open-mmlab/mmdetection · error · KeyError

`init_cfg` must contain the key "type"

Error message

`init_cfg` must contain the key "type"

What it means

DetectoRS_ResNet requires that any dict init_cfg containing a 'checkpoint' key also contains type='Pretrained' (the only supported init mode via this legacy path). An init_cfg dict without a 'type' key raises this KeyError.

Source

Thrown at mmdet/models/backbones/detectors_resnet.py:252

    def __init__(self,
                 sac=None,
                 stage_with_sac=(False, False, False, False),
                 rfp_inplanes=None,
                 output_img=False,
                 pretrained=None,
                 init_cfg=None,
                 **kwargs):
        assert not (init_cfg and pretrained), \
            'init_cfg and pretrained cannot be specified at the same time'
        self.pretrained = pretrained
        if init_cfg is not None:
            assert isinstance(init_cfg, dict), \
                f'init_cfg must be a dict, but got {type(init_cfg)}'
            if 'type' in init_cfg:
                assert init_cfg.get('type') == 'Pretrained', \
                    'Only can initialize module by loading a pretrained model'
            else:
                raise KeyError('`init_cfg` must contain the key "type"')
            self.pretrained = init_cfg.get('checkpoint')
        self.sac = sac
        self.stage_with_sac = stage_with_sac
        self.rfp_inplanes = rfp_inplanes
        self.output_img = output_img
        super(DetectoRS_ResNet, self).__init__(**kwargs)

        self.inplanes = self.stem_channels
        self.res_layers = []
        for i, num_blocks in enumerate(self.stage_blocks):
            stride = self.strides[i]
            dilation = self.dilations[i]
            dcn = self.dcn if self.stage_with_dcn[i] else None
            sac = self.sac if self.stage_with_sac[i] else None
            if self.plugins is not None:
                stage_plugins = self.make_stage_plugins(self.plugins, i)
            else:
                stage_plugins = None

View on GitHub (pinned to cfd5d3a985)

Solutions

1. Always include the type key: init_cfg=dict(type='Pretrained', checkpoint='...')
2. Use only type='Pretrained' when supplying a checkpoint in this branch
3. Validate init_cfg keys before constructing: assert 'type' in init_cfg

Example fix

# before
backbone=dict(type='DetectoRS_ResNet', init_cfg=dict(checkpoint='resnet50.pth'))
# after
backbone=dict(type='DetectoRS_ResNet', init_cfg=dict(type='Pretrained', checkpoint='resnet50.pth'))

if isinstance(init_cfg, dict) and 'checkpoint' in init_cfg:
    assert init_cfg.get('type') == 'Pretrained', 'init_cfg needs type="Pretrained"'

def valid_init_cfg(cfg) -> bool:
    return not isinstance(cfg, dict) or ('type' in cfg and (cfg.get('type') != 'Pretrained' or 'checkpoint' in cfg))

When it happens

Trigger: Passing init_cfg={'checkpoint':'resnet50.pth'} (missing 'type') to DetectoRS_ResNet; passing init_cfg with type='Constant' together with a checkpoint key through this legacy branch.

Common situations: Hand-writing init_cfg and forgetting the type field; merging configs where 'type' gets overridden or dropped; mixing new-style init_cfg with legacy pretrained handling.

Related errors

• pretrained must be a str or None
• pretrained must be a str or None
• pretrained must be a str or None
• pretrained must be a str or None
• pretrained must be a str or None

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/dba1fd3c66965781. Report an issue: GitHub.