o #i @sddlmZddlZddlmZddlZGdddejZGdddejZGdddejZ d e d e d e d e d e f ddZ d e fddZ dS))TupleNc@eZdZdZdedeffdd Zdejdejfdd ZZ S) AttPoolzAttention-Pooling module that estimates the attention score. Args: input_dim (int): Input feature dimension. att_dim (int): Attention Tensor dimension. input_dimatt_dimcs.tt|t|d|_t|||_dS)N)superr__init__nnLinearlinear1linear2selfrr __class___/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/torchaudio/models/squim/subjective.pyr szAttPool.__init__xreturncCsF||}|dd}tjj|dd}t||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)`. rdim) r transposer functionalsoftmaxtorchmatmulsqueezer )rrattrrrforwards  zAttPool.forward __name__ __module__ __qualname____doc__intr rTensorr __classcell__rrrrrsrcr) 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. rrcs$tt|t|||_||_dSN)rr)r ratt_pool_layerrrrrrr 0s  zPredictor.__init__rrcCsD||}tjj|dd}tjdd|j|jd}||jdd}|S)aPredict subjective evaluation metric score. Args: x (torch.Tensor): Input Tensor with dimensions `(batch, time, feature_dim)`. Returns: (torch.Tensor): Subjective metric score. Tensor with dimensions `(batch,)`. rrr)stepsdevice) r+r rrrlinspacerr.sum)rrBrrrr 5s zPredictor.forwardr!rrrrr)(sr)cspeZdZdZdejdejdejffdd Zdejdejd e ejejffd d Z dejdejfd d Z Z S)SquimSubjectiveaPSpeech Quality and Intelligibility Measures (SQUIM) model that predicts **subjective** metric scores for speech enhancement (e.g., Mean Opinion Score (MOS)). The model is adopted from *NORESQA-MOS* :cite:`manocha2022speech` which predicts MOS scores given the input speech and a non-matching reference. Args: ssl_model (torch.nn.Module): The self-supervised learning model for feature extraction. projector (torch.nn.Module): Projection layer that projects SSL feature to a lower dimension. predictor (torch.nn.Module): Predict the subjective scores. ssl_model projector predictorcs$tt|||_||_||_dSr*)rr2r r3r4r5)rr3r4r5rrrr Ps zSquimSubjective.__init__waveform referencercs`|jd}jd}||kr$||d}tjfddt|Ddd|ddd|ffS)aCut or pad the reference Tensor to make it aligned with waveform Tensor. Args: waveform (torch.Tensor): Input waveform for evaluation. Tensor with dimensions `(batch, time)`. reference (torch.Tensor): Non-matching clean reference. Tensor with dimensions `(batch, time_ref)`. Returns: (torch.Tensor, torch.Tensor): The aligned waveform and reference Tensors with same dimensions `(batch, time)`. rcsg|]}qSrr).0_r7rr esz1SquimSubjective._align_shapes..rN)shapercatrange)rr6r7 T_waveform T_reference num_paddingrr;r _align_shapesVs   zSquimSubjective._align_shapescCsh|||\}}||j|dd}||j|dd}tj||fdd}||}d|S)aPredict subjective evaluation metric score. Args: waveform (torch.Tensor): Input waveform for evaluation. Tensor with dimensions `(batch, time)`. reference (torch.Tensor): Non-matching clean reference. Tensor with dimensions `(batch, time_ref)`. Returns: (torch.Tensor): Subjective metric score. Tensor with dimensions `(batch,)`. rr8rr)rCr4r3extract_featuresrr>r5)rr6r7concat score_diffrrrr hs   zSquimSubjective.forward) r"r#r$r%r Moduler rr'rrCr r(rrrrr2Es  &r2ssl_typefeat_dimproj_dimrrcCs4ttj|}t||}t|d|}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)r2)rIrJrKrr3r4r5rrrsquim_subjective_modelzs  rOcCstdddddS)zXBuild :class:`torchaudio.prototype.models.SquimSubjective` model with default arguments. wav2vec2_basei rD)rIrJrKr)rOrrrrsquim_subjective_bases rR)typingrrtorch.nnr rMrHrr)r2strr&rOrRrrrrs&   5