Spaces:
Running
Running
File size: 9,091 Bytes
1f4e6d7 |
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 |
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import yaml
from diffusion_onnx import GaussianDiffusion
class DotDict(dict):
def __getattr__(*args):
val = dict.get(*args)
return DotDict(val) if type(val) is dict else val
__setattr__ = dict.__setitem__
__delattr__ = dict.__delitem__
def load_model_vocoder(
model_path,
device='cpu'):
config_file = os.path.join(os.path.split(model_path)[0], 'config.yaml')
with open(config_file, "r") as config:
args = yaml.safe_load(config)
args = DotDict(args)
# load model
model = Unit2Mel(
args.data.encoder_out_channels,
args.model.n_spk,
args.model.use_pitch_aug,
128,
args.model.n_layers,
args.model.n_chans,
args.model.n_hidden,
args.model.timesteps,
args.model.k_step_max)
print(' [Loading] ' + model_path)
ckpt = torch.load(model_path, map_location=torch.device(device))
model.to(device)
model.load_state_dict(ckpt['model'])
model.eval()
return model, args
class Unit2Mel(nn.Module):
def __init__(
self,
input_channel,
n_spk,
use_pitch_aug=False,
out_dims=128,
n_layers=20,
n_chans=384,
n_hidden=256,
timesteps=1000,
k_step_max=1000):
super().__init__()
self.unit_embed = nn.Linear(input_channel, n_hidden)
self.f0_embed = nn.Linear(1, n_hidden)
self.volume_embed = nn.Linear(1, n_hidden)
if use_pitch_aug:
self.aug_shift_embed = nn.Linear(1, n_hidden, bias=False)
else:
self.aug_shift_embed = None
self.n_spk = n_spk
if n_spk is not None and n_spk > 1:
self.spk_embed = nn.Embedding(n_spk, n_hidden)
self.timesteps = timesteps if timesteps is not None else 1000
self.k_step_max = k_step_max if k_step_max is not None and k_step_max>0 and k_step_max<self.timesteps else self.timesteps
# diffusion
self.decoder = GaussianDiffusion(out_dims, n_layers, n_chans, n_hidden,self.timesteps,self.k_step_max)
self.hidden_size = n_hidden
self.speaker_map = torch.zeros((self.n_spk,1,1,n_hidden))
def forward(self, units, mel2ph, f0, volume, g = None):
'''
input:
B x n_frames x n_unit
return:
dict of B x n_frames x feat
'''
decoder_inp = F.pad(units, [0, 0, 1, 0])
mel2ph_ = mel2ph.unsqueeze(2).repeat([1, 1, units.shape[-1]])
units = torch.gather(decoder_inp, 1, mel2ph_) # [B, T, H]
x = self.unit_embed(units) + self.f0_embed((1 + f0.unsqueeze(-1) / 700).log()) + self.volume_embed(volume.unsqueeze(-1))
if self.n_spk is not None and self.n_spk > 1: # [N, S] * [S, B, 1, H]
g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1)) # [N, S, B, 1, 1]
g = g * self.speaker_map # [N, S, B, 1, H]
g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
x = x.transpose(1, 2) + g
return x
else:
return x.transpose(1, 2)
def init_spkembed(self, units, f0, volume, spk_id = None, spk_mix_dict = None, aug_shift = None,
gt_spec=None, infer=True, infer_speedup=10, method='dpm-solver', k_step=300, use_tqdm=True):
'''
input:
B x n_frames x n_unit
return:
dict of B x n_frames x feat
'''
x = self.unit_embed(units) + self.f0_embed((1+ f0 / 700).log()) + self.volume_embed(volume)
if self.n_spk is not None and self.n_spk > 1:
if spk_mix_dict is not None:
spk_embed_mix = torch.zeros((1,1,self.hidden_size))
for k, v in spk_mix_dict.items():
spk_id_torch = torch.LongTensor(np.array([[k]])).to(units.device)
spk_embeddd = self.spk_embed(spk_id_torch)
self.speaker_map[k] = spk_embeddd
spk_embed_mix = spk_embed_mix + v * spk_embeddd
x = x + spk_embed_mix
else:
x = x + self.spk_embed(spk_id - 1)
self.speaker_map = self.speaker_map.unsqueeze(0)
self.speaker_map = self.speaker_map.detach()
return x.transpose(1, 2)
def OnnxExport(self, project_name=None, init_noise=None, export_encoder=True, export_denoise=True, export_pred=True, export_after=True):
hubert_hidden_size = 768
n_frames = 100
hubert = torch.randn((1, n_frames, hubert_hidden_size))
mel2ph = torch.arange(end=n_frames).unsqueeze(0).long()
f0 = torch.randn((1, n_frames))
volume = torch.randn((1, n_frames))
spk_mix = []
spks = {}
if self.n_spk is not None and self.n_spk > 1:
for i in range(self.n_spk):
spk_mix.append(1.0/float(self.n_spk))
spks.update({i:1.0/float(self.n_spk)})
spk_mix = torch.tensor(spk_mix)
spk_mix = spk_mix.repeat(n_frames, 1)
self.init_spkembed(hubert, f0.unsqueeze(-1), volume.unsqueeze(-1), spk_mix_dict=spks)
self.forward(hubert, mel2ph, f0, volume, spk_mix)
if export_encoder:
torch.onnx.export(
self,
(hubert, mel2ph, f0, volume, spk_mix),
f"{project_name}_encoder.onnx",
input_names=["hubert", "mel2ph", "f0", "volume", "spk_mix"],
output_names=["mel_pred"],
dynamic_axes={
"hubert": [1],
"f0": [1],
"volume": [1],
"mel2ph": [1],
"spk_mix": [0],
},
opset_version=16
)
self.decoder.OnnxExport(project_name, init_noise=init_noise, export_denoise=export_denoise, export_pred=export_pred, export_after=export_after)
def ExportOnnx(self, project_name=None):
hubert_hidden_size = 768
n_frames = 100
hubert = torch.randn((1, n_frames, hubert_hidden_size))
mel2ph = torch.arange(end=n_frames).unsqueeze(0).long()
f0 = torch.randn((1, n_frames))
volume = torch.randn((1, n_frames))
spk_mix = []
spks = {}
if self.n_spk is not None and self.n_spk > 1:
for i in range(self.n_spk):
spk_mix.append(1.0/float(self.n_spk))
spks.update({i:1.0/float(self.n_spk)})
spk_mix = torch.tensor(spk_mix)
self.orgforward(hubert, f0.unsqueeze(-1), volume.unsqueeze(-1), spk_mix_dict=spks)
self.forward(hubert, mel2ph, f0, volume, spk_mix)
torch.onnx.export(
self,
(hubert, mel2ph, f0, volume, spk_mix),
f"{project_name}_encoder.onnx",
input_names=["hubert", "mel2ph", "f0", "volume", "spk_mix"],
output_names=["mel_pred"],
dynamic_axes={
"hubert": [1],
"f0": [1],
"volume": [1],
"mel2ph": [1]
},
opset_version=16
)
condition = torch.randn(1,self.decoder.n_hidden,n_frames)
noise = torch.randn((1, 1, self.decoder.mel_bins, condition.shape[2]), dtype=torch.float32)
pndm_speedup = torch.LongTensor([100])
K_steps = torch.LongTensor([1000])
self.decoder = torch.jit.script(self.decoder)
self.decoder(condition, noise, pndm_speedup, K_steps)
torch.onnx.export(
self.decoder,
(condition, noise, pndm_speedup, K_steps),
f"{project_name}_diffusion.onnx",
input_names=["condition", "noise", "pndm_speedup", "K_steps"],
output_names=["mel"],
dynamic_axes={
"condition": [2],
"noise": [3],
},
opset_version=16
)
if __name__ == "__main__":
project_name = "dddsp"
model_path = f'{project_name}/model_500000.pt'
model, _ = load_model_vocoder(model_path)
# 分开Diffusion导出（需要使用MoeSS/MoeVoiceStudio或者自己编写Pndm/Dpm采样）
model.OnnxExport(project_name, export_encoder=True, export_denoise=True, export_pred=True, export_after=True)
# 合并Diffusion导出（Encoder和Diffusion分开，直接将Encoder的结果和初始噪声输入Diffusion即可）
# model.ExportOnnx(project_name)
|