Using trained models in a separate process for inference

[seamusjolly]

Hi, thank you for this great paper and implementation. I have a question about loading a trained model in a separate process and using it for inference.

I have been able to successfully train a BertForPromptFinetuning model using a BERT MLM, and it has been saved to a directory. I am attempting to load it for inference using the .from_pretrained function, like so:

# load models, config and tokenizers
config = AutoConfig.from_pretrained("../LM-BFF/result/tmp/config.json")
model = BertForPromptFinetuning.from_pretrained('../LM-BFF/result/tmp/', config=config)
tokenizer = AutoTokenizer.from_pretrained('../LM-BFF/result/tmp/')

This required loading the custom class BertForPromptFinetuning from your repo, along with some other configuration, e.g. initialising label_word_list to the list of vocabulary IDs used during training, which in my case was:

However, when attempting inference like this:

# predict
inputs = tokenizer("employment agreement", return_tensors="pt")
model(**inputs)

I get an error relating to prediction_mask_scores: IndexError: index 413 is out of bounds for dimension 1 with size 1

Now, I am unsure if what I attempting is even possible. I am pretty new to using transformers, so I could be doing something silly. But I wondered if you had any advice on using the model in this way.

[gaotianyu1350]

Hi,

The thing you're trying should be able to work. Can you provide more error information for debugging? For example, the full error message.