File size: 11,488 Bytes
bccbc5b | 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 | """
ein notation:
b - batch
n - sequence
nt - text sequence
nw - raw wave length
d - dimension
"""
from __future__ import annotations
from random import random
from typing import Callable
import torch
import torch.nn.functional as F
from torch import nn
from torch.nn.utils.rnn import pad_sequence
from torchdiffeq import odeint
from f5_tts.model.modules import MelSpec
from f5_tts.model.utils import (
default,
exists,
get_epss_timesteps,
lens_to_mask,
list_str_to_idx,
list_str_to_tensor,
mask_from_frac_lengths,
)
class CFM(nn.Module):
def __init__(
self,
transformer: nn.Module,
sigma=0.0,
odeint_kwargs: dict = dict(
# atol = 1e-5,
# rtol = 1e-5,
method="euler" # 'midpoint'
),
audio_drop_prob=0.3,
cond_drop_prob=0.2,
num_channels=None,
mel_spec_module: nn.Module | None = None,
mel_spec_kwargs: dict = dict(),
frac_lengths_mask: tuple[float, float] = (0.7, 1.0),
vocab_char_map: dict[str:int] | None = None,
):
super().__init__()
self.frac_lengths_mask = frac_lengths_mask
# mel spec
self.mel_spec = default(mel_spec_module, MelSpec(**mel_spec_kwargs))
num_channels = default(num_channels, self.mel_spec.n_mel_channels)
self.num_channels = num_channels
# classifier-free guidance
self.audio_drop_prob = audio_drop_prob
self.cond_drop_prob = cond_drop_prob
# transformer
self.transformer = transformer
dim = transformer.dim
self.dim = dim
# conditional flow related
self.sigma = sigma
# sampling related
self.odeint_kwargs = odeint_kwargs
# vocab map for tokenization
self.vocab_char_map = vocab_char_map
@property
def device(self):
return next(self.parameters()).device
@torch.no_grad()
def sample(
self,
cond: float["b n d"] | float["b nw"], # noqa: F722
text: int["b nt"] | list[str], # noqa: F722
duration: int | int["b"], # noqa: F821
*,
lens: int["b"] | None = None, # noqa: F821
steps=32,
cfg_strength=1.0,
sway_sampling_coef=None,
seed: int | None = None,
max_duration=4096,
vocoder: Callable[[float["b d n"]], float["b nw"]] | None = None, # noqa: F722
use_epss=True,
no_ref_audio=False,
duplicate_test=False,
t_inter=0.1,
edit_mask=None,
):
try:
self.eval()
# raw wave
if cond.ndim == 2:
cond = self.mel_spec(cond)
cond = cond.permute(0, 2, 1)
assert cond.shape[-1] == self.num_channels
cond = cond.to(next(self.parameters()).dtype)
batch, cond_seq_len, device = *cond.shape[:2], cond.device
if not exists(lens):
lens = torch.full((batch,), cond_seq_len, device=device, dtype=torch.long)
# text
if isinstance(text, list):
if exists(self.vocab_char_map):
text = list_str_to_idx(text, self.vocab_char_map).to(device)
else:
text = list_str_to_tensor(text).to(device)
assert text.shape[0] == batch
# duration
cond_mask = lens_to_mask(lens)
if edit_mask is not None:
cond_mask = cond_mask & edit_mask
if isinstance(duration, int):
duration = torch.full((batch,), duration, device=device, dtype=torch.long)
duration = torch.maximum(
torch.maximum((text != -1).sum(dim=-1), lens) + 1, duration
) # duration at least text/audio prompt length plus one token, so something is generated
duration = duration.clamp(max=max_duration)
max_duration = duration.amax()
# duplicate test corner for inner time step oberservation
if duplicate_test:
test_cond = F.pad(cond, (0, 0, cond_seq_len, max_duration - 2 * cond_seq_len), value=0.0)
cond = F.pad(cond, (0, 0, 0, max_duration - cond_seq_len), value=0.0)
if no_ref_audio:
cond = torch.zeros_like(cond)
cond_mask = F.pad(cond_mask, (0, max_duration - cond_mask.shape[-1]), value=False)
cond_mask = cond_mask.unsqueeze(-1)
step_cond = torch.where(
cond_mask, cond, torch.zeros_like(cond)
) # allow direct control (cut cond audio) with lens passed in
if batch > 1:
mask = lens_to_mask(duration)
else: # save memory and speed up, as single inference need no mask currently
mask = None
# neural ode
def fn(t, x):
# at each step, conditioning is fixed
# step_cond = torch.where(cond_mask, cond, torch.zeros_like(cond))
# predict flow (cond)
if cfg_strength < 1e-5:
pred = self.transformer(
x=x,
cond=step_cond,
text=text,
time=t,
mask=mask,
drop_audio_cond=False,
drop_text=False,
cache=True,
skip_flash_attn=True
)
return pred
# predict flow (cond and uncond), for classifier-free guidance
pred_cfg = self.transformer(
x=x,
cond=step_cond,
text=text,
time=t,
mask=mask,
cfg_infer=True,
cache=True,
skip_flash_attn=True
)
pred, null_pred = torch.chunk(pred_cfg, 2, dim=0)
return pred + (pred - null_pred) * cfg_strength
# noise input
# to make sure batch inference result is same with different batch size, and for sure single inference
# still some difference maybe due to convolutional layers
y0 = []
for dur in duration:
if exists(seed):
torch.manual_seed(seed)
y0.append(torch.randn(dur, self.num_channels, device=self.device, dtype=step_cond.dtype))
y0 = pad_sequence(y0, padding_value=0, batch_first=True)
t_start = 0
# duplicate test corner for inner time step oberservation
if duplicate_test:
t_start = t_inter
y0 = (1 - t_start) * y0 + t_start * test_cond
steps = int(steps * (1 - t_start))
if t_start == 0 and use_epss: # use Empirically Pruned Step Sampling for low NFE
t = get_epss_timesteps(steps, device=self.device, dtype=step_cond.dtype)
else:
t = torch.linspace(t_start, 1, steps + 1, device=self.device, dtype=step_cond.dtype)
if sway_sampling_coef is not None:
t = t + sway_sampling_coef * (torch.cos(torch.pi / 2 * t) - 1 + t)
trajectory = odeint(fn, y0, t, **self.odeint_kwargs)
self.transformer.clear_cache()
sampled = trajectory[-1]
out = sampled
# out = torch.where(cond_mask, cond, out)
if exists(vocoder):
out = out.permute(0, 2, 1)
out = vocoder(out)
return out, trajectory
except Exception as e:
print(f"cond shape: {cond.shape}, text shape: {text.shape if torch.is_tensor(text) else 'N/A'}, duration: {duration.shape}, lens: {lens.shape}")
print(f"cond: {cond}, text: {text}, duration: {duration}, lens: {lens}")
raise e
def forward(
self,
inp: float["b n d"] | float["b nw"], # mel or raw wave # noqa: F722
text: int["b nt"] | list[str], # noqa: F722
*,
lens: int["b"] | None = None, # noqa: F821
noise_scheduler: str | None = None,
):
try:
# handle raw wave
# print(f"inp: {inp}, text: {text}, lens: {lens}")
if inp.ndim == 2:
inp = self.mel_spec(inp)
inp = inp.permute(0, 2, 1)
assert inp.shape[-1] == self.num_channels
batch, seq_len, dtype, device, _σ1 = *inp.shape[:2], inp.dtype, self.device, self.sigma
# handle text as string
if isinstance(text, list):
if exists(self.vocab_char_map):
text = list_str_to_idx(text, self.vocab_char_map).to(device)
else:
text = list_str_to_tensor(text).to(device)
assert text.shape[0] == batch
# lens and mask
if not exists(lens):
lens = torch.full((batch,), seq_len, device=device)
mask = lens_to_mask(lens, length=seq_len) # useless here, as collate_fn will pad to max length in batch
audio_frames_in_batch = mask.sum().item()
max_allowed_frames = mask.shape[0] * mask.shape[1]
perc_frame_utilised = audio_frames_in_batch / max_allowed_frames
# get a random span to mask out for training conditionally
frac_lengths = torch.zeros((batch,), device=self.device).float().uniform_(*self.frac_lengths_mask)
rand_span_mask = mask_from_frac_lengths(lens, frac_lengths)
if exists(mask):
rand_span_mask &= mask
# mel is x1
x1 = inp
# x0 is gaussian noise
x0 = torch.randn_like(x1)
# time step
time = torch.rand((batch,), dtype=dtype, device=self.device)
# TODO. noise_scheduler
# sample xt (φ_t(x) in the paper)
t = time.unsqueeze(-1).unsqueeze(-1)
φ = (1 - t) * x0 + t * x1
flow = x1 - x0
# only predict what is within the random mask span for infilling
cond = torch.where(rand_span_mask[..., None], torch.zeros_like(x1), x1)
# transformer and cfg training with a drop rate
drop_audio_cond = random() < self.audio_drop_prob # p_drop in voicebox paper
if random() < self.cond_drop_prob: # p_uncond in voicebox paper
drop_audio_cond = True
drop_text = True
else:
drop_text = False
# apply mask will use more memory; might adjust batchsize or batchsampler long sequence threshold
pred = self.transformer(
x=φ, cond=cond, text=text, time=time, drop_audio_cond=drop_audio_cond, drop_text=drop_text, mask=mask
)
# flow matching loss
loss = F.mse_loss(pred, flow, reduction="none")
loss = loss[rand_span_mask]
return loss.mean(), cond, pred, audio_frames_in_batch, max_allowed_frames, perc_frame_utilised, batch, seq_len
except Exception as e:
print(f"input shape: {inp.shape}, text shape: {text.shape if torch.is_tensor(text) else 'N/A'}")
print(f"input: {input}, text: {text}")
raise e |