_adapted, seg_pred.counts)
+
+        else:
+            # Otherwise, call the main process function directly
+            return self._process(clust_label, seg_label, seg_pred, ghost_pred)
+
+    def _process(self, clust_label, seg_label, seg_pred, ghost_pred=None):
+        """Adapts the cluster labels for one entry or a batch of entries.
+
+        Parameters
+        ----------
+        clust_label : Union[np.ndarray, torch.Tensor]
+            (N, N_l) Cluster label tensor
+        seg_label : Union[np.ndarray, torch.Tensor]
+            (M, 5) Segmentation label tensor
+        seg_pred : Union[np.ndarray, torch.Tensor]
+            (M/N_deghost) Segmentation predictions for each voxel
+        ghost_pred : Union[np.ndarray, torch.Tensor], optional
+            (M) Ghost predictions for each voxel
+
+        Returns
+        -------
+        Union[np.ndarray, torch.Tensor]
+            (N_deghost, N_l) Adapted cluster label tensor
+        """
+        # If there are no points in this event, nothing to do
+        coords = seg_label[:, :VALUE_COL]
+        num_cols = clust_label.shape[1]
+        if not len(coords):
+            return self._ones((0, num_cols))
+
+        # If there are no points after deghosting, nothing to do
+        if ghost_pred is not None:
+            deghost_index = self._where(ghost_pred == 0)[0]
+            if not len(deghost_index):
+                return self._ones((0, num_cols))
+
+        # If there are no label points in this event, return dummy labels
+        if not len(clust_label):
+            if ghost_pred is None:
+                shape = (len(coords), num_cols)
+                dummy_labels = -self._ones(shape)
+                dummy_labels[:, :VALUE_COL] = coords
+
+            else:
+                shape = (len(deghost_index), num_cols)
+                dummy_labels = -self._ones(shape)
+                dummy_labels[:, :VALUE_COL] = coords[deghost_index]
+
+            return dummy_labels
+
+        # Build a tensor of predicted segmentation that includes ghost points
+        seg_label = self._to_long(seg_label[:, SHAPE_COL])
+        if ghost_pred is not None and (len(ghost_pred) != len(seg_pred)):
+            seg_pred_long = self._to_long(GHOST_SHP*self._ones(len(coords)))
+            seg_pred_long[deghost_index] = seg_pred
+            seg_pred = seg_pred_long
+
+        # Prepare new labels
+        new_label = -self._ones((len(coords), num_cols))
+        new_label[:, :VALUE_COL] = coords
+
+        # Check if the segment labels and predictions are compatible. If they are
+        # compatible, store the cluster labels as is. Track points do not mix
+        # with other classes, but EM classes are allowed to.
+        compat_mat = self._eye(GHOST_SHP + 1)
+        compat_mat[([SHOWR_SHP, SHOWR_SHP, MICHL_SHP, DELTA_SHP],
+                    [MICHL_SHP, DELTA_SHP, SHOWR_SHP, SHOWR_SHP])] = True
+
+        true_deghost = seg_label < GHOST_SHP
+        seg_mismatch = ~compat_mat[(seg_pred, seg_label)]
+        new_label[true_deghost] = clust_label
+        new_label[true_deghost & seg_mismatch, VALUE_COL:] = -self._ones(1)
+
+        # For mismatched predictions, attempt to find a touching instance of the
+        # same class to assign it sensible cluster labels.
+        for s in self._unique(seg_pred):
+            # Skip predicted ghosts (they keep their invalid labels)
+            if s == GHOST_SHP:
+                continue
+
+            # Restrict to points in this class that have incompatible segment
+            # labels. Track points do not mix, EM points are allowed to.
+            bad_index = self._where(
+                    (seg_pred == s) & (~true_deghost | seg_mismatch))[0]
+            if len(bad_index) == 0:
+                continue
+
+            # Find points in clust_label that have compatible segment labels
+            seg_clust_mask = compat_mat[s][self._to_long(clust_label[:, SHAPE_COL])]
+            X_true = clust_label[seg_clust_mask]
+            if len(X_true) == 0:
+                continue
+
+            # Loop over the set of unlabeled predicted points
+            X_pred = coords[bad_index]
+            tagged_voxels_count = 1
+            while tagged_voxels_count > 0 and len(X_pred) > 0:
+                # Find the nearest neighbor to each predicted point
+                closest_ids = self._compute_neighbor(X_pred, X_true)
+
+                # Compute Chebyshev distance between predicted and closest true.
+                distances = self._compute_distances(X_pred, X_true[closest_ids])
+
+                # Label unlabeled voxels that touch a compatible true voxel
+                select_mask = distances < 1.1
+                select_index = self._where(select_mask)[0]
+                tagged_voxels_count = len(select_index)
+                if tagged_voxels_count > 0:
+                    # Use the label of the touching true voxel
+                    additional_clust_label = self._cat(
+                            [X_pred[select_index],
+                             X_true[closest_ids[select_index], VALUE_COL:]], 1)
+                    new_label[bad_index[select_index]] = additional_clust_label
+
+                    # Update the mask to not include the new assigned points
+                    leftover_index = self._where(~select_mask)[0]
+                    bad_index = bad_index[leftover_index]
+
+                    # The new true available points are the ones we just added.
+                    # The new pred points are those not yet labeled
+                    X_true = additional_clust_label
+                    X_pred = X_pred[leftover_index]
+
+        # Remove predicted ghost points from the labels, set the shape
+        # column of the label to the segmentation predictions.
+        if ghost_pred is not None:
+            new_label = new_label[deghost_index]
+            new_label[:, SHAPE_COL] = seg_pred[deghost_index]
+        else:
+            new_label[:, SHAPE_COL] = seg_pred
+
+        # Build a list of cluster indexes to break
+        new_label_np = new_label
+        if torch.is_tensor(new_label):
+            new_label_np = new_label.detach().cpu().numpy()
+
+        clusts = []
+        labels = new_label_np[:, CLUST_COL]
+        shapes = new_label_np[:, SHAPE_COL]
+        for break_class in self.break_classes:
+            index_s = np.where(shapes == break_class)[0]
+            labels_s = labels[index_s]
+            for c in np.unique(labels_s):
+                # If the cluster ID is invalid, skip
+                if c < 0:
+                    continue
+
+                # Append cluster
+                clusts.append(index_s[labels_s == c])
+
+        # Now if an instance was broken up, assign it different cluster IDs
+        new_label[:, CLUST_COL] = break_clusters(
+                new_label, clusts, self.break_eps, self.break_metric)
+
+        return new_label
+
+    def _where(self, x):
+        if self.torch:
+            return torch.where(x)
+        else:
+            return np.where(x)
+
+    def _cat(self, x, axis):
+        if self.torch:
+            return torch.cat(x, axis)
+        else:
+            return np.concatenate(x, axis)
+
+    def _ones(self, x):
+        if self.torch:
+            return torch.ones(x, dtype=self.dtype, device=self.device)
+        else:
+            return np.ones(x)
+
+    def _eye(self, x):
+        if self.torch:
+            return torch.eye(x, dtype=torch.bool, device=self.device)
+        else:
+            return np.eye(x, dtype=bool)
+
+    def _unique(self, x):
+        if self.torch:
+            return torch.unique(x).long()
+        else:
+            return np.unique(x).astype(np.int64)
+
+    def _to_long(self, x):
+        if self.torch:
+            return x.long()
+        else:
+            return x.astype(int64)
+
+    def _compute_neighbor(self, x, y):
+        if self.torch:
+            return knn(y[:, COORD_COLS], x[:, COORD_COLS], 1)[1]
+        else:
+            return cdist(x[:, COORD_COLS], y[:, COORD_COLS]).argmin(axis=1)
+
+    def _compute_distances(self, x, y):
+        if self.torch:
+            return torch.amax(torch.abs(y[:, COORD_COLS] - x[:, COORD_COLS]), dim=1)
+        else:
+            return np.amax(np.abs(x[:, COORD_COLS] - y[:, COORD_COLS]), axis=1)