File size: 13,759 Bytes
eca55dc | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 | import os
from functools import partial
from typing import Any, Dict, List, Optional, Union
import lightning as L
import numpy as np
import pandas as pd
import torch
import torchaudio
from torch.utils.data import DataLoader, Dataset
from src.data.audio_utils import DatasetResamplerCropper, collate_audio_batch
class YT1BDataset(Dataset):
"""
Dataset for YT-Temporal-1B data using Parquet metadata files.
Args:
parquet_path (str): Path to the parquet file containing metadata (must include 'file_path', 'video_id', 'duration_sec').
If a 'sample_rate' column exists, it is used to avoid probing files for source sample rate.
min_duration (Optional[float]): Minimum duration in seconds to include a file.
max_duration (Optional[float]): Maximum duration in seconds to include a file.
transform (Optional[callable]): Optional transform to apply to the waveform.
max_length (Optional[int]): Maximum length of the waveform in samples (at target_sample_rate).
target_sample_rate (int): Target sample rate for the waveform. Defaults to 16000.
decode_window_sec (Optional[float]): Optional decode window length in seconds. If None,
defaults to max_length / target_sample_rate (when max_length is set).
"""
def __init__(
self,
parquet_path: str,
min_duration: Optional[float] = None,
max_duration: Optional[float] = 30.0,
transform: Optional[Any] = None,
max_length: Optional[int] = None,
target_sample_rate: int = 16000,
decode_window_sec: Optional[float] = None,
):
print(f"Loading metadata from {parquet_path}...")
self.transform = transform
self.max_length = max_length
self.target_sample_rate = target_sample_rate
self.decode_window_sec = decode_window_sec
# --- Metadata Loading ---
if not os.path.exists(parquet_path):
raise FileNotFoundError(f"Parquet file not found at: {parquet_path}")
# Pyarrow is required for read_parquet
try:
df = pd.read_parquet(parquet_path)
except ImportError:
raise ImportError(
"Please install pyarrow to read parquet files: `uv add pyarrow`"
)
required_cols = {"file_path", "video_id", "duration_sec"}
if not required_cols.issubset(df.columns):
# Check if we have compatible columns or raise error
# Some datasets might use different names, strictly enforcing for now based on user prompt
raise ValueError(
f"Parquet file must contain columns: {required_cols}. Found: {df.columns.tolist()}"
)
if min_duration is not None and min_duration < 0:
raise ValueError(f"min_duration must be >= 0, got {min_duration}")
if max_duration is not None and max_duration < 0:
raise ValueError(f"max_duration must be >= 0, got {max_duration}")
if (
min_duration is not None
and max_duration is not None
and min_duration > max_duration
):
raise ValueError(
"min_duration must be <= max_duration; "
f"got min_duration={min_duration}, max_duration={max_duration}"
)
if min_duration is not None:
df = df[df["duration_sec"] >= min_duration]
if max_duration is not None:
df = df[df["duration_sec"] <= max_duration]
self.ids = df["video_id"].tolist()
self.paths = df["file_path"].tolist()
self.durations_sec = df["duration_sec"].tolist()
if "sample_rate" in df.columns:
sample_rates = pd.to_numeric(df["sample_rate"], errors="coerce").to_numpy(
dtype=np.float64
)
self.source_sample_rates: Optional[list[Optional[int]]] = [
int(sr) if np.isfinite(sr) and sr > 0 else None for sr in sample_rates
]
else:
self.source_sample_rates = None
self.length = len(self.ids)
# --- Resampler ---
# Uses the optimized class that caches resamplers
self.resampler = DatasetResamplerCropper(
target_sr=target_sample_rate, max_length=max_length
)
print(f"Dataset loaded. Length: {self.length:,}")
def __len__(self) -> int:
return self.length
def __getitem__(self, idx: int) -> Dict[str, Union[torch.Tensor, str, int]]:
audio_path = self.paths[idx]
audio_id = self.ids[idx]
# Load waveform
try:
decode_window_sec = self.decode_window_sec
if decode_window_sec is None and self.max_length is not None:
decode_window_sec = self.max_length / self.target_sample_rate
if decode_window_sec is None:
waveform, sr = torchaudio.load(audio_path)
else:
duration_sec = float(self.durations_sec[idx])
if duration_sec <= 0:
waveform, sr = torchaudio.load(audio_path)
else:
source_sr: Optional[int]
if self.source_sample_rates is not None:
source_sr = self.source_sample_rates[idx]
else:
source_sr = None
if source_sr is None:
_, source_sr = torchaudio.load(
audio_path, frame_offset=0, num_frames=1
)
total_frames = max(1, int(duration_sec * source_sr))
max_decode_frames = max(1, int(decode_window_sec * source_sr))
decode_frames = min(max_decode_frames, total_frames)
if total_frames > decode_frames:
max_start = total_frames - decode_frames
frame_offset = int(np.random.randint(0, max_start + 1))
else:
frame_offset = 0
waveform, sr = torchaudio.load(
audio_path,
frame_offset=frame_offset,
num_frames=decode_frames,
)
except Exception as e:
print(f"Error loading {audio_path}: {e}")
# Return a dummy silent waveform to prevent crash
len_samples = (
self.max_length if self.max_length else self.target_sample_rate
)
return {
"waveform": torch.zeros(1, len_samples),
"audio_name": audio_id,
"index": idx,
"error": True,
}
# Mix down to mono if necessary
if waveform.shape[0] > 1:
waveform = torch.mean(waveform, dim=0, keepdim=True)
# Resample and crop
waveform = self.resampler(waveform, source_sr=sr)
# Ensure channel dim exists [1, T] if resampler stripped it or returned [T]
if waveform.ndim == 1:
waveform = waveform.unsqueeze(0)
if self.transform:
waveform = self.transform(waveform)
return {
"waveform": waveform,
"audio_name": audio_id,
"index": idx,
}
class YT1BDataModule(L.LightningDataModule):
"""
LightningDataModule for YT-Temporal-1B.
Args:
data_dir (str): Root directory for data.
train_parquet (str): Filename of training parquet file.
val_parquet (str): Filename of validation parquet file.
test_parquet (str): Filename of test parquet file.
batch_size (int): Batch size for dataloaders.
num_workers (int): Number of workers for dataloaders.
pin_memory (bool): Whether to pin memory in dataloaders.
max_audio_length_sec (Optional[float]): Maximum audio length in seconds.
min_duration_sec (Optional[float]): Minimum audio duration in seconds to filter.
max_duration_sec (Optional[float]): Maximum audio duration in seconds to filter.
target_sample_rate (int): Target sample rate.
collate_mode (str): 'pad' or 'truncate'.
decode_window_sec (Optional[float]): Optional decode window length in seconds. If None,
defaults to max_audio_length_sec.
"""
def __init__(
self,
data_dir: str = "data/YT-Temporal-1B",
train_parquet: str = "train_metadata.parquet",
val_parquet: str = "val_metadata.parquet",
test_parquet: str = "val_metadata.parquet",
batch_size: int = 64,
num_workers: int = 4,
pin_memory: bool = True,
max_audio_length_sec: Optional[float] = 10.0,
min_duration_sec: Optional[float] = None,
max_duration_sec: Optional[float] = 30.0,
target_sample_rate: int = 16000,
collate_mode: str = "pad",
decode_window_sec: Optional[float] = None,
):
super().__init__()
self.save_hyperparameters()
self.data_dir = data_dir
self.train_parquet_path = os.path.join(data_dir, train_parquet)
self.val_parquet_path = os.path.join(data_dir, val_parquet)
self.test_parquet_path = os.path.join(data_dir, test_parquet)
if max_audio_length_sec is not None:
self.max_audio_length = int(max_audio_length_sec * target_sample_rate)
else:
self.max_audio_length = None
self.train_dataset: Optional[YT1BDataset] = None
self.val_dataset: Optional[YT1BDataset] = None
self.test_dataset: Optional[YT1BDataset] = None
def setup(self, stage: Optional[str] = None) -> None:
if stage == "fit" or stage is None:
if os.path.exists(self.train_parquet_path):
self.train_dataset = YT1BDataset(
self.train_parquet_path,
min_duration=self.hparams["min_duration_sec"],
max_duration=self.hparams["max_duration_sec"],
max_length=self.max_audio_length,
target_sample_rate=self.hparams["target_sample_rate"],
decode_window_sec=self.hparams["decode_window_sec"],
)
if os.path.exists(self.val_parquet_path):
self.val_dataset = YT1BDataset(
self.val_parquet_path,
min_duration=self.hparams["min_duration_sec"],
max_duration=self.hparams["max_duration_sec"],
max_length=self.max_audio_length,
target_sample_rate=self.hparams["target_sample_rate"],
decode_window_sec=self.hparams["decode_window_sec"],
)
if stage == "test":
if os.path.exists(self.test_parquet_path):
self.test_dataset = YT1BDataset(
self.test_parquet_path,
min_duration=self.hparams["min_duration_sec"],
max_duration=self.hparams["max_duration_sec"],
max_length=self.max_audio_length,
target_sample_rate=self.hparams["target_sample_rate"],
decode_window_sec=self.hparams["decode_window_sec"],
)
def train_dataloader(self) -> DataLoader:
if not self.train_dataset:
raise RuntimeError(
f"Train dataset not initialized. File not found: {self.train_parquet_path}"
)
return DataLoader(
self.train_dataset,
batch_size=self.hparams["batch_size"],
shuffle=True,
num_workers=self.hparams["num_workers"],
pin_memory=self.hparams["pin_memory"],
persistent_workers=self.hparams["num_workers"] > 0,
collate_fn=partial(self.collate_fn, mode=self.hparams["collate_mode"]),
)
def val_dataloader(self) -> DataLoader:
if not self.val_dataset:
# Often validation sets are missing in large scale pretraining or we use a subset of train
# For now, raise strict error or return empty list (lightning supports empty list for no val)
# Raising error is safer to debug configuration issues.
raise RuntimeError(
f"Val dataset not initialized. File not found: {self.val_parquet_path}"
)
return DataLoader(
self.val_dataset,
batch_size=self.hparams["batch_size"],
shuffle=False,
num_workers=self.hparams["num_workers"],
pin_memory=self.hparams["pin_memory"],
persistent_workers=self.hparams["num_workers"] > 0,
collate_fn=partial(self.collate_fn, mode=self.hparams["collate_mode"]),
)
def test_dataloader(self) -> DataLoader:
if not self.test_dataset:
raise RuntimeError(
f"Test dataset not initialized. File not found: {self.test_parquet_path}"
)
return DataLoader(
self.test_dataset,
batch_size=self.hparams["batch_size"],
shuffle=False,
num_workers=self.hparams["num_workers"],
pin_memory=self.hparams["pin_memory"],
collate_fn=partial(self.collate_fn, mode=self.hparams["collate_mode"]),
)
@staticmethod
def collate_fn(batch: List[Dict[str, Any]], mode: str = "pad") -> Dict[str, Any]:
# Filter out errors
batch = [x for x in batch if not x.get("error", False)]
if len(batch) == 0:
raise RuntimeError("All items in batch failed to load.")
return collate_audio_batch(
batch=batch,
waveform_key="waveform",
mode=mode,
)
|