By analyzing the code, I found that, during inference, the network is conditioned on the same input batch:
epsilon_theta = net((x, cond, mask, diffusionsteps,)) # predict \epsilon according to \epsilon\theta
Where cond is just the batch without applying any mask. Do I miss something? In a real imputation case scenario, should I replace the missing segment with zeros, similar to applying the mask to the cond variable?
Hi, thank you for sharing the code.
By analyzing the code, I found that, during inference, the network is conditioned on the same input batch:
epsilon_theta = net((x, cond, mask, diffusionsteps,)) # predict \epsilon according to \epsilon\theta
Where cond is just the batch without applying any mask. Do I miss something? In a real imputation case scenario, should I replace the missing segment with zeros, similar to applying the mask to the cond variable?
Thank you in advance for your support.