0;ji"*ddlZddlZddlmcmZddlmZddlm Z ddl m Z dZ GddeZ Gd d e Z dS) N)BASECFM) CFM_PARAMS)tqdmc fd|DS)Nclg|]0}|jjr |jkr|n|1S)dtypeis_floating_pointto).0ar s _/root/voice-cloning/.venv/lib/python3.11/site-packages/chatterbox/models/s3gen/flow_matching.py zcast_all..s> d d dZ[ag/ UAGu4D4DAA!$$u++ d d dr )r argss` rcast_allrs d d d d_c d d ddrc eZdZddejjffd Zejddddejdddd fd Z dd Z dd Z xZ S)ConditionalCFMr@N estimatorct|||||j|_|j|_|j|_||dkr|ndz}||_dS)N)n_feats cfm_paramsn_spks spk_emb_dimr)super__init__ t_schedulertraining_cfg_rateinference_cfg_raterself in_channelsrrrr __class__s rrzConditionalCFM.__init__st !#    &1!+!=","?!FQJJ[[AF "r?rPc  td)aForward diffusion Args: mu (torch.Tensor): output of encoder shape: (batch_size, n_feats, mel_timesteps) mask (torch.Tensor): output_mask shape: (batch_size, 1, mel_timesteps) n_timesteps (int): number of diffusion steps temperature (float, optional): temperature for scaling noise. Defaults to 1.0. spks (torch.Tensor, optional): speaker ids. Defaults to None. shape: (batch_size, spk_emb_dim) cond: Not used but kept for future purposes Returns: sample: generated mel-spectrogram shape: (batch_size, n_feats, mel_timesteps) z)unused, needs updating for meanflow model)NotImplementedErrortorch randn_liker devicer shapeconcatstacklinspacercospi solve_euler)r#mumask n_timesteps temperaturespkscond prompt_len flow_cachez cache_sizez_cachemu_cachet_spans rforwardzConditionalCFM.forward)s(""MNNNrFc |j}t|||||||jj\}}}}}}|d|d} } t jd| zd| g|j|j} t jd| zd| g|j|j} t jd| zd| g|j|j} t jd| zg|j|j}t jd| zdg|j|j}t jd| zd| g|j|j}t jd| zg|j|j}t|dd|ddD]\}}|d }|d }|x| d| <| | d<|x| d| <| | d<|| d| <|x|d| <|| d<||d| <||d| <|x|d| <|| d<|j | | | ||||r|nd }t j || | gd \}}d |j z|z|j |zz }||z }|||zz}| |S) at Fixed euler solver for ODEs. Args: x (torch.Tensor): random noise t_span (torch.Tensor): n_timesteps interpolated shape: (n_timesteps + 1,) mu (torch.Tensor): output of encoder shape: (batch_size, n_feats, mel_timesteps) mask (torch.Tensor): output_mask shape: (batch_size, 1, mel_timesteps) spks (torch.Tensor, optional): speaker ids. Defaults to None. shape: (batch_size, spk_emb_dim) cond: Not used but kept for future purposes meanflow: meanflow mode r rr(r'r-r rN)dim)xr6r5tr9r:rr&) r rrsizer+zerosr-zip unsqueezerBsplitr!r )r#rHrAr5r6r9r:meanflowin_dtypeBTx_inmask_inmu_int_inspks_incond_inr_inrIrJdxdtcfg_dxdtdts rr4zConditionalCFM.solve_eulerNs 7*21fb$dZ^ZhZn*o*o*o'62tT4wwqzz166!991+q1ub!nQXQWMMM+q1uq!nQXQWMMM+q1ub!nQXQWMMM+q1unQXQWMMM+q1ubnQXQWMMM+q1ub!nQXQWMMM+q1unQXQWMMMss VABBZ00! ! DAq  ""A  ""A"#$ #D!HtABBx(, ,GBQBK'!""+E"1"I"# #D!HtABBxGBQBKGBQBK"# #D!HtABBx>))W$W7",$$*D#[1v1===ND(422d:T=TW_=__DQBBI AAttH~~rc|j\}}}tj|ddg|j|j}|jdkr'dtj|dztjzz }tj|} dd|j z |zz | z||zz} |d|j z | zz } |j dkrqtj||j|j k} || dddz}|| ddz}|| dddz}| | ||| ||} tj| |z| |zd tj|| jdzz }|| fS) aComputes diffusion loss Args: x1 (torch.Tensor): Target shape: (batch_size, n_feats, mel_timesteps) mask (torch.Tensor): target mask shape: (batch_size, 1, mel_timesteps) mu (torch.Tensor): output of encoder shape: (batch_size, n_feats, mel_timesteps) spks (torch.Tensor, optional): speaker embedding. Defaults to None. shape: (batch_size, spk_emb_dim) Returns: loss: conditional flow matching loss y: conditional flow shape: (batch_size, n_feats, mel_timesteps) rrEcosine?r)r-rFsum) reduction)r.r+randr-r rr2r3r, sigma_minr viewrsqueezeFmse_lossra)r#x1r6r5r9r:b_rIr=yucfg_maskpredlosss r compute_losszConditionalCFM.compute_losss$(1a J1ay"( C C C  x ' 'EIa#g0111A  R  !dn$) )Q .R 7 !dn$) )  !A % %z!BI6669OOHhmmB1---B(--A...D(--Aq111D~~ar199;;dCCz$+q4x5AAAUYt__WXW^_`WaEabQwr)rrN)F)NN) __name__ __module__ __qualname__r+nnModulerinference_moderLrBr4rq __classcell__r%s@rrrs # #UZU]Ud # # # # # #U9<4d_`mxmrmxyz|~ABDEnFnF"f"f"f"fHCCCCJ''''''''rrcbeZdZdedddffd Zejd dZd ZxZ S) CausalConditionalCFMrr'Nc`t|||||d|_dS)N)rr rand_noiser"s rrzCausalConditionalCFM.__init__s. j&+yQQQrr&Fc |d} tj|} |4|d|dz } || d| df<tjdd|dz|j|j} |s2|jdkr'dtj| dztjzz } |r| | | |||| dfS| | | ||||| dfS) aForward diffusion Args: mu (torch.Tensor): output of encoder shape: (batch_size, n_feats, mel_timesteps) mask (torch.Tensor): output_mask shape: (batch_size, 1, mel_timesteps) n_timesteps (int): number of diffusion steps temperature (float, optional): temperature for scaling noise. Defaults to 1.0. spks (torch.Tensor, optional): speaker ids. Defaults to None. shape: (batch_size, spk_emb_dim) cond: Not used but kept for future purposes noised_mels: gt mels noised a time t Returns: sample: generated mel-spectrogram shape: (batch_size, n_feats, mel_timesteps) rNr(.rrEr_r`)rAr5r6r9r:)rAr5r6r9r:rP) rKr+r,r1r-r rr2r3 basic_eulerr4) r#r5r6r7r8r9r: noised_melsrPrRr=r;rAs rrBzCausalConditionalCFM.forwards( GGAJJ  R   "k&6&6q&9&99J"-Ac:;; 1kAobirxXXX Mel Inference....rFr)total)r6r5rIr9r:rJ) r rrprintrrMr.rBr ) r#rHrAr5r6r9r:rQrIrJr[r]s rrz CausalConditionalCFM.basic_eulers7*21fb$dZ^ZhZn*o*o*o'62tT4 ,---VC"H-vc122g??v|TVGWZ[G[\\\  DAqT7AdGqA>))!$2TX\])^^DQBBI AAttH~~r)r&NNNF) rrrsrtrrr+rwrBrrxrys@rr{r{s#&:aUWcg U%s%s%s%sN       rr{) threadingr+torch.nn.functionalru functionalrgmatcha.flow_matchingrconfigsrrrrr{r rrrs ))))))eee`````W```F99999>99999r