open-mmlab/mmdetection · error · ValueError

Invalid text mode "{self.text_mode}".

Error message

Invalid text mode "{self.text_mode}".

What it means

RefCoco dataset in mmdet supports only specific text_mode values for turning the multi-sentence referring expressions of a region into training text. After handling 'select_first' and 'original', any other string falls through to this ValueError listing the offending mode.

Source

Thrown at mmdet/datasets/refcoco.py:145

            instances = []
            sentences = []
            for grounding_anno in grounding_dict[img_id]:
                texts = [x['raw'].lower() for x in grounding_anno['sentences']]
                # random select one text
                if self.text_mode == 'random':
                    idx = random.randint(0, len(texts) - 1)
                    text = [texts[idx]]
                # concat all texts
                elif self.text_mode == 'concat':
                    text = [''.join(texts)]
                # select the first text
                elif self.text_mode == 'select_first':
                    text = [texts[0]]
                # use all texts
                elif self.text_mode == 'original':
                    text = texts
                else:
                    raise ValueError(f'Invalid text mode "{self.text_mode}".')
                ins = [{
                    'mask': grounding_anno['segmentation'],
                    'ignore_flag': 0
                }] * len(text)
                instances.extend(ins)
                sentences.extend(text)
            data_info = {
                'img_path': join_path(img_prefix, image['file_name']),
                'img_id': img_id,
                'instances': instances,
                'text': sentences
            }
            data_list.append(data_info)

        if len(data_list) == 0:
            raise ValueError(f'No sample in split "{self.split}".')

        return data_list

View on GitHub (pinned to cfd5d3a985)

Solutions

1. Set text_mode='select_first' to use only the first referring expression per instance
2. Set text_mode='original' to keep all referring expressions
3. Check the installed mmdet version's refcoco.py for the exact list of supported modes and use one of those

Example fix

# before
dataset = dict(type='RefCoco', ann_file=..., data_prefix=..., text_mode='random')
# after
dataset = dict(type='RefCoco', ann_file=..., data_prefix=..., text_mode='select_first')

from mmdet.datasets import RefCoco
VALID_TEXT_MODES = {'select_first', 'original'}
text_mode = 'select_first' if text_mode not in VALID_TEXT_MODES else text_mode
dataset = RefCoco(ann_file=..., data_prefix=..., split='train', text_mode=text_mode)

def is_valid_text_mode(mode: str) -> bool:
    return mode in {'select_first', 'original'}

When it happens

Trigger: Constructing RefCoco (or building its transform pipeline in a config) with text_mode set to anything other than 'select_first' or 'original' — e.g. 'random', 'last', a typo like 'orginal', or None.

Common situations: Porting configs from older mmdet versions or other referring-expression codebases where text_mode='random' existed; typos in config strings; assuming all modes listed in a paper are implemented.

Understand the failure class

Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.

Related errors

• No sample in split "{self.split}".
• The type of frame_range must be int or list.
• results does not contain masks.
• {metric} is not in results
• metric must be a list or a str.

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/7554f65eb8377e04. Report an issue: GitHub.