{metric.class.name} got empty `self.results`.Please

Error message

{metric.__class__.__name__} got empty `self.results`.Please ensure that the processed results are properly added into `self.results` in `process` method.

What it means

Warning from MultiDatasetsEvaluator.evaluate: one of the per-dataset metrics has an empty self.results list at evaluate time, meaning process() never added results for that dataset — its metrics will be meaningless/zero.

Source

Thrown at mmdet/evaluation/evaluator/multi_datasets_evaluator.py:73

            size (int): Length of the entire validation dataset. When batch
                size > 1, the dataloader may pad some data samples to make
                sure all ranks have the same length of dataset slice. The
                ``collect_results`` function will drop the padded data based on
                this size.

        Returns:
            dict: Evaluation results of all metrics. The keys are the names
            of the metrics, and the values are corresponding results.
        """
        metrics_results = OrderedDict()
        dataset_slices = self._get_cumulative_sizes()
        assert len(dataset_slices) == len(self.dataset_prefixes)

        for dataset_prefix, start, end, metric in zip(
                self.dataset_prefixes, [0] + dataset_slices[:-1],
                dataset_slices, self.metrics):
            if len(metric.results) == 0:
                warnings.warn(
                    f'{metric.__class__.__name__} got empty `self.results`.'
                    'Please ensure that the processed results are properly '
                    'added into `self.results` in `process` method.')

            results = collect_results(metric.results, size,
                                      metric.collect_device)

            if is_main_process():
                # cast all tensors in results list to cpu
                results = _to_cpu(results)
                _metrics = metric.compute_metrics(
                    results[start:end])  # type: ignore

                if metric.prefix:
                    final_prefix = '/'.join((dataset_prefix, metric.prefix))
                else:
                    final_prefix = dataset_prefix
                print(f'================{final_prefix}================')

View on GitHub (pinned to cfd5d3a985)

Solutions

Check that every dataset in the multi-evaluator has data flowing through process() (dataloader not empty for that subset)
Verify metric class implements process() appending to self.results for your datasample type
Ensure dataset_prefixes/metrics lists in the evaluator config align with the dataloader datasets
Log len(metric.results) per prefix before evaluate to pinpoint the empty one

Example fix

# before
val_evaluator = dict(type='MultiDatasetsEvaluator', metrics=[dict(type='CocoMetric')], dataset_prefixes=['coco','voc'])
# after: ensure each subset yields data, e.g. fix empty voc split / batch remainder handling

Defensive patterns

Strategy: validation

Validate before calling

for prefix, metric in zip(evaluator.dataset_prefixes, evaluator.metrics):
    if len(metric.results) == 0:
        print(f'warning: no results collected for {prefix}/{type(metric).__name__}')

Prevention

Sanity-check that each dataset subset yields batches
Verify metric.process() handles your datasample type
Align dataset_prefixes with the multi-dataset dataloader config

When it happens

Trigger: Using MultiDatasetsEvaluator where a metric bound to a dataset_prefix received no process() calls (e.g. batch size not divisible causing dropped samples, a metric whose process filters everything, or a mis-matched evaluator/dataset wiring).

Common situations: Evaluating on multiple concatenated datasets (e.g. COCO+VOC) with metrics that skip results; leads to empty mAP or divide-by-zero downstream.

Related errors

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/1dc540387ae973ca. Report an issue: GitHub.