o
    óôj8)  ã                   @   sV   d dl Z d dlm  mZ d dlmZ d dlmZ G dd„ deƒZ	G dd„ de	ƒZ
dS )é    N)ÚBASECFM)Úset_all_random_seedc                       sn   e Zd Zddejjf‡ fdd„Ze ¡ dddde dd	dd
¡fdd„ƒZ	ddd„Z
ddd„Zddd„Z‡  ZS )ÚConditionalCFMé   é@   NÚ	estimatorc                    sJ   t ƒ j||||d� |j| _|j| _|j| _||dkr|nd }|| _d S )N)Zn_featsÚ
cfm_paramsÚn_spksÚspk_emb_dimr   )ÚsuperÚ__init__Út_schedulerÚtraining_cfg_rateÚinference_cfg_rater   ©ÚselfÚin_channelsr   r	   r
   r   ©Ú	__class__© úR/var/www/html/academy/tools/voices/clone/CosyVoice/cosyvoice/flow/flow_matching.pyr      s   ü
zConditionalCFM.__init__ç      ð?r   éP   é   c	                 C   sz  t  |¡ |j¡ |j¡| }	|jd }
|
dkrJ|dd…dd…dd…df |	dd…dd…d|
…f< |dd…dd…dd…df |dd…dd…d|
…f< t j|	dd…dd…d|…f |	dd…dd…dd…f gdd�}t j|dd…dd…d|…f |dd…dd…dd…f gdd�}t j||gdd�}t jdd|d |j|jd�}| j	d	kr°dt  
|d
 t j ¡ }| j|	|||||d�|fS )á¼  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   r   Nr   iÞÿÿÿ©Údiméÿÿÿÿ©ÚdeviceÚdtypeÚcosineç      à?)Út_spanÚmuÚmaskÚspksÚcond)ÚtorchÚ
randn_likeÚtor   r    ÚshapeÚconcatÚstackÚlinspacer   ÚcosÚpiÚsolve_euler)r   r$   r%   Ún_timestepsÚtemperaturer&   r'   Ú
prompt_lenÚcacheÚzÚ
cache_sizeZz_cacheZmu_cacher#   r   r   r   Úforward$   s   
22>>
zConditionalCFM.forwardFc              
   C   sÔ  |d |d |d |d  }}	}
|j dd�}g }tjdd| d¡g|j|jd�}tjdd| d¡g|j|jd�}tjdd| d¡g|j|jd�}tjdg|j|jd�}tjddg|j|jd�}tjdd| d¡g|j|jd�}tdt|ƒƒD ]k}||dd…< ||dd…< ||d< |  d¡|dd…< ||d< ||d< |  |||||||¡}tj	|| d¡| d¡gdd�\}}d	| j
 | | j
|  }||
|  }||
 }| |¡ |t|ƒd k rã||d  | }
qx|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   r   r   r   r   r   r   Nr   )Ú	unsqueezer(   ÚzerosÚsizer   r    ÚrangeÚlenÚforward_estimatorÚsplitr   ÚappendÚfloat)r   Úxr#   r$   r%   r&   r'   Ú	streamingÚtÚ_ÚdtÚsolÚx_inZmask_inZmu_inZt_inZspks_inZcond_inÚstepZdphi_dtZcfg_dphi_dtr   r   r   r1   G   s@   $    û$
€zConditionalCFM.solve_eulerc              	   C   sˆ  t | jtjjƒr| j|||||||d�S | j ¡ \\}}	}
tj ¡  ¡  |	�� | 	ddd| 
d¡f¡ | 	ddd| 
d¡f¡ | 	ddd| 
d¡f¡ | 	dd	¡ | 	d
d¡ | 	ddd| 
d¡f¡ | ¡  ¡ | ¡  ¡ | ¡  ¡ | ¡  ¡ | ¡  ¡ | ¡  ¡ | ¡ g}t|ƒD ]\}}| |
 |¡|¡ qŠ| tj ¡ j¡du s¥J ‚tj ¡  ¡  W d   ƒ n1 s¶w   Y  | j ||	¡ |S )N©rC   rB   r   r   r%   r   r$   rD   )r   r&   )r   r   r'   T)Ú
isinstancer   r(   ÚnnÚModuleÚacquire_estimatorÚcudaÚcurrent_streamÚsynchronizeZset_input_shaper;   Ú
contiguousÚdata_ptrÚ	enumerateZset_tensor_addressZget_tensor_nameZexecute_async_v3Úcuda_streamÚrelease_estimator)r   rB   r%   r$   rD   r&   r'   rC   r   ÚstreamÚ
trt_engineÚ	data_ptrsÚiÚjr   r   r   r>   ~   s4   





úîz ConditionalCFM.forward_estimatorc              	   C   s  |j \}}}	tj|ddg|j|jd�}	t |¡}
dd| j |	  |
 |	|  }|d| j |
  }| jdkrZtj||jd�| jk}|| ddd¡ }|| dd¡ }|| ddd¡ }| j	||||	 
¡ |||d�}tj|| || dd�t |¡|j d   }||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   r   )r   r   rJ   Úsum)Ú	reduction)r+   r(   Úrandr   r    r)   Ú	sigma_minr   Úviewr   ÚsqueezeÚFÚmse_lossr\   )r   Úx1r%   r$   r&   r'   rC   ÚbrE   rD   r6   ÚyÚuZcfg_maskÚpredÚlossr   r   r   Úcompute_loss›   s   

,zConditionalCFM.compute_loss©r   r   N)F)NNF)Ú__name__Ú
__module__Ú__qualname__r(   rL   rM   r   Úinference_moder:   r8   r1   r>   rj   Ú__classcell__r   r   r   r   r      s    "
"
7r   c                       s:   e Zd Zddejjf‡ fdd„Ze ¡ dd	d
„ƒZ‡  Z	S )ÚCausalConditionalCFMr   r   Nr   c                    s0   t ƒ  |||||¡ tdƒ t g d¢¡| _d S )Nr   )r   r   i˜:  )r   r   r   r(   ÚrandnÚ
rand_noiser   r   r   r   r   Å   s   zCausalConditionalCFM.__init__r   Fc           
   	   C   s�   | j dd…dd…d| d¡…f  |j¡ |j¡| }tjdd|d |j|jd�}	| jdkr:dt |	d tj	 ¡ }	| j
||	|||||d�dfS )	r   Nr   r   r   r   r!   r"   )r#   r$   r%   r&   r'   rC   )rs   r;   r*   r   r    r(   r.   r   r/   r0   r1   )
r   r$   r%   r2   r3   r&   r'   rC   r6   r#   r   r   r   r8   Ê   s
   6
zCausalConditionalCFM.forwardrk   )r   NNF)
rl   rm   rn   r(   rL   rM   r   ro   r8   rp   r   r   r   r   rq   Ä   s    rq   )r(   Ztorch.nn.functionalrL   Ú
functionalrb   Z&matcha.models.components.flow_matchingr   Úcosyvoice.utils.commonr   r   rq   r   r   r   r   Ú<module>   s    0