Unused Weights When Loading Model #108

AshOlogn · 2020-10-31T23:04:30Z

I am using the newest version of Transformers (i.e. version 3.0) and try to load a Scibert model using the following code

`scibert_tokenizer = AutoTokenizer.from_pretrained('allenai/scibert_scivocab_uncased')

scibert_model = BertForMaskedLM.from_pretrained('allenai/scibert_scivocab_uncased')
`

When I do this I get the following message indicating unused weights:

`Some weights of the model checkpoint at allenai/scibert_scivocab_uncased were not used when initializing BertForMaskedLM: ['cls.seq_relationship.weight', 'cls.seq_relationship.bias']

This IS expected if you are initializing BertForMaskedLM from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).
This IS NOT expected if you are initializing BertForMaskedLM from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).`

I do not get a similar error when using an earlier version of Transformers (i.e. 2.something).

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Unused Weights When Loading Model #108

Unused Weights When Loading Model #108

Unused Weights When Loading Model #108

Unused Weights When Loading Model #108

Comments

Uh oh!