Closed XZWY closed 1 year ago
Hi,have you figure out the causality of the model,whether for the reason that in the diffusion iteration(T),the model utilize all the frames? Thanks
Hi there, you are correct that all models in this repo are currently noncausal. However, it can easily be made causal by just making the underlying DNN causal, as everything else (the SDE and sampler) is already causal by construction. Details are described in our accepted ICASSP 2023 paper https://arxiv.org/abs/2303.08674, which scored Top 5 in the ICASSP 2023 Signal Improvement Challenge. There we implemented it with a causal model (but for different corruptions). There we also included a causal AGC to deal with input and output normalization without accessing all samples at once.
Hi there, you are correct that all models in this repo are currently noncausal. However, it can easily be made causal by just making the underlying DNN causal, as everything else (the SDE and sampler) is already causal by construction. Details are described in our accepted ICASSP 2023 paper https://arxiv.org/abs/2303.08674, which scored Top 5 in the ICASSP 2023 Signal Improvement Challenge. There we implemented it with a causal model (but for different corruptions). There we also included a causal AGC to deal with input and output normalization without accessing all samples at once.
Thanks for you reply, nice to read you paper mentioned above. And have you tested the difference between causal and uncausal backbone(total diffusion process)'s performance, may you add some indicator's value?
Hi, just want to make sure, all the backbone models used here are not causal and uses all the future frames right? Thanks!