This is the version of the code that reproduces the result in the paper.
With this version you can use the pre-trained weights from Zenodo. After upgrading to PyTorch 1.1.0, there was a refactor of the code. Specifically, the layers/modules existing in the code for the paper merged to one module. This has as an effect the change in the names of the layers. Therefore, the pre-trained weights cannot be used as is.
Furthermore, after manually substituting the pre-trained weights to the new system, it seems that the results of the paper cannot be used. I hypothesise that this happens because at the code for the paper the GRUCell was used but in the new code the GRU is used.
Thus, if you want to use the pre-trained weights then you have to get this release of the code.
I hope it helps!