espnet3.systems.f5tts.src.solvers.odeint

source

espnet3.systems.f5tts.src.solvers.odeint(func: Callable, y0: Tensor, t: Tensor, method: str = 'euler', **kwargs)

Fixed-step ODE integration over a caller-supplied time grid.

• Parameters:
  • func – Derivative func(t_scalar, y) -> dy/dt.
  • y0 – Initial state.
  • t – 1-D time grid. Steps use dt = t[i + 1] - t[i], so non-uniform grids (EPSS, sway sampling) work unchanged.
  • method – "euler" or "midpoint". No other method is supported.
  • **kwargs – Accepted and ignored, so the torchdiffeq.odeint call signature stays a drop-in.
• Returns: Trajectory stacked as [len(t), *y0.shape] with solution[0] == y0, matching torchdiffeq.odeint, so solution[-1] is the final state.
• Raises:ValueError – If method is not "euler" or "midpoint".

Example

>>> sol = odeint(lambda t, y: y, torch.tensor([1.0]),
...              torch.linspace(0, 1, 3))
>>> sol.shape
torch.Size([3, 1])

NOTE

Both methods reproduce torchdiffeq’s fixed-step solvers exactly, so dropping that dependency changes no output. Returning the whole trajectory rather than just the endpoint is deliberate: it keeps the torchdiffeq signature, so this stays a drop-in replacement.