FQF logit computation about di-engine

[PaParaZz1]

I think we have already implemented the operation you mentioned line 763-768 (link), i.e.:

        q_quantiles_hats = (q_quantiles[:, 1:] + q_quantiles[:, :-1]).detach() / 2.  # (batch_size, num_quantiles)

        # NOTE(rjy): reparameterize q_quantiles_hats
        q_quantile_net = self.quantile_net(q_quantiles_hats)  # [batch_size, num_quantiles, hidden_size(64)]
        # x.view[batch_size, 1, hidden_size(64)]
        q_x = (x.view(batch_size, 1, -1) * q_quantile_net)  # [batch_size, num_quantiles, hidden_size(64)]

Could you please confirm again? If you have other problems, you can continue to reply in this issue.

[dmartinezbaselga]

Hi,

Thanks for the response. What you are referring is the computation of $\hat{\tau_i}=\frac{\tau_{i+1}-\tau{i}}{2}$ (line 763) and the embedding computation of $\hat{\tau}$ (764-768), which is in the paper in section 3.4:

These refer to the quantile values computation:

The part that I am missing is the $\tau_{i+1}-\tau_{i}$ that measures the width of the quantile fractions of:

Instead of a mean, it's a weighted mean.

[PaParaZz1]

Thanks for your feedback, I have checked this part of the original paper and our implementation.

• On the one hand, the operation you mentioned is implemented in line 773 (link). This variable is only used for the computation of target value.
• On the other hand, the variable logit is only used to select action in collect/eval mode, it is not used in training (learn mode).