modules/flow_matching.py

from abc import ABC

import torch
import torch.nn.functional as F

from modules.diffusion_transformer import DiT
from modules.commons import sequence_mask

class BASECFM(torch.nn.Module, ABC):
    def __init__(
        self,
        args,
    ):
        super().__init__()
        self.sigma_min = 1e-6

        self.estimator = None

        self.in_channels = args.DiT.in_channels

        self.criterion = torch.nn.MSELoss() if args.reg_loss_type == "l2" else torch.nn.L1Loss()

        if hasattr(args.DiT, 'zero_prompt_speech_token'):
            self.zero_prompt_speech_token = args.DiT.zero_prompt_speech_token
        else:
            self.zero_prompt_speech_token = False

    @torch.inference_mode()
    def inference(self, mu, x_lens, prompt, style, f0, n_timesteps, temperature=1.0, inference_cfg_rate=0.5):
        """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)
        """
        B, T = mu.size(0), mu.size(1)
        z = torch.randn([B, self.in_channels, T], device=mu.device) * temperature
        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
        return self.solve_euler(z, x_lens, prompt, mu, style, f0, t_span, inference_cfg_rate)

    def solve_euler(self, x, x_lens, prompt, mu, style, f0, t_span, inference_cfg_rate=0.5):
        """
        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
        """
        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]

        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
        # Or in future might add like a return_all_steps flag
        sol = []
        # apply prompt
        prompt_len = prompt.size(-1)
        prompt_x = torch.zeros_like(x)
        prompt_x[..., :prompt_len] = prompt[..., :prompt_len]
        x[..., :prompt_len] = 0
        if self.zero_prompt_speech_token:
            mu[..., :prompt_len] = 0
        for step in range(1, len(t_span)):
            dphi_dt = self.estimator(x, prompt_x, x_lens, t.unsqueeze(0), style, mu, f0)
            # Classifier-Free Guidance inference introduced in VoiceBox
            if inference_cfg_rate > 0:
                cfg_dphi_dt = self.estimator(
                    x, torch.zeros_like(prompt_x), x_lens, t.unsqueeze(0),
                    torch.zeros_like(style),
                    torch.zeros_like(mu), None
                )
                dphi_dt = ((1.0 + inference_cfg_rate) * dphi_dt -
                           inference_cfg_rate * cfg_dphi_dt)
            x = x + dt * dphi_dt
            t = t + dt
            sol.append(x)
            if step < len(t_span) - 1:
                dt = t_span[step + 1] - t
            x[:, :, :prompt_len] = 0

        return sol[-1]

    def forward(self, x1, x_lens, prompt_lens, mu, style, f0=None):
        """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)
        """
        b, _, t = x1.shape

        # random timestep
        t = torch.rand([b, 1, 1], device=mu.device, dtype=x1.dtype)
        # sample noise p(x_0)
        z = torch.randn_like(x1)

        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
        u = x1 - (1 - self.sigma_min) * z

        prompt = torch.zeros_like(x1)
        for bib in range(b):
            prompt[bib, :, :prompt_lens[bib]] = x1[bib, :, :prompt_lens[bib]]
            # range covered by prompt are set to 0
            y[bib, :, :prompt_lens[bib]] = 0
            if self.zero_prompt_speech_token:
                mu[bib, :, :prompt_lens[bib]] = 0

        estimator_out = self.estimator(y, prompt, x_lens, t.squeeze(), style, mu, f0)
        loss = 0
        for bib in range(b):
            loss += self.criterion(estimator_out[bib, :, prompt_lens[bib]:x_lens[bib]], u[bib, :, prompt_lens[bib]:x_lens[bib]])
        loss /= b

        return loss, y



class CFM(BASECFM):
    def __init__(self, args):
        super().__init__(
            args
        )
        if args.dit_type == "DiT":
            self.estimator = DiT(args)
        else:
            raise NotImplementedError(f"Unknown diffusion type {args.dit_type}")