"Some weights were not used" message with AutoModel

Beginners
1

Hi, I’m a newbie trying to learn Hugginface transformers api.
The line

model = AutoModel.from_pretrained(“bert-base-cased”)

give the following message:

Some weights of the model checkpoint at bert-base-cased were not used when initializing BertModel: [‘cls.predictions.transform.dense.bias’, ‘cls.predictions.bias’, ‘cls.seq_relationship.weight’, ‘cls.predictions.transform.LayerNorm.bias’, ‘cls.seq_relationship.bias’, ‘cls.predictions.transform.LayerNorm.weight’, ‘cls.predictions.decoder.weight’, ‘cls.predictions.transform.dense.weight’]

I didn’t expect that. I know you can obtain such a message if you init a ModelForSomething using a checkpoint for a basic model, but this is not the case.

Any help?
Thanks

2

Probably I understood the problem: in the checkpoint there are some weights related to training layers that are dropped and are not present in BertModel class. So it shouln’t be a problem

3

Yup, exactly!

4

Thank you for replying… When I read more carefully the warning I realized that there were not uninitialized weights in the model (it can happen if you use a base model checkpoint for initializing a ModelForSomething) but the opposite: there were unused checkpoint weights

5

I am having the same problem but with TinyLlama, I fine-tuned a model for a specific task (question answering) and when I am loading the fine-tuned model it shows this warning in addition to:
Some weights of LlamaForCausalLM were not initialized from the model checkpoint at and are newly initialized: [‘lm_head.weight’, ‘model.embed_tokens.weight’, ‘model.norm.weight’]

what could be the problem?

Notes:

  • I am using multu-gpu training and here is how I save the model:
save_dir = os.path.join(project_dir, new_model)
print(f"Saving model to {save_dir}...") 
# save the tokenizer
tokenizer.save_pretrained(save_dir)
# save the model
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(save_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save)
  • here how I load the model:
model = AutoModelForCausalLM.from_pretrained(model_path, local_files_only=True)