o }oi@sxddlmZddlZddlmZddlmZzddlZdZWn e y)dZYnwGdddeZ Gd d d eZ dS) )AnyN)Metric)loggingTFcseZdZUdZdZeed<ejed<ejed<dZ e ed<ded e ffd d Z d ejd ejddfddZ dejfddZfddZZS)SquimMOSMetriczuA metric calculating the average Torchaudio Squim MOS. Args: fs: sampling rate of the input signals > sample_ratemos_sum num_examplesThigher_is_betterfskwargsc stjdi|tst|jjd||jkr/tjj ||jddddd|_ t d||j||_ tjj|_|jdtd d d |jd td d d t d|jj|j dS)Nz metric needs `torchaudio`.@bR?sinc_interp_kaiserQaTi-@ orig_freqnew_freqlowpass_filter_widthrolloffresampling_methodbeta3Input signals will be resampled from fs=%d to %d Hzrsumdefaultdist_reduce_fxr rz Setup metric %s with input fs=%s)super__init__HAVE_TORCHAUDIOModuleNotFoundError __class____name__r torchaudio transformsResample_squim_mos_metric_resamplerrwarningr pipelinesSQUIM_SUBJECTIVE get_model_squim_mos_metric_model add_statetorchtensordebug)selfr r r#rX/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/audio/metrics/squim.pyr )s$ zSquimMOSMetric.__init__predstargetreturnNcCs|j|jkr||}||}|jdkr |d}|d}n|jdkr.td|jd|||}|j|7_|j | 7_ dS)zUpdate the metric by calculating the MOS score for the current batch. Args: preds: tensor with predictions, shape (B, T) target: tensor with target signals, shape (B, T). Target can be a non-matching reference. rExpected 1D or 2D signals, got D signalsN) r rr(ndim unsqueeze ValueErrorr-rrr numel)r2r5r6 mos_batchrrr4updateCs        zSquimMOSMetric.updatecC |j|jSzCompute the underlying metric.)rr r2rrr4computeZ zSquimMOSMetric.computec8tj|i|}dd|D}|D]}||=q|S):Do not save the MOS model and resampler in the state dict.cS g|] }d|vs d|vr|qS)r(r-r.0keyrrr4 b z-SquimMOSMetric.state_dict..r state_dictkeysr2argsr rP remove_keysrLr3rr4rP^zSquimMOSMetric.state_dict)r$ __module__ __qualname____doc__rint__annotations__r/Tensorr boolrr rArErP __classcell__rrr3r4rs     rcseZdZUdZdZeed<ejed<ejed<dZ e ed<ded e d e ffd d Z ddejde dd fddZdejfddZfddZZS)SquimObjectiveMetriczA metric calculating the average Torchaudio Squim objective metric. Args: fs: sampling rate of the input signals metric: the objective metric to calculate. One of 'stoi', 'pesq', 'si_sdr' rr metric_sumr Tr r metricr c stjdi|tst|jjd||jkr/tjj ||jddddd|_ t d||j||_ |dvr>td |d ||_tjj|_|jd td d d|jdtdd dt d|jj|j|j dS)Nz needs `torchaudio`.r rrrrr)stoipesqsi_sdrzUnsupported metric z1. Supported metrics are "stoi", "pesq", "si_sdr".r_rrrr rz$Setup %s with metric=%s, input fs=%sr)rr r!r"r#r$rr%r&r'!_squim_objective_metric_resamplerrr)r r>r`r*SQUIM_OBJECTIVEr,_squim_objective_metric_modelr.r/r0r1)r2r r`r r3rr4r ys* zSquimObjectiveMetric.__init__Nr5r6r7cCs|j|jkr ||}|jdkr|d}n|jdkr$td|jd||\}}}|jdkr4|}n|jdkr<|}n|jdkrD|}ntd |j|j| 7_|j | 7_ d S) zUpdate the metric by calculating the selected metric score for the current batch. Args: preds: tensor with predictions, shape (B, T) target: None, not used. Keeping for interfacfe compatibility with other metrics. r8rr9r:r;rarbrczUnknown metric N) r rrdr<r=r>rfr`r_rr r?)r2r5r6 stoi_batch pesq_batch si_sdr_batch metric_batchrrr4rAs        zSquimObjectiveMetric.updatecCrBrC)r_r rDrrr4rErFzSquimObjectiveMetric.computecrG)rHcSrI)rdrfrrJrrr4rMrNz3SquimObjectiveMetric.state_dict..rOrRr3rr4rPrUzSquimObjectiveMetric.state_dict)N)r$rVrWrXrrYrZr/r[r r\strrr rArErPr]rrr3r4r^ls     r^) typingrr/ torchmetricsr nemo.utilsrr%r!r"rr^rrrr4s    O