ineube_separator.py

from collections import OrderedDict
from typing import Dict, List, Optional, Tuple, Union

import torch
from packaging.version import parse as V
from torch_complex.tensor import ComplexTensor

from espnet2.enh.decoder.stft_decoder import STFTDecoder
from espnet2.enh.encoder.stft_encoder import STFTEncoder
from espnet2.enh.layers.beamformer import tik_reg, to_double
from espnet2.enh.layers.tcndenseunet import TCNDenseUNet
from espnet2.enh.separator.abs_separator import AbsSeparator

is_torch_1_9_plus = V(torch.__version__) >= V("1.9.0")


class iNeuBe(AbsSeparator):
    """iNeuBe, iterative neural/beamforming enhancement

    Reference:
    Lu, Y. J., Cornell, S., Chang, X., Zhang, W., Li, C., Ni, Z., ... & Watanabe, S.
    Towards Low-Distortion Multi-Channel Speech Enhancement:
    The ESPNET-Se Submission to the L3DAS22 Challenge. ICASSP 2022 p. 9201-9205.

    NOTES:
    As outlined in the Reference, this model works best when coupled with
    the MultiResL1SpecLoss defined in criterions/time_domain.py.
    The model is trained with variance normalized mixture input and target.
    e.g. with mixture of shape [batch, microphones, samples] you normalize
    it by dividing with torch.std(mixture, (1, 2)). You must do the same
    for the target signal.
    In the Reference, the variance normalization was performed offline
    (we normalized by the std computed on the entire training set and not
    for each input separately). However we found out that also normalizing
    each input and target separately works well.


    Args:
        n_spk: number of output sources/speakers.
        n_fft: stft window size.
        stride: stft stride.
        window: stft window type choose between 'hamming', 'hanning' or None.
        mic_channels: number of microphones channels
            (only fixed-array geometry supported).
        hid_chans: number of channels in the subsampling/upsampling conv layers.
        hid_chans_dense: number of channels in the densenet layers
            (reduce this to reduce VRAM requirements).
        ksz_dense: kernel size in the densenet layers thorough iNeuBe.
        ksz_tcn: kernel size in the TCN submodule.
        tcn_repeats: number of repetitions of blocks in the TCN submodule.
        tcn_blocks: number of blocks in the TCN submodule.
        tcn_channels: number of channels in the TCN submodule.
        activation: activation function to use in the whole iNeuBe model,
            you can use any torch supported activation e.g. 'relu' or 'elu'.
        output_from: output the estimate from 'dnn1', 'mfmcwf' or 'dnn2'.
        n_chunks: number of future and past frames to consider for mfMCWF computation.
        freeze_dnn1: whether or not freezing dnn1 parameters during training of dnn2.
        tik_eps: diagonal loading in the mfMCWF computation.
    """

    def __init__(
        self,
        n_spk=1,
        n_fft=512,
        stride=128,
        window="hann",
        mic_channels=1,
        hid_chans=32,
        hid_chans_dense=32,
        ksz_dense=(3, 3),
        ksz_tcn=3,
        tcn_repeats=4,
        tcn_blocks=7,
        tcn_channels=384,
        activation="elu",
        output_from="dnn1",
        n_chunks=3,
        freeze_dnn1=False,
        tik_eps=1e-8,
    ):
        super().__init__()
        self.n_spk = n_spk
        self.output_from = output_from
        self.n_chunks = n_chunks
        self.freeze_dnn1 = freeze_dnn1
        self.tik_eps = tik_eps
        fft_c_channels = n_fft // 2 + 1
        assert is_torch_1_9_plus, (
            "iNeuBe model requires torch>=1.9.0, "
            "please install latest torch version."
        )

        self.enc = STFTEncoder(n_fft, n_fft, stride, window=window)
        self.dec = STFTDecoder(n_fft, n_fft, stride, window=window)

        self.dnn1 = TCNDenseUNet(
            n_spk,
            fft_c_channels,
            mic_channels,
            hid_chans,
            hid_chans_dense,
            ksz_dense,
            ksz_tcn,
            tcn_repeats,
            tcn_blocks,
            tcn_channels,
            activation=activation,
        )

        self.dnn2 = TCNDenseUNet(
            1,
            fft_c_channels,
            mic_channels + 2,
            hid_chans,
            hid_chans_dense,
            ksz_dense,
            ksz_tcn,
            tcn_repeats,
            tcn_blocks,
            tcn_channels,
            activation=activation,
        )

    @staticmethod
    def unfold(tf_rep, chunk_size):
        """unfolding STFT representation to add context in the mics channel.

        Args:
            mixture (torch.Tensor): 3D tensor (monaural complex STFT)
                of shape [B, T, F] batch, frames, microphones, frequencies.
            n_chunks (int): number of past and future to consider.

        Returns:
            est_unfolded (torch.Tensor): complex 3D tensor STFT with context channel.
                shape now is [B, T, C, F] batch, frames, context, frequencies.
                Basically same shape as a multi-channel STFT with C microphones.

        """
        bsz, freq, _ = tf_rep.shape
        if chunk_size == 0:
            return tf_rep

        est_unfolded = torch.nn.functional.unfold(
            torch.nn.functional.pad(
                tf_rep, (chunk_size, chunk_size), mode="constant"
            ).unsqueeze(-1),
            kernel_size=(2 * chunk_size + 1, 1),
            padding=(0, 0),
            stride=(1, 1),
        )
        n_chunks = est_unfolded.shape[-1]
        est_unfolded = est_unfolded.reshape(bsz, freq, 2 * chunk_size + 1, n_chunks)
        est_unfolded = est_unfolded.transpose(1, 2)
        return est_unfolded

    @staticmethod
    def mfmcwf(mixture, estimate, n_chunks, tik_eps):
        """multi-frame multi-channel wiener filter.

        Args:
            mixture (torch.Tensor): multi-channel STFT complex mixture tensor,
                of shape [B, T, C, F] batch, frames, microphones, frequencies.
            estimate (torch.Tensor): monaural STFT complex estimate
                of target source [B, T, F] batch, frames, frequencies.
            n_chunks (int): number of past and future mfMCWF frames.
                If 0 then standard MCWF.
            tik_eps (float): diagonal loading for matrix inversion in MCWF computation.

        Returns:
            beamformed (torch.Tensor): monaural STFT complex estimate
                of target source after MFMCWF [B, T, F] batch, frames, frequencies.
        """
        mixture = mixture.permute(0, 2, 3, 1)
        estimate = estimate.transpose(-1, -2)

        bsz, mics, _, frames = mixture.shape

        mix_unfolded = iNeuBe.unfold(
            mixture.reshape(bsz * mics, -1, frames), n_chunks
        ).reshape(bsz, mics * (2 * n_chunks + 1), -1, frames)

        mix_unfolded = to_double(mix_unfolded)
        estimate1 = to_double(estimate)
        zeta = torch.einsum("bmft, bft->bmf", mix_unfolded, estimate1.conj())
        scm_mix = torch.einsum("bmft, bnft->bmnf", mix_unfolded, mix_unfolded.conj())
        inv_scm_mix = torch.inverse(
            tik_reg(scm_mix.permute(0, 3, 1, 2), tik_eps)
        ).permute(0, 2, 3, 1)
        bf_vector = torch.einsum("bmnf, bnf->bmf", inv_scm_mix, zeta)

        beamformed = torch.einsum("...mf,...mft->...ft", bf_vector.conj(), mix_unfolded)
        beamformed = beamformed.to(mixture)

        return beamformed.transpose(-1, -2)

    @staticmethod
    def pad2(input_tensor, target_len):
        input_tensor = torch.nn.functional.pad(
            input_tensor, (0, target_len - input_tensor.shape[-1])
        )
        return input_tensor

    def forward(
        self,
        input: Union[torch.Tensor, ComplexTensor],
        ilens: torch.Tensor,
        additional: Optional[Dict] = None,
    ) -> Tuple[List[Union[torch.Tensor, ComplexTensor]], torch.Tensor, OrderedDict]:
        """Forward.

        Args:
            input (torch.Tensor/ComplexTensor): batched multi-channel audio tensor with
                    C audio channels and T samples [B, T, C]
            ilens (torch.Tensor): input lengths [Batch]
            additional (Dict or None): other data, currently unused in this model.

        Returns:
            enhanced (List[Union[torch.Tensor, ComplexTensor]]):
                    [(B, T), ...] list of len n_spk
                    of mono audio tensors with T samples.
            ilens (torch.Tensor): (B,)
            additional (Dict or None): other data, currently unused in this model,
                    we return it also in output.
        """
        # B, T, C
        bsz, mixture_len, mics = input.shape

        mix_stft = self.enc(input, ilens)[0]
        # B, T, C, F
        est_dnn1 = self.dnn1(mix_stft)
        if self.freeze_dnn1:
            est_dnn1 = est_dnn1.detach()
        _, _, frames, freq = est_dnn1.shape
        output_dnn1 = self.dec(
            est_dnn1.reshape(bsz * self.num_spk, frames, freq), ilens
        )[0]
        output_dnn1 = self.pad2(output_dnn1.reshape(bsz, self.num_spk, -1), mixture_len)
        output_dnn1 = [output_dnn1[:, src] for src in range(output_dnn1.shape[1])]
        others = OrderedDict()
        if self.output_from == "dnn1":
            return output_dnn1, ilens, others
        elif self.output_from in ["mfmcwf", "dnn2"]:
            others["dnn1"] = output_dnn1
            est_mfmcwf = iNeuBe.mfmcwf(
                mix_stft,
                est_dnn1.reshape(bsz * self.n_spk, frames, freq),
                self.n_chunks,
                self.tik_eps,
            ).reshape(bsz, self.n_spk, frames, freq)
            output_mfmcwf = self.dec(
                est_mfmcwf.reshape(bsz * self.num_spk, frames, freq), ilens
            )[0]
            output_mfmcwf = self.pad2(
                output_mfmcwf.reshape(bsz, self.num_spk, -1), mixture_len
            )
            if self.output_from == "mfmcwf":
                return (
                    [output_mfmcwf[:, src] for src in range(output_mfmcwf.shape[1])],
                    ilens,
                    others,
                )
            elif self.output_from == "dnn2":
                others["dnn1"] = output_dnn1
                others["beam"] = output_mfmcwf
                est_dnn2 = self.dnn2(
                    torch.cat(
                        (
                            mix_stft.repeat(self.num_spk, 1, 1, 1),
                            est_dnn1.reshape(
                                bsz * self.num_spk, frames, freq
                            ).unsqueeze(2),
                            est_mfmcwf.reshape(
                                bsz * self.num_spk, frames, freq
                            ).unsqueeze(2),
                        ),
                        2,
                    )
                )

                output_dnn2 = self.dec(est_dnn2[:, 0], ilens)[0]
                output_dnn2 = self.pad2(
                    output_dnn2.reshape(bsz, self.num_spk, -1), mixture_len
                )
                return (
                    [output_dnn2[:, src] for src in range(output_dnn2.shape[1])],
                    ilens,
                    others,
                )
            else:
                raise NotImplementedError
        else:
            raise NotImplementedError

    @property
    def num_spk(self):
        return self.n_spk