TF's alternative for HuggingFace's deprecated `TF`<foo> models?

TensorFlow
,
1

HuggingFace has deprecated TF models. What is the TF alternative for me to use the transformers tokenizer and BERT models? Specifically, I am looking for the following:

from transformers import AutoTokenizer, TFBertForMaskedLM
MODEL = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = TFBertForMaskedLM.from_pretrained(MODEL)
3

Hi @khteh, The recommended approach is to switching from TF to PyTorch.

from transformers import AutoTokenizer, AutoModelForMaskedLM
MODEL = “bert-base-uncased”
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForMaskedLM.from_pretrained(MODEL)
4

Do I need to install PyTorch library for this to work? If “yes”, wouldn’t that introduce too many and sometimes conflicting dependencies in my project? Is there no solution at all from TF?

5

Could you tried with older version of transformers??

6

No I can’t and don’t do that!

8

Hi @khteh, Could you please refer to this document to switch TFBertForMaskedLM model to PyTorch.

10

What you mean is that using tensorflow/keras models is discouraged / a dead end with HuggingFace?