loading xmlr from local without using torch.hub.load

[ljj7975]

Is there any way to load XLMRModel class from fairseq with Pytorch 1.3?

I cannot use torch.hub.load because I do not have any internet connection In order to bypass the downloading, I have the model on local but I cannot figure out how I can initialize the model and load the model.pt.

For PyTorch 1.0, the example in the repo has

from fairseq.models.roberta import XLMRModel
xlmr = XLMRModel.from_pretrained('/path/to/xlmr.large.v0', checkpoint_file='model.pt')
xlmr.eval()  # disable dropout (or leave in train mode to finetune)

but it seems like it is no longer availble for PyTorch 1.3

Can anyone help me? Thanks

[myleott]

The XLMRModel.from_pretrained version should work in PyTorch 1.3 as well. Did you try it?

[ljj7975]

yes, I was getting following error so I thought it was no longer supported

>>> from fairseq.models.roberta import XLMRModel
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: cannot import name 'XLMRModel' from 'fairseq.models.roberta' (/home/xxx/anaconda3/envs/xxx/lib/python3.7/site-packages/fairseq/models/roberta/__init__.py)

[myleott]

Ah, did you install fairseq via pip? The pip package is quite outdated so doesn't have XLMRModel in it yet.

You need to instead clone fairseq and pip install --editable .. See "Installing from source" here: https://github.com/pytorch/fairseq#requirements-and-installation

[ljj7975]

confirmed it works. Thanks!