How can I get the file 'vocab.txt'? about spert HOT 8 CLOSED

Hi ZengYoufeng,
the corresponding vocabulary file is already included for the models we trained on the CoNLL04, ADE and SciERC datasets (under "./data/models" after executing "./scripts/fetch_models.sh"). If you want to train a model from-scratch, you should have a look at the "transformers" (https://github.com/huggingface/transformers) library we are using. You can either specify one of the pre-trained BERT models included in "transformers" as both the "model_path" and "tokenizer_path" (e.g. "bert-base-cased" as in "configs/example_train.conf") or point "model_path" and "tokenizer_path" to a directory that contains model weights or a vocabulary file respectively.

from spert.

Did you train your own model on the self-built dataset or where is 'pytorch_model.bin' from? You should use the same 'vocab.txt' file that was used while training 'pytorch_model.bin'.

from spert.

You should use the same 'vocab.txt' file that was used during training. In your case, this is probably the one located in 'data/models/conll04'. So just set 'tokenizer_path' in 'example_eval.conf' to point to 'data/models/conll04'. This is not a vocabulary file specifically created for the CoNLL04 dataset, but the rather generic 'bert-base-cased' vocabulary (see the transformers library documentation).

from spert.

Did you train your own model on the self-built dataset or where is 'pytorch_model.bin' from? You should use the same 'vocab.txt' file that was used while training 'pytorch_model.bin'.

I have trained my own model on the self-built dataset on the same training settings of the CoNLL04 dataset. Then, I got three files, 'pytorch_model.bin', 'config.json' and 'extra.state', respectively. However, I can't find the file 'vocab.txt'. Therefore, I cannot execute the command 'python ./spert.py eval --config configs/example_eval.conf'. I want to know how to get the file 'vocab.txt' that fits self-built dataset.

Please refer to the use of huggingface/transformers!!!!!!

from spert.

Hi ZengYoufeng,
the corresponding vocabulary file is already included for the models we trained on the CoNLL04, ADE and SciERC datasets (under "./data/models" after executing "./scripts/fetch_models.sh"). If you want to train a model from-scratch, you should have a look at the "transformers" (https://github.com/huggingface/transformers) library we are using. You can either specify one of the pre-trained BERT models included in "transformers" as both the "model_path" and "tokenizer_path" (e.g. "bert-base-cased" as in "configs/example_train.conf") or point "model_path" and "tokenizer_path" to a directory that contains model weights or a vocabulary file respectively.

Thanks for your answer, I have got the trained model 'pytorch_model.bin' with self-built dataset, but I cannot execute the command 'python ./spert.py eval --config configs/example_eval.conf' successfully due to the lack of the file 'vocab.txt'. Now I want to know how to generate a vocabulary file that fits self-built dataset for evaluation. Hope you can give me some detailed guidance, I will be grateful!

from spert.

Did you train your own model on the self-built dataset or where is 'pytorch_model.bin' from? You should use the same 'vocab.txt' file that was used while training 'pytorch_model.bin'.

I have trained my own model on the self-built dataset on the same training settings of the CoNLL04 dataset. Then, I got three files, 'pytorch_model.bin', 'config.json' and 'extra.state', respectively. However, I can't find the file 'vocab.txt'. Therefore, I cannot execute the command 'python ./spert.py eval --config configs/example_eval.conf'. I want to know how to get the file 'vocab.txt' that fits self-built dataset.

from spert.

I understand that the file 'vocab.txt' is related to the language representation model, not to the dataset. I observed that the vocabulary files corresponding to the three datasets were different, so I misunderstood that the vocabulary file corresponds to the dataset and was generated by it. I'm so sorry for the misunderstanding and thank you very much for your kind help.

from spert.

No need to be sorry, I'm glad I could help you. The vocabulary files differ because we used SciBERT, which is accompanied by its own vocabulary, instead of BERT-Base-Cased for the SciERC dataset. The ADE and CoNLL04 models were trained using the same BERT-Base-Cased vocabulary.

from spert.