o si$Q@s^ddlZddlZddlmZddlZd ddZGdddeZGd d d Z Gd ddZ dS) N) rearrangemetrics OutputMetricscCsv|}|d7}|d|jjd7}|d|jjd7}|jdur*|d|jjd7}|jdur9|d|jjd7}|S) Nz$ Shapes: (is shape) (symbolic shape)z speech_prediction: z (K_target, N)z speech_source: z (K_source, N)z speech_contribution: z (K_source, K_target, N)z noise_contribution: )speech_predictionshape speech_sourcespeech_contributionnoise_contribution)msgrr R/home/ubuntu/.local/lib/python3.10/site-packages/pb_bss_eval/evaluation/wrapper.py _get_err_msgs   r cseZdZfddZZS)VerboseKeyErrorcst|jdkr |j\}}ddl}|j||ddd}|d|St|jdkrD|j\}}}ddl}|j||ddd}|d|d|StS)Nrd)cutoffnz. Close matches:  )lenargsdifflibget_close_matchessuper__str__)selfitemkeysr suggestionsr  __class__r r rs*   zVerboseKeyError.__str__)__name__ __module__ __qualname__r __classcell__r r rr rsrc @s&eZdZ    d6ddddddd dd ed ef d d ZddZddZejddZ ejddZ ejddZ ejddZ ejddZ ejddZejddZejd d!Zejd"d#Zejd$d%Zejd&d'Zd(d)Zd*d+Zd,d-Zed.d/Zed0d1Zed2d3Zd4d5ZdS)7 InputMetricsNF observationz Shape(D, N)rShape(K_source, N) speech_imagezShape(K_source, D, N) noise_image sample_rate enable_si_sdrcCst||_||_||_||_||_||duo|du|_|jjd|_|jjd|_ |jjd|_ ||_ | dS)a Args: observation: When you pass D channels, you get D metrics per speaker. If you want to select a reference channel, you need to slice the input to just have a singleton channel dimension. speech_source: speech_image: noise_image: sample_rate: enable_si_sdr: Since SI-SDR is only well defined for non-reverb single-channel data, it is disabled by default. Nr) r&rr(r)r*enable_1d_inputs_has_image_signalsrsampleschannelsK_sourcer+ check_inputs)rr&rr(r)r*r+r r r __init__7s zInputMetrics.__init__cC8|jjdkr |jd|_|jjdkr|jd|_dSdSN)r&ndimrrr r r r._   zInputMetrics.enable_1d_inputscCs4|jjdks J|jj|jjdksJ|jjdS)Nr)r&r8rrr9r r r r3eszInputMetrics.check_inputscCs4tjjt|jg|jdt|jg|jddddS)N4channels sources samples -> sources channels samples4sources channels samples -> sources channels samplesTF) reference estimation return_dictcompute_permutation) pb_bss_eval evaluationmir_eval_sourcesrrr1r&r2r9r r r mir_evalis  zInputMetrics.mir_evalcC |jdSNsdrrDr9r r r mir_eval_sdrx zInputMetrics.mir_eval_sdrcCrENsirrHr9r r r mir_eval_sir|rJzInputMetrics.mir_eval_sircCrENsarrHr9r r r mir_eval_sarrJzInputMetrics.mir_eval_sarcCs.tjjt|jg|jd|jg|j|jdS)Nr;)r*) rArBpesqrrr1r&r2r*r9r r r rQs  zInputMetrics.pesqcCs2ddlm}|t|jdt|jddddd}|S)Nr) input_sxrz2sources sensors samples -> sources sensors samplesz"sensors samples -> sensors samplesFT)average_sourcesaverage_channelsr?)!pb_bss_eval.evaluation.sxr_modulerRrr(r))rrR invasive_sxrr r r rVs   zInputMetrics.invasive_sxrcCrErFrVr9r r r invasive_sdrrJzInputMetrics.invasive_sdrcCrErKrWr9r r r invasive_sirrJzInputMetrics.invasive_sircCrENsnrrWr9r r r invasive_snrrJzInputMetrics.invasive_snrcCs8tjjt|jg|jdt|jg|jd|jd}|S)Nr;r<r=r>r*) rArBstoirrr1r&r2r*)rscoresr r r r^s   zInputMetrics.stoicCsF|jrtjj|jdddddf|jdddddfdStdN)r=r>zSI-SDR is disabled by default since it is only well-defined for non-reverberant single-channel data. Enable it with `enable_si_sdr=True`.)r+rArBsi_sdrrr& ValueErrorr9r r r raszInputMetrics.si_sdrcCsDgd}|jr |d|jr|d|d|dt|S)NrQr^rIrMrPrarXr\rY)r+appendr/tupler metric_namesr r r _available_metric_namess    z$InputMetrics._available_metric_namescCs<g}|js |d|js|d|d|d|S)NrarXr\rY)r+rdr/rdisabledr r r _disabled_metric_namess    z#InputMetrics._disabled_metric_namescfddDS)Nci|]}||qSr r .0namer9r r z(InputMetrics.as_dict..rhr9r r9r as_dictzInputMetrics.as_dictcC|jSNrIr9r r r rGzInputMetrics.sdrcCrvrwrMr9r r r rLryzInputMetrics.sircCrvrwrPr9r r r rOryzInputMetrics.sarcpt|trfdd|DSt|tsJt||fzt|WSty*Ynwt|d)Ncrmr r rnr9r r rqrrz,InputMetrics.__getitem__.. Disabled: isinstanceliststrtypegetattrAttributeErrorrrhrkrrr r9r __getitem__    zInputMetrics.__getitem__)NNNF)r!r"r#intboolr4r.r3cached_propertyrDrIrMrPrQrVrXrYr\r^rarhrkrtpropertyrGrLrOrr r r r r%6sh (                r%c@sHeZdZ     d>> from IPython.lib.pretty import pprint >>> metrics = OutputMetrics( ... speech_prediction=np.array([[1., 2., 3., 4.] * 1000, ... [4., 3., 2., 1.] * 1000]), ... speech_source=np.array([[1., 2., 2., 3., 2.] * 800, ... [4., 3., 3., 2., 3.] * 800]), ... sample_rate=8000, ... ) # Obtain all metrics (recommended) >>> with np.printoptions(precision=4): ... pprint(metrics.as_dict()) {'pesq': array([1.2235, 1.225 ]), 'stoi': array([0.0503, 0.0638]), 'mir_eval_sdr': array([7.2565, 7.3303]), 'mir_eval_sir': array([25.6896, 46.638 ]), 'mir_eval_sar': array([7.3309, 7.3309]), 'mir_eval_selection': array([0, 1])} # Obtain particular metric (e.g. pesq) >>> metrics.pesq array([1.22345543, 1.2250005 ]) # Obtain multiple metrics (e.g. pesq and stoi) >>> pprint({m: metrics[m] for m in ['pesq', 'stoi']}) {'pesq': array([1.22345543, 1.2250005 ]), 'stoi': array([0.05026565, 0.06377457])} Nr,r)rrrr r*r@r._has_contribution_signalsrr0r2K_targetr+r3)rrrrr r*r+r@r r r r4sC zOutputMetrics.__init__cCr5r6)rr8rr9r r r r.]r:zOutputMetrics.enable_1d_inputsc Cs|jjdks J|jj|jjdksJ|jj|jdks)Jtd|jd||jdks:Jtd|jd||j|j|jdfvsLJtd||jjd|jks\Jtd||jdur|j dur|j duspJ|j |jj\}}}|j|ksJtd ||j|ksJtd ||jdksJtd ||j j\}}|j|ksJtd ||j|ksJtd |t t |jt j |jdd|j }|dksJd|dS|jdur|j dusJd|jd|j dS)Nrz9Number of source speakers (K_source) of speech_source is z). Expect a reasonable value of 5 or less.z=Number of target speakers (K_target) of speech_prediction is r7zlNumber of target speakers (K_target) should be equal to number of source speakers (K_source) or K_target + 1zPNum samples (N) of speech_source does not fit to theshape from speech_predictionzVNum samples (N) of speech_contribution does not fit to theshape from speech_predictionzfNum target speakers (K_target) of speech_contribution does not fit to the shape from speech_predictionzbNum source speakers (K_source) of speech_contribution does not fit to the shape from speech_sourcezVNum samples (N) of noise_contribution does not fit to the shape from speech_predictionzeNum target speakers (K_target) of noise_contribution does not fit to the shape from speech_predictionr)axisgMbP?zbThe deviation of speech prediction and the sum of individual contributions is expected to be low: ziExpect that speech_contribution and noise_contribution are both None or given. Got: speech_contribution: z noise_contribution: )rr8rrr2r rr0rr npstdabssum)r K_source_ K_target_samples_ deviationr r r r3cs      zOutputMetrics.check_inputscCrE)N selectionrHr9r r r mir_eval_selectionrJz OutputMetrics.mir_eval_selectioncCs||jjdks J|jj|jjddksJ|jj|js |jS|jjdt|jt|jdfvs8J|jj|j|jS)Nrr r7)rr8rr@rrr9r r r speech_prediction_selections  z)OutputMetrics.speech_prediction_selectioncCstjj|j|jddS)NT)r=r>r?)rArBrCrrr9r r r rDs zOutputMetrics.mir_evalcCrErFrHr9r r r rIrJzOutputMetrics.mir_eval_sdrcCrErKrHr9r r r rMrJzOutputMetrics.mir_eval_sircCrErNrHr9r r r rPrJzOutputMetrics.mir_eval_sarcCtjj|j|j|jdSNr])rArBrQrrr*r9r r r rQ zOutputMetrics.pesqcCsRddlm}|t|jddd|jddft|jd|jddfddd}|S)Nr) output_sxrz2sources targets samples -> sources targets samplesz"targets samples -> targets samplesFT)rSr?)rUrrrrr )rrrVr r r rVs"   zOutputMetrics.invasive_sxrcCrErFrWr9r r r rXrJzOutputMetrics.invasive_sdrcCrErKrWr9r r r rYrJzOutputMetrics.invasive_sircCrErZrWr9r r r r\rJzOutputMetrics.invasive_snrcCrr)rArBr^rrr*r9r r r r^rzOutputMetrics.stoicCs"|jr tjj|j|jdStdr`)r+rArBrarrrbr9r r r raszOutputMetrics.si_sdrcCsTgd}|jr |d|jr|d|jr&|d|d|dt|S)NrcrrarXr\rY)r@rdr+rrerfr r r rhs     z%OutputMetrics._available_metric_namescCsLg}|js |d|js|d|js$|d|d|d|S)NrrarXr\rY)r@rdr+rrir r r rk#s     z$OutputMetrics._disabled_metric_namescrl)Ncrmr r rnr9r r rq0rrz)OutputMetrics.as_dict..rsr9r r9r rt/ruzOutputMetrics.as_dictcCrvrwrxr9r r r rG3ryzOutputMetrics.sdrcCrvrwrzr9r r r rL7ryzOutputMetrics.sircCrvrwr{r9r r r rO;ryzOutputMetrics.sarcr|)Ncrmr r rnr9r r rqArrz-OutputMetrics.__getitem__..r}r~rr r9r r?rzOutputMetrics.__getitem__)NNNFT)r!r"r#rrr4r.r3rrrrDrIrMrPrQrVrXrYr\r^rarhrkrtrrGrLrOrr r r r rsv YN                   )rr) rnumpyreinopsrrAr KeyErrorrr%rr r r r s  N