mnoukhov / ecn Goto Github PK

one possible way to increase speed is by creating a dataset with utilities, item pools, and termination lengths (N) then load it with torch.data
would also be good for direct reproducibility

currently agents do masking communication in a multi-step process but this isn't fully right given that, in the real world, it doesn't make sense to later mask something you've previously unmasked.

perhaps masking communication should be additive (i.e. you should only be able to unmask items you previously masked)

I think you can greatly improve speed here

matches_argmax over time is just measuring the highest probability in a roundabout way, a better measure of the distribution is the entropy

as per https://arxiv.org/pdf/1706.05125.pdf:

• 10 turns total (5 turns each)
• communication at every step
• simultaneous proposal after 10 turns
• reward = 0 if proposals are not the same

replace type_constr with to(device)
remove Variable for requires_grad=True
change Tensor(...)._fill(0) to torch.zeros

change dataset generation to try and take advantage of caching to speed things up