To get n-best (score) results from decoding with HLG

[jty016]

I am using one_best_decoding function to get the best path of lattice. Can I know what will be the most light way to get the n-best (score) path from the lattice in k2-based decoding scheme?

Because when I tried this with nbest_decoding function, it works. But it consumes much memory and reports OOM when the batch size is large.

What I'd like to try is utilizing k best list for other rescoring algorithm

[csukuangfj]

But it consumes much memory and reports OOM when the batch size is large.

How about decreasing batch size?

[jty016]

I am utilizing k2 decoding scheme in real-time sense, so decreasing batch size means capacity shrinking ... so.

[danpovey]

You could try decreasing the num_paths, and/or decreasing the lattice depth by decreasing the beam or the max_active options.

[jty016]

@danpovey Okay will try. Maybe I have to reduce also T size in N x T x C. Thank you