chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,18 @@
|
||||
from utils import hparams
|
||||
|
||||
from .pm import ParselmouthPE
|
||||
from .pw import HarvestPE
|
||||
from .rmvpe import RMVPE
|
||||
|
||||
|
||||
def initialize_pe():
|
||||
pe = hparams['pe']
|
||||
pe_ckpt = hparams['pe_ckpt']
|
||||
if pe == 'parselmouth':
|
||||
return ParselmouthPE()
|
||||
elif pe == 'rmvpe':
|
||||
return RMVPE(pe_ckpt)
|
||||
elif pe == 'harvest':
|
||||
return HarvestPE()
|
||||
else:
|
||||
raise ValueError(f" [x] Unknown f0 extractor: {pe}")
|
||||
@@ -0,0 +1,15 @@
|
||||
from basics.base_pe import BasePE
|
||||
from utils.binarizer_utils import get_pitch_parselmouth
|
||||
|
||||
|
||||
class ParselmouthPE(BasePE):
|
||||
def get_pitch(
|
||||
self,waveform, samplerate, length,
|
||||
*, hop_size, f0_min=65, f0_max=1100,
|
||||
speed=1, interp_uv=False
|
||||
):
|
||||
return get_pitch_parselmouth(
|
||||
waveform, samplerate=samplerate, length=length,
|
||||
hop_size=hop_size, f0_min=f0_min, f0_max=f0_max,
|
||||
speed=speed, interp_uv=interp_uv
|
||||
)
|
||||
@@ -0,0 +1,29 @@
|
||||
from basics.base_pe import BasePE
|
||||
import numpy as np
|
||||
import pyworld as pw
|
||||
from utils.pitch_utils import interp_f0
|
||||
|
||||
|
||||
class HarvestPE(BasePE):
|
||||
def get_pitch(
|
||||
self, waveform, samplerate, length,
|
||||
*, hop_size, f0_min=65, f0_max=1100,
|
||||
speed=1, interp_uv=False
|
||||
):
|
||||
hop_size = int(np.round(hop_size * speed))
|
||||
time_step = 1000 * hop_size / samplerate
|
||||
|
||||
f0, _ = pw.harvest(
|
||||
waveform.astype(np.float64), samplerate,
|
||||
f0_floor=f0_min, f0_ceil=f0_max, frame_period=time_step
|
||||
)
|
||||
f0 = f0.astype(np.float32)
|
||||
|
||||
if f0.size < length:
|
||||
f0 = np.pad(f0, (0, length - f0.size))
|
||||
f0 = f0[:length]
|
||||
uv = f0 == 0
|
||||
|
||||
if interp_uv:
|
||||
f0, uv = interp_f0(f0, uv)
|
||||
return f0, uv
|
||||
@@ -0,0 +1,5 @@
|
||||
from .constants import *
|
||||
from .model import E2E0
|
||||
from .utils import to_local_average_f0, to_viterbi_f0
|
||||
from .inference import RMVPE
|
||||
from .spec import MelSpectrogram
|
||||
@@ -0,0 +1,9 @@
|
||||
SAMPLE_RATE = 16000
|
||||
|
||||
N_CLASS = 360
|
||||
|
||||
N_MELS = 128
|
||||
MEL_FMIN = 30
|
||||
MEL_FMAX = 8000
|
||||
WINDOW_LENGTH = 1024
|
||||
CONST = 1997.3794084376191
|
||||
@@ -0,0 +1,173 @@
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from .constants import N_MELS
|
||||
|
||||
|
||||
class ConvBlockRes(nn.Module):
|
||||
def __init__(self, in_channels, out_channels, momentum=0.01):
|
||||
super(ConvBlockRes, self).__init__()
|
||||
self.conv = nn.Sequential(
|
||||
nn.Conv2d(in_channels=in_channels,
|
||||
out_channels=out_channels,
|
||||
kernel_size=(3, 3),
|
||||
stride=(1, 1),
|
||||
padding=(1, 1),
|
||||
bias=False),
|
||||
nn.BatchNorm2d(out_channels, momentum=momentum),
|
||||
nn.ReLU(),
|
||||
|
||||
nn.Conv2d(in_channels=out_channels,
|
||||
out_channels=out_channels,
|
||||
kernel_size=(3, 3),
|
||||
stride=(1, 1),
|
||||
padding=(1, 1),
|
||||
bias=False),
|
||||
nn.BatchNorm2d(out_channels, momentum=momentum),
|
||||
nn.ReLU(),
|
||||
)
|
||||
if in_channels != out_channels:
|
||||
self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
|
||||
self.is_shortcut = True
|
||||
else:
|
||||
self.is_shortcut = False
|
||||
|
||||
def forward(self, x):
|
||||
if self.is_shortcut:
|
||||
return self.conv(x) + self.shortcut(x)
|
||||
else:
|
||||
return self.conv(x) + x
|
||||
|
||||
|
||||
class ResEncoderBlock(nn.Module):
|
||||
def __init__(self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01):
|
||||
super(ResEncoderBlock, self).__init__()
|
||||
self.n_blocks = n_blocks
|
||||
self.conv = nn.ModuleList()
|
||||
self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
|
||||
for i in range(n_blocks - 1):
|
||||
self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
|
||||
self.kernel_size = kernel_size
|
||||
if self.kernel_size is not None:
|
||||
self.pool = nn.AvgPool2d(kernel_size=kernel_size)
|
||||
|
||||
def forward(self, x):
|
||||
for i in range(self.n_blocks):
|
||||
x = self.conv[i](x)
|
||||
if self.kernel_size is not None:
|
||||
return x, self.pool(x)
|
||||
else:
|
||||
return x
|
||||
|
||||
|
||||
class ResDecoderBlock(nn.Module):
|
||||
def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
|
||||
super(ResDecoderBlock, self).__init__()
|
||||
out_padding = (0, 1) if stride == (1, 2) else (1, 1)
|
||||
self.n_blocks = n_blocks
|
||||
self.conv1 = nn.Sequential(
|
||||
nn.ConvTranspose2d(in_channels=in_channels,
|
||||
out_channels=out_channels,
|
||||
kernel_size=(3, 3),
|
||||
stride=stride,
|
||||
padding=(1, 1),
|
||||
output_padding=out_padding,
|
||||
bias=False),
|
||||
nn.BatchNorm2d(out_channels, momentum=momentum),
|
||||
nn.ReLU(),
|
||||
)
|
||||
self.conv2 = nn.ModuleList()
|
||||
self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
|
||||
for i in range(n_blocks-1):
|
||||
self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
|
||||
|
||||
def forward(self, x, concat_tensor):
|
||||
x = self.conv1(x)
|
||||
x = torch.cat((x, concat_tensor), dim=1)
|
||||
for i in range(self.n_blocks):
|
||||
x = self.conv2[i](x)
|
||||
return x
|
||||
|
||||
|
||||
class Encoder(nn.Module):
|
||||
def __init__(self, in_channels, in_size, n_encoders, kernel_size, n_blocks, out_channels=16, momentum=0.01):
|
||||
super(Encoder, self).__init__()
|
||||
self.n_encoders = n_encoders
|
||||
self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
|
||||
self.layers = nn.ModuleList()
|
||||
self.latent_channels = []
|
||||
for i in range(self.n_encoders):
|
||||
self.layers.append(ResEncoderBlock(in_channels, out_channels, kernel_size, n_blocks, momentum=momentum))
|
||||
self.latent_channels.append([out_channels, in_size])
|
||||
in_channels = out_channels
|
||||
out_channels *= 2
|
||||
in_size //= 2
|
||||
self.out_size = in_size
|
||||
self.out_channel = out_channels
|
||||
|
||||
def forward(self, x):
|
||||
concat_tensors = []
|
||||
x = self.bn(x)
|
||||
for i in range(self.n_encoders):
|
||||
_, x = self.layers[i](x)
|
||||
concat_tensors.append(_)
|
||||
return x, concat_tensors
|
||||
|
||||
|
||||
class Intermediate(nn.Module):
|
||||
def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
|
||||
super(Intermediate, self).__init__()
|
||||
self.n_inters = n_inters
|
||||
self.layers = nn.ModuleList()
|
||||
self.layers.append(ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum))
|
||||
for i in range(self.n_inters-1):
|
||||
self.layers.append(ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum))
|
||||
|
||||
def forward(self, x):
|
||||
for i in range(self.n_inters):
|
||||
x = self.layers[i](x)
|
||||
return x
|
||||
|
||||
|
||||
class Decoder(nn.Module):
|
||||
def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
|
||||
super(Decoder, self).__init__()
|
||||
self.layers = nn.ModuleList()
|
||||
self.n_decoders = n_decoders
|
||||
for i in range(self.n_decoders):
|
||||
out_channels = in_channels // 2
|
||||
self.layers.append(ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum))
|
||||
in_channels = out_channels
|
||||
|
||||
def forward(self, x, concat_tensors):
|
||||
for i in range(self.n_decoders):
|
||||
x = self.layers[i](x, concat_tensors[-1-i])
|
||||
return x
|
||||
|
||||
|
||||
class TimbreFilter(nn.Module):
|
||||
def __init__(self, latent_rep_channels):
|
||||
super(TimbreFilter, self).__init__()
|
||||
self.layers = nn.ModuleList()
|
||||
for latent_rep in latent_rep_channels:
|
||||
self.layers.append(ConvBlockRes(latent_rep[0], latent_rep[0]))
|
||||
|
||||
def forward(self, x_tensors):
|
||||
out_tensors = []
|
||||
for i, layer in enumerate(self.layers):
|
||||
out_tensors.append(layer(x_tensors[i]))
|
||||
return out_tensors
|
||||
|
||||
|
||||
class DeepUnet0(nn.Module):
|
||||
def __init__(self, kernel_size, n_blocks, en_de_layers=5, inter_layers=4, in_channels=1, en_out_channels=16):
|
||||
super(DeepUnet0, self).__init__()
|
||||
self.encoder = Encoder(in_channels, N_MELS, en_de_layers, kernel_size, n_blocks, en_out_channels)
|
||||
self.intermediate = Intermediate(self.encoder.out_channel // 2, self.encoder.out_channel, inter_layers, n_blocks)
|
||||
self.tf = TimbreFilter(self.encoder.latent_channels)
|
||||
self.decoder = Decoder(self.encoder.out_channel, en_de_layers, kernel_size, n_blocks)
|
||||
|
||||
def forward(self, x):
|
||||
x, concat_tensors = self.encoder(x)
|
||||
x = self.intermediate(x)
|
||||
x = self.decoder(x, concat_tensors)
|
||||
return x
|
||||
@@ -0,0 +1,78 @@
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torchaudio.transforms import Resample
|
||||
|
||||
from basics.base_pe import BasePE
|
||||
from utils.infer_utils import resample_align_curve
|
||||
from utils.pitch_utils import interp_f0
|
||||
from .constants import *
|
||||
from .model import E2E0
|
||||
from .spec import MelSpectrogram
|
||||
from .utils import to_local_average_f0, to_viterbi_f0
|
||||
|
||||
|
||||
class RMVPE(BasePE):
|
||||
def __init__(self, model_path, hop_length=160):
|
||||
self.resample_kernel = {}
|
||||
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
||||
self.model = E2E0(4, 1, (2, 2)).eval().to(self.device)
|
||||
ckpt = torch.load(model_path, map_location=self.device)
|
||||
self.model.load_state_dict(ckpt['model'], strict=False)
|
||||
self.hop_length = hop_length
|
||||
self.seg_length = 32 * hop_length
|
||||
self.mel_extractor = MelSpectrogram(
|
||||
N_MELS, SAMPLE_RATE, WINDOW_LENGTH, hop_length, None, MEL_FMIN, MEL_FMAX
|
||||
).to(self.device)
|
||||
|
||||
@torch.no_grad()
|
||||
def mel2hidden(self, mel):
|
||||
n_frames = mel.shape[-1]
|
||||
mel = F.pad(mel, (0, 32 * ((n_frames - 1) // 32 + 1) - n_frames), mode='reflect')
|
||||
hidden = self.model(mel)
|
||||
return hidden[:, :n_frames]
|
||||
|
||||
def decode(self, hidden, thred=0.03, use_viterbi=False):
|
||||
if use_viterbi:
|
||||
f0 = to_viterbi_f0(hidden, thred=thred)
|
||||
else:
|
||||
f0 = to_local_average_f0(hidden, thred=thred)
|
||||
return f0
|
||||
|
||||
def infer_from_audio(self, audio, sample_rate=16000, thred=0.03, use_viterbi=False):
|
||||
audio = torch.from_numpy(audio).float().unsqueeze(0).to(self.device)
|
||||
if sample_rate == 16000:
|
||||
audio_res = audio
|
||||
else:
|
||||
key_str = str(sample_rate)
|
||||
if key_str not in self.resample_kernel:
|
||||
self.resample_kernel[key_str] = Resample(sample_rate, 16000, lowpass_filter_width=128)
|
||||
self.resample_kernel[key_str] = self.resample_kernel[key_str].to(self.device)
|
||||
audio_res = self.resample_kernel[key_str](audio)
|
||||
B, T = audio_res.shape
|
||||
n_frames = T // self.hop_length + 1
|
||||
T1 = T + self.hop_length
|
||||
T_pad = self.seg_length * ((T1 - 1) // self.seg_length + 1) - T1
|
||||
audio_res = F.pad(audio_res, (0, T_pad))
|
||||
mel = self.mel_extractor(audio_res, center=True)
|
||||
with torch.no_grad():
|
||||
hidden = self.model(mel)
|
||||
f0 = self.decode(hidden[:, :n_frames], thred=thred, use_viterbi=use_viterbi)
|
||||
return f0
|
||||
|
||||
def get_pitch(
|
||||
self, waveform, samplerate, length,
|
||||
*, hop_size, f0_min=65, f0_max=1100,
|
||||
speed=1, interp_uv=False
|
||||
):
|
||||
f0 = self.infer_from_audio(waveform, sample_rate=samplerate)
|
||||
uv = f0 == 0
|
||||
f0, uv = interp_f0(f0, uv)
|
||||
|
||||
hop_size = int(np.round(hop_size * speed))
|
||||
time_step = hop_size / samplerate
|
||||
f0_res = resample_align_curve(f0, 0.01, time_step, length)
|
||||
uv_res = resample_align_curve(uv.astype(np.float32), 0.01, time_step, length) > 0.5
|
||||
if not interp_uv:
|
||||
f0_res[uv_res] = 0
|
||||
return f0_res, uv_res
|
||||
@@ -0,0 +1,32 @@
|
||||
from torch import nn
|
||||
|
||||
from .constants import *
|
||||
from .deepunet import DeepUnet0
|
||||
from .seq import BiGRU
|
||||
|
||||
|
||||
class E2E0(nn.Module):
|
||||
def __init__(self, n_blocks, n_gru, kernel_size, en_de_layers=5, inter_layers=4, in_channels=1,
|
||||
en_out_channels=16):
|
||||
super(E2E0, self).__init__()
|
||||
self.unet = DeepUnet0(kernel_size, n_blocks, en_de_layers, inter_layers, in_channels, en_out_channels)
|
||||
self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
|
||||
if n_gru:
|
||||
self.fc = nn.Sequential(
|
||||
BiGRU(3 * N_MELS, 256, n_gru),
|
||||
nn.Linear(512, N_CLASS),
|
||||
nn.Dropout(0.25),
|
||||
nn.Sigmoid()
|
||||
)
|
||||
else:
|
||||
self.fc = nn.Sequential(
|
||||
nn.Linear(3 * N_MELS, N_CLASS),
|
||||
nn.Dropout(0.25),
|
||||
nn.Sigmoid()
|
||||
)
|
||||
|
||||
def forward(self, mel):
|
||||
mel = mel.transpose(-1, -2).unsqueeze(1)
|
||||
x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
|
||||
x = self.fc(x)
|
||||
return x
|
||||
@@ -0,0 +1,10 @@
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
class BiGRU(nn.Module):
|
||||
def __init__(self, input_features, hidden_features, num_layers):
|
||||
super(BiGRU, self).__init__()
|
||||
self.gru = nn.GRU(input_features, hidden_features, num_layers=num_layers, batch_first=True, bidirectional=True)
|
||||
|
||||
def forward(self, x):
|
||||
return self.gru(x)[0]
|
||||
@@ -0,0 +1,68 @@
|
||||
import torch
|
||||
import numpy as np
|
||||
import torch.nn.functional as F
|
||||
from librosa.filters import mel
|
||||
|
||||
|
||||
class MelSpectrogram(torch.nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
n_mel_channels,
|
||||
sampling_rate,
|
||||
win_length,
|
||||
hop_length,
|
||||
n_fft=None,
|
||||
mel_fmin=0,
|
||||
mel_fmax=None,
|
||||
clamp=1e-5
|
||||
):
|
||||
super().__init__()
|
||||
n_fft = win_length if n_fft is None else n_fft
|
||||
self.hann_window = {}
|
||||
mel_basis = mel(
|
||||
sr=sampling_rate,
|
||||
n_fft=n_fft,
|
||||
n_mels=n_mel_channels,
|
||||
fmin=mel_fmin,
|
||||
fmax=mel_fmax,
|
||||
htk=True)
|
||||
mel_basis = torch.from_numpy(mel_basis).float()
|
||||
self.register_buffer("mel_basis", mel_basis)
|
||||
self.n_fft = win_length if n_fft is None else n_fft
|
||||
self.hop_length = hop_length
|
||||
self.win_length = win_length
|
||||
self.sampling_rate = sampling_rate
|
||||
self.n_mel_channels = n_mel_channels
|
||||
self.clamp = clamp
|
||||
|
||||
def forward(self, audio, keyshift=0, speed=1, center=True):
|
||||
factor = 2 ** (keyshift / 12)
|
||||
n_fft_new = int(np.round(self.n_fft * factor))
|
||||
win_length_new = int(np.round(self.win_length * factor))
|
||||
hop_length_new = int(np.round(self.hop_length * speed))
|
||||
|
||||
keyshift_key = str(keyshift) + '_' + str(audio.device)
|
||||
if keyshift_key not in self.hann_window:
|
||||
self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(audio.device)
|
||||
|
||||
fft = torch.stft(
|
||||
audio,
|
||||
n_fft=n_fft_new,
|
||||
hop_length=hop_length_new,
|
||||
win_length=win_length_new,
|
||||
window=self.hann_window[keyshift_key],
|
||||
center=center,
|
||||
return_complex=True
|
||||
)
|
||||
magnitude = fft.abs()
|
||||
|
||||
if keyshift != 0:
|
||||
size = self.n_fft // 2 + 1
|
||||
resize = magnitude.size(1)
|
||||
if resize < size:
|
||||
magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
|
||||
magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
|
||||
|
||||
mel_output = torch.matmul(self.mel_basis, magnitude)
|
||||
log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
|
||||
return log_mel_spec
|
||||
@@ -0,0 +1,43 @@
|
||||
import librosa
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
from .constants import *
|
||||
|
||||
|
||||
def to_local_average_f0(hidden, center=None, thred=0.03):
|
||||
idx = torch.arange(N_CLASS, device=hidden.device)[None, None, :] # [B=1, T=1, N]
|
||||
idx_cents = idx * 20 + CONST # [B=1, N]
|
||||
if center is None:
|
||||
center = torch.argmax(hidden, dim=2, keepdim=True) # [B, T, 1]
|
||||
start = torch.clip(center - 4, min=0) # [B, T, 1]
|
||||
end = torch.clip(center + 5, max=N_CLASS) # [B, T, 1]
|
||||
idx_mask = (idx >= start) & (idx < end) # [B, T, N]
|
||||
weights = hidden * idx_mask # [B, T, N]
|
||||
product_sum = torch.sum(weights * idx_cents, dim=2) # [B, T]
|
||||
weight_sum = torch.sum(weights, dim=2) # [B, T]
|
||||
cents = product_sum / (weight_sum + (weight_sum == 0)) # avoid dividing by zero, [B, T]
|
||||
f0 = 10 * 2 ** (cents / 1200)
|
||||
uv = hidden.max(dim=2)[0] < thred # [B, T]
|
||||
f0 = f0 * ~uv
|
||||
return f0.squeeze(0).cpu().numpy()
|
||||
|
||||
|
||||
def to_viterbi_f0(hidden, thred=0.03):
|
||||
# Create viterbi transition matrix
|
||||
if not hasattr(to_viterbi_f0, 'transition'):
|
||||
xx, yy = np.meshgrid(range(N_CLASS), range(N_CLASS))
|
||||
transition = np.maximum(30 - abs(xx - yy), 0)
|
||||
transition = transition / transition.sum(axis=1, keepdims=True)
|
||||
to_viterbi_f0.transition = transition
|
||||
|
||||
# Convert to probability
|
||||
prob = hidden.squeeze(0).cpu().numpy()
|
||||
prob = prob.T
|
||||
prob = prob / prob.sum(axis=0)
|
||||
|
||||
# Perform viterbi decoding
|
||||
path = librosa.sequence.viterbi(prob, to_viterbi_f0.transition).astype(np.int64)
|
||||
center = torch.from_numpy(path).unsqueeze(0).unsqueeze(-1).to(hidden.device)
|
||||
|
||||
return to_local_average_f0(hidden, center=center, thred=thred)
|
||||
Reference in New Issue
Block a user