Spaces:
Running on Zero
Running on Zero
File size: 8,915 Bytes
b522cc2 | 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 | import os
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import Qwen2Config, Qwen2Model
from .istft import ISTFTHead
_AUDIO_VAE_ATTN_ENV = "MING_TINY_AUDIO_VAE_ATTN"
_AUDIO_VAE_ATTN_BACKENDS = {
"config",
"eager",
"sdpa",
"flash_attention_2",
}
def _build_qwen2_config(config_dict):
"""Build the AudioVAE backbone config with a portable attention backend.
The training config records ``flash_attention_2`` for both AudioVAE
backbones. Requiring that optional binary extension just to construct the
HF model makes the standalone checkpoint unusable on an otherwise valid
Torch/SDPA environment. SDPA is the safe inference default; an explicitly
prepared runtime can opt back into FlashAttention with the environment
variable below.
"""
backend = os.getenv(_AUDIO_VAE_ATTN_ENV, "config").strip().lower()
if backend not in _AUDIO_VAE_ATTN_BACKENDS:
raise ValueError(
f"{_AUDIO_VAE_ATTN_ENV}={backend!r} is invalid; choose one of "
f"{sorted(_AUDIO_VAE_ATTN_BACKENDS)}"
)
config = Qwen2Config.from_dict(config_dict=config_dict)
if backend != "config":
config._attn_implementation = backend
return config
class StreamingLinearUpsample(nn.Module):
def __init__(self, scale_factor=4):
super().__init__()
self.scale_factor = scale_factor
self.upsampler = nn.Upsample(scale_factor=scale_factor, mode='linear', align_corners=False)
def forward(self, x, state=None, is_last=False):
# 初始化状态
if state is None:
state = {
'prev_chunk': None,
'history_last': None,
'is_first': True
}
if x is None and not is_last:
return None, state
if state['is_first'] and is_last:
out = self.upsampler(x.transpose(1, 2)).transpose(1, 2)
return out, None # 结束后清除状态
output_chunks = []
if state['is_first']:
state['prev_chunk'] = x
state['is_first'] = False
if not is_last:
return None, state
if state['prev_chunk'] is not None:
p = state['prev_chunk'].transpose(1, 2)
if state['history_last'] is None:
lookahead = x[:, :1, :].transpose(1, 2)
inp = torch.cat([p, lookahead], dim=2)
up = self.upsampler(inp)
out_prev = up[:, :, :p.size(2) * self.scale_factor]
else:
lookahead = x[:, :1, :].transpose(1, 2)
inp = torch.cat([state['history_last'], p, lookahead], dim=2)
up = self.upsampler(inp)
start = self.scale_factor
end = start + p.size(2) * self.scale_factor
out_prev = up[:, :, start:end]
output_chunks.append(out_prev.transpose(1, 2))
state['history_last'] = p[:, :, -1:]
state['prev_chunk'] = x
if is_last:
p = state['prev_chunk'].transpose(1, 2)
inp = torch.cat([state['history_last'], p], dim=2)
up = self.upsampler(inp)
out_last = up[:, :, self.scale_factor:]
output_chunks.append(out_last.transpose(1, 2))
state = None # 结束
final_out = torch.cat(output_chunks, dim=1) if output_chunks else None
return final_out, state
class Encoder(nn.Module):
def __init__(self, encoder_args, input_dim=320, hop_size=320, latent_dim=64, patch_size=-1):
super().__init__()
config = _build_qwen2_config(encoder_args)
self.encoder = Qwen2Model(config)
self.input_dim = input_dim
self.hop_size = hop_size
self.latent_dim = latent_dim
self.fc1 = nn.Linear(input_dim, config.hidden_size, bias=False)
self.fc2 = nn.Linear(config.hidden_size, config.hidden_size)
self.fc3 = nn.Linear(config.hidden_size, latent_dim*2)
self.norm = nn.LayerNorm(config.hidden_size)
self.patch_size = patch_size
if patch_size != -1:
config.num_hidden_layers = 4
self.aggregator = Qwen2Model(config)
self.cls_embed = nn.Parameter(torch.rand(1, 1, config.hidden_size))
self.cls_embed.data.normal_(0, 0.02)
def get_frames(self, x):
num_frames_total = (x.size(-1) + self.hop_size - 1) // self.hop_size # 向上取整的帧数
expected_len = (num_frames_total - 1) * self.hop_size + self.input_dim
padding_needed = expected_len - x.size(-1)
waveform = F.pad(x, (0, padding_needed), value=0.0)
frames = waveform.unfold(dimension=-1, size=self.input_dim, step=self.hop_size) # [B, T, d]
return frames
def pad_patch_insert_cls(self, x):
bsz, _, dim = x.size()
num_frame = x.size(1)
r = num_frame % self.patch_size
pad_num = self.patch_size-r if r else 0
x = F.pad(x, (0, 0, 0, pad_num), value=0.0) # 帧数对齐到patch_size倍数
x = x.reshape(-1, self.patch_size, dim)
x = torch.cat((x, self.cls_embed.expand(x.size(0), -1, -1)), dim=1) # 每个patch后插入一个cls
x = x.reshape(bsz, -1, dim)
return x
def forward(self, waveform):
x = self.get_frames(waveform)
x = self.fc1(x)
x = self.fc2(x)
x = self.encoder(inputs_embeds=x)
x = x.last_hidden_state
# downsample
if self.patch_size != -1:
x = self.pad_patch_insert_cls(x)
x = self.aggregator(inputs_embeds=x)
x = x.last_hidden_state
bsz, _, dim = x.size()
x = x.reshape(-1, self.patch_size+1, dim)
x = x[:, -1:, :].reshape(bsz, -1, dim)
x = self.fc3(x)
return x, waveform.unsqueeze(1)
class Decoder(nn.Module):
def __init__(self, decoder_args, output_dim=320, latent_dim=64, semantic_model=None, patch_size=-1):
super().__init__()
config = _build_qwen2_config(decoder_args)
self.decoder = Qwen2Model(config)
self.output_dim = output_dim
self.latent_dim = latent_dim
self.fc1 = nn.Linear(latent_dim, config.hidden_size)
if semantic_model is not None:
self.gelu = nn.GELU()
self.fc2 = nn.Linear(config.hidden_size, semantic_model.audio_emb_dim)
self.semantic_model = semantic_model
self.fc3 = nn.Linear(semantic_model.audio_emb_dim, config.hidden_size)
else:
self.semantic_model = None
self.hop_length = output_dim
self.head = ISTFTHead(dim=config.hidden_size, n_fft=self.hop_length * 4, hop_length=self.hop_length, padding="same")
self.patch_size = patch_size
if self.patch_size != -1:
self.upsampling = StreamingLinearUpsample(scale_factor=patch_size)
def forward(self, x, only_semantic_emb=False, past_key_values=None, use_cache=False):
x = self.fc1(x)
if self.semantic_model is not None:
x = self.fc2(self.gelu(x))
x, past_key_values = self.semantic_model(whipser_feats=x, past_key_values=past_key_values, use_cache=use_cache)
unified_emb = x
if only_semantic_emb:
return unified_emb, past_key_values
x = self.fc3(x)
else:
unified_emb = None
if self.patch_size != -1:
x = self.upsampling(x.transpose(1, 2)).transpose(1, 2)
x = self.decoder(inputs_embeds=x)
x = x.last_hidden_state
x, _ = self.head(x)
return x, unified_emb
def low_level_reconstruct(self, x, past_key_values=None, use_cache=False, stream_state=None, last_chunk=False):
upsample_state, audio_buffer, window_buffer = stream_state
bsz, device, dtype = x.size(0), x.device, x.dtype
x = self.fc1(x)
if self.patch_size != -1:
if use_cache:
# streaming
x, upsample_state = self.upsampling(x, state=upsample_state, is_last=last_chunk)
if x is None:
stream_state = (upsample_state, audio_buffer, window_buffer)
return torch.empty(bsz, 1, 0, device=device, dtype=dtype), stream_state, past_key_values
else:
x = self.upsampling.upsampler(x.transpose(1, 2)).transpose(1, 2)
outputs = self.decoder(inputs_embeds=x, past_key_values=past_key_values, use_cache=use_cache)
past_key_values = outputs.past_key_values
x = outputs.last_hidden_state
x, _, audio_buffer, window_buffer = self.head(x, streaming=use_cache, audio_buffer=audio_buffer, window_buffer=window_buffer, last_chunk=last_chunk)
stream_state = (upsample_state, audio_buffer, window_buffer)
return x, stream_state, past_key_values
|