WZMIAOMIAO/deep-learning-for-image-processing · error · KeyError

not support iou_type: {self.iou_type}

Error message

not support iou_type: {self.iou_type}

What it means

CocoEval.update() only implements result preparation for iou_type == 'keypoints'; any other configured iou_type (e.g. 'bbox' or 'segm') hits the else branch and raises KeyError with the unsupported type. The wrapper class is a keypoint-specific adaptation of pycocotools' COCOeval, so it deliberately rejects other evaluation types.

Source

Thrown at pytorch_keypoint/HRNet/train_utils/coco_eval.py:99

            keypoints = np.concatenate([keypoints, scores], axis=1)
            keypoints = np.reshape(keypoints, -1)

            # We recommend rounding coordinates to the nearest tenth of a pixel
            # to reduce resulting JSON file size.
            keypoints = [round(k, 2) for k in keypoints.tolist()]

            res = {"image_id": target["image_id"],
                   "category_id": 1,  # person
                   "keypoints": keypoints,
                   "score": target["score"] * k_score}

            self.results.append(res)

    def update(self, targets, outputs):
        if self.iou_type == "keypoints":
            self.prepare_for_coco_keypoints(targets, outputs)
        else:
            raise KeyError(f"not support iou_type: {self.iou_type}")

    def synchronize_results(self):
        # 同步所有进程中的数据
        eval_ids, eval_results = merge(self.obj_ids, self.results)
        self.aggregation_results = {"obj_ids": eval_ids, "results": eval_results}

        # 主进程上保存即可
        if is_main_process():
            # results = []
            # [results.extend(i) for i in eval_results]
            # write predict results into json file
            json_str = json.dumps(eval_results, indent=4)
            with open(self.results_file_name, 'w') as json_file:
                json_file.write(json_str)

    def evaluate(self):
        # 只在主进程上评估即可
        if is_main_process():

View on GitHub (pinned to 1ec3fe6f37)

Solutions

1. Set iou_type="keypoints" where the evaluator is constructed for human-pose validation.
2. If you need bbox/segm eval, use pycocotools COCOeval directly or the torchvision coco_eval wrapper that supports those types.
3. Use the generic coco_eval.py from pytorch_object_detection/ references (supports bbox/segm) for detection tasks.
4. Guard the construction site to assert the supported value before training starts.

Example fix

# before
coco_evaluator = CocoEvaluator(base_dataset, iou_type="bbox")
# after
coco_evaluator = CocoEvaluator(base_dataset, iou_type="keypoints")

SUPPORTED_IOU_TYPES = {"keypoints"}
assert iou_type in SUPPORTED_IOU_TYPES, f"HRNet CocoEval only supports {SUPPORTED_IOU_TYPES}, got {iou_type}"

def is_keypoint_eval(iou_type: str) -> bool:
    return iou_type == "keypoints"

try:
    coco_evaluator.update(targets, outputs)
except KeyError as e:
    logging.error("Unsupported iou_type configured: %s", e)
    raise SystemExit("Set iou_type='keypoints' for pose evaluation")

When it happens

Trigger: Constructing the evaluator (or passing a config) with iou_type set to anything other than "keypoints", then calling update(targets, outputs) during validation.

Common situations: Copying the HRNet eval utils into a detection project and leaving/setting iou_type='bbox'; a config file shared between bbox and keypoint tasks; refactoring where the default iou_type changed.

Related errors

• Please run accumulate() first
• Unknown iou type {}
• Please run accumulate() first
• illegal stride value.
• The inverted_residual_setting should not be empty.

AI-assisted analysis of WZMIAOMIAO/deep-learning-for-image-processing@1ec3fe6f37 (2026-08-30). Data as JSON: /api/errors/7f1a3ea6df39288c. Report an issue: GitHub.