TF mT5 model is not adding new tokens into it's vocabulary.

[laddhakrishna]

Environment info

• transformers version: 4.11.2
• Platform: google colab
• Python version: 3.7.12
• PyTorch version (GPU?):
• Tensorflow version (GPU?): 2.6.0
• Using GPU in script?: Yes
• Using distributed or parallel set-up in script?: No

Who can help @patrickvonplaten, @patil-suraj

Information

Model I am using (Bert, XLNet ...):
I am using mT5 model. The problem arised when I am trying to resize the token embeddings. I added new special tokens to my tokenizer's vocabulary, when I resized the model's token embeddings according to the length of the tokenizer, I am facing the issue.

Link for the notebook: https://colab.research.google.com/drive/1ooKa5aQ_FAEnicxBL8bJnNAO4H9vFuv-?usp=sharing
Code from the notebook:
!pip install transformers
!pip install sentencepiece
from transformers import TFMT5ForConditionalGeneration, T5Tokenizer
model = TFMT5ForConditionalGeneration.from_pretrained("google/mt5-base")
tokenizer = T5Tokenizer.from_pretrained("google/mt5-base")
tokenizer.add_special_tokens({'bos_token':'','eos_token':''})
model._resize_token_embeddings(len(tokenizer))

The error message is:
ValueError: Attempt to convert a value (None) with an unsupported type (<class 'NoneType'>) to a Tensor.

The problem arises when using:

• the official example scripts: (give details below)
• my own modified scripts: (give details below)

The tasks I am working on is:

• an official GLUE/SQUaD task: (give the name)
• my own task or dataset: (give details below)

To reproduce

Steps to reproduce the behavior:

1. You can directly run the code & check the error coming from the last line.
2. Also you can go the colab notebook & have a look at the error as well.

Expected behavior

[patil-suraj]

This seems to be an issue with TF model cc @Rocketknight1

@patrickvonplaten it seems there are no extra tokens in mt5 tokenizer and there's a mismatch between tokenizer.vocab_size and config.vocab_size. Is this a known issue?

[patrickvonplaten]

Actually, everything looks fine for me here...

TFMT5Model and MT5Tokenier have a tokenizer mismatch because of the same reason as T5, see: #4875

Apart from this the following code works:

from transformers import TFMT5ForConditionalGeneration, T5Tokenizer
model = TFMT5ForConditionalGeneration.from_pretrained("google/mt5-base")
tokenizer = T5Tokenizer.from_pretrained("google/mt5-base")
tokenizer.add_special_tokens({'bos_token':'','eos_token':''})
model.resize_token_embeddings(len(tokenizer))

@laddhakrishna - is there any reason you used model._resize_token_embeddings(...) instead of model.resize_token_embeddings(...) ?

[laddhakrishna]

Sir even if we use --> model.resize_token_embeddings(...) ,still we are getting the same error.
What can we do??

[patrickvonplaten]

Hey @laddhakrishna,

could you provide a google colab that reproduces the error with model.resize_token_embeddings(...)? I didn't manage to reproduce the error. Thanks!

[laddhakrishna]

Colab notebook link:
https://colab.research.google.com/drive/1xSB7XlIgA7PrGTUqThZl-rkBknxLYc87?usp=sharing

Please have a look at it sir.
Thanks!

[patrickvonplaten]

Hey @laddhakrishna,

Thanks for the colab I can reproduce!

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[LysandreJik]

Maybe @Rocketknight1 or @gante can take a look?