[Question] Can not import BERT model because config lacks Task field

[Gelassen]

What is your question?

I need Chinese BERT-like classifier to split Chinese sentence into separate words on the mobile client. Mobile client allows to use custom models, but only in the tensorflow lite format. It is possible to convert pytorch models into tensorflow lite format, but I have to know input shape of the model.

I wrote a simple python script to do this, but faced with several issues I haven't solved yet.

How can I get shape for SequenceLabelingModel ? Seems SequenceLabelingModel is used for this babert model

What have you tried?

1. Load model as a plain pytorch model and get its parameters:

def loadModel():
    model = torch.load("/home/gelassen/Downloads/chinese_babert-base/pytorch_model.bin")
    model_shape = list(model.parameters())[0].shape 
    print(model_shape)
    print("Model shape" + model_shape)

loadModel()

2. Load model via Model.from_config() factory method:

def buildAdaseqModel() -> nn.Module:
    return Model.from_config("/home/gelassen/Downloads/chinese_babert-base/config.json")

Code (if necessary)

No response

What's your environment?

• AdaSeq Version (master):
• ModelScope Version (master):
• PyTorch Version (default one, supplied with adaseq):
• OS (e.g., Ubuntu 22.04):
• Python version: Python 3.10.12
• CUDA/cuDNN version: Driver Version: 470.199.02 CUDA Version: 11.4
• GPU models and configuration: NVIDIA GeForce GTX 860M
• Any other relevant information:

Code of Conduct

• I agree to follow this project's Code of Conduct