open-mmlab/mmdetection · error · ValueError

new classes {new_classes} is not a subset of classes {old_cl

Error message

new classes {new_classes} is not a subset of classes {old_classes} in METAINFO.

What it means

BaseSemanticSegmentationDataset.get_label_map builds a mapping when the user-supplied metainfo['classes'] differs from the dataset class's default METAINFO['classes']. The new classes must be a subset of the original classes; otherwise (renamed or entirely new labels) it raises this ValueError. The mapping only supports dropping/reordering existing classes and mapping dropped ones to background (0).

Source

Thrown at mmdet/datasets/base_semseg_dataset.py:169

        is not equal to new classes in self._metainfo and nether of them is not
        None, `label_map` is not None.

        Args:
            new_classes (list, tuple, optional): The new classes name from
                metainfo. Default to None.


        Returns:
            dict, optional: The mapping from old classes in cls.METAINFO to
                new classes in self._metainfo
        """
        old_classes = cls.METAINFO.get('classes', None)
        if (new_classes is not None and old_classes is not None
                and list(new_classes) != list(old_classes)):

            label_map = {}
            if not set(new_classes).issubset(cls.METAINFO['classes']):
                raise ValueError(
                    f'new classes {new_classes} is not a '
                    f'subset of classes {old_classes} in METAINFO.')
            for i, c in enumerate(old_classes):
                if c not in new_classes:
                    # 0 is background
                    label_map[i] = 0
                else:
                    label_map[i] = new_classes.index(c)
            return label_map
        else:
            return None

    def _update_palette(self) -> list:
        """Update palette after loading metainfo.

        If length of palette is equal to classes, just return the palette.
        If palette is not defined, it will randomly generate a palette.
        If classes is updated by customer, it will return the subset of

View on GitHub (pinned to cfd5d3a985)

Solutions

Use exactly the original class names (copy spelling/case from the dataset class METAINFO) and only drop entries to define your subset
If you need genuinely new classes, subclass the dataset and override METAINFO = dict(classes=(...)) instead of passing metainfo at construction
Check for typos/case differences between your classes list and METAINFO['classes'] programmatically before constructing

Example fix

# before
ds = CityscapesDataset(
  data_root='data/', data_prefix=dict(img_path='leftImg8bit', seg_map_path='gtFine'),
  metainfo=dict(classes=['road', 'sidewalk', 'vehicle']))  # 'vehicle' not in METAINFO
# after
from mmdet.datasets import CityscapesDataset
# option A: subset with exact names
ds = CityscapesDataset(..., metainfo=dict(classes=['road', 'sidewalk', 'car']))
# option B: new taxonomy -> subclass
class MyDataset(CityscapesDataset):
  METAINFO = dict(classes=('road', 'sidewalk', 'vehicle'))
ds = MyDataset(...)

Defensive patterns

Strategy: validation

Validate before calling

import inspect
ds_cls = CocoDataset  # whatever class you use
old = list(ds_cls.METAINFO['classes'])
new = my_classes
unknown = set(map(str, new)) - set(map(str, old))
assert not unknown, f'classes not in METAINFO (typo/subset violation): {sorted(unknown)}'

Type guard

def is_valid_class_subset(cls, new_classes) -> bool:
    old = list(cls.METAINFO.get('classes') or [])
    return set(map(str, new_classes)).issubset(set(map(str, old)))

Try / catch

try:
    ds = DatasetClass(..., metainfo=dict(classes=my_classes))
except ValueError as e:
    if 'not a subset of classes' in str(e):
        raise SystemExit(f'Fix class names {my_classes} to match METAINFO {cls.METAINFO["classes"]} or subclass with new METAINFO')
    raise

Prevention

Copy class names verbatim from the dataset class source or METAINFO
For new taxonomies, subclass and override METAINFO instead of metainfo at init
Diff your classes list against METAINFO['classes'] in a unit test, catching case/whitespace mismatches

When it happens

Trigger: Passing metainfo=dict(classes=[...]) to a semseg dataset where the list contains class names absent from the class's METAINFO['classes'], e.g. custom names for a model fine-tuned on a different label taxonomy, or reordered-with-renames.

Common situations: Fine-tuning a semseg model on a subset of ADE20K/Cityscapes but changing label spellings; loading a custom dataset by extending a built-in class; case mismatches ('Road' vs 'road').

Related errors

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/105972a5fb1229c3. Report an issue: GitHub.