语义分割时，单卡预训练权重加载

[creater-zq]

您好！我使用upernet_uniformer做语义分割，只使用了一个GPU。

• 使用预训练权重uniformer_base_in1k.pth，加载时出现如下错误：

  
  unexpected key in source state_dict: norm.weight, norm.bias, norm.running_mean, norm.running_var, norm.num_batches_tracked, head.weight, head.bias

missing keys in source state_dict: norm1.weight, norm1.bias, norm2.weight, norm2.bias, norm3.weight, norm3.bias, norm4.weight, norm4.bias

* 使用预训练权重`uniformer_base_ls_in1k.pth`时，加载出现如下错误：
```bash
unexpected key in source state_dict: norm.weight, norm.bias, norm.running_mean, norm.running_var, norm.num_batches_tracked, head.weight, head.bias, blocks3.0.gamma_1, blocks3.0.gamma_2, blocks3.1.gamma_1, blocks3.1.gamma_2, blocks3.2.gamma_1, blocks3.2.gamma_2, blocks3.3.gamma_1, blocks3.3.gamma_2, blocks3.4.gamma_1, blocks3.4.gamma_2, blocks3.5.gamma_1, blocks3.5.gamma_2, blocks3.6.gamma_1, blocks3.6.gamma_2, blocks3.7.gamma_1, blocks3.7.gamma_2, blocks3.8.gamma_1, blocks3.8.gamma_2, blocks3.9.gamma_1, blocks3.9.gamma_2, blocks3.10.gamma_1, blocks3.10.gamma_2, blocks3.11.gamma_1, blocks3.11.gamma_2, blocks3.12.gamma_1, blocks3.12.gamma_2, blocks3.13.gamma_1, blocks3.13.gamma_2, blocks3.14.gamma_1, blocks3.14.gamma_2, blocks3.15.gamma_1, blocks3.15.gamma_2, blocks3.16.gamma_1, blocks3.16.gamma_2, blocks3.17.gamma_1, blocks3.17.gamma_2, blocks3.18.gamma_1, blocks3.18.gamma_2, blocks3.19.gamma_1, blocks3.19.gamma_2, blocks4.0.gamma_1, blocks4.0.gamma_2, blocks4.1.gamma_1, blocks4.1.gamma_2, blocks4.2.gamma_1, blocks4.2.gamma_2, blocks4.3.gamma_1, blocks4.3.gamma_2, blocks4.4.gamma_1, blocks4.4.gamma_2, blocks4.5.gamma_1, blocks4.5.gamma_2, blocks4.6.gamma_1, blocks4.6.gamma_2

missing keys in source state_dict: norm1.weight, norm1.bias, norm2.weight, norm2.bias, norm3.weight, norm3.bias, norm4.weight, norm4.bias

• 我使用的backbone为(https://github.com/Sense-X/UniFormer/blob/main/semantic_segmentation/mmseg/models/backbones/uniformer.py).
• _base_中的配置使用了(https://github.com/Sense-X/UniFormer/blob/main/semantic_segmentation/configs/_base_/models/upernet_uniformer.py). 并将norm_cfg进行修改。

  # norm_cfg = dict(type='SyncBN', requires_grad=True)
  norm_cfg = dict(type='BN', requires_grad=True)

• 在配置文件中添加了pretrained_path:

  model = dict(
  backbone=dict(
      type='UniFormer',
      embed_dim=[64, 128, 320, 512],
      layers=[5, 8, 20, 7],
      head_dim=64,
      drop_path_rate=0.4,
      use_checkpoint=True,  # 避免显存溢出
      checkpoint_num=[0, 0, 2, 0],  # 使用checkpoint的位置
      # 可替换为uniformer_base_ls_in1k.pth
      pretrained_path='../uniformer_pretrained/uniformer_base_in1k.pth',
      windows=False,
      hybrid=False  # 测试时使用，True
  ),
  decode_head=dict(
      in_channels=[64, 128, 320, 512],
      num_classes=150
  ),
  auxiliary_head=dict(
      in_channels=320,
      num_classes=150
  ))

• 请问，我应该如何解决该问题，谢谢！

[Andy1621]

It's normal to print this error because of the difference between the backbones of pre-training and downstream tasks. However, this is just a warning as the strict=False: https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/a90acdda2156c93a447e16a2fd8befc63bd534de/mmcv_custom/checkpoint.py#L97-L106

[creater-zq]

哦哦哦，也就是不用管，直接train就完事了