[Enhance] Remove useless param in 3D Seg Dataset classes (#607)

* remove useless label_weight args in dataset class

* modify unit tests

* update compatibility docs

configs/_base_/datasets/s3dis_seg-3d-13class.py

@@ -115,10 +115,6 @@ data = dict(
         scene_idxs=[
             data_root + f'seg_info/Area_{i}_resampled_scene_idxs.npy'
             for i in train_area
-        ],
-        label_weights=[
-            data_root + f'seg_info/Area_{i}_label_weight.npy'
-            for i in train_area
         ]),
     val=dict(
         type=dataset_type,

configs/_base_/datasets/scannet_seg-3d-20class.py

@@ -110,8 +110,7 @@ data = dict(
         classes=class_names,
         test_mode=False,
         ignore_index=len(class_names),
-        scene_idxs=data_root + 'seg_info/train_resampled_scene_idxs.npy',
-        label_weight=data_root + 'seg_info/train_label_weight.npy'),
+        scene_idxs=data_root + 'seg_info/train_resampled_scene_idxs.npy'),
     val=dict(
         type=dataset_type,
         data_root=data_root,

docs/compatibility.md

@@ -2,6 +2,12 @@
 
 This document provides detailed descriptions of the BC-breaking changes in MMDetection3D.
 
+## MMDetection3D 0.15.0
+
+### Dataset class for 3D segmentation task
+
+We remove a useless parameter `label_weight` from segmentation datasets including `Custom3DSegDataset`, `ScanNetSegDataset` and `S3DISSegDataset` since this weight is utilized in the loss function of model class. Please modify the code as well as the config files accordingly if you use or inherit from these codes.
+
 ## MMDetection3D 0.14.0
 
 ### ScanNet data pre-processing

mmdet3d/datasets/custom_3d_seg.py

@@ -38,9 +38,6 @@ class Custom3DSegDataset(Dataset):
         scene_idxs (np.ndarray | str, optional): Precomputed index to load
             data. For scenes with many points, we may sample it several times.
             Defaults to None.
-        label_weight (np.ndarray | str, optional): Precomputed weight to \
-            balance loss calculation. If None is given, use equal weighting.
-            Defaults to None.
     """
     # names of all classes data used for the task
     CLASSES = None
@@ -63,8 +60,7 @@ class Custom3DSegDataset(Dataset):
                  modality=None,
                  test_mode=False,
                  ignore_index=None,
-                 scene_idxs=None,
-                 label_weight=None):
+                 scene_idxs=None):
         super().__init__()
         self.data_root = data_root
         self.ann_file = ann_file
@@ -79,8 +75,7 @@ class Custom3DSegDataset(Dataset):
         self.ignore_index = len(self.CLASSES) if \
             ignore_index is None else ignore_index
 
-        self.scene_idxs, self.label_weight = \
-            self.get_scene_idxs_and_label_weight(scene_idxs, label_weight)
+        self.scene_idxs = self.get_scene_idxs(scene_idxs)
         self.CLASSES, self.PALETTE = \
             self.get_classes_and_palette(classes, palette)
 
@@ -250,26 +245,16 @@ class Custom3DSegDataset(Dataset):
             for cls_name in class_names
         ]
 
-        # also need to modify self.label_weight
-        self.label_weight = np.array([
-            self.label_weight[self.CLASSES.index(cls_name)]
-            for cls_name in class_names
-        ]).astype(np.float32)
-
         return class_names, palette
 
-    def get_scene_idxs_and_label_weight(self, scene_idxs, label_weight):
-        """Compute scene_idxs for data sampling and label weight for loss \
-        calculation.
+    def get_scene_idxs(self, scene_idxs):
+        """Compute scene_idxs for data sampling.
 
-        We sample more times for scenes with more points. Label_weight is
-        inversely proportional to number of class points.
+        We sample more times for scenes with more points.
         """
         if self.test_mode:
             # when testing, we load one whole scene every time
-            # and we don't need label weight for loss calculation
-            return np.arange(len(self.data_infos)).astype(np.int32), \
-                np.ones(len(self.CLASSES)).astype(np.float32)
+            return np.arange(len(self.data_infos)).astype(np.int32)
 
         # we may need to re-sample different scenes according to scene_idxs
         # this is necessary for indoor scene segmentation such as ScanNet
@@ -280,15 +265,7 @@ class Custom3DSegDataset(Dataset):
         else:
             scene_idxs = np.array(scene_idxs)
 
-        if label_weight is None:
-            # we don't used label weighting in training
-            label_weight = np.ones(len(self.CLASSES))
-        elif isinstance(label_weight, str):
-            label_weight = np.load(label_weight)
-        else:
-            label_weight = np.array(label_weight)
-
-        return scene_idxs.astype(np.int32), label_weight.astype(np.float32)
+        return scene_idxs.astype(np.int32)
 
     def format_results(self,
                        outputs,

mmdet3d/datasets/s3dis_dataset.py

@@ -41,9 +41,6 @@ class _S3DISSegDataset(Custom3DSegDataset):
         scene_idxs (np.ndarray | str, optional): Precomputed index to load
             data. For scenes with many points, we may sample it several times.
             Defaults to None.
-        label_weight (np.ndarray | str, optional): Precomputed weight to \
-            balance loss calculation. If None is given, compute from data.
-            Defaults to None.
     """
     CLASSES = ('ceiling', 'floor', 'wall', 'beam', 'column', 'window', 'door',
                'table', 'chair', 'sofa', 'bookcase', 'board', 'clutter')
@@ -66,8 +63,7 @@ class _S3DISSegDataset(Custom3DSegDataset):
                  modality=None,
                  test_mode=False,
                  ignore_index=None,
-                 scene_idxs=None,
-                 label_weight=None):
+                 scene_idxs=None):
 
         super().__init__(
             data_root=data_root,
@@ -78,8 +74,7 @@ class _S3DISSegDataset(Custom3DSegDataset):
             modality=modality,
             test_mode=test_mode,
             ignore_index=ignore_index,
-            scene_idxs=scene_idxs,
-            label_weight=label_weight)
+            scene_idxs=scene_idxs)
 
     def get_ann_info(self, index):
         """Get annotation info according to the given index.
@@ -153,21 +148,17 @@ class _S3DISSegDataset(Custom3DSegDataset):
                             pred_sem_mask, out_dir, file_name,
                             np.array(self.PALETTE), self.ignore_index, show)
 
-    def get_scene_idxs_and_label_weight(self, scene_idxs, label_weight):
-        """Compute scene_idxs for data sampling and label weight for loss \
-        calculation.
+    def get_scene_idxs(self, scene_idxs):
+        """Compute scene_idxs for data sampling.
 
-        We sample more times for scenes with more points. Label_weight is
-        inversely proportional to number of class points.
+        We sample more times for scenes with more points.
         """
         # when testing, we load one whole scene every time
-        # and we don't need label weight for loss calculation
         if not self.test_mode and scene_idxs is None:
             raise NotImplementedError(
                 'please provide re-sampled scene indexes for training')
 
-        return super().get_scene_idxs_and_label_weight(scene_idxs,
-                                                       label_weight)
+        return super().get_scene_idxs(scene_idxs)
 
 
 @DATASETS.register_module()
@@ -178,7 +169,7 @@ class S3DISSegDataset(_S3DISSegDataset):
     This class serves as the API for experiments on the S3DIS Dataset.
     It wraps the provided datasets of different areas.
     We don't use `mmdet.datasets.dataset_wrappers.ConcatDataset` because we
-    need to concat the `scene_idxs` and `label_weights` of different areas.
+    need to concat the `scene_idxs` of different areas.
 
     Please refer to the `google form <https://docs.google.com/forms/d/e/1FAIpQL
     ScDimvNMCGhy_rmBA2gHfDu3naktRm6A8BPwAWWDv-Uhm6Shw/viewform?c=0&w=1>`_ for
@@ -203,9 +194,6 @@ class S3DISSegDataset(_S3DISSegDataset):
         scene_idxs (list[np.ndarray] | list[str], optional): Precomputed index
             to load data. For scenes with many points, we may sample it several
             times. Defaults to None.
-        label_weights (list[np.ndarray] | list[str], optional): Precomputed
-            weight to balance loss calculation. If None is given, compute from
-            data. Defaults to None.
     """
 
     def __init__(self,
@@ -217,14 +205,11 @@ class S3DISSegDataset(_S3DISSegDataset):
                  modality=None,
                  test_mode=False,
                  ignore_index=None,
-                 scene_idxs=None,
-                 label_weights=None):
+                 scene_idxs=None):
 
-        # make sure that ann_files, scene_idxs and label_weights have same len
+        # make sure that ann_files and scene_idxs have same length
         ann_files = self._check_ann_files(ann_files)
         scene_idxs = self._check_scene_idxs(scene_idxs, len(ann_files))
-        label_weights = self._check_label_weights(label_weights,
-                                                  len(ann_files))
 
         # initialize some attributes as datasets[0]
         super().__init__(
@@ -236,8 +221,7 @@ class S3DISSegDataset(_S3DISSegDataset):
             modality=modality,
             test_mode=test_mode,
             ignore_index=ignore_index,
-            scene_idxs=scene_idxs[0],
-            label_weight=label_weights[0])
+            scene_idxs=scene_idxs[0])
 
         datasets = [
             _S3DISSegDataset(
@@ -249,14 +233,12 @@ class S3DISSegDataset(_S3DISSegDataset):
                 modality=modality,
                 test_mode=test_mode,
                 ignore_index=ignore_index,
-                scene_idxs=scene_idxs[i],
-                label_weight=label_weights[i]) for i in range(len(ann_files))
+                scene_idxs=scene_idxs[i]) for i in range(len(ann_files))
         ]
 
-        # data_infos, scene_idxs, label_weight need to be concat
+        # data_infos and scene_idxs need to be concat
         self.concat_data_infos([dst.data_infos for dst in datasets])
         self.concat_scene_idxs([dst.scene_idxs for dst in datasets])
-        self.concat_label_weight([dst.label_weight for dst in datasets])
 
         # set group flag for the sampler
         if not self.test_mode:
@@ -287,15 +269,6 @@ class S3DISSegDataset(_S3DISSegDataset):
                 [self.scene_idxs, one_scene_idxs + offset]).astype(np.int32)
             offset = np.unique(self.scene_idxs).max() + 1
 
-    def concat_label_weight(self, label_weights):
-        """Concat label_weight from several datasets to form self.label_weight.
-
-        Args:
-            label_weights (list[np.ndarray])
-        """
-        # TODO: simply average them?
-        self.label_weight = np.array(label_weights).mean(0).astype(np.float32)
-
     @staticmethod
     def _duplicate_to_list(x, num):
         """Repeat x `num` times to form a list."""
@@ -321,17 +294,3 @@ class S3DISSegDataset(_S3DISSegDataset):
             return scene_idx
         # single idx
         return self._duplicate_to_list(scene_idx, num)
-
-    def _check_label_weights(self, label_weight, num):
-        """Make label_weights as list/tuple."""
-        if label_weight is None:
-            return self._duplicate_to_list(label_weight, num)
-        # label_weight could be str, np.ndarray, list or tuple
-        if isinstance(label_weight, str):  # str
-            return self._duplicate_to_list(label_weight, num)
-        if isinstance(label_weight[0], str):  # list of str
-            return label_weight
-        if isinstance(label_weight[0], (list, tuple, np.ndarray)):  # list of w
-            return label_weight
-        # single weight
-        return self._duplicate_to_list(label_weight, num)

mmdet3d/datasets/scannet_dataset.py

@@ -225,9 +225,6 @@ class ScanNetSegDataset(Custom3DSegDataset):
         scene_idxs (np.ndarray | str, optional): Precomputed index to load
             data. For scenes with many points, we may sample it several times.
             Defaults to None.
-        label_weight (np.ndarray | str, optional): Precomputed weight to \
-            balance loss calculation. If None is given, compute from data.
-            Defaults to None.
     """
     CLASSES = ('wall', 'floor', 'cabinet', 'bed', 'chair', 'sofa', 'table',
                'door', 'window', 'bookshelf', 'picture', 'counter', 'desk',
@@ -271,8 +268,7 @@ class ScanNetSegDataset(Custom3DSegDataset):
                  modality=None,
                  test_mode=False,
                  ignore_index=None,
-                 scene_idxs=None,
-                 label_weight=None):
+                 scene_idxs=None):
 
         super().__init__(
             data_root=data_root,
@@ -283,8 +279,7 @@ class ScanNetSegDataset(Custom3DSegDataset):
             modality=modality,
             test_mode=test_mode,
             ignore_index=ignore_index,
-            scene_idxs=scene_idxs,
-            label_weight=label_weight)
+            scene_idxs=scene_idxs)
 
     def get_ann_info(self, index):
         """Get annotation info according to the given index.
@@ -358,21 +353,17 @@ class ScanNetSegDataset(Custom3DSegDataset):
                             pred_sem_mask, out_dir, file_name,
                             np.array(self.PALETTE), self.ignore_index, show)
 
-    def get_scene_idxs_and_label_weight(self, scene_idxs, label_weight):
-        """Compute scene_idxs for data sampling and label weight for loss \
-        calculation.
+    def get_scene_idxs(self, scene_idxs):
+        """Compute scene_idxs for data sampling.
 
-        We sample more times for scenes with more points. Label_weight is
-        inversely proportional to number of class points.
+        We sample more times for scenes with more points.
         """
         # when testing, we load one whole scene every time
-        # and we don't need label weight for loss calculation
         if not self.test_mode and scene_idxs is None:
             raise NotImplementedError(
                 'please provide re-sampled scene indexes for training')
 
-        return super().get_scene_idxs_and_label_weight(scene_idxs,
-                                                       label_weight)
+        return super().get_scene_idxs(scene_idxs)
 
     def format_results(self, results, txtfile_prefix=None):
         r"""Format the results to txt file. Refer to `ScanNet documentation

tests/test_data/test_datasets/test_s3dis_dataset.py

@@ -17,11 +17,6 @@ def test_seg_getitem():
                [170, 120, 200], [255, 0, 0], [200, 100, 100], [10, 200, 100],
                [200, 200, 200], [50, 50, 50]]
     scene_idxs = [0 for _ in range(20)]
-    label_weight = [
-        3.0441623, 3.3606708, 2.6408234, 4.5086737, 4.8403897, 4.7637715,
-        4.4685297, 4.7051463, 4.9190116, 5.3899403, 4.6436925, 5.0669650,
-        3.6270046
-    ]
 
     pipelines = [
         dict(
@@ -65,8 +60,7 @@ def test_seg_getitem():
         modality=None,
         test_mode=False,
         ignore_index=None,
-        scene_idxs=scene_idxs,
-        label_weights=label_weight)
+        scene_idxs=scene_idxs)
 
     data = s3dis_dataset[0]
     points = data['points']._data
@@ -102,8 +96,6 @@ def test_seg_getitem():
     assert original_palette == palette
     assert s3dis_dataset.scene_idxs.dtype == np.int32
     assert np.all(s3dis_dataset.scene_idxs == np.array(scene_idxs))
-    assert np.allclose(s3dis_dataset.label_weight, np.array(label_weight),
-                       1e-5)
 
     # test dataset with selected classes
     s3dis_dataset = S3DISSegDataset(
@@ -122,7 +114,6 @@ def test_seg_getitem():
     assert s3dis_dataset.VALID_CLASS_IDS == [3, 5]
     assert s3dis_dataset.label_map == label_map
     assert s3dis_dataset.label2cat == {0: 'beam', 1: 'window'}
-    assert np.all(s3dis_dataset.label_weight == np.ones(2))
 
     # test load classes from file
     import tempfile
@@ -160,7 +151,6 @@ def test_seg_getitem():
         test_mode=True,
         scene_idxs=scene_idxs)
     assert np.all(s3dis_dataset.scene_idxs == np.array([0]))
-    assert np.all(s3dis_dataset.label_weight == np.ones(len(class_names)))
 
 
 def test_seg_evaluate():
@@ -232,33 +222,23 @@ def test_multi_areas():
                [170, 120, 200], [255, 0, 0], [200, 100, 100], [10, 200, 100],
                [200, 200, 200], [50, 50, 50]]
     scene_idxs = [0 for _ in range(20)]
-    label_weight = [
-        3.0441623, 3.3606708, 2.6408234, 4.5086737, 4.8403897, 4.7637715,
-        4.4685297, 4.7051463, 4.9190116, 5.3899403, 4.6436925, 5.0669650,
-        3.6270046
-    ]
 
     # repeat
     repeat_num = 3
     s3dis_dataset = S3DISSegDataset(
         data_root=root_path,
         ann_files=[ann_file for _ in range(repeat_num)],
-        scene_idxs=scene_idxs,
-        label_weights=label_weight)
+        scene_idxs=scene_idxs)
     assert s3dis_dataset.CLASSES == class_names
     assert s3dis_dataset.PALETTE == palette
     assert len(s3dis_dataset.data_infos) == repeat_num
     assert np.all(s3dis_dataset.scene_idxs == np.concatenate(
         [np.array(scene_idxs) + i for i in range(repeat_num)]))
-    assert np.allclose(s3dis_dataset.label_weight, np.array(label_weight))
 
-    # different scene_idxs and label_weight input
-    label_weights = np.random.rand(repeat_num, len(class_names))
+    # different scene_idxs input
     s3dis_dataset = S3DISSegDataset(
         data_root=root_path,
         ann_files=[ann_file for _ in range(repeat_num)],
-        scene_idxs=[[0, 0, 1, 2, 2], [0, 1, 2, 3, 3, 4], [0, 1, 1, 2, 2, 2]],
-        label_weights=label_weights)
+        scene_idxs=[[0, 0, 1, 2, 2], [0, 1, 2, 3, 3, 4], [0, 1, 1, 2, 2, 2]])
     assert np.all(s3dis_dataset.scene_idxs == np.array(
         [0, 0, 1, 2, 2, 3, 4, 5, 6, 6, 7, 8, 9, 9, 10, 10, 10]))
-    assert np.allclose(s3dis_dataset.label_weight, label_weights.mean(0))

tests/test_data/test_datasets/test_scannet_dataset.py

@@ -310,12 +310,6 @@ def test_seg_getitem():
         [82, 84, 163],
     ]
     scene_idxs = [0 for _ in range(20)]
-    label_weight = [
-        2.389689, 2.7215734, 4.5944676, 4.8543367, 4.096086, 4.907941,
-        4.690836, 4.512031, 4.623311, 4.9242644, 5.358117, 5.360071, 5.019636,
-        4.967126, 5.3502126, 5.4023647, 5.4027233, 5.4169416, 5.3954206,
-        4.6971426
-    ]
 
     # test network inputs are (xyz, rgb, normalized_xyz)
     pipelines = [
@@ -361,8 +355,7 @@ def test_seg_getitem():
         modality=None,
         test_mode=False,
         ignore_index=None,
-        scene_idxs=scene_idxs,
-        label_weight=label_weight)
+        scene_idxs=scene_idxs)
 
     data = scannet_dataset[0]
     points = data['points']._data
@@ -407,8 +400,6 @@ def test_seg_getitem():
     assert original_palette == palette
     assert scannet_dataset.scene_idxs.dtype == np.int32
     assert np.all(scannet_dataset.scene_idxs == np.array(scene_idxs))
-    assert np.allclose(scannet_dataset.label_weight, np.array(label_weight),
-                       1e-5)
 
     # test network inputs are (xyz, rgb)
     np.random.seed(0)
@@ -496,7 +487,6 @@ def test_seg_getitem():
     assert scannet_dataset.VALID_CLASS_IDS == [3, 5]
     assert scannet_dataset.label_map == label_map
     assert scannet_dataset.label2cat == {0: 'cabinet', 1: 'chair'}
-    assert np.all(scannet_dataset.label_weight == np.ones(2))
 
     # test load classes from file
     import tempfile
@@ -534,7 +524,6 @@ def test_seg_getitem():
         test_mode=True,
         scene_idxs=scene_idxs)
     assert np.all(scannet_dataset.scene_idxs == np.array([0]))
-    assert np.all(scannet_dataset.label_weight == np.ones(len(class_names)))
 
 
 def test_seg_evaluate():