Older pytorch-transformer versions error

[JJMinton]

With pytorch-transformer==1.0.0 fails to run the evaluate command with the following error.

  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.8/site-packages/nubia_score/nubia.py", line 190, in score
    nubia, gpt_ref = self.nubia(ref, hyp, get_features=True, six_dim=six_dim,
  File "/usr/local/lib/python3.8/site-packages/nubia_score/nubia.py", line 128, in nubia
    gpt_ref = self._gpt_score(ref)
  File "/usr/local/lib/python3.8/site-packages/nubia_score/nubia.py", line 113, in _gpt_score
    tensor_input = torch.tensor([[self.tokenizer. eos_token_id] +
AttributeError: 'GPT2Tokenizer' object has no attribute 'eos_token_id'

Upgrading to pytorch-transformer==1.2.0 solves this so the requirements in setup.py should be constrained to avoid this.