File size: 11,583 Bytes
7fa2003 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 |
"""
ein notation:
b - batch
n - sequence
nt - text sequence
nw - raw wave length
d - dimension
"""
# ruff: noqa: F722 F821
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import nn
from torch.nn.utils.rnn import pad_sequence
from x_transformers.x_transformers import RotaryEmbedding
from f5_tts.model.modules import (
AdaLayerNorm_Final,
ConvNeXtV2Block,
ConvPositionEmbedding,
DiTBlock,
TimestepEmbedding,
precompute_freqs_cis,
)
# Text embedding
class TextEmbedding(nn.Module):
def __init__(
self, text_num_embeds, text_dim, mask_padding=True, average_upsampling=False, conv_layers=0, conv_mult=2
):
super().__init__()
self.text_embed = nn.Embedding(text_num_embeds + 1, text_dim) # use 0 as filler token
self.mask_padding = mask_padding # mask filler and batch padding tokens or not
self.average_upsampling = average_upsampling # zipvoice-style text late average upsampling (after text encoder)
if average_upsampling:
assert mask_padding, "text_embedding_average_upsampling requires text_mask_padding to be True"
if conv_layers > 0:
self.extra_modeling = True
self.precompute_max_pos = 4096 # ~44s of 24khz audio
self.register_buffer("freqs_cis", precompute_freqs_cis(text_dim, self.precompute_max_pos), persistent=False)
self.text_blocks = nn.Sequential(
*[ConvNeXtV2Block(text_dim, text_dim * conv_mult) for _ in range(conv_layers)]
)
else:
self.extra_modeling = False
def average_upsample_text_by_mask(self, text, text_mask, audio_mask):
batch, text_len, text_dim = text.shape
if audio_mask is None:
audio_mask = torch.ones_like(text_mask, dtype=torch.bool)
valid_mask = audio_mask & text_mask
audio_lens = audio_mask.sum(dim=1) # [batch]
valid_lens = valid_mask.sum(dim=1) # [batch]
upsampled_text = torch.zeros_like(text)
for i in range(batch):
audio_len = audio_lens[i].item()
valid_len = valid_lens[i].item()
if valid_len == 0:
continue
valid_ind = torch.where(valid_mask[i])[0]
valid_data = text[i, valid_ind, :] # [valid_len, text_dim]
base_repeat = audio_len // valid_len
remainder = audio_len % valid_len
indices = []
for j in range(valid_len):
repeat_count = base_repeat + (1 if j >= valid_len - remainder else 0)
indices.extend([j] * repeat_count)
indices = torch.tensor(indices[:audio_len], device=text.device, dtype=torch.long)
upsampled = valid_data[indices] # [audio_len, text_dim]
upsampled_text[i, :audio_len, :] = upsampled
return upsampled_text
def forward(self, text: int["b nt"], seq_len, drop_text=False, audio_mask: bool["b n"] | None = None):
text = text + 1 # use 0 as filler token. preprocess of batch pad -1, see list_str_to_idx()
text = text[:, :seq_len] # curtail if character tokens are more than the mel spec tokens
text = F.pad(text, (0, seq_len - text.shape[1]), value=0) # (opt.) if not self.average_upsampling:
if self.mask_padding:
text_mask = text == 0
if drop_text: # cfg for text
text = torch.zeros_like(text)
text = self.text_embed(text) # b n -> b n d
# possible extra modeling
if self.extra_modeling:
# sinus pos emb
text = text + self.freqs_cis[:seq_len, :]
# convnextv2 blocks
if self.mask_padding:
text = text.masked_fill(text_mask.unsqueeze(-1).expand(-1, -1, text.size(-1)), 0.0)
for block in self.text_blocks:
text = block(text)
text = text.masked_fill(text_mask.unsqueeze(-1).expand(-1, -1, text.size(-1)), 0.0)
else:
text = self.text_blocks(text)
if self.average_upsampling:
text = self.average_upsample_text_by_mask(text, ~text_mask, audio_mask)
return text
# noised input audio and context mixing embedding
class InputEmbedding(nn.Module):
def __init__(self, mel_dim, text_dim, out_dim):
super().__init__()
self.proj = nn.Linear(mel_dim * 2 + text_dim, out_dim)
self.conv_pos_embed = ConvPositionEmbedding(dim=out_dim)
def forward(
self,
x: float["b n d"],
cond: float["b n d"],
text_embed: float["b n d"],
drop_audio_cond=False,
audio_mask: bool["b n"] | None = None,
):
if drop_audio_cond: # cfg for cond audio
cond = torch.zeros_like(cond)
x = self.proj(torch.cat((x, cond, text_embed), dim=-1))
x = self.conv_pos_embed(x, mask=audio_mask) + x
return x
# Transformer backbone using DiT blocks
class DiT(nn.Module):
def __init__(
self,
*,
dim,
depth=8,
heads=8,
dim_head=64,
dropout=0.1,
ff_mult=4,
mel_dim=100,
text_num_embeds=256,
text_dim=None,
text_mask_padding=True,
text_embedding_average_upsampling=False,
qk_norm=None,
conv_layers=0,
pe_attn_head=None,
attn_backend="torch", # "torch" | "flash_attn"
attn_mask_enabled=False,
long_skip_connection=False,
checkpoint_activations=False,
):
super().__init__()
self.time_embed = TimestepEmbedding(dim)
if text_dim is None:
text_dim = mel_dim
self.text_embed = TextEmbedding(
text_num_embeds,
text_dim,
mask_padding=text_mask_padding,
average_upsampling=text_embedding_average_upsampling,
conv_layers=conv_layers,
)
self.text_cond, self.text_uncond = None, None # text cache
self.input_embed = InputEmbedding(mel_dim, text_dim, dim)
self.rotary_embed = RotaryEmbedding(dim_head)
self.dim = dim
self.depth = depth
self.transformer_blocks = nn.ModuleList(
[
DiTBlock(
dim=dim,
heads=heads,
dim_head=dim_head,
ff_mult=ff_mult,
dropout=dropout,
qk_norm=qk_norm,
pe_attn_head=pe_attn_head,
attn_backend=attn_backend,
attn_mask_enabled=attn_mask_enabled,
)
for _ in range(depth)
]
)
self.long_skip_connection = nn.Linear(dim * 2, dim, bias=False) if long_skip_connection else None
self.norm_out = AdaLayerNorm_Final(dim) # final modulation
self.proj_out = nn.Linear(dim, mel_dim)
self.checkpoint_activations = checkpoint_activations
self.initialize_weights()
def initialize_weights(self):
# Zero-out AdaLN layers in DiT blocks:
for block in self.transformer_blocks:
nn.init.constant_(block.attn_norm.linear.weight, 0)
nn.init.constant_(block.attn_norm.linear.bias, 0)
# Zero-out output layers:
nn.init.constant_(self.norm_out.linear.weight, 0)
nn.init.constant_(self.norm_out.linear.bias, 0)
nn.init.constant_(self.proj_out.weight, 0)
nn.init.constant_(self.proj_out.bias, 0)
def ckpt_wrapper(self, module):
# https://github.com/chuanyangjin/fast-DiT/blob/main/models.py
def ckpt_forward(*inputs):
outputs = module(*inputs)
return outputs
return ckpt_forward
def get_input_embed(
self,
x, # b n d
cond, # b n d
text, # b nt
drop_audio_cond: bool = False,
drop_text: bool = False,
cache: bool = True,
audio_mask: bool["b n"] | None = None,
):
if self.text_uncond is None or self.text_cond is None or not cache:
if audio_mask is None:
text_embed = self.text_embed(text, x.shape[1], drop_text=drop_text, audio_mask=audio_mask)
else:
batch = x.shape[0]
seq_lens = audio_mask.sum(dim=1)
text_embed_list = []
for i in range(batch):
text_embed_i = self.text_embed(
text[i].unsqueeze(0),
seq_lens[i].item(),
drop_text=drop_text,
audio_mask=audio_mask,
)
text_embed_list.append(text_embed_i[0])
text_embed = pad_sequence(text_embed_list, batch_first=True, padding_value=0)
if cache:
if drop_text:
self.text_uncond = text_embed
else:
self.text_cond = text_embed
if cache:
if drop_text:
text_embed = self.text_uncond
else:
text_embed = self.text_cond
x = self.input_embed(x, cond, text_embed, drop_audio_cond=drop_audio_cond, audio_mask=audio_mask)
return x
def clear_cache(self):
self.text_cond, self.text_uncond = None, None
def forward(
self,
x: float["b n d"], # nosied input audio
cond: float["b n d"], # masked cond audio
text: int["b nt"], # text
time: float["b"] | float[""], # time step
mask: bool["b n"] | None = None,
drop_audio_cond: bool = False, # cfg for cond audio
drop_text: bool = False, # cfg for text
cfg_infer: bool = False, # cfg inference, pack cond & uncond forward
cache: bool = False,
):
batch, seq_len = x.shape[0], x.shape[1]
if time.ndim == 0:
time = time.repeat(batch)
# t: conditioning time, text: text, x: noised audio + cond audio + text
t = self.time_embed(time)
if cfg_infer: # pack cond & uncond forward: b n d -> 2b n d
x_cond = self.get_input_embed(
x, cond, text, drop_audio_cond=False, drop_text=False, cache=cache, audio_mask=mask
)
x_uncond = self.get_input_embed(
x, cond, text, drop_audio_cond=True, drop_text=True, cache=cache, audio_mask=mask
)
x = torch.cat((x_cond, x_uncond), dim=0)
t = torch.cat((t, t), dim=0)
mask = torch.cat((mask, mask), dim=0) if mask is not None else None
else:
x = self.get_input_embed(
x, cond, text, drop_audio_cond=drop_audio_cond, drop_text=drop_text, cache=cache, audio_mask=mask
)
rope = self.rotary_embed.forward_from_seq_len(seq_len)
if self.long_skip_connection is not None:
residual = x
for block in self.transformer_blocks:
if self.checkpoint_activations:
# https://pytorch.org/docs/stable/checkpoint.html#torch.utils.checkpoint.checkpoint
x = torch.utils.checkpoint.checkpoint(self.ckpt_wrapper(block), x, t, mask, rope, use_reentrant=False)
else:
x = block(x, t, mask=mask, rope=rope)
if self.long_skip_connection is not None:
x = self.long_skip_connection(torch.cat((x, residual), dim=-1))
x = self.norm_out(x, t)
output = self.proj_out(x)
return output
|