No pre-trained weights for {self.class.name}, traini

Error message

No pre-trained weights for {self.__class__.__name__}, training start from scratch

What it means

Swin's init_weights logs (via logger.warn) that no checkpoint was configured, so weights are randomly initialized and training starts from scratch. This happens when init_cfg is None. It is expected for training-from-scratch runs but a red flag when you intended to load a pretrained model.

Source

Thrown at mmdet/models/backbones/swin.py:674

            self.drop_after_pos.eval()

        for i in range(1, self.frozen_stages + 1):

            if (i - 1) in self.out_indices:
                norm_layer = getattr(self, f'norm{i-1}')
                norm_layer.eval()
                for param in norm_layer.parameters():
                    param.requires_grad = False

            m = self.stages[i - 1]
            m.eval()
            for param in m.parameters():
                param.requires_grad = False

    def init_weights(self):
        logger = MMLogger.get_current_instance()
        if self.init_cfg is None:
            logger.warn(f'No pre-trained weights for '
                        f'{self.__class__.__name__}, '
                        f'training start from scratch')
            if self.use_abs_pos_embed:
                trunc_normal_(self.absolute_pos_embed, std=0.02)
            for m in self.modules():
                if isinstance(m, nn.Linear):
                    trunc_normal_init(m, std=.02, bias=0.)
                elif isinstance(m, nn.LayerNorm):
                    constant_init(m, 1.0)
        else:
            assert 'checkpoint' in self.init_cfg, f'Only support ' \
                                                  f'specify `Pretrained` in ' \
                                                  f'`init_cfg` in ' \
                                                  f'{self.__class__.__name__} '
            ckpt = CheckpointLoader.load_checkpoint(
                self.init_cfg.checkpoint, logger=logger, map_location='cpu')
            if 'state_dict' in ckpt:
                _state_dict = ckpt['state_dict']

View on GitHub (pinned to cfd5d3a985)

Solutions

Verify you intended to train from scratch; if not, add init_cfg=dict(type='Pretrained', checkpoint='<path-or-url>') to the backbone config
Check that the checkpoint path/URL exists and is readable
If from-scratch is intended, ignore the message

Example fix

// before
backbone=dict(type='SwinTransformer', embed_dims=96, ...)
// after
backbone=dict(type='SwinTransformer', embed_dims=96, init_cfg=dict(type='Pretrained', checkpoint='https://download.openmmlab.com/...'), ...)

Defensive patterns

Strategy: validation

Validate before calling

backbone_cfg = cfg['model']['backbone']
if backbone_cfg.get('init_cfg') is None:
    logging.warning('Swin will train from scratch; add init_cfg if a checkpoint was intended')

Prevention

Always verify backbone.init_cfg before training when fine-tuning is intended
Log the effective init_cfg at startup

When it happens

Trigger: Building a SwinTransformer backbone without init_cfg/pretrained and calling .init_weights() (directly or via detector init).

Common situations: Checkpoint path typo or missing init_cfg in config; user assumed default weights load automatically; fine-tuning config copied but init_cfg removed.

Related errors

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/27b45c57643f00de. Report an issue: GitHub.