i ddlmZddlZddlmZddlZGddej ZGddej ZGddej Z d e d e d e d e d e f dZ d e fdZ y))TupleNcdeZdZdZdedeffd 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|||_y)N)superr__init__nnLinearlinear1linear2selfrr __class__s w/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/torchaudio/models/squim/subjective.pyr zAttPool.__init__s4 gt%'yyA. yyG4 xreturnc|j|}|jdd}tjj |d}t j ||jd}|j|}|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.forwardsgll1ommAq!mm##CQ#/ LLa ( ( + LLOr __name__ __module__ __qualname____doc__intr rTensorr" __classcell__rs@rrrs65#55 %,,rrcdeZdZdZdedeffd 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. rrcZtt| t|||_||_yN)r r-r ratt_pool_layerrrs rr zPredictor.__init__0s& i')%i9 rrrc|j|}tjj|d}t j dd|j |j}||zjd}|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.forward5sb    " MM ! !! ! + NN1at||AHH E UKKAK rr#r+s@rr-r-(s6#  %,, rr-ceZdZdZdej dej dej ffd 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 predictorcTtt| ||_||_||_yr/)r r9r r:r;r<)rr:r;r<rs rr zSquimSubjective.__init__Ps% ot-/"""rwaveform referencerc|jd}|jd}||kr6||zdz}tjt|Dcgc]}|c}d}||ddd|ffScc}w)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)shapercatrange)rr>r? T_waveform T_reference num_padding_s r _align_shapeszSquimSubjective._align_shapesVs{^^B' oob)  #$ 3a7K eK6H"I6H96H"IqQI1kzk>222#Js A'cJ|j||\}}|j|jj|dd}|j|jj|dd}t j ||fd}|j |}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)rIr;r:extract_featuresrrCr<)rr>r?concat score_diffs rr"zSquimSubjective.forwardhs#009E)>>$.."A"A("KA"Nr"RSNN4>>#B#B9#Ma#PQS#TU Ix0a8^^F+ :~r) r$r%r&r'r Moduler rr)rrIr"r*r+s@rr9r9Es#"))# #bii# 3ell3u||3PUV[VbVbdidpdpVpPq3$ 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)rPrQrRrr:r;r<s rsquim_subjective_modelrWzsJ5 ))846I (H-I(Q,0I 9i ;;rc tddddS)zXBuild :class:`torchaudio.prototype.models.SquimSubjective` model with default arguments. wav2vec2_basei rK)rPrQrRr)rWrrsquim_subjective_baser\s !   r)typingrrtorch.nnr rUrOrr-r9strr(rWr\r[rrr`s bii@ :2bii2j<<<< <  <*r