Spaces:
Runtime error
Runtime error
File size: 4,005 Bytes
0047e35 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 |
from typing import Union
import numpy as np
import torch
import torch.nn.functional as F
from modules.F0Predictor.F0Predictor import F0Predictor
from .fcpe.model import FCPEInfer
class FCPEF0Predictor(F0Predictor):
def __init__(self, hop_length=512, f0_min=50, f0_max=1100, dtype=torch.float32, device=None, sampling_rate=44100,
threshold=0.05):
self.fcpe = FCPEInfer(model_path="pretrain/fcpe.pt", device=device, dtype=dtype)
self.hop_length = hop_length
self.f0_min = f0_min
self.f0_max = f0_max
if device is None:
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
else:
self.device = device
self.threshold = threshold
self.sampling_rate = sampling_rate
self.dtype = dtype
self.name = "fcpe"
def repeat_expand(
self, content: Union[torch.Tensor, np.ndarray], target_len: int, mode: str = "nearest"
):
ndim = content.ndim
if content.ndim == 1:
content = content[None, None]
elif content.ndim == 2:
content = content[None]
assert content.ndim == 3
is_np = isinstance(content, np.ndarray)
if is_np:
content = torch.from_numpy(content)
results = torch.nn.functional.interpolate(content, size=target_len, mode=mode)
if is_np:
results = results.numpy()
if ndim == 1:
return results[0, 0]
elif ndim == 2:
return results[0]
def post_process(self, x, sampling_rate, f0, pad_to):
if isinstance(f0, np.ndarray):
f0 = torch.from_numpy(f0).float().to(x.device)
if pad_to is None:
return f0
f0 = self.repeat_expand(f0, pad_to)
vuv_vector = torch.zeros_like(f0)
vuv_vector[f0 > 0.0] = 1.0
vuv_vector[f0 <= 0.0] = 0.0
# 去掉0频率, 并线性插值
nzindex = torch.nonzero(f0).squeeze()
f0 = torch.index_select(f0, dim=0, index=nzindex).cpu().numpy()
time_org = self.hop_length / sampling_rate * nzindex.cpu().numpy()
time_frame = np.arange(pad_to) * self.hop_length / sampling_rate
vuv_vector = F.interpolate(vuv_vector[None, None, :], size=pad_to)[0][0]
if f0.shape[0] <= 0:
return torch.zeros(pad_to, dtype=torch.float, device=x.device).cpu().numpy(), vuv_vector.cpu().numpy()
if f0.shape[0] == 1:
return (torch.ones(pad_to, dtype=torch.float, device=x.device) * f0[
0]).cpu().numpy(), vuv_vector.cpu().numpy()
# 大概可以用 torch 重写?
f0 = np.interp(time_frame, time_org, f0, left=f0[0], right=f0[-1])
# vuv_vector = np.ceil(scipy.ndimage.zoom(vuv_vector,pad_to/len(vuv_vector),order = 0))
return f0, vuv_vector.cpu().numpy()
def compute_f0(self, wav, p_len=None):
x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
if p_len is None:
p_len = x.shape[0] // self.hop_length
else:
assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
f0 = self.fcpe(x, sr=self.sampling_rate, threshold=self.threshold)[0,:,0]
if torch.all(f0 == 0):
rtn = f0.cpu().numpy() if p_len is None else np.zeros(p_len)
return rtn, rtn
return self.post_process(x, self.sampling_rate, f0, p_len)[0]
def compute_f0_uv(self, wav, p_len=None):
x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
if p_len is None:
p_len = x.shape[0] // self.hop_length
else:
assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
f0 = self.fcpe(x, sr=self.sampling_rate, threshold=self.threshold)[0,:,0]
if torch.all(f0 == 0):
rtn = f0.cpu().numpy() if p_len is None else np.zeros(p_len)
return rtn, rtn
return self.post_process(x, self.sampling_rate, f0, p_len) |