File size: 13,528 Bytes
52519d5 057b47e 52519d5 057b47e 52519d5 057b47e 52519d5 057b47e 52519d5 057b47e 52519d5 057b47e | 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 | from pathlib import Path
import cv2
import librosa
import mediapipe as mp
import numpy as np
import requests
import torch
import torch.nn.functional as F
import torchvision.transforms.v2 as transforms
from numpy.typing import NDArray
from packaging.version import Version
from python_speech_features import logfbank
from transformers import FeatureExtractionMixin
from transformers.feature_extraction_utils import BatchFeature
use_legacy_mp = False
if Version(mp.__version__) <= Version("0.10.21"):
mp_face_mesh = mp.solutions.face_mesh
use_legacy_mp = True
else:
BaseOptions = mp.tasks.BaseOptions
FaceLandmarker = mp.tasks.vision.FaceLandmarker
FaceLandmarkerOptions = mp.tasks.vision.FaceLandmarkerOptions
VisionRunningMode = mp.tasks.vision.RunningMode
class AVHubertFeatureExtractor(FeatureExtractionMixin):
model_input_names = ["input_values", "pixel_values"]
def __init__(
self,
max_sample_size: int | None = None,
normalize: bool = True,
stack_order_audio: int = 4,
image_crop_size: int = 88,
image_mean: float = 0.421,
image_std: float = 0.165,
sr: int = 16_000,
static_image_mode: bool = False,
refine_landmarks: bool = False,
min_detection_confidence: float = 0.5,
min_tracking_confidence: float = 0.5,
landmark_indices: tuple[int, ...] = (5, 411, 199, 187), # (top, right, bottom, left) of mouth
**kwargs,
) -> None:
super().__init__(**kwargs)
self.max_sample_size = max_sample_size
self.normalize = normalize
self.stack_order_audio = stack_order_audio
self.image_crop_size = image_crop_size
self.transforms = transforms.Compose(
[
transforms.ToImage(),
transforms.CenterCrop(image_crop_size),
transforms.ToDtype(torch.float32, scale=True),
transforms.Normalize([image_mean], [image_std]),
]
)
self.sr = sr
self.static_image_mode = static_image_mode
self.refine_landmarks = refine_landmarks
self.min_detection_confidence = min_detection_confidence
self.min_tracking_confidence = min_tracking_confidence
self.landmark_indices = landmark_indices
def _load_video(self, video: str | NDArray[np.uint8], extract_mouth: bool = False) -> torch.FloatTensor:
"""Input video must be in RGB format if type is numpy array."""
if isinstance(video, str):
cap = cv2.VideoCapture(video)
frames = []
for _ in range(int(cap.get(cv2.CAP_PROP_FRAME_COUNT))):
ret, frame = cap.read()
if not ret:
break
if not extract_mouth: # Already extracted mouth
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY))
else:
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
frames_np = np.stack(frames, axis=0)
else:
frames_np = video
if not extract_mouth: # Already extracted mouth
frames_np = np.stack([cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) for frame in frames_np], axis=0)
if extract_mouth:
frames_np = self._extract_mouth_legacy(frames_np) if use_legacy_mp else self._extract_mouth(frames_np)
return torch.from_numpy(frames_np).unsqueeze(dim=1)
def _extract_mouth(self, frames: NDArray[np.uint8]) -> NDArray[np.uint8]:
mouth_frames = []
top_idx, right_idx, bottom_idx, left_idx = self.landmark_indices
model_path = Path.home() / ".cache" / "reazonspeech" / "mediapipe---models--face_landmarker.task"
model_path.parent.mkdir(parents=True, exist_ok=True)
if not model_path.exists():
with open(model_path, "wb") as f:
f.write(requests.get("https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/latest/face_landmarker.task").content)
with FaceLandmarker.create_from_options(
FaceLandmarkerOptions(
base_options=BaseOptions(model_asset_path=model_path.as_posix()),
running_mode=VisionRunningMode.IMAGE,
num_faces=1,
min_face_detection_confidence=self.min_detection_confidence,
min_tracking_confidence=self.min_tracking_confidence,
)
) as face_mesh:
for frame in frames:
res = face_mesh.detect(
mp.Image(image_format=mp.ImageFormat.SRGB, data=frame)
)
if res.face_landmarks is None or len(res.face_landmarks) == 0:
mouth_frames.append(np.zeros([self.image_crop_size, self.image_crop_size], dtype=np.uint8))
continue
landmarks = res.face_landmarks[0]
top = landmarks[top_idx]
left = landmarks[left_idx]
right = landmarks[right_idx]
bottom = landmarks[bottom_idx]
H, W = frame.shape[:2]
xmax = max(top.x, left.x, right.x, bottom.x)
ymax = max(top.y, left.y, right.y, bottom.y)
xmin = min(top.x, left.x, right.x, bottom.x)
ymin = min(top.y, left.y, right.y, bottom.y)
patch_size = max((xmax - xmin) * W, (ymax - ymin) * H) # To extract square region
half = int(patch_size / 2)
y_center = int(ymin * H) + int(((ymax - ymin) / 2) * H)
x_center = int(xmin * W) + int(((xmax - xmin) / 2) * W)
lip = frame[
y_center - half : y_center + half,
x_center - half : x_center + half,
:,
]
try:
lip = cv2.resize(lip, (self.image_crop_size, self.image_crop_size))
except Exception:
lip = np.zeros([self.image_crop_size, self.image_crop_size, 3], dtype=np.uint8)
mouth_frames.append(cv2.cvtColor(lip, cv2.COLOR_RGB2GRAY))
return np.stack(mouth_frames, axis=0)
def _extract_mouth_legacy(self, frames: NDArray[np.uint8]) -> NDArray[np.uint8]:
mouth_frames = []
top_idx, right_idx, bottom_idx, left_idx = self.landmark_indices
with mp_face_mesh.FaceMesh(
static_image_mode=self.static_image_mode,
max_num_faces=1,
refine_landmarks=self.refine_landmarks,
min_detection_confidence=self.min_detection_confidence,
min_tracking_confidence=self.min_tracking_confidence,
) as face_mesh:
for frame in frames:
res = face_mesh.process(frame)
if res.multi_face_landmarks is None or len(res.multi_face_landmarks) == 0:
mouth_frames.append(np.zeros([self.image_crop_size, self.image_crop_size], dtype=np.uint8))
continue
landmarks = res.multi_face_landmarks[0].landmark
top = landmarks[top_idx]
left = landmarks[left_idx]
right = landmarks[right_idx]
bottom = landmarks[bottom_idx]
H, W = frame.shape[:2]
xmax = max(top.x, left.x, right.x, bottom.x)
ymax = max(top.y, left.y, right.y, bottom.y)
xmin = min(top.x, left.x, right.x, bottom.x)
ymin = min(top.y, left.y, right.y, bottom.y)
patch_size = max((xmax - xmin) * W, (ymax - ymin) * H) # To extract square region
half = int(patch_size / 2)
y_center = int(ymin * H) + int(((ymax - ymin) / 2) * H)
x_center = int(xmin * W) + int(((xmax - xmin) / 2) * W)
lip = frame[
y_center - half : y_center + half,
x_center - half : x_center + half,
:,
]
try:
lip = cv2.resize(lip, (self.image_crop_size, self.image_crop_size))
except Exception:
lip = np.zeros([self.image_crop_size, self.image_crop_size, 3], dtype=np.uint8)
mouth_frames.append(cv2.cvtColor(lip, cv2.COLOR_RGB2GRAY))
return np.stack(mouth_frames, axis=0)
def _load_audio(self, audio: str | NDArray[np.float32]) -> torch.FloatTensor:
def stacker(feats, stack_order):
feat_dim = feats.shape[1]
if len(feats) % stack_order != 0:
res = stack_order - len(feats) % stack_order
res = np.zeros([res, feat_dim]).astype(feats.dtype)
feats = np.concatenate([feats, res], axis=0)
feats = feats.reshape((-1, stack_order, feat_dim)).reshape(-1, stack_order * feat_dim)
return feats
sr = None
if isinstance(audio, str):
audio, sr = librosa.load(audio, sr=16_000)
if sr is None:
sr = self.sr
fbank = logfbank(audio, samplerate=sr).astype(np.float32)
fbank = stacker(fbank, self.stack_order_audio)
return torch.from_numpy(fbank)
def _align_time_steps(
self, audio: list[torch.FloatTensor], video: list[torch.FloatTensor]
) -> tuple[list[torch.FloatTensor], list[torch.FloatTensor]]:
aligned_indices = []
for sample_audio, sample_video in zip(audio, video):
diff = len(sample_audio) - len(sample_video)
if diff != 0:
aligned_indices.append(
torch.arange(0, len(sample_audio)).float() * len(sample_video) / len(sample_audio)
)
else:
aligned_indices.append(torch.arange(0, len(sample_audio)))
return (
audio,
[
sample[torch.clamp(torch.floor(indices), max=sample.shape[0] - 1).long()]
for sample, indices in zip(video, aligned_indices)
],
)
def __call__(
self,
raw_audio: NDArray[np.float32] | str | list[NDArray[np.float32]] | list[str] | None = None,
raw_video: NDArray[np.uint8] | str | list[NDArray[np.uint8]] | list[str] | None = None,
extract_mouth: bool = False,
**kwargs,
) -> BatchFeature:
if not isinstance(raw_audio, list):
raw_audio = [raw_audio]
if not isinstance(raw_video, list):
raw_video = [raw_video]
audio = [self._load_audio(sample) if sample is not None else None for sample in raw_audio]
video = [self._load_video(sample, extract_mouth) if sample is not None else None for sample in raw_video]
for batch_idx in range(len(audio)):
sample_a = audio[batch_idx]
sample_v = video[batch_idx]
assert sample_a is not None or sample_v is not None
if sample_a is None:
sample_a = torch.zeros((sample_v.shape[0], 26 * self.stack_order_audio))
audio[batch_idx] = sample_a
elif sample_v is None: # 25 fps
sample_v = torch.zeros((sample_a.shape[0], 1, self.image_crop_size, self.image_crop_size))
video[batch_idx] = sample_v
audio, video = self._align_time_steps(audio, video)
max_length = max(len(data) for data in audio)
input_values = []
pixel_values = []
padding_mask = []
for feat_audio, feat_video in zip(audio, video):
remainder_length = max_length - len(feat_audio)
audio_remainder = torch.zeros(
size=(remainder_length,) + feat_audio.size()[1:],
dtype=feat_audio.dtype,
)
video_remainder = torch.zeros(
size=(remainder_length,) + feat_video.size()[1:],
dtype=feat_video.dtype,
)
feat_audio = torch.cat((feat_audio, audio_remainder))
feat_video = torch.cat((feat_video, video_remainder))
if self.max_sample_size:
feat_audio = feat_audio[: self.max_sample_size]
feat_video = feat_video[: self.max_sample_size]
pad_mask = torch.zeros(max_length)
pad_mask[max_length - remainder_length :] = 1
input_values.append(feat_audio)
pixel_values.append(feat_video)
padding_mask.append(pad_mask)
input_values = torch.stack(input_values)
batch = BatchFeature(
{
"input_values": (
F.layer_norm(input_values, input_values.shape[2:]) if self.normalize else input_values
),
"pixel_values": self.transforms(torch.stack(pixel_values)),
"padding_mask": torch.stack(padding_mask),
}
)
return batch
def to_dict(self):
output = super().to_dict()
output["transforms"] = self._transforms_to_dict(output["transforms"])
return output
def _transforms_to_dict(self, transforms: transforms.Compose):
output = []
for component in transforms.__dict__["transforms"]:
name = component.__class__.__name__
component_dict = {"transforms_type": name}
for k, v in component.__dict__.items():
if k.startswith("_"):
continue
component_dict[k] = str(v)
output.append(component_dict)
return output
|