File size: 13,507 Bytes
1ef5ba8 | 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 | """PSM: Phoneme-Synchrony Manifold.
Hypothesis
----------
Phonemes are physical instructions on the vocal tract. Articulation of /b/
requires bilabial closure, /f/ requires labiodental contact, etc. The
mapping { phoneme -> lip shape } is largely *speaker-independent*; the same
phoneme by different speakers yields similar lip shapes up to a small
identity-style variation.
If we learn an embedding space where (phoneme, lip_shape) pairs from REAL
videos lie on a compact manifold, then fake videos — whose generators
memorize audio->lip correlations but may violate the physical constraint
in subtle ways — will lie farther off-manifold on average.
Advantages over 'one-class real distribution learning'
------------------------------------------------------
We are NOT learning the full real joint. We are learning only the
*phoneme-conditional* distribution. Quality / domain differences across
CelebV-HQ / DFDC / HDTF mostly leak through identity and texture, not
through the phoneme-lip relation, so the conditional is far more robust.
Implementation
--------------
phoneme feats: wav2vec2-lv-60-espeak-cv-ft (CTC phoneme model, frozen)
lip feats: MediaPipe lip crop -> DINO ViT-S (half trainable)
projection head (MLP) -> 256-D manifold
Contrastive loss: anchor = phoneme token; positive = temporally-aligned
lip token; negatives = lip tokens from the same batch from other clips
AND a momentum memory bank.
Classifier: at inference we score the average per-frame (phoneme, lip)
compatibility. High compatibility => real; low => fake.
"""
from __future__ import annotations
from typing import Any, Dict, List, Optional
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from .base import BaseMethod
# -----------------------------------------------------------------------------
# phoneme encoder (wav2vec2 frozen)
# -----------------------------------------------------------------------------
class PhonemeEncoder(nn.Module):
def __init__(self, hf_id: str, freeze: bool = True):
super().__init__()
from transformers import Wav2Vec2Model
self.model = Wav2Vec2Model.from_pretrained(hf_id)
self.feature_dim = self.model.config.hidden_size
if freeze:
for p in self.model.parameters():
p.requires_grad_(False)
self.model.eval()
def forward(self, wav: torch.Tensor) -> torch.Tensor:
with torch.no_grad() if not any(p.requires_grad for p in self.model.parameters()) else torch.enable_grad():
out = self.model(wav)
return out.last_hidden_state # (B, T_a, D)
# -----------------------------------------------------------------------------
# lip crop (MediaPipe lazily, CPU) + ViT image encoder
# -----------------------------------------------------------------------------
class MediaPipeLipCropper:
"""Run MediaPipe once per forward to get lip bbox per frame.
Falls back to center crop if MediaPipe unavailable or no face found.
Lazy-initialized per worker to be DataLoader-safe.
"""
_LIP_LANDMARKS = [61, 146, 91, 181, 84, 17, 314, 405, 321, 375, 291,
308, 324, 318, 402, 317, 14, 87, 178, 88, 95]
def __init__(self, output_size: int = 96):
self.output_size = output_size
self._mp = None
def _ensure(self):
if self._mp is None:
try:
import mediapipe as mp
self._mp = mp.solutions.face_mesh.FaceMesh(
static_image_mode=False, max_num_faces=1, refine_landmarks=False,
)
except Exception:
self._mp = False # mark as unavailable
def crop_frame(self, frame: np.ndarray) -> np.ndarray:
"""frame: (H, W, 3) uint8 RGB. Returns (out, out, 3) uint8 RGB."""
self._ensure()
H, W = frame.shape[:2]
if self._mp is False:
return self._center(frame)
res = self._mp.process(frame)
if not res.multi_face_landmarks:
return self._center(frame)
lm = res.multi_face_landmarks[0].landmark
xs = np.array([lm[i].x for i in self._LIP_LANDMARKS]) * W
ys = np.array([lm[i].y for i in self._LIP_LANDMARKS]) * H
x0, x1 = xs.min(), xs.max(); y0, y1 = ys.min(), ys.max()
cx, cy = (x0 + x1) / 2, (y0 + y1) / 2
# square, with some margin
side = max(x1 - x0, y1 - y0) * 1.8
x0 = int(max(0, cx - side / 2)); x1 = int(min(W, cx + side / 2))
y0 = int(max(0, cy - side / 2)); y1 = int(min(H, cy + side / 2))
crop = frame[y0:y1, x0:x1]
if crop.size == 0:
return self._center(frame)
import cv2
return cv2.resize(crop, (self.output_size, self.output_size))
def _center(self, frame: np.ndarray) -> np.ndarray:
import cv2
H, W = frame.shape[:2]
s = min(H, W) // 2
cy, cx = H // 2, W // 2
crop = frame[cy - s:cy + s, cx - s:cx + s]
return cv2.resize(crop, (self.output_size, self.output_size))
class LipEncoder(nn.Module):
"""MediaPipe lip crop -> DINO ViT-S. Trainable top half."""
def __init__(self, vit_hf_id: str, freeze_ratio: float = 0.5, lip_size: int = 96):
super().__init__()
from transformers import AutoModel
self.vit = AutoModel.from_pretrained(vit_hf_id)
self.feature_dim = self.vit.config.hidden_size
self.lip_size = lip_size
if freeze_ratio > 0:
layers = self.vit.encoder.layer
n = int(len(layers) * freeze_ratio)
for layer in layers[:n]:
for p in layer.parameters():
p.requires_grad_(False)
def forward(self, lip_imgs: torch.Tensor) -> torch.Tensor:
"""lip_imgs: (B, T, 3, lip_size, lip_size) already normalized.
Returns (B, T, D)."""
B, T, C, H, W = lip_imgs.shape
x = lip_imgs.reshape(B * T, C, H, W)
# ViT expects 224; upsample if needed
if H != 224:
x = F.interpolate(x, size=(224, 224), mode="bilinear", align_corners=False)
out = self.vit(pixel_values=x)
tok = out.last_hidden_state[:, 0] # CLS token
return tok.view(B, T, -1)
# -----------------------------------------------------------------------------
# manifold projection + classifier
# -----------------------------------------------------------------------------
class ProjectionHead(nn.Module):
def __init__(self, in_dim: int, out_dim: int):
super().__init__()
self.net = nn.Sequential(
nn.Linear(in_dim, in_dim), nn.GELU(), nn.Linear(in_dim, out_dim),
)
def forward(self, x): return F.normalize(self.net(x), dim=-1)
# -----------------------------------------------------------------------------
# Lightning module
# -----------------------------------------------------------------------------
class PSMLitModule(BaseMethod):
def __init__(self, method_cfg, backbone_cfg, data_cfg):
super().__init__(method_cfg=method_cfg, backbone_cfg=backbone_cfg, data_cfg=data_cfg)
self.phoneme = PhonemeEncoder(
hf_id=method_cfg.phoneme_encoder.hf_id,
freeze=method_cfg.phoneme_encoder.freeze,
)
self.lip = LipEncoder(
vit_hf_id=method_cfg.lip_encoder.vit_hf_id,
freeze_ratio=method_cfg.lip_encoder.freeze_ratio,
)
pd = method_cfg.manifold.projection_dim
self.proj_p = ProjectionHead(self.phoneme.feature_dim, pd)
self.proj_l = ProjectionHead(self.lip.feature_dim, pd)
# classifier on [compat, lip_pool, phon_pool]
self.cls = nn.Sequential(
nn.Linear(pd * 2 + 1, method_cfg.classifier.hidden),
nn.GELU(),
nn.Dropout(method_cfg.classifier.dropout),
nn.Linear(method_cfg.classifier.hidden, 1),
)
self.register_buffer(
"mem_bank",
F.normalize(torch.randn(method_cfg.manifold.num_negatives, pd), dim=-1),
)
self.register_buffer("mem_ptr", torch.zeros(1, dtype=torch.long))
# MediaPipe cropper runs on tensors in _prepare_lip_tensor below;
# for simplicity we use center-crop in this initial version and
# expose a hook to plug in MediaPipe at dataloader level.
# ------------------------------------------------------------------
@staticmethod
def _resample_to_match(src: torch.Tensor, target_T: int) -> torch.Tensor:
"""Linear interpolation of token sequence along T dim."""
B, T, D = src.shape
x = src.transpose(1, 2) # (B, D, T)
x = F.interpolate(x, size=target_T, mode="linear", align_corners=False)
return x.transpose(1, 2) # (B, target_T, D)
def _lip_from_video(self, video: torch.Tensor) -> torch.Tensor:
"""Produce lip crops from a normalized video tensor.
Initial version: center crop the bottom-half of the frame (lip area
heuristic). A more accurate version using MediaPipe runs in the
dataloader; see TODO below.
"""
B, T, C, H, W = video.shape
# crop bottom half then rightmost 80% of width
crop = video[..., H // 2:, :]
# resize to 96 for lip encoder-friendly size
crop = F.interpolate(
crop.reshape(B * T, C, crop.shape[-2], crop.shape[-1]),
size=(96, 96), mode="bilinear", align_corners=False,
).view(B, T, C, 96, 96)
return crop
# ------------------------------------------------------------------
@torch.no_grad()
def _enqueue(self, feats: torch.Tensor):
K = self.mem_bank.size(0)
n = feats.size(0)
ptr = int(self.mem_ptr.item())
if ptr + n > K:
first = K - ptr
self.mem_bank[ptr:] = feats[:first]
self.mem_bank[:n - first] = feats[first:]
self.mem_ptr[0] = (n - first) % K
else:
self.mem_bank[ptr:ptr + n] = feats
self.mem_ptr[0] = (ptr + n) % K
# ------------------------------------------------------------------
def forward_feats(self, video, audio):
ph = self.phoneme(audio) # (B, T_a, Dp)
lip_imgs = self._lip_from_video(video) # (B, T_v, 3, 96, 96)
li = self.lip(lip_imgs) # (B, T_v, Dl)
# resample phoneme to video's T
T_v = li.shape[1]
ph_r = self._resample_to_match(ph, T_v)
zp = self.proj_p(ph_r) # (B, T_v, pd) unit
zl = self.proj_l(li)
return zp, zl
# ------------------------------------------------------------------
def _compatibility(self, zp, zl) -> torch.Tensor:
"""Per-sample mean cosine similarity between aligned phoneme & lip
embeddings. Higher => more on-manifold => more real."""
return (zp * zl).sum(dim=-1).mean(dim=1) # (B,)
def training_step(self, batch, batch_idx):
if batch is None:
return None
video = batch["video"]; audio = batch["audio"]
labels = batch["label"].long()
zp, zl = self.forward_feats(video, audio)
B, T, D = zp.shape
zp_flat = zp.reshape(B * T, D)
zl_flat = zl.reshape(B * T, D)
# contrastive on REAL only
is_real = (labels == 0)
if is_real.any():
rmask = is_real.unsqueeze(1).expand(-1, T).reshape(-1)
zp_r = zp_flat[rmask]; zl_r = zl_flat[rmask]
neg_bank = self.mem_bank # (K, D)
logits = torch.cat([
(zp_r * zl_r).sum(-1, keepdim=True), # positives
zp_r @ neg_bank.t(), # negatives
], dim=1) # (N, 1+K)
logits = logits / self.method_cfg.manifold.temperature
target = torch.zeros(logits.size(0), dtype=torch.long, device=logits.device)
loss_con = F.cross_entropy(logits, target)
with torch.no_grad():
self._enqueue(zl_r.detach())
else:
loss_con = zp.new_zeros([])
# consistency: neighbouring frames' compatibility should be smooth
compat = (zp * zl).sum(dim=-1) # (B, T)
loss_cons = ((compat[:, 1:] - compat[:, :-1]) ** 2).mean()
# classifier
feat_cls = torch.cat([
zp.mean(1), zl.mean(1), compat.mean(1, keepdim=True),
], dim=-1)
logits_cls = self.cls(feat_cls)
loss_cls = F.binary_cross_entropy_with_logits(
logits_cls.squeeze(-1), labels.float(),
)
loss = (
self.method_cfg.loss.contrastive_weight * loss_con
+ self.method_cfg.loss.consistency_weight * loss_cons
+ self.method_cfg.loss.cls_weight * loss_cls
)
self.log_dict({
"train/loss": loss,
"train/loss_con": loss_con,
"train/loss_cons": loss_cons,
"train/loss_cls": loss_cls,
}, on_step=True, on_epoch=True, sync_dist=True)
return loss
@torch.no_grad()
def score(self, batch: Dict[str, Any]) -> torch.Tensor:
zp, zl = self.forward_feats(batch["video"], batch["audio"])
compat = (zp * zl).sum(dim=-1)
feat_cls = torch.cat([
zp.mean(1), zl.mean(1), compat.mean(1, keepdim=True),
], dim=-1)
logits = self.cls(feat_cls)
return torch.sigmoid(logits.squeeze(-1))
|