NLP Question Answering - How to use another pretrained model?

[geoionescu]

Hello everyone,

Playing with the Question Answering scenario, I noticed that you're using pretrained_Roberta_encoder.tsm file; I guess this is the English model (but cannot tell for sure).

My question is twofold:

1. What is that tsm file (I didn't see any tsm files in the 'usual' models from Hugging Face, for example)
2. How can one produce a tsm file from a RoBERTa model (e.g. the one from here: https://huggingface.co/FacebookAI/xlm-roberta-large)?

[michaelgsharp]

Currently we don't support using another model for these. The main problem is that with each model, you have to have the associated code that exactly matches that model. So if you swapped the model and not the code it would break on you, and having the user swap the code is not something we have in place right now.

it is something we have talked about for the long run, but currently we have no concrete plans around that scenario.

I wish PyTorch/TorchSharp had an easier way of sharing the models that didn't require the direct python/c# code.