size match

[Richeerard]

作者您好！我使用了4scale进行训练，使用了预训练模型，结果出现了以下错误：size mismatch for transformer.decoder.class_embed.0.weight: copying a param with shape torch.Size([91, 256]) from checkpoint, the shape in current model is torch.Size([4, 256]). 请问我应该如何修改代码

[SlongLiu]

您是在个人数据上训练吗，这个错误是由于类别数被设置成了4和原始预训练模型不同导致的。

[Richeerard]

您是在个人数据上训练吗，这个错误是由于类别数被设置成了4和原始预训练模型不同导致的。

是的，我设置的类别数为4，但是我仍然想要使用预训练模型进行训练。我看了异常位置的代码_load_output = model_without_ddp.load_state_dict（_tmp_st， strict=False），strict设置为False应该能跳过size不匹配问题吧，但是实际上它未能解决该问题，希望得到您的帮助，谢谢！

[Richeerard]

我十分想知道如何在个人数据集上使用类别为91的预训练模型进行预训练，谢谢您了！

我十分想知道如何在个人数据集上使用类别为91的预训练模型进行预训练，谢谢您了！

[SlongLiu]

请参考这里： https://github.com/IDEA-Research/DINO#run Training/Fine-tuning a DINO on your custom dataset

To leverage our pre-trained models for model fine-tuning, we suggest adding two more commands in bash:

--pretrain_model_path /path/to/a/pretrianed/model. specify a pre-trained model. --finetune_ignore label_enc.weight class_embed ignore some inconsistent parameters.

[SlongLiu]

具体来说，添加finetune_ignore忽略掉不一致的参数即可。

--finetune_ignore label_enc.weight class_embed

[anweshb]

Hi @SlongLiu, I tried following the guide to finetune model on a custom dataset with specifically 1 class, changed the number of classes in the corresponding config file and ensured dn_labelbook_size >= num_class + 1 and ran the command !bash "scripts/DINO_train.sh" "../Data" --pretrain_model_path "checkpoint0033_4scale.pth" --finetune_ignore label_enc.weight class_embed. I get the error RuntimeError: Error(s) in loading state_dict for DINO: size mismatch for transformer.decoder.class_embed.0.weight: copying a param with shape torch.Size([91, 256]) from checkpoint, the shape in current model is torch.Size([1, 256]) and similar for all other layers.

Do I need to make some other changed before I try to finetune?

Thanks

[Raneem-MT]

Hi @anweshb we are having a similar problem, did you manage to solve this?