ENH MuReNNDirect.to_conv1d()

[xir4n]

What's Changed:

1. Added a new feature to_conv1d as described in MuReNNDirect.to_conv1d() #45.
2. MuReNNDirect now supports a variable number of Conv1d filters per scale. To specify the number of filters for each scale, use a list or dictionary instead of an integer.
3. The Conv1d kernel sizes in MuReNNDirect are no longer equal to the T; they are now equal to T * Q[j] where Q[j] is the number of filters per octave at octave j.
4. Added J_phi to specify the number of levels of downsampling in MuReNNDirect.
5. The output tensor shape of MuReNNDirect has been changed from a 5-dimensional tensor with shape (B, C, Q, J, T) to a 4-dimensional tensor with shape (B, C, sum(Q_j), T//2**J_phi). This is a breaking change.
6. alternate_gh is now false

[lostanlen]

Another breaking change is that MuReNNDirect now uses alternate_gh=False. @xir4n says it's better-conditioned in terms of frame bounds, particularly when the input length is not a power of two.

[lostanlen]

👏🏻 👏🏻 👏🏻 👏🏻 👏🏻

[lostanlen]

i think supporting dict is not necessary. list is sufficient

[3, 4, 5, 2] is more explicit than {0: 3, 1: 3, 3: 5, 4: 2}

[xir4n]

remove properties, conflict revolved

[lostanlen]

Excellent work.

This will require detailed release notes !