o
    èôj¹  ã                   @   st   d dl mZ d dlZd dlm  mZ d dlmZ d dl	m
Z
 d dlmZ G dd„ dejjeƒZG dd	„ d	eƒZdS )
é    )ÚABCN)ÚDiT)Úsequence_mask)Útqdmc                       s@   e Zd Z‡ fdd„Ze ¡ ddd„ƒZddd„Zd	d
„ Z‡  Z	S )ÚBASECFMc                    sd   t ƒ  ¡  d| _d | _|jj| _|jdkrtj 	¡ ntj 
¡ | _t|jdƒr-|jj| _d S d| _d S )Ng�íµ ÷Æ°>Úl2Úzero_prompt_speech_tokenF)ÚsuperÚ__init__Ú	sigma_minÚ	estimatorr   Úin_channelsZreg_loss_typeÚtorchÚnnÚMSELossÚL1LossÚ	criterionÚhasattrr   ©ÚselfÚargs©Ú	__class__© úI/var/www/html/academy/tools/voices/clone/seed-vc/modules/flow_matching.pyr
      s   

 
zBASECFM.__init__ç      ð?ç      à?c	              
   C   sb   |  d¡|  d¡}	}
tj|	| j|
g|jd�| }tjdd|d |jd�}|  ||||||||¡S )a¼  Forward 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)
        r   é   )Údevice)Úsizer   Úrandnr   r   ÚlinspaceÚsolve_euler)r   ÚmuÚx_lensÚpromptÚstyleÚf0Zn_timestepsÚtemperatureÚinference_cfg_rateÚBÚTÚzÚt_spanr   r   r   Ú	inference   s   zBASECFM.inferencec	              	   C   sà  |d |d |d |d  }	}
}
g }|  d¡}t |¡}|dd|…f |dd|…f< d|dd|…f< | jr?d|dd|…f< ttdt|ƒƒƒD ]£}|| ||d   }|dkr²tj|t |¡gdd�}tj|t |¡gdd�}tj|t |¡gdd�}tj||gdd�}tj|	 d¡|	 d¡gdd�}|  	||||||¡}|j
ddd�\}}d| | ||  }n|  	||||	 d¡||¡}|||  }|	| }	| |¡ |t|ƒd k rÞ||d  |	 }d|dd…dd…d|…f< qH|d S )	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
        r   éÿÿÿÿr   .N)Údimé   r   )r   r   Ú
zeros_liker   r   ÚrangeÚlenÚcatÚ	unsqueezer   ÚchunkÚappend)r   Úxr$   r%   r#   r&   r'   r-   r)   ÚtÚ_ZsolZ
prompt_lenZprompt_xÚstepÚdtZstacked_prompt_xZstacked_styleZ
stacked_muZ	stacked_xZ	stacked_tZstacked_dphi_dtZdphi_dtZcfg_dphi_dtr   r   r   r"   7   s:   $

ÿ
zBASECFM.solve_eulerc                 C   sx  |j \}}}tj|ddg|j|jd�}t |¡}	dd| j |  |	 ||  }
|d| j |	  }t |¡}t|ƒD ]7}||dd…d|| …f ||dd…d|| …f< d|
|dd…d|| …f< | j	rpd||dd…d|| …f< q9|  
|
||| d¡ d¡|||¡}d}t|ƒD ]$}||  ||dd…|| || …f ||dd…|| || …f ¡7 }qˆ|| }||d| j |	  fS )a�  Computes 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   )r   ÚdtypeNr   )Úshaper   Úrandr   r>   Ú
randn_liker   r2   r3   r   r   Úsqueezer   )r   Úx1r$   Zprompt_lensr#   r&   Úbr;   r:   r,   ÚyÚur%   ÚbibZestimator_outÚlossr   r   r   Úforwardq   s$   

0€"FzBASECFM.forward)r   r   )r   )
Ú__name__Ú
__module__Ú__qualname__r
   r   Úinference_moder.   r"   rI   Ú__classcell__r   r   r   r   r      s    
:r   c                       s   e Zd Z‡ fdd„Z‡  ZS )ÚCFMc                    s4   t ƒ  |¡ |jdkrt|ƒ| _d S td|j› �ƒ‚)Nr   zUnknown diffusion type )r	   r
   Údit_typer   r   ÚNotImplementedErrorr   r   r   r   r
       s   ÿ
zCFM.__init__)rJ   rK   rL   r
   rN   r   r   r   r   rO   Ÿ   s    rO   )Úabcr   r   Ztorch.nn.functionalr   Ú
functionalÚFZmodules.diffusion_transformerr   Úmodules.commonsr   r   ÚModuler   rO   r   r   r   r   Ú<module>   s     