Linear assignment layer training

[PhanTask]

Hi, thanks so much for open sourcing this solid work! I try to understand how the linear assignment layer is trained and would appreciate your help.

In the class AttentionalGNN, the linear assignment score is computed and processed as this:

        scores = log_optimal_transport(scores.log_softmax(dim=-2), self.bin_score, iters=5)[:, :-1, :-1].view(unreachable.shape)
        score_min = scores.min() - scores.max()
        scores = scores + (score_min - 40) * invalid.float() + (score_min - 20) * unreachable.float()

        return scores * 15

Could you please explain a bit more about this code? I am not pretty sure how to interpret invalid.float(), unreachable.float() and numbers used here (e.g., 40, 20, 15).

Also I try to use such a differentiable linear assignment layer in my work, where I do the bipartite matching to match the same robots across different frames (t0 <-> t1, t1 <-> t2, t2 <-> t3, ...) to get matching traces, and use a discriminator to evaluate the overall quality of the traces and get a loss. Overall it is more like a GAN structure. How should I use this loss to update the linear assignment layer? Thanks!

[illusive-chase]

As scores represents "log probability of frontiers being sampled", we use a bias like (score_min - 40) * invalid.float() to make it exp(40)=2E+17 times less likely that these INVALID frontiers will be sampled.

Similarly, we add (score_min - 20) * unreachable.float(). Here we use 20 instead of 40 because we expect that if all frontiers are either INVALID or UNREACHABLE, we would rather sample the UNREACHABLE ones than the INVALID ones.

And we think that the linear assignment layer is non-learning and does not need to be trained. It is considered as a differentiable version of the Hungarian algorithm. We use GNN to evaluate the cost between nodes and use the linear assignment layer to do assignment based on the cost.

Hope that helps :)

[JiayunjieJYJ]

As scores represents "log probability of frontiers being sampled", we use a bias like (score_min - 40) * invalid.float() to make it exp(40)=2E+17 times less likely that these INVALID frontiers will be sampled.

Similarly, we add (score_min - 20) * unreachable.float(). Here we use 20 instead of 40 because we expect that if all frontiers are either INVALID or UNREACHABLE, we would rather sample the UNREACHABLE ones than the INVALID ones.

And we think that the linear assignment layer is non-learning and does not need to be trained. It is considered as a differentiable version of the Hungarian algorithm. We use GNN to evaluate the cost between nodes and use the linear assignment layer to do assignment based on the cost.

Hope that helps :)

Could you please further explain the meaning of :
return scores * 15

And how this matrix influence the final output action?

Thank you very much!

[illusive-chase]

Similarly, scores * 15 is to sharpen the distribution of output actions.

This code may help you understand better.

>>> import torch
>>> C = torch.distributions.Categorical
>>> torch.randn(10)
tensor([ 0.2402,  0.9646, -0.0617,  0.5511,  0.1720, -0.5559, -0.3629,  0.3931,
        -0.7448, -0.2066])
>>> a=torch.randn(10)
>>> a
tensor([-0.9821,  2.3382, -0.2364, -1.0046, -1.2925,  0.1279,  0.4886, -1.5084,
        -0.8465,  0.5353])
>>> C(logits=a).probs
tensor([0.0217, 0.5993, 0.0457, 0.0212, 0.0159, 0.0657, 0.0943, 0.0128, 0.0248,
        0.0988])
>>> C(logits=a*15).probs
tensor([2.3463e-22, 1.0000e+00, 1.6899e-17, 1.6732e-22, 2.2271e-24, 3.9920e-15,
        8.9233e-13, 8.7412e-26, 1.7918e-21, 1.8004e-12])