o i4@s ddlZddlZddlmZddlZddlZddlZddlmZ ddl m Z ddl m Z ddlmZe eje dkZGdd d e eZeejZGd d d eZGd d d eZGdddeZGdddeZGdddeZGdddeZGdddeZdS)N)ABC)parse) ComplexTensor) AbsEnhLoss)Stftz1.9.0cspeZdZdZedefddZedefddZedefddZ edefd d Z dfd d Z Z S)TimeDomainLossz8Base class for all time-domain Enhancement loss modules.returncC|jSN)_nameselfr[/home/ubuntu/.local/lib/python3.10/site-packages/espnet2/enh/loss/criterions/time_domain.pynamezTimeDomainLoss.namecCr r )_only_for_testr rrr only_for_testrzTimeDomainLoss.only_for_testcCr r )_is_noise_lossr rrr is_noise_lossrzTimeDomainLoss.is_noise_losscCr r )_is_dereverb_lossr rrris_dereverb_loss rzTimeDomainLoss.is_dereverb_lossFcs^t||_||_||_|r|rtd|r d|vr |d}|r*d|vr*|d}||_dS)NzF`is_noise_loss` and `is_dereverb_loss` cannot be True at the same timenoise_noisedereverb _dereverb)super__init__rrr ValueErrorr )r rrrr __class__rrr$s    zTimeDomainLoss.__init__)FFF) __name__ __module__ __qualname____doc__propertystrrboolrrrr __classcell__rrrrrsrcsHeZdZdZ     d fdd Zdejdejd ejfd d ZZS) CISDRLossaCI-SDR loss Reference: Convolutive Transfer Function Invariant SDR Training Criteria for Multi-Channel Reverberant Speech Separation; C. Boeddeker et al., 2021; https://arxiv.org/abs/2011.15003 Args: ref: (Batch, samples) inf: (Batch, samples) filter_length (int): a time-invariant filter that allows slight distortion via filtering Returns: loss: (Batch,) NFcs.|durdn|}tj||||d||_dS)N ci_sdr_lossrrr)rr filter_length)r r-rrrrr rrrrQs zCISDRLoss.__init__refinfrcCs2|j|jksJ|j|jftjj||d|jdS)NF)compute_permutationr-)shapeci_sdrptr+r-)r r.r/rrrforwardcs zCISDRLoss.forward)r*NFFF r!r"r#r$rtorchTensorr4r(rrrrr)@sr)csDeZdZeddddffdd Zdejdejdejfdd ZZS) SNRLossNFcs2|durdn|}tj||||dt||_dS)Nsnr_lossr,)rrfloateps)r r;rrrrr rrrrqszSNRLoss.__init__r.r/rc CsR||}dttj|dddj|jdttj|dddj|jd}| S)N)pdim)min)r6log10normclampr;)r r.r/rsnrrrrr4szSNRLoss.forward) r!r"r#EPSrr6r7r4r(rrrrr8ps$r8csPeZdZdZ         d fdd Zdejd ejd ejfd d ZZS)SDRLossaSDR loss. filter_length: int The length of the distortion filter allowed (default: ``512``) use_cg_iter: If provided, an iterative method is used to solve for the distortion filter coefficients instead of direct Gaussian elimination. This can speed up the computation of the metrics in case the filters are long. Using a value of 10 here has been shown to provide good accuracy in most cases and is sufficient when using this loss to train neural separation networks. clamp_db: float clamp the output value in [-clamp_db, clamp_db] zero_mean: bool When set to True, the mean of all signals is subtracted prior. load_diag: If provided, this small value is added to the diagonal coefficients of the system metrices when solving for the filter coefficients. This can help stabilize the metric in the case where some of the reference signals may sometimes be zero r*NTFc sF|durdn|} tj| ||| d||_||_||_||_||_dS)Nsdr_lossr,)rrr- use_cg_iterclamp_db zero_mean load_diag) r r-rIrJrKrLrrrrr rrrrs  zSDRLoss.__init__r.estrc Cs(tj|||j|j|j|j|jdd}|S)aSDR forward. Args: ref: Tensor, (..., n_samples) reference signal est: Tensor (..., n_samples) estimated signal Returns: loss: (...,) the SDR loss (negative sdr) F)rMr.r-rIrKrJrLpairwise) fast_bss_evalrHr-rIrKrJrL)r r.rMrHrrrr4s zSDRLoss.forward) r*NNTNNFFFr5rrrrrGs$rGcsLeZdZdZ       d fdd Zdejdejd ejfd d ZZS) SISNRLossanSI-SNR (or named SI-SDR) loss A more stable SI-SNR loss with clamp from `fast_bss_eval`. Attributes: clamp_db: float clamp the output value in [-clamp_db, clamp_db] zero_mean: bool When set to True, the mean of all signals is subtracted prior. eps: float Deprecated. Kept for compatibility. NTFc sr|durdn|}tj||||d||_||_|dur5td|jdur7t|d| d|_dSdSdS)N si_snr_lossr,z7Eps is deprecated in si_snr loss, set clamp_db instead.r> )rrrJrKloggingwarningmathrB) r rJrKr;rrrrr rrrrs   zSISNRLoss.__init__r.rMrcCs8t|r t|sJ|tj|||j|jdd}|S)aSI-SNR forward. Args: ref: Tensor, (..., n_samples) reference signal est: Tensor (..., n_samples) estimated signal Returns: loss: (...,) the SI-SDR loss (negative si-sdr) F)rMr.rKrJrN)r6 is_tensorrO si_sdr_lossrKrJ)r r.rMsi_snrrrrr4szSISNRLoss.forward)NTNNFFFr5rrrrrPs$rPc6eZdZ    dfdd ZdejfddZZS) TimeDomainMSENFc(|durdn|}tj||||ddS)N TD_MSE_lossr,rrr rrrrr rrrr zTimeDomainMSE.__init__rcCsz|j|jksJ|j|jf||d}|dkr%|jddgd}|S|dkr3|jdd}|Std|j|j)zTime-domain MSE loss forward. Args: ref: (Batch, T) or (Batch, T, C) inf: (Batch, T) or (Batch, T, C) Returns: loss: (Batch,) r=r>r@#Invalid input shape: ref={}, inf={})r1powr@meanrformat)r r.r/mselossrrrr4.s   zTimeDomainMSE.forwardNFFFr!r"r#rr6r7r4r(rrrrrZrZcrY) TimeDomainL1NFcr[)N TD_L1_lossr,r]r^rrrrFr_zTimeDomainL1.__init__rcCsx|j|jksJ|j|jft||}|dkr$|jddgd}|S|dkr2|jdd}|Std|j|j)zTime-domain L1 loss forward. Args: ref: (Batch, T) or (Batch, T, C) inf: (Batch, T) or (Batch, T, C) Returns: loss: (Batch,) r`r>r=rarb)r1absr@rdrre)r r.r/l1lossrrrr4Us   zTimeDomainL1.forwardrgrhrrrrrjErirjcs`eZdZdZdgdddddffdd Zed efd d Zd d Zde j de j fddZ Z S)MultiResL1SpecLossaYMulti-Resolution L1 time-domain + STFT mag loss 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. Attributes: window_sz: (list) list of STFT window sizes. hop_sz: (list, optional) list of hop_sizes, default is each window_sz // 2. eps: (float) stability epsilon time_domain_weight: (float) weight for time domain loss. r*Ng:0yE>g?Fc s|durdn|}tt|j||dtdd|DsJ||_|dur-dd|D|_n||_||_||_tj g|_ t |j|jD]\}} t ||| ddddd} |j | qDdS) Nrk)rcSsg|]}|ddkqS)r=rr.0xrrr sz/MultiResL1SpecLoss.__init__..cSsg|]}|dqS)r=rrorrrrrsTF)n_fft win_length hop_lengthwindowcenter normalizedonesided)rrnrall window_szhop_sztime_domain_weightr;r6nn ModuleList stft_encodersziprappend) r r{r|r;r}rrr whstft_encrrrrs,  zMultiResL1SpecLoss.__init__rcCsdS)Nzl1_timedomain+magspec_lossrr rrrrszMultiResL1SpecLoss.namecCs:trt|d|d}|St|d|d}|S)N).r).r>)is_torch_1_9_plusr6complexrrl)r stftrrr get_magnitudes z MultiResL1SpecLoss.get_magnitudetargetestimatec Cs|j|jksJ|j|jftj||dddtj|dddd|j}tj|||dd}t|jdkr;|St|}|jD]%}|||d}||||d}tj||dd} || 7}qC||j d|j |t|jS) zforward. Args: target: (Batch, T) estimate: (Batch, T) Returns: loss: (Batch,) T)keepdimr=rar)r>r=r>) r1r6sumr;rllenr zeros_likerr}) r rrscaling_factortime_domain_loss spectral_lossr target_mag estimate_magc_lossrrrr4s*    zMultiResL1SpecLoss.forward) r!r"r#r$rr%r&rrr6r7r4r(rrrrrnls"#rn)rSrUabcrr2rOr6packaging.versionrVtorch_complex.tensorr$espnet2.enh.loss.criterions.abs_lossrespnet2.layers.stftr __version__rrfinfoget_default_dtyper;rFr)r8rGrPrZrjrnrrrrs(     ,0MB''