jbcdnr / gretel-path-extrapolation Goto Github PK

I read your paper and I'm impressed about your work.
I'm trying to run the source you share but getting some errors:

RuntimeError: Unknown keyword argument 'fill_value' for operator 'scatter_max'. Schema: scatter_max(Tensor src, Tensor index, int dim=-1, Tensor? out=None, int? dim_size=None) -> ((Tensor, Tensor))

scatter_max function has no argument fill_value but some of your source codes (exp. graph.py line 328 or metrics.py line 196 and 207) uses this argument.

if I remove the fill_value argument then another error occurs:

possible_targets_mask = right_distance_mask & given_as_target.unsqueeze(0)
RuntimeError: result type Byte can't be cast to the desired output type Bool

Is there particular reason to use those or am i missing something.

Thanks.

I am trying to understand how data files in gps folder are produced but I could not figure it out. There are files like "lengths.txt", "observations.txt" and "paths.txt" that are produced in the function write_to_directory. However, the class variables self._lengths, self._indices, self._weights, and self._traversed_edges are needed which I do not see how we can compute them. In addition, I have tried to open hmm_edges.pickle but the error module '__main__' has no attribute 'Leg' is returned. Could you please provide some guidelines or code examples on how the files in the /workspace/gps folder are produced?