Closed philgzl closed 2 days ago
Thanks for pointing out the issue. @Emrys365, can you check this issue?
Hi @philgzl, to make the model fully causal, you will also need to set the norm_type
to either "cLN" or "cfLN" in addition to causal=True
.
Thanks @Emrys365. Can you confirm that the configurations marked as causal in your paper have this modification? The norm_type
option does not seem to be specified in the configuration files provided in Emrys365/se-scaling.
No, the paper only employed the causal=True
argument at that time (as we overlooked the normalization part unfortunately), so the models are not strictly causal (all operations but the normalization layers are causal). The conclusions should be similar though. But the performance of a truly causal model should be slightly worse than reported.
Cool, thanks for clarifying!
It seems to me that BSRNN with
causal=True
is not causal. Forward passing a tensor with NaNs in the last time frame results in a tensor with NaNs only.The first non-causal operation seems to be the normalization layers in the band-split module, which calculate statistics over time frames.