Ability to use cdqa on CPU

The self.device parameter controls wether to use gpu or cpu. To enable cpu only, it seems we need to use the parameter no_cuda of the model.

[andrelmfarias]

I just released the CPU version of the model in both pytorch binary version and joblib version with sklearn wrapper.

[andrelmfarias]

The error occurs when we try to load a model that was trained using apex, using the following

model = load('../models/bert_qa_squad_v1.1_sklearn.joblib')

What I understand is that when we save a model that was trained using apex, it will also save the apex configurations (eg. the use of apex’s BertLayerNorm in the model architecture instead of PyTorch’s standard LayerNorm, as explained on the issue referenced above). So when we load it, it will look for apex configurations, which cannot be found in a machine with no GPU.

After searching for a solution and reading the issues about it, the only solution I can think of for now is to retrain a model without apex and save it.