msvq#

class diffsptk.MultiStageVectorQuantization(order: int, codebook_size: int, n_stage: int, **kwargs)[source]#

See this page for details.

Parameters:

orderint >= 0: The order of the input vector, \(M\).
codebook_sizeint >= 1: The codebook size, \(K\).
n_stageint >= 1: The number of stages (quantizers), \(Q\).
**kwargsadditional keyword arguments: See this page for details.

References

[1]

A. v. d. Oord et al., “Neural discrete representation learning,” Advances in Neural Information Processing Systems, pp. 6309-6318, 2017.

forward(x: Tensor, codebooks: Tensor | None = None, **kwargs) → tuple[Tensor, Tensor, Tensor][source]#

Perform residual vector quantization.

Parameters:

xTensor [shape=(…, M+1)]: The input vectors.
codebooksTensor [shape=(Q, K, M+1)] or None: The external codebook. If None, use the internal codebook.
**kwargsadditional keyword arguments: See this page for details.

Returns:

xqTensor [shape=(…, M+1)]: The quantized vectors.
indicesTensor [shape=(…, Q)]: The codebook indices.
lossesTensor [shape=(Q,)]: The commitment losses.

Examples

>>> x = diffsptk.nrand(4)
>>> x
tensor([-0.5206,  1.0048, -0.3370,  1.3364, -0.2933])
>>> msvq = diffsptk.MultiStageVectorQuantization(4, 3, 2).eval()
>>> xq, indices, _ = msvq(x)
>>> xq
tensor([-0.4561,  0.9835, -0.3787, -0.1488, -0.8025])
>>> indices
tensor([0, 2])