o iN]@slddlZddlmZmZddlZddlmZddlm Z m Z ddl m Z m Z mZmZddlmZmZ   d2d ejd ejd eed eedeedeedeejdffddZ  d3dejdejdeed eedeejejff ddZ d4dejdejdeedeejejffddZ      d5d ejd ejdeedeed eedeedeed eedejfdd Z     d6d ejd ejdeedeed eedeedeedejfd!d"Z     d6d ejd ejdeedeed eedeedeedejfd#d$Z       d7d ejd ejdeedeed eedeedeed%eed&eedejfd'd(Z    d8d ejd ejdeedeed eed eedeedeejejffd)d*Z d4d+ejd,ejdeedeejejejffd-d.Z     d9d ejd ejdeedeed eedeed/eedeedejfd0d1Z dS):N)OptionalTuple)fft)!block_toeplitz_conjugate_gradienttoeplitz_conjugate_gradient)_coherence_to_neg_sdr _normalize _remove_mean_solve_permutation)toeplitzblock_toeplitzFTrefest zero_meanpairwise load_diag with_coh_sarreturn.c Cs|rt|dd}t|dd}t|dd}t|dd}|s|r%td||}|r*|}ntd||}t|}|rrtd||}|durO|t|jd|}tj||} td|| } |rnt || ddddf\}} || fS|S)zJ Special case of computing cosine metrics with filter length of 1 axisz...cn,...dn->...cdz...n,...n->...N...lc,...lc->...c.) r r npeinsumsquareeyeshapelinalgsolvebroadcast_arrays) rrrrrrxcorrcoh_sdracmsolcoh_sarr&O/home/ubuntu/.local/lib/python3.10/site-packages/fast_bss_eval/numpy/metrics.py'square_cosine_metrics_length_one_filter%s*     r(xylengthc Cs|dur |jd}t|jd|jd}dtt|jd|d}tj||dd}tj||dd}tj|jd|j d|d}|rjt d| |} tj| |dd} |d d|f| d d|ddffS| |} tj| |dd} |d d|f| d d|ffS) a Compute the auto correlation function of x and its cross-correlation with y. This function is specialized for when only the SDR is needed. In this case, only the auto-correlation of each source is required. Parameters ---------- x: numpy.ndarray, (..., n_chan_x, n_samples) Usually the reference signals y: numpy.ndarray, (..., n_chan_y, n_samples) Usually the estimated signals length: int, optional The length at which to truncate the statistics pairwise: bool, optional When this flag is true, statistics are computed for all pairs of source/estimate. This only affects the cross-correlation. Returns ------- acf: numpy.ndarray, (..., n_chan_x, length) The auto-correlation functions of x xcorr: numpy.ndarray The cross-correlation functions of x and y. The shape is (..., n_chan_x, length, n_chan_y) when pairwise is True, and (..., n_chan_x, length) when pairwise is False. Nrrnr)r.z...cn,...dn->...cnd.) rmaxmathceillog2rrfftirfftrealimagrrconj) r)r*r+rmax_lenn_fftXYacfXYr!r&r&r' compute_statsQs "&  r?c Cs|dur |jd}t|jd|jd}dtt|jd|d}tj||dd}tj||dd}td|| }tj ||dd}tj |dd|ddddf|d| dddddfgdd }td| |} tj | |dd} | dd|ddddf} || fS) a Compute the auto correlation function of x and its cross-correlation with y. This function computes the full auto-correlation matrix for all reference signals. This is only needed when both the SDR and the SIR have to be computed. Parameters ---------- x: np.ndarray, (..., n_chan_x, n_samples) Usually the reference signals y: np.ndarray, (..., n_chan_y, n_samples) Usually the estimated signals length: int, optional The length at which to truncate the statistics Returns ------- acf: (..., length, n_chan_x, n_chan_y) The auto-correlation functions of x xcorr: np.ndarray, (..., length, n_chan_x, n_chan_y) The cross-correlation functions of y Nrr,rr-z...cn,...dn->...ncd.r) rr0r1r2r3rr4rrr8r5 concatenate) r)r*r+r9r:r;r<prodr=r>r!r&r&r'compute_stats_2s "DrC filter_length use_cg_iterclamp_dbcCs|dkrt|||d|dd}nT|rt|}t|}t|dd}t|dd}t||||d\} } |dur;| d|7<|durKt| } tj| | } nt| | |d } |r\t d | | }nt d | | }t ||d } | S) a Compute the negative signal-to-distortion ratio (SDR). The order of est/ref follows the convention of pytorch loss functions (i.e., est first). Parameters ---------- est: np.ndarray (..., n_channels_est, n_samples) The estimated signals ref: np.ndarray (..., n_channels_ref, n_samples) The groundtruth reference signals filter_length: int, optional The length of the distortion filter allowed (default: ``512``) use_cg_iter: int, optional 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. zero_mean: bool, optional When set to True, the mean of all signals is subtracted prior to computation of the metrics (default: ``False``) clamp_db: bool, optional If provided, the resulting metrics are clamped to be in the range ``[-clamp_db, clamp_db]`` pairwise: bool, optional If set to True, the metrics are computed for every est/ref signals pair (default: False). When set to False, it is expected that ``n_channels_ref == n_channels_est``. load_diag: float, optional If provided, this small value is added to the diagonal coefficients of the system metrics 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 Returns ------- neg_cisdr: np.ndarray The negative SDR of the input signal. The returned tensor has shape (..., n_channels_est) if ``pairwise == False`` and (..., n_channels_ref, n_channels_est) if ``pairwise == True``. rNF)rrrrrr)r+r).rn_iterr...l,...l->...rG) r(r r r?r rrrrrr)rrrErFrrGrrcohr=r!R_matr$ neg_cisdrr&r&r'sdr_losss45    rOc Cst|||||||ddd S)NTF)rErFrrGr change_sign return_perm)sdrrrrErFrrGrr&r&r' sdr_pit_losss rTc Cst|||||||ddS)a Compute the negative signal-to-distortion ratio (SDR). This function computes the SDR for all pairs of est/ref signals. The order of est/ref follows the convention of pytorch loss functions (i.e., est first). Parameters ---------- est: np.ndarray (..., n_channels_est, n_samples) The estimated signals ref: np.ndarray (..., n_channels_ref, n_samples) The groundtruth reference signals filter_length: int, optional The length of the distortion filter allowed (default: ``512``) use_cg_iter: int, optional 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. zero_mean: bool, optional When set to True, the mean of all signals is subtracted prior to computation of the metrics (default: ``False``) clamp_db: bool, optional If provided, the resulting metrics are clamped to be in the range ``[-clamp_db, clamp_db]`` load_diag: float, optional If provided, this small value is added to the diagonal coefficients of the system metrics 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 Returns ------- neg_cisdr: np.ndarray, (..., n_channels_ref, n_channels_est) The negative SDR of the input signal T)rErFrrGrr)rOrSr&r&r'pairwise_sdr_loss5s0rUrQrPc CsFt||||d|||d} t| dd\} } |r| n| } |r!| | fS| S)ah Compute the signal-to-distortion ratio (SDR). This function computes the SDR for all pairs of est/ref signals and finds the permutation maximizing the SDR. The order of ref/est follows the convention of bss_eval (i.e., ref first). Parameters ---------- ref: np.ndarray (..., n_channels_est, n_samples) The estimated signals est: np.ndarray (..., n_channels_ref, n_samples) The groundtruth reference signals filter_length: int, optional The length of the distortion filter allowed (default: ``512``) use_cg_iter: int, optional 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. zero_mean: bool, optional When set to True, the mean of all signals is subtracted prior to computation of the metrics (default: ``False``) clamp_db: bool, optional If provided, the resulting metrics are clamped to be in the range ``[-clamp_db, clamp_db]`` load_diag: float, optional If provided, this small value is added to the diagonal coefficients of the system metrics 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 return_perm: bool, optional If set to True, the optimal permutation of the estimated signals is also returned (default: ``False``) change_sign: bool, optional If set to True, the sign is flipped and the negative SDR is returned (default: ``False``) Returns ------- cisdr: np.ndarray, (..., n_channels_est) The SDR of the input signal perm: np.ndarray, (..., n_channels_est), optional The index of the corresponding reference signal. Only returned if ``return_perm == True`` T)rErFrrrGrrQ)rOr ) rrrErFrrGrrQrPneg_sdrpermrRr&r&r'rRqs< rRcCs|dkrt|||||ddS|rt|}t|}t|dd}t|dd}t|||d\}}tt|jd} |durF|d d | | f|7<|d | | f} | dd} | d d|f} |rd|d d} n|d dd| | fdd} |durt| } t j | | } nt | | |d } |rt d | | }nt d| | }||jdd d|jdd}|durt|} t j | |} n(|r| }| d d}||jdd d|jdd}nd}t||||d} t d || }|rt ||d dddf\}}||fS)a  This function computes the square cosines of the angles between 1. the estimate and the subspaces of shifts of each of the reference signals 2. the estimate and the subspace of shifts of all reference signals The order of ref/est follows the convention of bss_eval (i.e., ref first). Parameters ---------- ref: np.ndarray (..., n_channels_est, n_samples) The estimated signals est: np.ndarray (..., n_channels_ref, n_samples) The groundtruth reference signals filter_length: int, optional The length of the distortion filter allowed (default: ``512``) use_cg_iter: int, optional 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. zero_mean: bool, optional When set to True, the mean of all signals is subtracted prior to computation of the metrics (default: ``False``) pairwise: bool, optional When this flag is true, statistics are computed for all pairs of source/estimate. This only affects the cross-correlation. load_diag: float, optional If provided, this small value is added to the diagonal coefficients of the system metrics 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 Returns ------- neg_cisdr: np.ndarray, (..., n_channels_ref, n_channels_est) The negative SDR of the input signal rT)rrrrrr)r+r/N.rr@rHrrJ)r)rIr))r(r r rClistrangerswapaxesr rrrrrreshaper rr )rrrErFrrrr=r! diag_indicesacf_sdr xcorr_sdrrMr$r"x0r%r&r&r'square_cosine_metricssZ3    & (rar"r%cCs6||}t||d}t||d}t||d}|||fS)zP Helper function that converts the square cosine metrics to SDR/SIR/SAR rK)r)r"r%rGcoh_sirrWneg_sirneg_sarr&r&r'_base_metrics_bss_evalAs     recompute_permutationc Cs|s|jd|jdksJdt|||||||d\}} t|| |d\} } } |r>t| | | dd\} } } } | | | | fS| | | fS)av This function computes the SDR, SIR, and SAR for the input reference and estimated signals. The order of ref/est follows the convention of bss_eval (i.e., ref first). Parameters ---------- ref: np.ndarray (..., n_channels_est, n_samples) The estimated signals est: np.ndarray (..., n_channels_ref, n_samples) The groundtruth reference signals filter_length: int, optional The length of the distortion filter allowed (default: ``512``) use_cg_iter: int, optional 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. zero_mean: bool, optional When set to True, the mean of all signals is subtracted prior to computation of the metrics (default: ``False``) clamp_db: bool, optional If provided, the resulting metrics are clamped to be in the range ``[-clamp_db, clamp_db]`` compute_permutation: bool, optional When this flag is true, the optimal permutation of the estimated signals to maximize the SIR is computed (default: ``True``) load_diag: float, optional If provided, this small value is added to the diagonal coefficients of the system metrics 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 Returns ------- sdr: numpy.ndarray, (..., n_channels_est) The signal-to-distortion-ratio (SDR) sir: numpy.ndarray, (..., n_channels_est) The signal-to-interference-ratio (SIR) sar: numpy.ndarray, (..., n_channels_est) The signal-to-interference-ratio (SIR) perm: numpy.ndarray, (..., n_channels_est) The index of the corresponding reference signal (only when ``compute_permutation == True``) r/z\For non-pairwise computation, the number of reference and estimated signals must be the same)rErFrrrrKTrV)rrarer )rrrErFrrGrfrr"r%rWrcrdrXr&r&r'bss_eval_sourcesQs,:   rg)FTNT)NT)N)rDNFNNF)rDNFNN)rDNFNNFF)rDNFTN)rDNFNTN)!r1typingrrnumpyrscipyrcgdrrhelpersrr r r rr r ndarrayboolfloatr(intr?rCrOrTrUrRrarergr&r&r&r's   / ; 3 i  ?  T