KeyError: 'PALETTE' when use method init_segmentor()

[amadeusneko]

The error code:

Use load_from_local loader
~/workspace/mmcv-test/mmseg/models/backbones/hrnet.py:298: UserWarning: DeprecationWarning: pretrained is a deprecated, please use "init_cfg" instead
  warnings.warn('DeprecationWarning: pretrained is a deprecated, '
---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
<ipython-input-9-da74faa07110> in <module>
      7     pth_file = osp.join(work_dir, 'iter_30000.pth')
      8     print(torch.load(pth_file))
----> 9     model = init_segmentor(cfg, pth_file, device='cuda:0')
     10 else:
     11     model.cfg = cfg

~/workspace/mmcv-test/mmseg/apis/inference.py in init_segmentor(config, checkpoint, device)
     33         checkpoint = load_checkpoint(model, checkpoint, map_location='cpu')
     34         model.CLASSES = checkpoint['meta']['CLASSES']
---> 35         model.PALETTE = checkpoint['meta']['PALETTE']
     36     model.cfg = config  # save the config in the model for convenience
     37     model.to(device)

KeyError: 'PALETTE'

My CustomDataset:

@DATASETS.register_module()
class BDD100KLaneLabelDataset(CustomDataset):
  CLASSES = classes
  PALETTE = palette
  def __init__(self, split, **kwargs):
    super().__init__(img_suffix='.jpg', seg_map_suffix='.png', 
                     split=split, **kwargs)
    assert osp.exists(self.img_dir) and self.split is not None
print('done')

I solved it using a simple way:
modify model.PALETTE = checkpoint['meta']['PALETTE'] to model.PALETTE = my_paletee
at line 35 in mmseg/apis/inference.py.

But I want to know an official way to solve this error, please tell me.
thanks

[Junjun2016]

Hi @wgggSg
Maybe you have not saved the meta or PALETTE in the checkpoint.

[amadeusneko]

@Junjun2016
It is indeed the problem you said. I have solved it by add the code

cfg.checkpoint_config.meta = dict(
    CLASSES=classes,
    PALETTE=palette)

to my config file, thanks!

[459737087]

Good job!

[LansonChen]

@Junjun2016 It is indeed the problem you said. I have solved it by add the code

cfg.checkpoint_config.meta = dict(
    CLASSES=classes,
    PALETTE=palette)

to my config file, thanks!
@wgggSg
hi,did you use jupyter notebook to modify those configs? Could u share one copy of the ipynb to me, thanks!

[amadeusneko]

@Junjun2016 It is indeed the problem you said. I have solved it by add the code

cfg.checkpoint_config.meta = dict(
    CLASSES=classes,
    PALETTE=palette)

to my config file, thanks!
@wgggSg
hi,did you use jupyter notebook to modify those configs? Could u share one copy of the ipynb to me, thanks!

sorry, it have been a long time I didn't use mmsegmentation, so that my code file couldn't be found.....
I remembered i add the code above to train.py so that the error is solved, maybe just add PALETTE and CLASSES to your checkpoint could also solve this problem.