Beam search validation

[yellowjs0304]

Hi, Is this any way to test your model as beam search?

I saw the code in train.py and agent.py ("beam_valid", "beam_search_test")

Could you please share any way to use it?

[airsplay]

Thanks. Please check this issue for a reference.

[yellowjs0304]

thank u for reply. I'm sorry it's too late..
I already saw that and It seems you already updated the beam search code in github with named "_dijkstra"
I have confusing with the code "_dijkstra" function.
why is this named _dijkstra? i thought the beam search is only rely on agent's logits. but in your code it seems it's related with the dijkstra path.

It seems i haven't quite understood your system.

[airsplay]

why is this named _dijkstra? i thought the beam search is only rely on agent's logits. but in your code it seems it's related with the dijkstra path.

This algorithm still follows the beam-search approach mentioned in the Speaker-Follower paper. I modified it to a greedy approach (Dijkstra algorithm) that only expands the state with the lowest score at each time step. It would empirically have a better estimation of the best path than the beam-search approach without losing speed.

[yellowjs0304]

Okay I understood. If so, is it affected with the agent(listener) model's tokenizer?
Because I tried with my own models which has bert embedding layer, and it returns really low success rate with beam search...
In greedy search it doesn't that so low.
Could you please give me some opinion about this?