File size: 9,308 Bytes
574a515 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 |
import sys
import numpy as np
import torch
import torch.nn as nn
from vq.residual_vq import ResidualVQ
from vq.module import WNConv1d, DecoderBlock, ResLSTM
from vq.alias_free_torch import *
from vq import activations
import vq.blocks as blocks
from torch.nn import utils
from vq.bs_roformer5 import TransformerBlock
from torchtune.modules import RotaryPositionalEmbeddings
def init_weights(m):
if isinstance(m, nn.Conv1d):
nn.init.trunc_normal_(m.weight, std=0.02)
nn.init.constant_(m.bias, 0)
class CodecDecoder(nn.Module):
def __init__(self,
in_channels=1024,
upsample_initial_channel=1536,
ngf=48,
use_rnn=True,
rnn_bidirectional=False,
rnn_num_layers=2,
up_ratios=(5, 4, 4, 4, 2),
dilations=(1, 3, 9),
vq_num_quantizers=1,
vq_dim=2048,
vq_commit_weight=0.25,
vq_weight_init=False,
vq_full_commit_loss=False,
codebook_size=16384,
codebook_dim=32,
):
super().__init__()
self.hop_length = np.prod(up_ratios)
self.ngf = ngf
self.up_ratios = up_ratios
self.quantizer = ResidualVQ(
num_quantizers=vq_num_quantizers,
dim=vq_dim, # double the dim for acousitc and semantic
codebook_size=codebook_size,
codebook_dim=codebook_dim,
threshold_ema_dead_code=2,
commitment=vq_commit_weight,
weight_init=vq_weight_init,
full_commit_loss=vq_full_commit_loss,
)
channels = upsample_initial_channel
layers = [WNConv1d(in_channels, channels, kernel_size=7, padding=3)]
if use_rnn:
layers += [
ResLSTM(channels,
num_layers=rnn_num_layers,
bidirectional=rnn_bidirectional
)
]
for i, stride in enumerate(up_ratios):
input_dim = channels // 2**i
output_dim = channels // 2 ** (i + 1)
layers += [DecoderBlock(input_dim, output_dim, stride, dilations)]
layers += [
Activation1d(activation=activations.SnakeBeta(output_dim, alpha_logscale=True)),
WNConv1d(output_dim, 1, kernel_size=7, padding=3),
nn.Tanh(),
]
self.model = nn.Sequential(*layers)
self.reset_parameters()
def forward(self, x, vq=True):
if vq is True:
x, q, commit_loss = self.quantizer(x)
return x, q, commit_loss
x = self.model(x)
return x
def vq2emb(self, vq):
self.quantizer = self.quantizer.eval()
x = self.quantizer.vq2emb(vq)
return x
def get_emb(self):
self.quantizer = self.quantizer.eval()
embs = self.quantizer.get_emb()
return embs
def inference_vq(self, vq):
x = vq[None,:,:]
x = self.model(x)
return x
def inference_0(self, x):
x, q, loss, perp = self.quantizer(x)
x = self.model(x)
return x, None
def inference(self, x):
x = self.model(x)
return x, None
def remove_weight_norm(self):
"""Remove weight normalization module from all of the layers."""
def _remove_weight_norm(m):
try:
torch.nn.utils.remove_weight_norm(m)
except ValueError: # this module didn't have weight norm
return
self.apply(_remove_weight_norm)
def apply_weight_norm(self):
"""Apply weight normalization module from all of the layers."""
def _apply_weight_norm(m):
if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d):
torch.nn.utils.weight_norm(m)
self.apply(_apply_weight_norm)
def reset_parameters(self):
self.apply(init_weights)
class CodecDecoder_oobleck_Transformer(nn.Module):
def __init__(self,
ngf=32,
up_ratios=(5, 4, 4, 4, 2),
dilations=(1, 3, 9),
vq_num_quantizers=1,
vq_dim=1024,
vq_commit_weight=0.25,
vq_weight_init=False,
vq_full_commit_loss=False,
codebook_size=16384,
codebook_dim=16,
hidden_dim=1024,
depth=12,
heads=16,
pos_meb_dim=64,
):
super().__init__()
self.hop_length = np.prod(up_ratios)
self.capacity = ngf
self.up_ratios = up_ratios
self.hidden_dim = hidden_dim
self.quantizer = ResidualVQ(
num_quantizers=vq_num_quantizers,
dim=vq_dim, # double the dim for acousitc and semantic
codebook_size=codebook_size,
codebook_dim=codebook_dim,
threshold_ema_dead_code=2,
commitment=vq_commit_weight,
weight_init=vq_weight_init,
full_commit_loss=vq_full_commit_loss,
)
time_rotary_embed = RotaryPositionalEmbeddings(dim=pos_meb_dim)
transformer_blocks = [
TransformerBlock(dim=hidden_dim, n_heads=heads, rotary_embed=time_rotary_embed)
for _ in range(depth)
]
self.transformers = nn.Sequential(*transformer_blocks)
self.final_layer_norm = nn.LayerNorm(hidden_dim, eps=1e-6)
self.conv_blocks = blocks.DilatedResidualDecoder(
capacity=self.capacity,
dilated_unit=self.dilated_unit,
upsampling_unit=self.upsampling_unit,
ratios=up_ratios, # 逆转编码器的下采样比率
dilations=dilations,
pre_network_conv=self.pre_conv,
post_network_conv=self.post_conv,
)
self.reset_parameters()
def forward(self, x, vq=True):
if vq is True:
x, q, commit_loss = self.quantizer(x)
return x, q, commit_loss
x= self.transformers(x)
x = self.final_layer_norm(x)
x = x.permute(0, 2, 1)
x = self.conv_blocks(x)
return x
def vq2emb(self, vq):
self.quantizer = self.quantizer.eval()
x = self.quantizer.vq2emb(vq)
return x
def get_emb(self):
self.quantizer = self.quantizer.eval()
embs = self.quantizer.get_emb()
return embs
def inference_vq(self, vq):
x = vq[None,:,:]
x = self.model(x)
return x
def inference_0(self, x):
x, q, loss, perp = self.quantizer(x)
x = self.model(x)
return x, None
def inference(self, x):
x = self.model(x)
return x, None
def remove_weight_norm(self):
"""Remove weight normalization module from all of the layers."""
def _remove_weight_norm(m):
try:
torch.nn.utils.remove_weight_norm(m)
except ValueError: # this module didn't have weight norm
return
self.apply(_remove_weight_norm)
def apply_weight_norm(self):
"""Apply weight normalization module from all of the layers."""
def _apply_weight_norm(m):
if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d):
torch.nn.utils.weight_norm(m)
self.apply(_apply_weight_norm)
def reset_parameters(self):
self.apply(init_weights)
def pre_conv(self, out_channels):
return nn.Conv1d(in_channels=self.hidden_dim, out_channels=out_channels, kernel_size=1)
# 定义后处理卷积层，将模型的输出映射到最终的输出通道数
def post_conv(self,in_channels):
return nn.Conv1d(in_channels=in_channels, out_channels=1, kernel_size=1)
def dilated_unit(self, hidden_dim, dilation):
return blocks.DilatedConvolutionalUnit(
hidden_dim=hidden_dim,
dilation=dilation,
kernel_size=3,
activation=nn.ReLU ,
normalization=utils.weight_norm
)
# 定义上采样单元
def upsampling_unit(self,input_dim, output_dim, stride):
return blocks.UpsamplingUnit(
input_dim=input_dim,
output_dim=output_dim,
stride=stride,
activation=nn.ReLU ,
normalization=utils.weight_norm
)
def main():
# 设置设备
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")
# 初始化模型
model = CodecDecoder_oobleck_Transformer().to(device)
print("Model initialized.")
# 创建测试输入: batch_size x in_channels x sequence_length
batch_size = 2
in_channels = 1024
sequence_length = 100 # 示例长度，可以根据需要调整
dummy_input = torch.randn(batch_size, sequence_length, in_channels).to(device)
print(f"Dummy input shape: {dummy_input.shape}")
# 将模型设为评估模式
model.eval()
output_no_vq = model(dummy_input, vq=False)
c=1
if __name__ == "__main__":
main() |