File size: 8,674 Bytes
8f40dc9 | 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 | """
ein notation:
b - batch
n - sequence
nt - text sequence
nw - raw wave length
d - dimension
"""
# ruff: noqa: F722 F821
from __future__ import annotations
import threading
import torch
from torch import nn
from x_transformers.x_transformers import RotaryEmbedding
from f5_tts.model.modules import (
AdaLayerNorm_Final,
ConvPositionEmbedding,
MMDiTBlock,
TimestepEmbedding,
get_pos_embed_indices,
precompute_freqs_cis,
)
# text embedding
class TextEmbedding(nn.Module):
def __init__(self, out_dim, text_num_embeds, mask_padding=True):
super().__init__()
self.text_embed = nn.Embedding(text_num_embeds + 1, out_dim) # will use 0 as filler token
self.mask_padding = mask_padding # mask filler and batch padding tokens or not
self.precompute_max_pos = 1024
self.register_buffer("freqs_cis", precompute_freqs_cis(out_dim, self.precompute_max_pos), persistent=False)
def forward(self, text: int["b nt"], drop_text=False) -> int["b nt d"]:
text = text + 1 # use 0 as filler token. preprocess of batch pad -1, see list_str_to_idx()
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 nt -> b nt d
# sinus pos emb
batch_start = torch.zeros((text.shape[0],), dtype=torch.long)
batch_text_len = text.shape[1]
pos_idx = get_pos_embed_indices(batch_start, batch_text_len, max_pos=self.precompute_max_pos)
text_pos_embed = self.freqs_cis[pos_idx]
text = text + text_pos_embed
if self.mask_padding:
text = text.masked_fill(text_mask.unsqueeze(-1).expand(-1, -1, text.size(-1)), 0.0)
return text
# noised input & masked cond audio embedding
class AudioEmbedding(nn.Module):
def __init__(self, in_dim, out_dim):
super().__init__()
self.linear = nn.Linear(2 * in_dim, out_dim)
self.conv_pos_embed = ConvPositionEmbedding(out_dim)
def forward(self, x: float["b n d"], cond: float["b n d"], drop_audio_cond=False):
if drop_audio_cond:
cond = torch.zeros_like(cond)
x = torch.cat((x, cond), dim=-1)
x = self.linear(x)
x = self.conv_pos_embed(x) + x
return x
# Transformer backbone using MM-DiT blocks
class MMDiT(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_mask_padding=True,
qk_norm=None,
checkpoint_activations=False,
attn_backend="torch",
attn_mask_enabled=False,
):
super().__init__()
self.time_embed = TimestepEmbedding(dim)
self.text_embed = TextEmbedding(dim, text_num_embeds, mask_padding=text_mask_padding)
self.audio_embed = AudioEmbedding(mel_dim, dim)
self.rotary_embed = RotaryEmbedding(dim_head)
self.dim = dim
self.depth = depth
self.transformer_blocks = nn.ModuleList(
[
MMDiTBlock(
dim=dim,
heads=heads,
dim_head=dim_head,
dropout=dropout,
ff_mult=ff_mult,
context_pre_only=i == depth - 1,
qk_norm=qk_norm,
attn_backend=attn_backend,
attn_mask_enabled=attn_mask_enabled,
)
for i in range(depth)
]
)
self.norm_out = AdaLayerNorm_Final(dim) # final modulation
self.proj_out = nn.Linear(dim, mel_dim)
self.checkpoint_activations = checkpoint_activations
self.initialize_weights()
# `_cache_local` is lazily initialized on first inference-time cache write so that
# training models (which never touch the cache) stay deepcopy-friendly for EMA.
def _get_cache_local(self):
cache = self.__dict__.get("_cache_local")
if cache is None:
cache = threading.local()
self.__dict__["_cache_local"] = cache
return cache
@property
def text_cond(self):
cache = self.__dict__.get("_cache_local")
return getattr(cache, "text_cond", None) if cache is not None else None
@text_cond.setter
def text_cond(self, value):
self._get_cache_local().text_cond = value
@property
def text_uncond(self):
cache = self.__dict__.get("_cache_local")
return getattr(cache, "text_uncond", None) if cache is not None else None
@text_uncond.setter
def text_uncond(self, value):
self._get_cache_local().text_uncond = value
def initialize_weights(self):
# Zero-out AdaLN layers in MMDiT blocks:
for block in self.transformer_blocks:
nn.init.constant_(block.attn_norm_x.linear.weight, 0)
nn.init.constant_(block.attn_norm_x.linear.bias, 0)
nn.init.constant_(block.attn_norm_c.linear.weight, 0)
nn.init.constant_(block.attn_norm_c.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):
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,
):
if cache:
if drop_text:
if self.text_uncond is None:
self.text_uncond = self.text_embed(text, drop_text=True)
c = self.text_uncond
else:
if self.text_cond is None:
self.text_cond = self.text_embed(text, drop_text=False)
c = self.text_cond
else:
c = self.text_embed(text, drop_text=drop_text)
x = self.audio_embed(x, cond, drop_audio_cond=drop_audio_cond)
return x, c
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 = x.shape[0]
if time.ndim == 0:
time = time.repeat(batch)
# t: conditioning (time), c: context (text + masked cond audio), x: noised input audio
t = self.time_embed(time)
c_mask = (text + 1) != 0 # True = valid, False = padding (-1 tokens)
if cfg_infer: # pack cond & uncond forward: b n d -> 2b n d
x_cond, c_cond = self.get_input_embed(x, cond, text, drop_audio_cond=False, drop_text=False, cache=cache)
x_uncond, c_uncond = self.get_input_embed(x, cond, text, drop_audio_cond=True, drop_text=True, cache=cache)
x = torch.cat((x_cond, x_uncond), dim=0)
c = torch.cat((c_cond, c_uncond), dim=0)
t = torch.cat((t, t), dim=0)
mask = torch.cat((mask, mask), dim=0) if mask is not None else None
c_mask = torch.cat((c_mask, c_mask), dim=0)
else:
x, c = self.get_input_embed(
x, cond, text, drop_audio_cond=drop_audio_cond, drop_text=drop_text, cache=cache
)
seq_len = x.shape[1]
text_len = text.shape[1]
rope_audio = self.rotary_embed.forward_from_seq_len(seq_len)
rope_text = self.rotary_embed.forward_from_seq_len(text_len)
for block in self.transformer_blocks:
if self.checkpoint_activations:
c, x = torch.utils.checkpoint.checkpoint(
self.ckpt_wrapper(block), x, c, t, mask, rope_audio, rope_text, c_mask, use_reentrant=False
)
else:
c, x = block(x, c, t, mask=mask, rope=rope_audio, c_rope=rope_text, c_mask=c_mask)
x = self.norm_out(x, t)
output = self.proj_out(x)
return output
|