File size: 5,620 Bytes
ad16788 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 |
from distutils.version import LooseVersion
from typing import Optional
from typing import Tuple
from typing import Union
import torch
from torch_complex.tensor import ComplexTensor
from typeguard import check_argument_types
from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet2.layers.inversible_interface import InversibleInterface
class Stft(torch.nn.Module, InversibleInterface):
def __init__(
self,
n_fft: int = 512,
win_length: int = None,
hop_length: int = 128,
window: Optional[str] = "hann",
center: bool = True,
normalized: bool = False,
onesided: bool = True,
):
assert check_argument_types()
super().__init__()
self.n_fft = n_fft
if win_length is None:
self.win_length = n_fft
else:
self.win_length = win_length
self.hop_length = hop_length
self.center = center
self.normalized = normalized
self.onesided = onesided
if window is not None and not hasattr(torch, f"{window}_window"):
raise ValueError(f"{window} window is not implemented")
self.window = window
def extra_repr(self):
return (
f"n_fft={self.n_fft}, "
f"win_length={self.win_length}, "
f"hop_length={self.hop_length}, "
f"center={self.center}, "
f"normalized={self.normalized}, "
f"onesided={self.onesided}"
)
def forward(
self, input: torch.Tensor, ilens: torch.Tensor = None
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""STFT forward function.
Args:
input: (Batch, Nsamples) or (Batch, Nsample, Channels)
ilens: (Batch)
Returns:
output: (Batch, Frames, Freq, 2) or (Batch, Frames, Channels, Freq, 2)
"""
bs = input.size(0)
if input.dim() == 3:
multi_channel = True
# input: (Batch, Nsample, Channels) -> (Batch * Channels, Nsample)
input = input.transpose(1, 2).reshape(-1, input.size(1))
else:
multi_channel = False
# NOTE(kamo):
# The default behaviour of torch.stft is compatible with librosa.stft
# about padding and scaling.
# Note that it's different from scipy.signal.stft
# output: (Batch, Freq, Frames, 2=real_imag)
# or (Batch, Channel, Freq, Frames, 2=real_imag)
if self.window is not None:
window_func = getattr(torch, f"{self.window}_window")
window = window_func(
self.win_length, dtype=input.dtype, device=input.device
)
else:
window = None
output = torch.stft(
input,
n_fft=self.n_fft,
win_length=self.win_length,
hop_length=self.hop_length,
center=self.center,
window=window,
normalized=self.normalized,
onesided=self.onesided,
)
# output: (Batch, Freq, Frames, 2=real_imag)
# -> (Batch, Frames, Freq, 2=real_imag)
output = output.transpose(1, 2)
if multi_channel:
# output: (Batch * Channel, Frames, Freq, 2=real_imag)
# -> (Batch, Frame, Channel, Freq, 2=real_imag)
output = output.view(bs, -1, output.size(1), output.size(2), 2).transpose(
1, 2
)
if ilens is not None:
if self.center:
pad = self.win_length // 2
ilens = ilens + 2 * pad
olens = (ilens - self.win_length) // self.hop_length + 1
output.masked_fill_(make_pad_mask(olens, output, 1), 0.0)
else:
olens = None
return output, olens
def inverse(
self, input: Union[torch.Tensor, ComplexTensor], ilens: torch.Tensor = None
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
"""Inverse STFT.
Args:
input: Tensor(batch, T, F, 2) or ComplexTensor(batch, T, F)
ilens: (batch,)
Returns:
wavs: (batch, samples)
ilens: (batch,)
"""
if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
istft = torch.functional.istft
else:
try:
import torchaudio
except ImportError:
raise ImportError(
"Please install torchaudio>=0.3.0 or use torch>=1.6.0"
)
if not hasattr(torchaudio.functional, "istft"):
raise ImportError(
"Please install torchaudio>=0.3.0 or use torch>=1.6.0"
)
istft = torchaudio.functional.istft
if self.window is not None:
window_func = getattr(torch, f"{self.window}_window")
window = window_func(
self.win_length, dtype=input.dtype, device=input.device
)
else:
window = None
if isinstance(input, ComplexTensor):
input = torch.stack([input.real, input.imag], dim=-1)
assert input.shape[-1] == 2
input = input.transpose(1, 2)
wavs = istft(
input,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=self.win_length,
window=window,
center=self.center,
normalized=self.normalized,
onesided=self.onesided,
length=ilens.max() if ilens is not None else ilens,
)
return wavs, ilens
|