Skip to content

Latest commit

 

History

History
35 lines (25 loc) · 1.54 KB

huggingface_transformers.md

File metadata and controls

35 lines (25 loc) · 1.54 KB
Raw

Use the model with huggingface/transformers

FinBERT is included in transformers, so you can simply use the model as usual:

model = transformers.BertForMaskedLM.from_pretrained("TurkuNLP/bert-base-finnish-cased-v1") 
model.eval()
if torch.cuda.is_available():
    model = model.cuda()
tokenizer = transformers.BertTokenizer.from_pretrained("TurkuNLP/bert-base-finnish-cased-v1")

Use "bert-base-finnish-uncased-v1" if you want to use the uncased model.

Convert from tensorflow

This is mostly a "note-to-self" on how the model is converted from tensorflow checkpoint to the huggingface/transformers format.

python3 -m venv venv-transformers
source venv-transformers/bin/activate
pip3 install torch torchvision torchtext tensorflow-gpu

Then you can convert the model as follows:

Note: If you get the error 'BertForPreTraining' object has no attribute 'shape' then you may need to edit venv-transformers/lib64/python3.6/site-packages/transformers/modeling_bert.py as instructed in this issue.

mkdir bert-base-finnish-cased-transformers-v1
python3 -m transformers.convert_bert_original_tf_checkpoint_to_pytorch --tf_check bert-base-finnish-cased/bert-base-finnish-cased-v1 --bert_config bert-base-finnish-cased-v1/bert_config.json --pytorch bert-base-finnish-cased-transformers-v1/pytorch_model.bin
cp ./bert-base-finnish-cased-v1/bert_config.json bert-base-finnish-cased-transformers-v1/config.json
cp ./bert-base-finnish-cased-v1/vocab.txt bert-base-finnish-cased-transformers-v1/