o :i$ @sddlmZmZddlZGdddejjZGdddeZGdddZGd d d eZ d d d d e dfde de de de dej dej f ddZdS))OptionalUnionNcxeZdZdZ ddejjdeffdd Z  dd ej d ej d ej d ej d e ej de e ej fdej fddZ ZS)DiffusionModelzA wrapper of diffusion models for inference. Args: model: The diffusion model. func_name: The function name to call. forward_fm_decodermodel func_namecs(t||_||_t|j||_dS)N)super__init__rrgetattr model_funcselfrr __class__5/home/ubuntu/LuxTTS/zipvoice/models/modules/solver.pyr s zDiffusionModel.__init__Ntxtext_conditionspeech_condition padding_maskguidance_scalereturnc Kst|stj||j|jd}|dkr#|jd |||||d|S|dks+Jtj|gddd}tj|gddd}tjt ||gdd}|dkr\tjt ||gdd}n |d}tj||gdd}|jd |||||d|j ddd\}} d|| ||} | S) aK Forward function that Handles the classifier-free guidance. Args: t: The current timestep, a tensor of a tensor of a single float. x: The initial value, with the shape (batch, seq_len, emb_dim). text_condition: The text_condition of the diffision model, with the shape (batch, seq_len, emb_dim). speech_condition: The speech_condition of the diffision model, with the shape (batch, seq_len, emb_dim). padding_mask: The mask for padding; True means masked position, with the shape (batch, seq_len). guidance_scale: The scale of classifier-free guidance, a float or a tensor of shape (batch, 1, 1). Retrun: The prediction with the shape (batch, seq_len, emb_dim). dtypedevicer)rxtrrrr)dimg?Nr) torch is_tensortensorrrallr r cat zeros_likechunk) rrrrrrrkwargs data_uncond data_condresrrrforward(sR      zDiffusionModel.forwardrNr__name__ __module__ __qualname____doc__r"nnModulestrr Tensorrrfloatr- __classcell__rrrrrs2   rcr)DistillDiffusionModelzA wrapper of distilled diffusion models for inference. Args: model: The distilled diffusion model. func_name: The function name to call. rrrcstj||ddS)N)rr)r r r rrrr xszDistillDiffusionModel.__init__Nrrrrrrrrc Ks<t|stj||j|jd}|jd||||||d|S)a? Forward function that Handles the classifier-free guidance. Args: t: The current timestep, a tensor of a single float. x: The initial value, with the shape (batch, seq_len, emb_dim). text_condition: The text_condition of the diffision model, with the shape (batch, seq_len, emb_dim). speech_condition: The speech_condition of the diffision model, with the shape (batch, seq_len, emb_dim). padding_mask: The mask for padding; True means masked position, with the shape (batch, seq_len). guidance_scale: The scale of classifier-free guidance, a float or a tensor of shape (batch, 1, 1). Retrun: The prediction with the shape (batch, seq_len, emb_dim). r)rrrrrrNr)r"r#r$rrr )rrrrrrrr)rrrr-s  zDistillDiffusionModel.forwardr.r/r0rrrrr;qs2    r;c@sxeZdZ ddejjdefddZ     dd ejd ejd ejd ejd e de e ejfde de de dejfddZ dS) EulerSolverrrrcCt||d|_dS)zConstruct a Euler Solver Args: model: The diffusion model. func_name: The function name to call. rN)rrr rrrr s zEulerSolver.__init__ r?rrrrnum_steprt_startt_endt_shiftrc  Ks|j} t|tr t|tsJt|||| | d} t|D]<} | | }| | d}|jd||||||d| }|d||}|||}| |dkrVd||||}q|}q|S)N)rBrCrArDrr!)rrrrrrr@r)r isinstancer9get_time_stepsranger)rrrrrrArrBrCrDr)r timestepsstept_curt_nextvx_1_predx_0_predrrrsamples8     zEulerSolver.sampleNr.)r?rrr@r@) r1r2r3r"r5r6r7r r8intrr9rOrrrrr<sB     r<c@s&eZdZ ddejjdefddZdS)DistillEulerSolverrrrcCr=)zwConstruct a Euler Solver for distilled diffusion models. Args: model: The diffusion model. r>N)r;rr rrrr s zDistillEulerSolver.__init__Nr.)r1r2r3r"r5r6r7r rrrrrQsrQrr@r?cpurBrCrArDrrcCs4t|||d|}||d|d|}|S)a6Compute the intermediate time steps for sampling. Args: t_start: The starting time of the sampling (default is 0). t_end: The starting time of the sampling (default is 1). num_step: The number of sampling. t_shift: shift the t toward smaller numbers so that the sampling will emphasize low SNR region. Should be in the range of (0, 1]. The shifting will be more significant when the number is smaller. device: A torch device. Returns: The time step with the shape (num_step + 1,). r!)r"linspaceto)rBrCrArDrrHrrrrFsrF)typingrrr"r5r6rr;r<rQrr9rPr8rFrrrrs0Z7G