~Vji ddlmZddlZddlmZddlZGddejZGddejZGddejZ d e d e d e d e d e f dZ d e fdZ dS))TupleNcPeZdZdZdedeffd ZdejdejfdZxZ S)AttPoolzAttention-Pooling module that estimates the attention score. Args: input_dim (int): Input feature dimension. att_dim (int): Attention Tensor dimension. input_dimatt_dimctt|tj|d|_tj|||_dS)N)superr__init__nnLinearlinear1linear2selfrr __class__s \/root/voice-cloning/.venv/lib/python3.11/site-packages/torchaudio/models/squim/subjective.pyr zAttPool.__init__sI gt%%'''yA.. yG44 xreturnc||}|dd}tj|d}t j||d}||}|S)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 rforwardzAttPool.forwardssll1oommAq!!m##CQ#// La ( ( + + LLOOr __name__ __module__ __qualname____doc__intr rTensorr" __classcell__rs@rrrs{5#5555555 %,rrcPeZdZdZdedeffd ZdejdejfdZxZ S) PredictorzPrediction module that apply pooling and attention, then predict subjective metric scores. Args: input_dim (int): Input feature dimension. att_dim (int): Attention Tensor dimension. rrctt|t|||_||_dSN)r r-r ratt_pool_layerrrs rr zPredictor.__init__0s< i'')))%i99 rrrc||}tj|d}t jdd|j|j}||zd}|S)aPredict 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) r0r rrrlinspacerr4sum)rrBs rr"zPredictor.forward5si    " " M ! !! ! + + N1at|AH E E E UKKAK  rr#r+s@rr-r-(s{#  %,        rr-ceZdZdZdejdejdejffd Zdejdejde ejejffd Z dejdejfd Z xZ S) SquimSubjectiveaPSpeech 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 predictorctt|||_||_||_dSr/)r r9r r:r;r<)rr:r;r<rs rr zSquimSubjective.__init__Ps8 ot$$--///"""rwaveform referencerc|jd}jd}||kr7||zdz}tjfdt|Dd|ddd|ffS)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 cg|]}SrC).0_r?s r z1SquimSubjective._align_shapes..es"I"I"I9"I"I"IrrN)shapercatrange)rr>r? T_waveform T_reference num_paddings ` r _align_shapeszSquimSubjective._align_shapesVs^B' ob)  # #$ 3a7K "I"I"I"IeK6H6H"I"I"IqQQQI111kzk>222rc||||\}}||j|dd}||j|dd}t j||fd}||}d|z S)aPredict 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,)`. rrArr)rMr;r:extract_featuresrrHr<)rr>r?concat score_diffs rr"zSquimSubjective.forwardhs#009EE)>>$."A"A("K"KA"Nr"RSSNN4>#B#B9#M#Ma#PQS#TUU Ix0a888^^F++ :~r) r$r%r&r'r Moduler rr)rrMr"r*r+s@rr9r9Es#")# #bi###### 3el3u|3PUV[VbdidpVpPq3333$ rr9ssl_typefeat_dimproj_dimrrcttj|}tj||}t |dz|}t |||S)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 r r-r9)rTrUrVrr:r;r<s rsquim_subjective_modelr[zsT5 )84466I (H--I(Q,00I 9i ; ;;rc(tddddS)zXBuild :class:`torchaudio.prototype.models.SquimSubjective` model with default arguments. wav2vec2_basei rO)rTrUrVr)r[rCrrsquim_subjective_baser_s% !     r)typingrrtorch.nnr rYrSrr-r9strr(r[r_rCrrrcs+ bi@ :22222bi222j<<<< <  <<<<*r