Some weights of SMTModelForCausalLM were not initialized from the model checkpoint

[micuentadecasa]

Hi, I'm trying to use your model, I copied some of your files to add the model to my program, but it didn't worked, I got this error:

Some weights of SMTModelForCausalLM were not initialized from the model checkpoint at antoniorv6/smt-camera-grandstaff and are newly initialized: ['decoder.decoder.layers.0.cross_attn.k_proj.bias', ......
....
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
<p

so it seems that the model is not loaded with the weights, I don't want, and I cannot run your project given GPU limitations, what we need is the model with the weights, instructions on how to use it, what output it produces, etc.

Regards.

[zral0kh]

Hi, it seems the decoder weights are missing entirely.
@antoniorv6 could you please share them on huggingface as well?

[antoniorv6]

Hi!
Are you trying to use the system-level weights or the full-page weights? For the moment, the system-level weights are pending to be updated at huggingface, so the problem that you are reporting is mainly because of that. Therefore, the Issue is almost related to Issue #19 .
If this issue is happening with the full-page weights (full-page GrandStaff, Mozarteum or Polish Scores, please let me know because definietly there is something wrong then.

[micuentadecasa]

yes, I tried the GrandStaff model, and it doesn't have the weights

[antoniorv6]

With the current update to the 2.0 version of the SMT, we did not update the weights of GrandStaff and Camera GrandStaff. Nevertheless, these weights will be pushed as soon as possible. I will reply to this Issue once this is done.

[antoniorv6]

Hi! Can you please check out if this problem dissappears if you load from this weights? PRAIG/smt-grandstaff I want to double-check before updating the docs.

[Harounnn]

Hi Antonio, I tried it and it is working, the weights are loaded and initialized well!

[antoniorv6]

Great! I will try to upload asap the rest of the weights (Camera GrandStaff and, quartets if possible). Full-page weights are available, but not linked in the README.MD.

[DavidNystrom]

How did you get it working, @Harounnn?
I'm exploring this repo, and can't seem to get their sample code working. I just get the "Some weights of the model checkpoint at antoniorv6/smt-grandstaff were not used when initializing SMTModelForCausalLM:" output.

[DavidNystrom]

Nevermind, I got it working!

[dillon-0]

Hi @antoniorv6 I tried using PRAIG/smt-grandstaff and smt-fp-grandstaff and got the following error:

Edit: I was able to get around this by using transformers==4.40.2 following huggingface/transformers#30965 (comment)

Traceback (most recent call last):
  File "/root/SMT/run.py", line 8, in <module>
    model = SMTModelForCausalLM.from_pretrained("PRAIG/smt-fp-grandstaff").to(device)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py", line 315, in _wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py", line 4930, in from_pretrained
    model = cls(config, *model_args, **model_kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/root/SMT/smt_model/modeling_smt.py", line 344, in __init__
    super().__init__(config)
  File "/usr/local/lib/python3.11/site-packages/transformers/modeling_utils.py", line 2191, in __init__
    self.config._attn_implementation, is_init_check=True
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/transformers/configuration_utils.py", line 211, in __getattribute__
    return super().__getattribute__(key)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/transformers/configuration_utils.py", line 407, in _attn_implementation
    return self._attn_implementation_internal
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/transformers/configuration_utils.py", line 211, in __getattribute__
    return super().__getattribute__(key)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'SMTConfig' object has no attribute '_attn_implementation_internal'

[antoniorv6]

Oh, it seems that this is most likely because of the HuggingFace version. I will put it in another issue so we can track it specifically.