Spaces:
Runtime error
Runtime error
File size: 7,929 Bytes
c968fc3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 |
# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import numpy as np
import torch
import torch.nn as nn
from torchaudio.models import Conformer
from models.svc.transformer.transformer import PositionalEncoding
from utils.f0 import f0_to_coarse
class ContentEncoder(nn.Module):
def __init__(self, cfg, input_dim, output_dim):
super().__init__()
self.cfg = cfg
assert input_dim != 0
self.nn = nn.Linear(input_dim, output_dim)
# Introduce conformer or not
if (
"use_conformer_for_content_features" in cfg
and cfg.use_conformer_for_content_features
):
self.pos_encoder = PositionalEncoding(input_dim)
self.conformer = Conformer(
input_dim=input_dim,
num_heads=2,
ffn_dim=256,
num_layers=6,
depthwise_conv_kernel_size=3,
)
else:
self.conformer = None
def forward(self, x, length=None):
# x: (N, seq_len, input_dim) -> (N, seq_len, output_dim)
if self.conformer:
x = self.pos_encoder(x)
x, _ = self.conformer(x, length)
return self.nn(x)
class MelodyEncoder(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
self.input_dim = self.cfg.input_melody_dim
self.output_dim = self.cfg.output_melody_dim
self.n_bins = self.cfg.n_bins_melody
if self.input_dim != 0:
if self.n_bins == 0:
# Not use quantization
self.nn = nn.Linear(self.input_dim, self.output_dim)
else:
self.f0_min = cfg.f0_min
self.f0_max = cfg.f0_max
self.nn = nn.Embedding(
num_embeddings=self.n_bins,
embedding_dim=self.output_dim,
padding_idx=None,
)
self.uv_embedding = nn.Embedding(2, self.output_dim)
def forward(self, x, uv=None, length=None):
# x: (B, frame_len)
if self.n_bins == 0:
x = x.unsqueeze(-1)
else:
x = f0_to_coarse(x, self.n_bins, self.f0_min, self.f0_max)
x = self.nn(x)
if self.cfg.use_uv:
uv = self.uv_embedding(uv)
x = x + uv
return x
class LoudnessEncoder(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
self.input_dim = self.cfg.input_loudness_dim
self.output_dim = self.cfg.output_loudness_dim
self.n_bins = self.cfg.n_bins_loudness
if self.input_dim != 0:
if self.n_bins == 0:
# Not use quantization
self.nn = nn.Linear(self.input_dim, self.output_dim)
else:
# TODO: set empirically now
self.loudness_min = 1e-30
self.loudness_max = 1.5
self.energy_bins = nn.Parameter(
torch.exp(
torch.linspace(
np.log(self.loudness_min),
np.log(self.loudness_max),
self.n_bins - 1,
)
),
requires_grad=False,
)
self.nn = nn.Embedding(
num_embeddings=self.n_bins,
embedding_dim=self.output_dim,
padding_idx=None,
)
def forward(self, x):
# x: (N, frame_len)
if self.n_bins == 0:
x = x.unsqueeze(-1)
else:
x = torch.bucketize(x, self.energy_bins)
return self.nn(x)
class SingerEncoder(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
self.input_dim = 1
self.output_dim = self.cfg.output_singer_dim
self.nn = nn.Embedding(
num_embeddings=cfg.singer_table_size,
embedding_dim=self.output_dim,
padding_idx=None,
)
def forward(self, x):
# x: (N, 1) -> (N, 1, output_dim)
return self.nn(x)
class ConditionEncoder(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
self.merge_mode = cfg.merge_mode
### Semantic Features ###
if cfg.use_whisper:
self.whisper_encoder = ContentEncoder(
self.cfg, self.cfg.whisper_dim, self.cfg.content_encoder_dim
)
if cfg.use_contentvec:
self.contentvec_encoder = ContentEncoder(
self.cfg, self.cfg.contentvec_dim, self.cfg.content_encoder_dim
)
if cfg.use_mert:
self.mert_encoder = ContentEncoder(
self.cfg, self.cfg.mert_dim, self.cfg.content_encoder_dim
)
if cfg.use_wenet:
self.wenet_encoder = ContentEncoder(
self.cfg, self.cfg.wenet_dim, self.cfg.content_encoder_dim
)
### Prosody Features ###
if cfg.use_f0:
self.melody_encoder = MelodyEncoder(self.cfg)
if cfg.use_energy:
self.loudness_encoder = LoudnessEncoder(self.cfg)
### Speaker Features ###
if cfg.use_spkid:
self.singer_encoder = SingerEncoder(self.cfg)
def forward(self, x):
outputs = []
if self.cfg.use_f0:
if self.cfg.use_uv:
pitch_enc_out = self.melody_encoder(
x["frame_pitch"], uv=x["frame_uv"], length=x["target_len"]
)
else:
pitch_enc_out = self.melody_encoder(
x["frame_pitch"], uv=None, length=x["target_len"]
)
outputs.append(pitch_enc_out)
if self.cfg.use_energy:
loudness_enc_out = self.loudness_encoder(x["frame_energy"])
outputs.append(loudness_enc_out)
if self.cfg.use_whisper:
# whisper_feat: [b, T, 1024]
whiser_enc_out = self.whisper_encoder(
x["whisper_feat"], length=x["target_len"]
)
outputs.append(whiser_enc_out)
seq_len = whiser_enc_out.shape[1]
if self.cfg.use_contentvec:
contentvec_enc_out = self.contentvec_encoder(
x["contentvec_feat"], length=x["target_len"]
)
outputs.append(contentvec_enc_out)
seq_len = contentvec_enc_out.shape[1]
if self.cfg.use_mert:
mert_enc_out = self.mert_encoder(x["mert_feat"], length=x["target_len"])
outputs.append(mert_enc_out)
seq_len = mert_enc_out.shape[1]
if self.cfg.use_wenet:
wenet_enc_out = self.wenet_encoder(x["wenet_feat"], length=x["target_len"])
outputs.append(wenet_enc_out)
seq_len = wenet_enc_out.shape[1]
if self.cfg.use_spkid:
speaker_enc_out = self.singer_encoder(x["spk_id"]) # [b, 1, 384]
assert (
"whisper_feat" in x.keys()
or "contentvec_feat" in x.keys()
or "mert_feat" in x.keys()
or "wenet_feat" in x.keys()
)
singer_info = speaker_enc_out.expand(-1, seq_len, -1)
outputs.append(singer_info)
encoder_output = None
if self.merge_mode == "concat":
encoder_output = torch.cat(outputs, dim=-1)
if self.merge_mode == "add":
# (#modules, N, seq_len, output_dim)
outputs = torch.cat([out[None, :, :, :] for out in outputs], dim=0)
# (N, seq_len, output_dim)
encoder_output = torch.sum(outputs, dim=0)
return encoder_output
|