k-best propable paths are same when specify mask parameter

[wangjunji]

I am dealing with variable sequence length. So need to mask padding tokens. But the _viteribi_decode_nbest function produces same paths when the mask contain zeros. Here is the snippet to reproduce

import torch
from torchcrf import CRF

num_tags = 5  # number of tags is 5
model = CRF(num_tags)
seq_length = 3  # maximum sequence length in a batch
batch_size = 2  # number of samples in the batch
emissions = torch.randn(seq_length, batch_size, num_tags)

# Computing log likelihood
tags = torch.tensor([[2, 3], [1, 0], [3, 4]], dtype=torch.long)  # (seq_length, batch_size)
model(emissions, tags)

mask = torch.tensor([[1, 1], [1, 1], [0, 1]])
# Decoding
print(model.decode(emissions, mask=mask))  # decoding the best path
print(model.decode(emissions, mask=mask, nbest=3))  # decoding the top 3 paths

result:

tensor([[[1, 0, 0],
         [3, 4, 0]]])
tensor([[[1, 0, 0],
         [3, 4, 0]],

        [[1, 0, 0],
         [3, 4, 2]],

        [[1, 0, 0],
         [3, 4, 1]]])