Is it possible to convert the GloVe Word Embedding from torchtext to a torchscript model?

1

I am trying to utilize the GloVe word embedding for my RNN in DJL however after looking around one of the developers said “DJL uses PyTorch C++ API, it only support TorchScript model. You need trace your model.” I have a feeling when the dev mentions model in this context I think he is talking about a saved nn model after training and not a word embedding. Is it possible to take the glove.6B.50d.txt.pt word embedding file that is downloaded to my file system and convert it to TorchScript?

I got the word embedding by using the line:

torchtext.vocab.GloVe(name='6B', dim=50)
2

The word embedding sound be a tensor and you can directly load it in the C++ API. Tracing is used for models (or general function calls) as you’ve described.

3

I apologize I am not trying to get someone to do my work its just that some of this is going over my head.

So if I load the glove.6B.50d.txt.pt file and load it into the C++ API I should be able to trace it and get the torchscript model?

4

Of course and I did not have the impression you are trying to offload your work.

The glove embedding is a tensor, not a model. It just contains the data and you would have to create the model separately to be able to trace it.

5

Oooh ok. Ok then that is my next bit of research to do. Thanks!