h SSKJr SSKrSSKJr SSKr"SS\R 5r"SS\R 5r"SS\R 5r S \ S \ S \ S \ S \ 4 Sjr S \ 4Sjr g))TupleNcr^\rSrSrSrS\S\4U4SjjrS\RS\R4Sjr S r U=r $) AttPoolzAttention-Pooling module that estimates the attention score. Args: input_dim (int): Input feature dimension. att_dim (int): Attention Tensor dimension. input_dimatt_dimc>[[U] 5 [R"US5Ul[R"X5Ulg)N)superr__init__nnLinearlinear1linear2selfrr __class__s Z/var/www/auris/envauris/lib/python3.13/site-packages/torchaudio/models/squim/subjective.pyr AttPool.__init__s2 gt%'yyA. yy4 xreturncURU5nURSS5n[RR USS9n[ R "X!5RS5nURU5nU$)zApply attention and pooling. Args: x (torch.Tensor): Input Tensor with dimensions `(batch, time, feature_dim)`. Returns: (torch.Tensor): Attention score with dimensions `(batch, att_dim)`. r dim) r transposer functionalsoftmaxtorchmatmulsqueezer)rratts rforwardAttPool.forwardsell1ommAq!mm##CQ#/ LL ( ( + LLOr)rr __name__ __module__ __qualname____firstlineno____doc__intr r Tensorr$__static_attributes__ __classcell__rs@rrrs;5#55 %,,rrcr^\rSrSrSrS\S\4U4SjjrS\RS\R4Sjr S r U=r $) Predictor(zPrediction module that apply pooling and attention, then predict subjective metric scores. Args: input_dim (int): Input feature dimension. att_dim (int): Attention Tensor dimension. rrcV>[[U] 5 [X5UlX lgN)r r2r ratt_pool_layerrrs rr Predictor.__init__0s" i')%i9 rrrcURU5n[RRUSS9n[R "SSUR URS9nX-RSS9nU$)zPredict subjective evaluation metric score. Args: x (torch.Tensor): Input Tensor with dimensions `(batch, time, feature_dim)`. Returns: (torch.Tensor): Subjective metric score. Tensor with dimensions `(batch,)`. r rr)stepsdevice) r6r rrr linspacerr;sum)rrBs rr$Predictor.forward5s`    " MM ! !! ! + NN1at||AHH E UKKAK r)rr6r&r0s@rr2r2(s;#  %,,  rr2c"^\rSrSrSrS\R S\R S\R 4U4SjjrS\RS\RS \ \R\R44S jr S\RS\R4S jr S r U=r$) SquimSubjectiveEa4Speech Quality and Intelligibility Measures (SQUIM) model that predicts **subjective** metric scores for speech enhancement (e.g., Mean Opinion Score (MOS)). The model is adopted from *NORESQA-MOS* :cite:`manocha2022speech` which predicts MOS scores given the input speech and a non-matching reference. Args: ssl_model (torch.nn.Module): The self-supervised learning model for feature extraction. projector (torch.nn.Module): Projection layer that projects SSL feature to a lower dimension. predictor (torch.nn.Module): Predict the subjective scores. ssl_model projector predictorcN>[[U] 5 XlX lX0lgr5)r rAr rCrDrE)rrCrDrErs rr SquimSubjective.__init__Ps ot-/"""rwaveform referencercURSnURSnXC:a5X4-S-n[R"[U5Vs/sHobPM snSS9nXSS2SU244$s snf)aCut or pad the reference Tensor to make it aligned with waveform Tensor. Args: waveform (torch.Tensor): Input waveform for evaluation. Tensor with dimensions `(batch, time)`. reference (torch.Tensor): Non-matching clean reference. Tensor with dimensions `(batch, time_ref)`. Returns: (torch.Tensor, torch.Tensor): The aligned waveform and reference Tensors with same dimensions `(batch, time)`. r rN)shaper catrange)rrHrI T_waveform T_reference num_padding_s r _align_shapesSquimSubjective._align_shapesVsu^^B' oob)  #$3a7K eK6H"I6H96H"IqQI1kzk>222#Js A%c@URX5upURURRU5SS5nURURRU5SS5n[R "X!4SS9nUR U5nSU- $)aYPredict subjective evaluation metric score. Args: waveform (torch.Tensor): Input waveform for evaluation. Tensor with dimensions `(batch, time)`. reference (torch.Tensor): Non-matching clean reference. Tensor with dimensions `(batch, time_ref)`. Returns: (torch.Tensor): Subjective metric score. Tensor with dimensions `(batch,)`. rrKrr)rSrDrCextract_featuresr rMrE)rrHrIconcat score_diffs rr$SquimSubjective.forwardhs#00E>>$.."A"A("KA"Nr"RSNN4>>#B#B9#Ma#PQS#TU I0a8^^F+ :~r)rErDrC)r'r(r)r*r+r Moduler r r-rrSr$r.r/r0s@rrArAEs#"))# #bii# 3ell3u||3PUV[VbVbdidpdpVpPq3$ rrAssl_typefeat_dimproj_dimrrc[[RU5"5n[R"X5n[ US-U5n[ XEU5$)aBuild a custome :class:`torchaudio.prototype.models.SquimSubjective` model. Args: ssl_type (str): Type of self-supervised learning (SSL) models. Must be one of ["wav2vec2_base", "wav2vec2_large"]. feat_dim (int): Feature dimension of the SSL feature representation. proj_dim (int): Output dimension of projection layer. att_dim (int): Dimension of attention scores. r)getattr torchaudiomodelsr rr2rA)r\r]r^rrCrDrEs rsquim_subjective_modelrczsD ))846I (-I(Q,0I 9 ;;rc[SSSSS9$)zXBuild :class:`torchaudio.prototype.models.SquimSubjective` model with default arguments. wav2vec2_basei rV)r\r]r^r)rcrrsquim_subjective_baserhs !   r)typingrr torch.nnr rar[rr2rAstrr,rcrhrgrrrls bii@ :2bii2j<<<< <  <*r