0;jiddlmZddlZddlmcmZddlmZGddejj eZ Gdde Z dS) )ABCN)DecodercdeZdZ d fd Zejd dZdZd dZxZ S) BASECFMrct||_||_||_|j|_t |dr |j|_nd|_d|_dS)N sigma_ming-C6?) super__init__n_featsn_spks spk_emb_dimsolverhasattrr estimator)selfr cfm_paramsrr __class__s f/root/voice-cloning/.venv/lib/python3.11/site-packages/chatterbox/models/s3gen/matcha/flow_matching.pyr zBASECFM.__init__ sk   & ' :{ + + "'1DNN!DN?Nctj||z}tjdd|dz|j}|||||||S)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) rr)device)t_spanmumaskspkscond)torch randn_likelinspacer solve_euler) rrr n_timesteps temperaturerrzrs rforwardzBASECFM.forwardsY&  R ; .1kAobiHHH&RdTXYYYrc X|d|d|d|dz } }}g} tdt|D]b} |||||||} || | zz}|| z}| || t|dz kr|| dz|z } c| dS)aP 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 rr)rangelenrappend) rxrrrrrt_dtsolstepdphi_dts rr#zBASECFM.solve_euler4s!9fRj&)fQi*?b1!S[[)) * *DnnQb!T4@@GBL ABA JJqMMMc&kkAo%%D1H%)2wrc |j\}}}tj|ddg|j|j}tj|} dd|jz |zz | z||zz} |d|jz | zz } tj| | ||| || dtj || 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) r)rdtypesum) reduction) shaper randrr5r!r Fmse_lossrsqueezer6) rx1rrrrbr/r.r&yulosss r compute_losszBASECFM.compute_lossTs$(1a J1ay"( C C C  R  !dn$) )Q .R 7 !dn$) )z$..D"aiikk4HH!W\]]] IdOOagaj ( Qwr)rr)rNN)NN) __name__ __module__ __qualname__r r inference_moder'r#rB __classcell__rs@rrr s  &UZZZZ,@rrc eZdZdfd ZxZS)CFMr@ct||||||dkr|ndz}td||d||_dS)N)r rrrrr) in_channels out_channels)r r rr)rrM out_channelrdecoder_paramsrrrs rr z CFM.__init__wsi !#    "FQJJ[[AF  e[{eeVdeer)rrK)rCrDrEr rGrHs@rrJrJvsG f f f f f f f f f frrJ) abcrr torch.nn.functionalnn functionalr:decoderrModulerrJrOrrrXs jjjjjehosjjjZ f f f f f' f f f f fr