byotrack 1.2.0.dev0__tar.gz → 1.2.0.dev1__tar.gz

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: byotrack
-Version: 1.2.0.dev0
+Version: 1.2.0.dev1
 Summary: Biological particle tracking with Python
 Home-page: https://github.com/raphaelreme/byotrack
 Author: Raphael Reme
@@ -28,7 +28,7 @@ Requires-Dist: platformdirs
 Requires-Dist: pylapy[scipy]
 Requires-Dist: tifffile[all]
 Requires-Dist: torch
-Requires-Dist: torch_tps
+Requires-Dist: torch-tps
 Requires-Dist: tqdm
 Provides-Extra: full
 Requires-Dist: matplotlib; extra == "full"

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/__init__.py

@@ -83,4 +83,4 @@ from byotrack.api.tracks import Track
 from byotrack.video import Video, VideoTransformConfig
 
 
-__version__ = "1.2.0.dev0"
+__version__ = "1.2.0.dev1"

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/api/detector/detections.py

@@ -38,7 +38,7 @@ def _check_confidence(confidence: torch.Tensor) -> None:
 
 @numba.njit(parallel=False)
 def _position_from_segmentation(segmentation: np.ndarray) -> np.ndarray:
-    """Return the centre of each instance in the segmentation"""
+    """Return the center (mean) of each instance in the segmentation"""
     # A bit slower than previous version in 2D, but still fine
 
     n = segmentation.max()
@@ -57,6 +57,42 @@ def _position_from_segmentation(segmentation: np.ndarray) -> np.ndarray:
     return m_1.astype(np.float32) / m_0.reshape(-1, 1)
 
 
+@numba.njit(parallel=False)
+def _median_from_segmentation(segmentation: np.ndarray) -> np.ndarray:
+    """Return the center (median) of each instance in the segmentation"""
+
+    # Flatten space axes
+    flat_segmentation = segmentation.reshape(-1)
+
+    n = segmentation.max()
+    counts = np.zeros(n, dtype=np.uint)
+
+    for i in range(flat_segmentation.shape[0]):
+        instance = flat_segmentation[i] - 1
+        if instance != -1:
+            counts[instance] += 1
+
+    m = np.max(counts)
+
+    # Reset counts and allocate position
+    counts[:] = 0
+    positions = np.empty((n, m, len(segmentation.shape)), dtype=np.uint)
+
+    for index in np.ndindex(*segmentation.shape):
+        instance = segmentation[index] - 1
+        if instance != -1:
+            positions[instance, counts[instance]] = index
+            counts[instance] += 1
+
+    # Compute medians
+    median = np.zeros((n, len(segmentation.shape)), dtype=np.float32)
+    for instance in range(n):
+        for axis in range(len(segmentation.shape)):
+            median[instance, axis] = np.median(positions[instance, : counts[instance], axis])
+
+    return median
+
+
 @numba.njit
 def _bbox_from_segmentation(segmentation: np.ndarray) -> np.ndarray:
     # A bit slower than previous version in 2D, but still fine
@@ -242,12 +278,15 @@ class Detections:
             Shape: ([D, ]H, W), dtype: int32
         confidence (torch.Tensor): Confidence for each instance
             Shape: (N,), dtype: float32
+        use_median_position (bool): Use median instead of mean to compute positions from segmentation.
+            Default: True (Usually more robust)
 
     """
 
-    def __init__(self, data: Dict[str, torch.Tensor], frame_id: int = -1) -> None:
+    def __init__(self, data: Dict[str, torch.Tensor], frame_id: int = -1, use_median_position=True) -> None:
         self.length = -1
         self.dim = -1
+        self._use_median_position = use_median_position
 
         if "position" in data:
             _check_position(data["position"])
@@ -321,6 +360,20 @@ class Detections:
 
         return confidence
 
+    @property
+    def use_median_position(self) -> bool:
+        return self._use_median_position
+
+    @use_median_position.setter
+    def use_median_position(self, value: bool) -> None:
+        if value is self._use_median_position:
+            return
+
+        self._use_median_position = value
+
+        # Invalidate computed positions
+        self._lazy_extrapolated_data.pop("position")
+
     def _extrapolate_shape(self) -> Tuple[int, ...]:
         """Extrapolate shape from data
 
@@ -354,6 +407,8 @@ class Detections:
 
         """
         if "segmentation" in self.data:
+            if self.use_median_position:
+                return torch.tensor(_median_from_segmentation(self.data["segmentation"].numpy()))
             return torch.tensor(_position_from_segmentation(self.data["segmentation"].numpy()))
 
         return self.data["bbox"][:, : self.dim] + (self.data["bbox"][:, self.dim :] - 1) / 2

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/dataset/ctc.py

@@ -40,6 +40,8 @@ class GroundTruthDetector(byotrack.BatchDetector):
 
     """
 
+    progress_bar_description = "Detections (Load from CTC format)"
+
     def detect(self, batch: np.ndarray) -> List[byotrack.Detections]:
         assert batch.shape[-1] == 1, "Multichannel segmentation are not supported"
         assert np.issubdtype(batch.dtype, np.integer)
@@ -131,7 +133,7 @@ def load_tracks(  # pylint: disable=too-many-locals,too-many-branches,too-many-s
     segmentation_paths = path.glob("mask*.tif") if is_res else path.glob("man_*.tif")
     loader = byotrack.video.reader.FrameTiffLoader()
 
-    for path_ in tqdm.tqdm(byotrack.utils.sorted_alphanumeric(segmentation_paths)):
+    for path_ in tqdm.tqdm(byotrack.utils.sorted_alphanumeric(segmentation_paths), desc="Loading CTC tracks"):
         if is_res:
             frame_id = int(path_.stem[len("mask") :])
         elif "seg" in path.stem:
@@ -225,7 +227,7 @@ def save_detections(
 
     os.makedirs(path, exist_ok=True)
 
-    for frame_id, detections in enumerate(tqdm.tqdm(detections_sequence)):
+    for frame_id, detections in enumerate(tqdm.tqdm(detections_sequence, desc="Saving Detections to CTC")):
         segmentation = detections.segmentation.numpy().astype(np.uint16)
 
         if as_res:
@@ -472,7 +474,7 @@ def save_tracks(  # pylint: disable=too-many-branches,too-many-locals,too-many-s
     else:
         assert shape is not None, "Without detections_sequence, you need to provide the shape argument"
 
-    for frame_id in tqdm.trange(last + 1):
+    for frame_id in tqdm.trange(last + 1, desc="Saving tracks to CTC"):
         has_detections = len(detections_sequence) > frame_id
 
         disk_positions = []

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/implementation/linker/frame_by_frame/base.py

@@ -221,6 +221,10 @@ class FrameByFrameLinkerParameters:  # pylint: disable=too-many-instance-attribu
         association_method (AssociationMethod): The frame-by-frame association to use. See `AssociationMethod`.
             It can be provided as a string. (Choice: GREEDY, OPT_HARD, OPT_SMOOTH)
             Default: OPT_SMOOTH
+        anisotropy (Tuple[float, float, float]): Anisotropy of images (Ratio of the pixel sizes
+            for each axis, depth first). This will be used to scale distances.
+            Default: (1., 1., 1.)
+
     """
 
     def __init__(
@@ -230,6 +234,7 @@ class FrameByFrameLinkerParameters:  # pylint: disable=too-many-instance-attribu
         n_valid=3,
         n_gap=3,
         association_method: Union[str, AssociationMethod] = AssociationMethod.OPT_SMOOTH,
+        anisotropy: Tuple[float, float, float] = (1.0, 1.0, 1.0),
     ):
         self.association_threshold = association_threshold
         self.n_valid = n_valid
@@ -239,11 +244,13 @@ class FrameByFrameLinkerParameters:  # pylint: disable=too-many-instance-attribu
             if isinstance(association_method, AssociationMethod)
             else AssociationMethod[association_method.upper()]
         )
+        self.anisotropy = anisotropy
 
     association_threshold: float = 5.0
     n_valid: int = 3
     n_gap: int = 3
     association_method: AssociationMethod = AssociationMethod.OPT_SMOOTH
+    anisotropy: Tuple[float, float, float] = (1.0, 1.0, 1.0)
 
 
 class FrameByFrameLinker(byotrack.OnlineLinker):

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/implementation/linker/frame_by_frame/kalman_linker.py

@@ -1,6 +1,7 @@
 import dataclasses
 import enum
 from typing import List, Optional, Tuple, Union
+import warnings
 
 import numpy as np
 import torch
@@ -86,6 +87,11 @@ class KalmanLinkerParameters(FrameByFrameLinkerParameters):
         association_method (AssociationMethod): The frame-by-frame association to use. See `AssociationMethod`.
             It can be provided as a string. (Choice: GREEDY, OPT_HARD, OPT_SMOOTH)
             Default: OPT_SMOOTH
+        anisotropy (Tuple[float, float, float]): Anisotropy of images (Ratio of the pixel sizes
+            for each axis, depth first). This will be used to scale distances. It will only impact
+            EUCLIDEAN[_SQ] costs. For probabilistic cost, anisotropy should be already integrated
+            in the stds of the kalman filter (providing one std for each dimension).
+            Default: (1., 1., 1.)
         cost_method (CostMethod): The cost method to use. It can be provided as a string.
             See `CostMethod`. It also indicates what is the correct unit of `association_threshold`.
             Default: EUCLIDEAN
@@ -96,7 +102,7 @@ class KalmanLinkerParameters(FrameByFrameLinkerParameters):
 
     """
 
-    def __init__(
+    def __init__(  # pylint: disable=too-many-arguments
         self,
         association_threshold: float = 5.0,
         *,
@@ -106,6 +112,7 @@ class KalmanLinkerParameters(FrameByFrameLinkerParameters):
         n_valid=3,
         n_gap=3,
         association_method: Union[str, AssociationMethod] = AssociationMethod.OPT_SMOOTH,
+        anisotropy: Tuple[float, float, float] = (1.0, 1.0, 1.0),
         cost: Union[str, Cost] = Cost.EUCLIDEAN,
         track_building: Union[str, TrackBuilding] = TrackBuilding.FILTERED,
     ):
@@ -113,9 +120,19 @@ class KalmanLinkerParameters(FrameByFrameLinkerParameters):
             association_threshold=association_threshold,
             n_valid=n_valid,
             n_gap=n_gap,
+            anisotropy=anisotropy,
             association_method=association_method,
         )
 
+        if isinstance(detection_std, float) and min(anisotropy) != max(anisotropy):
+            warnings.warn(
+                "A single `detection_std` is provided, but images are anisotrope. Consider giving one std by dimension."
+            )
+        if isinstance(process_std, float) and min(anisotropy) != max(anisotropy):
+            warnings.warn(
+                "A single `process_std` is provided, but images are anisotrope. Consider giving one std by dimension."
+            )
+
         self.detection_std = detection_std
         self.process_std = process_std
         self.kalman_order = kalman_order
@@ -282,12 +299,17 @@ class KalmanLinker(FrameByFrameLinker):
         if self.projections is None:
             raise RuntimeError("Projections should already be initialized.")
 
+        anisotropy = torch.tensor(self.specs.anisotropy)[: detections.dim]
+
         if self.specs.cost == Cost.EUCLIDEAN:
-            return torch.cdist(self.projections.mean[..., 0], detections.position), self.specs.association_threshold
+            return (
+                torch.cdist(self.projections.mean[..., 0] * anisotropy, detections.position * anisotropy),
+                self.specs.association_threshold,
+            )
 
         if self.specs.cost == Cost.EUCLIDEAN_SQ:
             return (
-                torch.cdist(self.projections.mean[..., 0], detections.position).pow_(2),
+                torch.cdist(self.projections.mean[..., 0] * anisotropy, detections.position * anisotropy).pow_(2),
                 self.specs.association_threshold**2,
             )
         if self.specs.cost == Cost.MAHALANOBIS:

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/implementation/linker/frame_by_frame/koft.py

@@ -2,6 +2,7 @@
 
 import dataclasses
 from typing import Optional, Tuple, Union
+import warnings
 
 import numpy as np
 import torch
@@ -44,6 +45,11 @@ class KOFTLinkerParameters(KalmanLinkerParameters):
         association_method (AssociationMethod): The frame-by-frame association to use. See `AssociationMethod`.
             It can be provided as a string. (Choice: GREEDY, OPT_HARD, OPT_SMOOTH)
             Default: OPT_SMOOTH
+        anisotropy (Tuple[float, float, float]): Anisotropy of images (Ratio of the pixel sizes
+            for each axis, depth first). This will be used to scale distances. It will only impact
+            EUCLIDEAN[_SQ] costs. For probabilistic cost, anisotropy should be already integrated
+            in the different std of the kalman filter.
+            Default: (1., 1., 1.)
         cost_method (CostMethod): The cost method to use. It can be provided as a string.
             See `CostMethod`. It also indicates what is the correct unit of `association_threshold`.
             Default: EUCLIDEAN
@@ -71,6 +77,7 @@ class KOFTLinkerParameters(KalmanLinkerParameters):
         n_valid=3,
         n_gap=3,
         association_method: Union[str, AssociationMethod] = AssociationMethod.OPT_SMOOTH,
+        anisotropy: Tuple[float, float, float] = (1.0, 1.0, 1.0),
         cost: Union[str, Cost] = Cost.EUCLIDEAN,
         track_building: Union[str, TrackBuilding] = TrackBuilding.FILTERED,
         extract_flows_on_detections=False,
@@ -84,10 +91,16 @@ class KOFTLinkerParameters(KalmanLinkerParameters):
             n_valid=n_valid,
             n_gap=n_gap,
             association_method=association_method,
+            anisotropy=anisotropy,
             cost=cost,
             track_building=track_building,
         )
 
+        if isinstance(flow_std, float) and min(anisotropy) != max(anisotropy):
+            warnings.warn(
+                "A single flow_std is provided, but the images are anisotrope. Consider giving one std by dimension."
+            )
+
         self.flow_std = flow_std
         self.extract_flows_on_detections = extract_flows_on_detections
         self.always_measure_velocity = always_measure_velocity
@@ -226,15 +239,21 @@ class KOFTLinker(KalmanLinker):
         if self.projections is None:
             raise RuntimeError("Projections should already be initialized.")
 
+        anisotropy = torch.tensor(self.specs.anisotropy)[: detections.dim]
+
         if self.specs.cost == Cost.EUCLIDEAN:
             return (
-                torch.cdist(self.projections.mean[:, : detections.dim, 0], detections.position),
+                torch.cdist(
+                    self.projections.mean[:, : detections.dim, 0] * anisotropy, detections.position * anisotropy
+                ),
                 self.specs.association_threshold,
             )
 
         if self.specs.cost == Cost.EUCLIDEAN_SQ:
             return (
-                torch.cdist(self.projections.mean[:, : detections.dim, 0], detections.position).pow_(2),
+                torch.cdist(
+                    self.projections.mean[:, : detections.dim, 0] * anisotropy, detections.position * anisotropy
+                ).pow_(2),
                 self.specs.association_threshold**2,
             )
 

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/implementation/linker/frame_by_frame/nearest_neighbor.py

@@ -45,6 +45,7 @@ class NearestNeighborParameters(FrameByFrameLinkerParameters):
         n_valid=3,
         n_gap=3,
         association_method: Union[str, AssociationMethod] = AssociationMethod.OPT_SMOOTH,
+        anisotropy: Tuple[float, float, float] = (1.0, 1.0, 1.0),
         ema=0.0,
         fill_gap=False,
     ):
@@ -53,6 +54,7 @@ class NearestNeighborParameters(FrameByFrameLinkerParameters):
             n_valid=n_valid,
             n_gap=n_gap,
             association_method=association_method,
+            anisotropy=anisotropy,
         )
         self.ema = ema
         self.fill_gap = fill_gap
@@ -110,7 +112,12 @@ class NearestNeighborLinker(FrameByFrameLinker):
         if self.active_positions is None:
             self.active_positions = torch.empty((0, detections.position.shape[1]))
 
-        return torch.cdist(self.active_positions, detections.position), self.specs.association_threshold
+        anisotropy = torch.tensor(self.specs.anisotropy)[: -detections.dim]
+
+        return (
+            torch.cdist(self.active_positions * anisotropy, detections.position * anisotropy),
+            self.specs.association_threshold,
+        )
 
     def post_association(self, _: np.ndarray, detections: byotrack.Detections, links: torch.Tensor):
         if self.active_positions is None:

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack/metrics/ctc.py

@@ -125,8 +125,14 @@ class CTCMetrics(CTCSoftwareRunner):
     and may add points to tracks to fill missing detections. Usually, this leads to better DET/SEG metrics
     as it reduces false positive and false negative.
 
+    It will store the logs of the last called metric in `self.last_log`
+
     """
 
+    def __init__(self, ctc_software):
+        super().__init__(ctc_software)
+        self.last_log = ""
+
     def compute_tracking_metric(
         self,
         metric: str,
@@ -210,7 +216,10 @@ class CTCMetrics(CTCSoftwareRunner):
                     **kwargs,
                 )
 
-            return self.run(metric, output_path, 1)
+            results = self.run(metric, output_path, 1)
+            self.last_log = (output_path / "01_RES" / f"{metric}_log.txt").read_text("utf-8")
+
+            return results
 
     def compute_detection_metric(
         self,
@@ -259,7 +268,10 @@ class CTCMetrics(CTCSoftwareRunner):
                     ground_truth_path, ground_truth_detections_sequence, as_res=False, as_seg=metric == "SEG"
                 )
 
-            return self.run(metric, output_path, 1)
+            results = self.run(metric, output_path, 1)
+            self.last_log = (output_path / "01_RES" / f"{metric}_log.txt").read_text("utf-8")
+
+            return results
 
     @staticmethod
     def copy_ground_truth(

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: byotrack
-Version: 1.2.0.dev0
+Version: 1.2.0.dev1
 Summary: Biological particle tracking with Python
 Home-page: https://github.com/raphaelreme/byotrack
 Author: Raphael Reme
@@ -28,7 +28,7 @@ Requires-Dist: platformdirs
 Requires-Dist: pylapy[scipy]
 Requires-Dist: tifffile[all]
 Requires-Dist: torch
-Requires-Dist: torch_tps
+Requires-Dist: torch-tps
 Requires-Dist: tqdm
 Provides-Extra: full
 Requires-Dist: matplotlib; extra == "full"

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/byotrack.egg-info/requires.txt

@@ -6,7 +6,7 @@ platformdirs
 pylapy[scipy]
 tifffile[all]
 torch
-torch_tps
+torch-tps
 tqdm
 
 [full]

{byotrack-1.2.0.dev0 → byotrack-1.2.0.dev1}/setup.cfg

@@ -34,7 +34,7 @@ install_requires =
 	pylapy[scipy]
 	tifffile[all]
 	torch
-	torch_tps
+	torch-tps
 	tqdm
 include_package_data = False