o ´‹ jàã@srddlZddlmmZddlmZGdd„dejƒZGdd„dejƒZGdd„dejƒZGd d „d ejƒZ dS) éN)ÚnncsHeZdZd‡fdd„ Zdd„Zdd „Zd d „Zd d „Zddd„Z‡Z S)ÚGE2ELossç$@çÀÚsoftmaxcsvtƒ ¡t t |¡¡|_t t |¡¡|_||_t dƒ|jdvs%J‚|jdkr.|j |_ |jdkr9|j |_ dSdS)a Implementation of the Generalized End-to-End loss defined in https://arxiv.org/abs/1710.10467 [1] Accepts an input of size (N, M, D) where N is the number of speakers in the batch, M is the number of utterances per speaker, and D is the dimensionality of the embedding vector (e.g. d-vector) Args: - init_w (float): defines the initial value of w in Equation (5) of [1] - init_b (float): definies the initial value of b in Equation (5) of [1] z* > Initialized Generalized End-to-End loss)rÚcontrastrrN) ÚsuperÚ__init__rÚ ParameterÚtorchÚtensorÚwÚbÚ loss_methodÚprintÚembed_loss_softmaxÚ embed_lossÚembed_loss_contrast)ÚselfÚinit_wÚinit_br©Ú __class__©úE/home/kuhnn/.local/lib/python3.10/site-packages/TTS/encoder/losses.pyr s    ÿzGE2ELoss.__init__c Cstt ||d|…f|||dd…ff¡}t |d¡}g}t|ƒD]\}}||kr/| |¡q!| |¡q!t |¡S)zP Calculates the new centroids excluding the reference utterance Nér)r ÚcatÚmeanÚ enumerateÚappendÚstack) rÚdvecsÚ centroidsÚspkrÚuttÚexclÚ new_centroidsÚiÚcentroidrrrÚcalc_new_centroids$s*    zGE2ELoss.calc_new_centroidsc Cs¦g}t|ƒD]G\}}g}t|ƒD]0\}}| ||||¡} | t t | d¡ dd¡|  dd¡¡t |¡tj| ddd¡¡qtj |dd}| |¡qt  |¡S)zE Make the cosine similarity matrix with dims (N,M,N) rr©Údimçíµ ÷Æ°>) rr)rr ÚclampÚmmÚ unsqueezeÚ transposeÚnormrr ) rr!r"Úcos_sim_matrixÚspkr_idxÚspeakerÚcs_rowÚutt_idxÚ utterancer&rrrÚcalc_cosine_sim2s& þüúÿ  zGE2ELoss.calc_cosine_simc Csn|j\}}}g}t|ƒD]%}g}t|ƒD]} | t ||| fd¡| ¡qt |¡}| |¡q t |¡S)zU Calculates the loss on each embedding $L(e_{ji})$ by taking softmax r)ÚshapeÚrangerÚFÚ log_softmaxr r ) rr!r2ÚNÚMÚ_ÚLÚjÚL_rowr'rrrrKs   "   zGE2ELoss.embed_loss_softmaxc Csª|j\}}}g}t|ƒD]C}g}t|ƒD]0} t ||| f¡} t | d|…| |dd…f¡} | dt ||| |f¡t | ¡¡qt |¡}| |¡q t |¡S)zj Calculates the loss on each embedding $L(e_{ji})$ by contrast loss with closest centroid Nrgð?)r9r:r ÚsigmoidrrÚmaxr ) rr!r2r=r>r?r@rArBr'Úcentroids_sigmoidsÚexcl_centroids_sigmoidsrrrrZs   "*   zGE2ELoss.embed_loss_contrastNcCs^| ¡ddks J‚t |d¡}| ||¡}t |jd¡|j||j}| ||¡}| ¡S)zv Calculates the GE2E loss for an input of dimensions (num_speakers, num_utts_per_speaker, dvec_feats) rér,)Úsizer rr8r-r rr)rÚxÚ_labelr"r2r@rrrÚforwardjs   zGE2ELoss.forward)rrr©N) Ú__name__Ú __module__Ú __qualname__r r)r8rrrKÚ __classcell__rrrrrsrcó,eZdZdZd ‡fdd„ Zd dd„Z‡ZS) ÚAngleProtoLossaÌ Implementation of the Angular Prototypical loss defined in https://arxiv.org/abs/2003.11982 Accepts an input of size (N, M, D) where N is the number of speakers in the batch, M is the number of utterances per speaker, and D is the dimensionality of the embedding vector Args: - init_w (float): defines the initial value of w - init_b (float): definies the initial value of b rrcsFtƒ ¡t t |¡¡|_t t |¡¡|_tj ¡|_ t dƒdS)Nz( > Initialized Angular Prototypical loss) rr rr r r r rÚCrossEntropyLossÚ criterionr)rrrrrrr †s   zAngleProtoLoss.__init__Nc CsÌ| ¡ddks J‚t |dd…dd…dd…fd¡}|dd…ddd…f}| ¡d}t | d¡ dd|¡| d¡ dd|¡ dd¡¡}t |j d¡||j |j }t  |¡  |j ¡}| ||¡}|S)z| Calculates the AngleProto loss for an input of dimensions (num_speakers, num_utts_per_speaker, dvec_feats) rrGNréÿÿÿÿr,)rHr rr;Úcosine_similarityr/Úexpandr0r-r rÚarangeÚtoÚdevicerT) rrIrJÚ out_anchorÚ out_positiveÚ num_speakersr2Úlabelr@rrrrKs" þ zAngleProtoLoss.forward©rrrL©rMrNrOÚ__doc__r rKrPrrrrrRzs  rRcs2eZdZdZ‡fdd„Zd dd„Zdd„Z‡ZS) Ú SoftmaxLosszÙ Implementation of the Softmax loss as defined in https://arxiv.org/abs/2003.11982 Args: - embedding_dim (float): speaker embedding dim - n_speakers (float): number of speakers cs0tƒ ¡tj ¡|_t ||¡|_tdƒdS)NzInitialised Softmax Loss) rr r rrSrTÚLinearÚfcr)rÚ embedding_dimÚ n_speakersrrrr ®s   zSoftmaxLoss.__init__NcCs8| d| ¡d¡}| d¡}| |¡}| ||¡}|S)NrU)ÚreshaperHrdrT)rrIr^r@rrrrK¶s    zSoftmaxLoss.forwardcCs0| |¡}tjjj|dd d¡}t |¡}|S)Nrr*r)rdr rÚ functionalrÚsqueezeÚargmax)rÚ embeddingrIÚ activationsÚclass_idrrrÚ inferenceÀs  zSoftmaxLoss.inferencerL)rMrNrOrar rKrnrPrrrrrb¦s    rbcrQ) ÚSoftmaxAngleProtoLossaf Implementation of the Softmax AnglePrototypical loss as defined in https://arxiv.org/abs/2009.14153 Args: - embedding_dim (float): speaker embedding dim - n_speakers (float): number of speakers - init_w (float): defines the initial value of w - init_b (float): definies the initial value of b rrcs.tƒ ¡t||ƒ|_t||ƒ|_tdƒdS)Nz)Initialised SoftmaxAnglePrototypical Loss)rr rbrrRÚ angleprotor)rrerfrrrrrr Ñs    zSoftmaxAngleProtoLoss.__init__NcCs| |¡}| ||¡}||S)zŠ Calculates the SoftmaxAnglePrototypical loss for an input of dimensions (num_speakers, num_utts_per_speaker, dvec_feats) )rpr)rrIr^ÚLpÚLsrrrrKÙs  zSoftmaxAngleProtoLoss.forwardr_rLr`rrrrroÇs ro) r Útorch.nn.functionalrrhr;ÚModulerrRrbrorrrrÚs s,!