antoniomae1234/Seed-VC-characters
6
1o
2 \9�f? � @ sh d dl mZ d dlZd dlm mZ d dlmZ d dl m3Z4 G dd� dejj e�Z G dd� de �Z
dS ) � )�ABCN)�DiT)�
sequence_maskc sB e Zd Z� fdd�Ze�� d dd��Zd
dd�Zdd5d �Z� Z S )�BASECFMc sd t � �� d| _d | _|jj| _|jdkrtj� � ntj�6� | _ t |jd�r-|jj
| _
d S d| _
d S )Ng�����ư>�l2�zero_prompt_speech_tokenF)�super�__init__� sigma_min� estimatorr � in_channelsZ
reg_loss_type�torch�nn�MSELoss�L1Loss� criterion�hasattrr ��self�args�� __class__� �#D:\seed-vc\modules\flow_matching.pyr 7 s 89 10zBASECFM.__init__� �?� �?c
11 C sb |� d�|� d�} }12tj| | j|13g|jd�| } tjdd|d |jd�} | �| |||||| |�S )a� Forward diffusion14 15 Args:16 mu (torch.Tensor): output of encoder17 shape: (batch_size, n_feats, mel_timesteps)18 mask (torch.Tensor): output_mask19 shape: (batch_size, 1, mel_timesteps)20 n_timesteps (int): number of diffusion steps21 temperature (float, optional): temperature for scaling noise. Defaults to 1.0.22 spks (torch.Tensor, optional): speaker ids. Defaults to None.23 shape: (batch_size, spk_emb_dim)24 cond: Not used but kept for future purposes25 26 Returns:27 sample: generated mel-spectrogram28 shape: (batch_size, n_feats, mel_timesteps)29 r � )�device)�sizer
�randnr r �linspace� solve_euler)
r �mu�x_lens�prompt�style�f0Z n_timesteps� temperature�inference_cfg_rate�B�T�z�t_spanr r r � inference s zBASECFM.inferencec C sX |d |d |d |d } }30} g } |� d�}
t�|�}|dd|
�f |dd|
�f< d|dd|
�f< | jr?d|dd|
�f< tdt|��D ]a}| �|||| �d�|||�}|dkr{| �|t�|�|| �d�t�|�t�|�d�}d| | || }|| | }| | } | �|� |t|�d k r�||d | } d|dd�dd�d|
�f< qF| d S )aP 31 Fixed euler solver for ODEs.32 Args:33 x (torch.Tensor): random noise34 t_span (torch.Tensor): n_timesteps interpolated35 shape: (n_timesteps + 1,)36 mu (torch.Tensor): output of encoder37 shape: (batch_size, n_feats, mel_timesteps)38 mask (torch.Tensor): output_mask39 shape: (batch_size, 1, mel_timesteps)40 spks (torch.Tensor, optional): speaker ids. Defaults to None.41 shape: (batch_size, spk_emb_dim)42 cond: Not used but kept for future purposes43 r �����r .Nr ) r r
�44zeros_liker �range�lenr � unsqueeze�append)r �xr# r$ r" r% r& r, r( �t�_�dt�solZ45prompt_lenZprompt_x�stepZdphi_dtZ cfg_dphi_dtr r r r! 4 s4 $4647 48�49� 50 zBASECFM.solve_eulerNc C sb |j \}}} tj|ddg|j|jd�} t�|�}51dd| j | |52 | | } |d| j |53 } t�|�}
t|�D ]7}||dd�d|| �f |
|dd�d|| �f< d| |dd�d|| �f< | j rpd||dd�d|| �f< q9| �54| |
|| � � |||�}d}t|�D ]$}|| � ||dd�|| || �f | |dd�|| || �f �7 }q�|| }|| fS )a� Computes diffusion loss55 56 Args:57 x1 (torch.Tensor): Target58 shape: (batch_size, n_feats, mel_timesteps)59 mask (torch.Tensor): target mask60 shape: (batch_size, 1, mel_timesteps)61 mu (torch.Tensor): output of encoder62 shape: (batch_size, n_feats, mel_timesteps)63 spks (torch.Tensor, optional): speaker embedding. Defaults to None.64 shape: (batch_size, spk_emb_dim)65 66 Returns:67 loss: conditional flow matching loss68 y: conditional flow69 shape: (batch_size, n_feats, mel_timesteps)70 r )r �dtypeNr )
�shaper
�randr r: �71randn_liker72 r/ r0 r r �squeezer )r �x1r# Z prompt_lensr" r% r& �br6 r5 r+ �y�ur$ �bibZ
estimator_out�lossr r r �forwardc s$ 73 74 0� FzBASECFM.forward)r r )r )N)75�__name__�76__module__� __qualname__r r
�inference_moder- r! rE �
__classcell__r r r r r s 77/r c s e Zd Z� fdd�Z� ZS )�CFMc s4 t � �|� |jdkrt|�| _d S td|j� ���)Nr zUnknown diffusion type )r r Zdit_typer r �NotImplementedErrorr r r r r � s �78z CFM.__init__)rF rG rH r rJ r r r r rK � s rK )�abcr r
Ztorch.nn.functionalr �79functional�FZ modules.diffusion_transformerr �modules.commonsr �Moduler rK r r r r �<module> s