Question about WTConv wavelet filters in the official pretrained checkpoints

[sincosx]

Thanks for the interesting work.
I have a question regarding the official pretrained checkpoints.

From the current source code, my understanding is that WTConv1d uses fixed db1/Haar-style wavelet filters (wt_filter / iwt_filter), shared consistently across channels. I also checked a fresh model (without loading any checkpoint), and this is indeed the case.

However, after loading the official checkpoints synth_only.pth and synth_and_ucr.pth, I found that in:

encoder.layer2.layer.0
encoder.layer3.layer.0
encoder.layer4.layer.0

the wt_filter / iwt_filter are no longer identical across channels.

For example, in encoder.layer2.layer.0.wt_filter, the first few rows in the checkpoint are:

[0] [0.8490355610847473, 0.7476450800895691]
[1] [0.6930041313171387, -0.6657928228378296]
[2] [0.6884918212890625, 0.7316277027130127]
[3] [0.380836546421051, -0.946887731552124]

If the current fresh initialization logic is followed, then [0] and [2] should be identical (low filters), and [1] and [3] should also be identical (high filters), but they are not.

I checked the raw state_dict directly, and these tensors are already like this before load_state_dict, so this seems to come from the checkpoint contents themselves.

Could you please clarify whether this is expected?
Are the official checkpoints intentionally using channel-wise different wavelet filters, or might there be an issue with the uploaded .pth files?

Thanks.