o si'@sddlZddlmZGdddeZGdddeZGdddeZed Zed Zed Zed Z ed Z ed Z ed Z ed Z ed ZdS) N)_Lossc*eZdZdZdfdd ZddZZS) PairwiseNegSDRaBase class for pairwise negative SI-SDR, SD-SDR and SNR on a batch. Args: sdr_type (str): choose between ``snr`` for plain SNR, ``sisdr`` for SI-SDR and ``sdsdr`` for SD-SDR [1]. zero_mean (bool, optional): by default it zero mean the target and estimate before computing the loss. take_log (bool, optional): by default the log10 of sdr is returned. Shape: - est_targets : :math:`(batch, nsrc, ...)`. - targets: :math:`(batch, nsrc, ...)`. Returns: :class:`torch.Tensor`: with shape :math:`(batch, nsrc, nsrc)`. Pairwise losses. Examples >>> import torch >>> from asteroid.losses import PITLossWrapper >>> targets = torch.randn(10, 2, 32000) >>> est_targets = torch.randn(10, 2, 32000) >>> loss_func = PITLossWrapper(PairwiseNegSDR("sisdr"), >>> pit_from='pairwise') >>> loss = loss_func(est_targets, targets) References [1] Le Roux, Jonathan, et al. "SDR half-baked or well done." IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2019. T:0yE>cs6tt||dvs J||_||_||_||_dS)Nsnrsisdrsdsdr)superr__init__sdr_type zero_meantake_logEPSselfr r rr __class__G/home/ubuntu/.local/lib/python3.10/site-packages/asteroid/losses/sdr.pyr %s   zPairwiseNegSDR.__init__c Cs`||ks |jdkrtd|d|d||ks&J|jrAtj|ddd}tj|ddd}||}||}tj|dd }tj|dd }|jd vrrtj||ddd}tj|dddd|j }|||} n | d|j ddd} |jd vr||} n|| } tj| ddd tj| ddd |j } |j rd t | |j } | S) N2Inputs must be of shape [batch, n_src, time], got  and  insteadTdimkeepdimrrr r r )sizendim TypeErrorr torchmean unsqueezer sumrrepeatshaperlog10) r est_targetstargets mean_source mean_estimates_target s_estimate pair_wise_dots_target_energypair_wise_proje_noise pair_wise_sdrrrrforward-s4   zPairwiseNegSDR.forwardTTr__name__ __module__ __qualname____doc__r r8 __classcell__rrrrrsrcs*eZdZdZd fdd ZddZZS) SingleSrcNegSDRaBase class for single-source negative SI-SDR, SD-SDR and SNR. Args: sdr_type (str): choose between ``snr`` for plain SNR, ``sisdr`` for SI-SDR and ``sdsdr`` for SD-SDR [1]. zero_mean (bool, optional): by default it zero mean the target and estimate before computing the loss. take_log (bool, optional): by default the log10 of sdr is returned. reduction (string, optional): Specifies the reduction to apply to the output: ``'none'`` | ``'mean'``. ``'none'``: no reduction will be applied, ``'mean'``: the sum of the output will be divided by the number of elements in the output. Shape: - est_targets : :math:`(batch, time)`. - targets: :math:`(batch, time)`. Returns: :class:`torch.Tensor`: with shape :math:`(batch)` if ``reduction='none'`` else [] scalar if ``reduction='mean'``. Examples >>> import torch >>> from asteroid.losses import PITLossWrapper >>> targets = torch.randn(10, 2, 32000) >>> est_targets = torch.randn(10, 2, 32000) >>> loss_func = PITLossWrapper(SingleSrcNegSDR("sisdr"), >>> pit_from='pw_pt') >>> loss = loss_func(est_targets, targets) References [1] Le Roux, Jonathan, et al. "SDR half-baked or well done." IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2019. TnonercsF|dksJttj|d|dvsJ||_||_||_d|_dS)Nr)) reductionrr)NotImplementedErrorr r r r rr)rr r rrBrrrrr zs  zSingleSrcNegSDR.__init__c Cs:||ks |jdkrtd|d|d|jr7tj|ddd}tj|ddd}||}||}|jdvrZtj||ddd}tj|dddd|j}|||}n|}|jd vrf||}n||}tj|ddd tj|ddd |j} |j rd t | |j} |j d kr| } | S| } | S) Nrz+Inputs must be of shape [batch, time], got rrrTrr r!rr"r') r#r$r%r r&r'r r)rrr,rB) r est_targettargetr/r0dotr4 scaled_targetr6lossesrrrr8s0   *zSingleSrcNegSDR.forward)TTrArr:rrrrr@Ts% r@cr) MultiSrcNegSDRaBase class for computing negative SI-SDR, SD-SDR and SNR for a given permutation of source and their estimates. Args: sdr_type (str): choose between ``snr`` for plain SNR, ``sisdr`` for SI-SDR and ``sdsdr`` for SD-SDR [1]. zero_mean (bool, optional): by default it zero mean the target and estimate before computing the loss. take_log (bool, optional): by default the log10 of sdr is returned. Shape: - est_targets : :math:`(batch, nsrc, time)`. - targets: :math:`(batch, nsrc, time)`. Returns: :class:`torch.Tensor`: with shape :math:`(batch)` if ``reduction='none'`` else [] scalar if ``reduction='mean'``. Examples >>> import torch >>> from asteroid.losses import PITLossWrapper >>> targets = torch.randn(10, 2, 32000) >>> est_targets = torch.randn(10, 2, 32000) >>> loss_func = PITLossWrapper(MultiSrcNegSDR("sisdr"), >>> pit_from='perm_avg') >>> loss = loss_func(est_targets, targets) References [1] Le Roux, Jonathan, et al. "SDR half-baked or well done." IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2019. Trcs2t|dvs J||_||_||_d|_dS)Nrr)r r r r rrrrrrr s   zMultiSrcNegSDR.__init__c Cs(||ks |jdkrtd|d|d|jr7tj|ddd}tj|ddd}||}||}|jdvrZtj||ddd}tj|dddd|j}|||}n|}|jd vrf||}n||}tj|ddd tj|ddd |j} |j rd t | |j} tj| d d  S) NrrrrrTrr r!rr") r#r$r%r r&r'r r)rrr,) rr-r.r/r0r3r4scaled_targetsr6r7rrrr8s.   zMultiSrcNegSDR.forwardr9r:rrrrrIs" rIrr r)r&torch.nn.modules.lossrrr@rIpairwise_neg_sisdrpairwise_neg_sdsdrpairwise_neg_snrsinglesrc_neg_sisdrsinglesrc_neg_sdsdrsinglesrc_neg_snrmultisrc_neg_sisdrmultisrc_neg_sdsdrmultisrc_neg_snrrrrrs ORO