DeepLabCut · C-Achard · May 20, 2026 · May 18, 2026 · May 18, 2026 · May 19, 2026
diff --git a/deeplabcut/generate_training_dataset/trainingsetmanipulation.py b/deeplabcut/generate_training_dataset/trainingsetmanipulation.py
@@ -493,11 +493,16 @@ def parse_video_filenames(videos: list[str]) -> list[str]:
     return filenames
 
 
-def _drop_likelihood_columns(df: pd.DataFrame) -> pd.DataFrame:
+def drop_likelihood_columns(df: pd.DataFrame) -> pd.DataFrame:
     """Drop any columns whose coord level is named 'likelihood'.
 
     This sanitizes annotation DataFrames coming from h5/csv files before they are
     used for training dataset generation.
+
+    # NOTE @C-Achard 2026-05-18: This is used in several places as a guard
+        Most call sites using this should instead go through a canonical, validated project loading function
+        AND THEN do any custom local processing they require. The current design is hard to maintain and error prone,
+        and lacks a clearly documented, centralized project I/O interface.
     """
     if not isinstance(df.columns, pd.MultiIndex):
         return df
@@ -507,7 +512,7 @@ def _drop_likelihood_columns(df: pd.DataFrame) -> pd.DataFrame:
 
     likelihood_mask = coord_values == "likelihood"
     if likelihood_mask.any():
-        logging.info("Detected likelihood columns in annotation data; dropping them.")
+        logging.warning("Detected likelihood columns in annotation data; dropping them.", stacklevel=2)
         df = df.drop(columns=df.columns[likelihood_mask])
 
     return df
@@ -569,7 +574,7 @@ def merge_annotateddatasets(cfg, trainingsetfolder_full):
     AnnotationData = AnnotationData.reindex(bodyparts, axis=1, level=AnnotationData.columns.names.index("bodyparts"))
     # Filter out any stray likelihood columns that may have been concatenated in
     # see napari-deeplabcut #204 and DeepLabCut #3319
-    AnnotationData = _drop_likelihood_columns(AnnotationData)
+    AnnotationData = drop_likelihood_columns(AnnotationData)
 
     if AnnotationData.empty:
         logging.warning(
@@ -701,23 +706,24 @@ def mergeandsplit(config, trainindex=0, uniform=True):
     fn = os.path.join(project_path, trainingsetfolder, "CollectedData_" + cfg["scorer"])
 
     try:
-        Data = pd.read_hdf(fn + ".h5")
+        data = pd.read_hdf(fn + ".h5")
+        data = drop_likelihood_columns(data)
     except FileNotFoundError:
-        Data = merge_annotateddatasets(
+        data = merge_annotateddatasets(
             cfg,
             Path(os.path.join(project_path, trainingsetfolder)),
         )
-        if Data is None:
+        if data is None:
             return [], []
 
-    conversioncode.guarantee_multiindex_rows(Data)
-    Data = Data[scorer]  # extract labeled data
+    conversioncode.guarantee_multiindex_rows(data)
+    data = data[scorer]  # extract labeled data
 
     if uniform:
         TrainingFraction = cfg["TrainingFraction"]
         trainFraction = TrainingFraction[trainindex]
         trainIndices, testIndices = SplitTrials(
-            range(len(Data.index)),
+            range(len(data.index)),
             trainFraction,
             True,
         )
@@ -726,7 +732,7 @@ def mergeandsplit(config, trainindex=0, uniform=True):
         test_video_name = [Path(i).stem for i in videos][trainindex]
         print("Excluding the following folder (from training):", test_video_name)
         trainIndices, testIndices = [], []
-        for index, name in enumerate(Data.index):
+        for index, name in enumerate(data.index):
             if test_video_name == name[1]:  # this is the video name
                 # print(name,test_video_name)
                 testIndices.append(index)
@@ -754,7 +760,7 @@ def to_matlab_cell(array):
         return outer
 
     # Again, remove likelihood if present
-    df = _drop_likelihood_columns(df)
+    df = drop_likelihood_columns(df)
 
     if isinstance(df.columns, pd.MultiIndex):
         coord_level = "coords" if "coords" in df.columns.names else df.columns.names[-1]

diff --git a/deeplabcut/modelzoo/generalized_data_converter/datasets/ma_dlc.py b/deeplabcut/modelzoo/generalized_data_converter/datasets/ma_dlc.py
@@ -13,6 +13,7 @@
 import numpy as np
 import pandas as pd
 
+from deeplabcut.generate_training_dataset.trainingsetmanipulation import drop_likelihood_columns
 from deeplabcut.modelzoo.generalized_data_converter.datasets.base_dlc import (
     BaseDLCPoseDataset,
 )
@@ -27,7 +28,7 @@ def __init__(self, proj_root, dataset_name, shuffle=1, modelprefix=""):
         super().__init__(proj_root, dataset_name, shuffle=shuffle, modelprefix=modelprefix)
 
     def _df2generic(self, df, image_id_offset=0):
-
+        df = drop_likelihood_columns(df)
         individuals = df.columns.get_level_values("individuals").unique().tolist()
 
         unique_bpts = []

diff --git a/deeplabcut/modelzoo/generalized_data_converter/datasets/ma_dlc_dataframe.py b/deeplabcut/modelzoo/generalized_data_converter/datasets/ma_dlc_dataframe.py
@@ -15,6 +15,7 @@
 import pandas as pd
 
 from deeplabcut.generate_training_dataset.trainingsetmanipulation import (
+    drop_likelihood_columns,
     parse_video_filenames,
 )
 from deeplabcut.modelzoo.generalized_data_converter.datasets.base import BasePoseDataset
@@ -79,6 +80,7 @@ def merge_annotateddatasets(cfg):
     else:
         bodyparts = cfg["bodyparts"]
     AnnotationData = AnnotationData.reindex(bodyparts, axis=1, level=AnnotationData.columns.names.index("bodyparts"))
+    AnnotationData = drop_likelihood_columns(AnnotationData)
 
     return AnnotationData
 
@@ -140,7 +142,7 @@ def populate_generic(self):
         self.whether_anno_image_match(self.generic_test_images, self.generic_test_annotations)
 
     def _df2generic(self, df, image_id_offset=0):
-
+        df = drop_likelihood_columns(df)
         individuals = df.columns.get_level_values("individuals").unique().tolist()
 
         unique_bpts = []

diff --git a/deeplabcut/pose_estimation_pytorch/data/dlcloader.py b/deeplabcut/pose_estimation_pytorch/data/dlcloader.py
@@ -23,6 +23,7 @@
 
 import deeplabcut.utils.auxiliaryfunctions as af
 from deeplabcut.core.engine import Engine
+from deeplabcut.generate_training_dataset.trainingsetmanipulation import drop_likelihood_columns
 from deeplabcut.pose_estimation_pytorch.data.base import Loader
 from deeplabcut.pose_estimation_pytorch.data.dataset import PoseDatasetParameters
 from deeplabcut.pose_estimation_pytorch.data.snapshots import Snapshot
@@ -373,6 +374,8 @@ def to_coco(
         Returns:
             the coco format data
         """
+        df = drop_likelihood_columns(df)
+
         with_individuals = "individuals" in df.columns.names
         if not with_individuals and (len(parameters.individuals) > 1 or len(parameters.unique_bpts) > 0):
             raise ValueError(

diff --git a/deeplabcut/utils/skeleton.py b/deeplabcut/utils/skeleton.py
@@ -31,6 +31,8 @@
 from scipy.spatial import KDTree
 from skimage import io
 
+from deeplabcut.generate_training_dataset.trainingsetmanipulation import drop_likelihood_columns
+
 
 # NOTE @C-Achard 2026-03-26 duplicate config read/write functions
 # should be addressed in config refactor
@@ -60,6 +62,7 @@ def __init__(self, config_path):
             folder = os.path.join(root, dir_)
             if os.path.isdir(folder) and not any(folder.endswith(s) for s in ("cropped", "labeled")):
                 self.df = pd.read_hdf(os.path.join(folder, f"CollectedData_{self.cfg['scorer']}.h5"))
+                self.df = drop_likelihood_columns(self.df)
                 row, col = self.pick_labeled_frame()
                 if "individuals" in self.df.columns.names:
                     self.df = self.df.xs(col, axis=1, level="individuals")

diff --git a/tests/pose_estimation_pytorch/data/test_dlc_dataloader.py b/tests/pose_estimation_pytorch/data/test_dlc_dataloader.py
@@ -0,0 +1,68 @@
+from types import SimpleNamespace
+
+import numpy as np
+import pandas as pd
+
+import deeplabcut.pose_estimation_pytorch.data.dlcloader as dlcloader_mod
+from deeplabcut.pose_estimation_pytorch.data.dlcloader import DLCLoader
+
+
+def test_to_coco_ignores_likelihood_columns(monkeypatch, tmp_path):
+    fake_shape = (3, 480, 640)
+    monkeypatch.setattr(
+        dlcloader_mod,
+        "read_image_shape_fast",
+        lambda _: fake_shape,
+    )
+
+    scorer = "testscorer"
+    bodyparts = ["nose", "tail"]
+
+    index = pd.MultiIndex.from_tuples(
+        [("labeled-data", "video1", "img0001.png")],
+        names=["set", "video", "image"],
+    )
+
+    # Baseline dataframe: x/y only
+    columns_xy = pd.MultiIndex.from_product(
+        [[scorer], bodyparts, ["x", "y"]],
+        names=["scorer", "bodyparts", "coords"],
+    )
+    df_xy = pd.DataFrame(
+        [[10.0, 20.0, 30.0, 40.0]],
+        index=index,
+        columns=columns_xy,
+    )
+
+    # Same data, but with likelihood columns added
+    columns_xyl = pd.MultiIndex.from_product(
+        [[scorer], bodyparts, ["x", "y", "likelihood"]],
+        names=["scorer", "bodyparts", "coords"],
+    )
+    df_xyl = pd.DataFrame(
+        [[10.0, 20.0, 0.9, 30.0, 40.0, 0.8]],
+        index=index,
+        columns=columns_xyl,
+    )
+
+    # to_coco only needs these attributes from parameters
+    params = SimpleNamespace(
+        bodyparts=bodyparts,
+        unique_bpts=[],
+        individuals=["animal"],
+    )
+
+    baseline = DLCLoader.to_coco(tmp_path, df_xy, params)
+    got = DLCLoader.to_coco(tmp_path, df_xyl, params)
+
+    assert len(got["images"]) == len(baseline["images"]) == 1
+    assert len(got["annotations"]) == len(baseline["annotations"]) == 1
+
+    got_ann = got["annotations"][0]
+    expected_ann = baseline["annotations"][0]
+
+    assert got_ann["image_id"] == expected_ann["image_id"]
+    assert got_ann["category_id"] == expected_ann["category_id"]
+    assert got_ann["num_keypoints"] == expected_ann["num_keypoints"] == 2
+    assert np.array_equal(got_ann["keypoints"], expected_ann["keypoints"])
+    assert np.allclose(got_ann["bbox"], expected_ann["bbox"])