Using tf-dcm/tf-scm independently

[834810269]

Thank you for your great work!
Regarding causal-rcm, can I independently use tf-dcm or tf-scm to distill my causal model to obtain a good few-step causal model? I already have a TF-based pipeline and a self-forcing-like DMD process built on our model, aiming to improve frame-wise performance. I have tried the causal consistency distillation（looks consistent with the logic of dCM）used in causal-forcing++, but the results were unsatisfactory for framewsie inference. Is this reasonable, and do you expect any improvement with sCM? I would very much like to test your approach.

[zhengkw18]

I would not expect TF-sCM to work if you cannot make TF-dCM work. But probably there are implementation issues for your causal consistency distillation.

You could let the code agent refer to the TF-dCM code in our Causal-rCM, and spot any possible bugs. I believe Causal-rCM's neat infrastructure and algorithm implementations are good references for verifying correctness and finding mismatches.