File size: 25,087 Bytes
9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 8fd1a55 9f7d425 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 | from __future__ import annotations
import json
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Mapping, Sequence
import torch
from torch.utils.data import DataLoader, Dataset
from data.encodec_utils import (
PRESNAP_TARGET_LAYOUT,
normalize_target_payload,
resolve_target_dim_from_cache_config,
resolve_target_layout_from_cache_config,
)
@dataclass(frozen=True)
class DiffusionExample:
example_path: Path
source_id: str
source_manifest_index: int
source_pt_rel: str
source_row_in_shard: int
beat_index: int
beat_index_end: int
split: str
kit_name: str
source_row_id: str
conditioning_mode: str
class_names: tuple[str, ...]
class_id_vocab_sizes: tuple[int, ...]
feature_row_names: tuple[str, ...]
grid_ft: torch.Tensor
grid_ids_ft: torch.Tensor
family_onsets_ft: torch.Tensor
family_onset_count_ft: torch.Tensor
grid_times_sec_t: torch.Tensor
token_times_sec_t: torch.Tensor
beat_boundaries_sec_rel: torch.Tensor
grid_num_frames: int
grid_frame_rate: float
bpm: float
duration_sec: float
target_sum_td: torch.Tensor
target_full_sum_td: torch.Tensor
target_dim: int
target_full_dim: int
target_layout: str
pca_basis_path: str
target_num_frames: int
source_codes_ct: torch.Tensor
target_sum_pool_d: torch.Tensor
@property
def target_sum_t128(self) -> torch.Tensor:
return self.target_sum_td
@property
def target_sum_pool_128(self) -> torch.Tensor:
return self.target_sum_pool_d
from data.grid_rate_downsample import downsample_grid_payload, uniform_grid_times
def _load_jsonl(path: Path) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
with path.open("r", encoding="utf-8") as handle:
for line in handle:
text = str(line).strip()
if not text:
continue
rows.append(dict(json.loads(text)))
return rows
def _require_payload_keys(payload: dict[str, Any], keys: Sequence[str], *, example_path: Path) -> None:
missing = [str(key) for key in list(keys) if str(key) not in payload]
if missing:
raise RuntimeError(
f"cache example {example_path} is missing required seconds-conditioning fields {missing}. "
"Rebuild the cache with scripts/build_diffusion_cache.py."
)
class DiffusionConditioningDataset(Dataset[DiffusionExample]):
def __init__(
self,
cache_root: str | Path,
*,
split: str = "train",
max_items: int = 0,
conditioning_mode: str = "seconds",
grid_rate_hz: float = 0.0,
grid_overlay_root: str | Path | None = None,
) -> None:
super().__init__()
self.cache_root = Path(cache_root).resolve()
self.split = str(split).strip().lower()
self.conditioning_mode = str(conditioning_mode).strip().lower()
if self.conditioning_mode not in {"seconds", "seconds_frontend"}:
raise ValueError(
f"unsupported conditioning_mode={conditioning_mode!r}; expected 'seconds' or 'seconds_frontend'"
)
self.target_dim = int(resolve_target_dim_from_cache_config(self.cache_root))
self.target_layout = str(resolve_target_layout_from_cache_config(self.cache_root))
self.grid_overlay_root = (
None
if grid_overlay_root is None or not str(grid_overlay_root).strip()
else Path(grid_overlay_root).expanduser().resolve()
)
if self.grid_overlay_root is not None:
overlay_config_path = self.grid_overlay_root / "config.json"
if not overlay_config_path.is_file():
raise FileNotFoundError(f"missing grid-overlay config: {overlay_config_path}")
overlay_config = json.loads(overlay_config_path.read_text(encoding="utf-8"))
if str(overlay_config.get("format")) != "grid2drum-grid-overlay-v1":
raise ValueError(f"unsupported grid-overlay format in {overlay_config_path}")
# 0 disables decimation entirely (cached rate is used as-is).
self.grid_rate_hz = float(max(0.0, float(grid_rate_hz)))
manifest_path = self.cache_root / "manifests" / f"{self.split}.jsonl"
self.rows = _load_jsonl(manifest_path)
if int(max_items) > 0:
self.rows = self.rows[: int(max_items)]
if not self.rows:
raise RuntimeError(f"no rows found for split={self.split!r} under {self.cache_root}")
def __len__(self) -> int:
return int(len(self.rows))
def __getitem__(self, index: int) -> DiffusionExample:
return self.example_from_row(self.rows[int(index)])
def example_from_row(self, row: Mapping[str, Any]) -> DiffusionExample:
row = dict(row)
example_path = (self.cache_root / str(row["out_pt"])).resolve()
payload = dict(torch.load(example_path, map_location="cpu", weights_only=False))
if self.grid_overlay_root is not None:
overlay_path = (self.grid_overlay_root / str(row["out_pt"])).resolve()
if not overlay_path.is_file():
raise FileNotFoundError(f"missing grid overlay for cache example: {overlay_path}")
overlay = dict(torch.load(overlay_path, map_location="cpu", weights_only=False))
if str(overlay.get("source_id") or "") != str(payload.get("source_id") or ""):
raise RuntimeError(f"grid-overlay source mismatch for {example_path}")
for key in (
"grid_ft",
"grid_ids_ft",
"family_onsets_ft",
"family_onset_count_ft",
"grid_times_sec_t",
"grid_num_frames",
"grid_frame_rate",
):
payload[key] = overlay[key]
_require_payload_keys(
payload,
(
"grid_ft",
"grid_ids_ft",
"family_onsets_ft",
"family_onset_count_ft",
"grid_times_sec_t",
"token_times_sec_t",
"beat_boundaries_sec_rel",
"grid_num_frames",
"grid_frame_rate",
"bpm",
"duration_sec",
"target_num_frames",
"source_codes_ct",
),
example_path=example_path,
)
target_payload = payload
if self.target_layout == PRESNAP_TARGET_LAYOUT:
target_payload = {**payload, "target_layout": PRESNAP_TARGET_LAYOUT, "pca_basis_path": ""}
target_sum_td, target_sum_pool_d, target_dim = normalize_target_payload(
target_payload,
fallback_target_dim=int(self.target_dim),
)
raw_target_sum_td = torch.as_tensor(
payload.get("target_sum_td", payload.get("target_sum_t128")),
dtype=torch.float32,
).contiguous()
raw_target_sum_pool_d = torch.as_tensor(
payload.get("target_sum_pool_d", payload.get("target_sum_pool_128")),
dtype=torch.float32,
).contiguous()
target_num_frames = int(payload["target_num_frames"])
if int(target_sum_td.shape[0]) != int(target_num_frames):
raise RuntimeError(
f"target_num_frames mismatch: target_sum_td has {target_sum_td.shape[0]} frames vs {target_num_frames}"
)
if int(raw_target_sum_td.shape[0]) != int(target_num_frames):
raise RuntimeError(
f"raw target_num_frames mismatch: target_sum_td has {raw_target_sum_td.shape[0]} frames vs {target_num_frames}"
)
grid_ft = torch.as_tensor(payload["grid_ft"], dtype=torch.float32).contiguous()
grid_ids_ft = torch.as_tensor(payload["grid_ids_ft"], dtype=torch.long).contiguous()
family_onsets_ft = torch.as_tensor(payload["family_onsets_ft"], dtype=torch.bool).contiguous()
family_onset_count_ft = torch.as_tensor(payload["family_onset_count_ft"], dtype=torch.uint8).contiguous()
grid_times_sec_t = torch.as_tensor(payload["grid_times_sec_t"], dtype=torch.float32).contiguous()
token_times_sec_t = torch.as_tensor(payload["token_times_sec_t"], dtype=torch.float32).contiguous()
beat_boundaries_sec_rel = torch.as_tensor(payload["beat_boundaries_sec_rel"], dtype=torch.float32).contiguous()
source_codes_ct = torch.as_tensor(payload["source_codes_ct"], dtype=torch.int16).contiguous()
grid_num_frames = int(payload["grid_num_frames"])
grid_frame_rate = float(payload["grid_frame_rate"])
bpm = float(payload["bpm"])
duration_sec = float(payload["duration_sec"])
if int(grid_ft.dim()) != 2:
raise RuntimeError(f"expected grid_ft [F,Tg], got {tuple(grid_ft.shape)}")
if int(grid_ids_ft.dim()) != 2:
raise RuntimeError(f"expected grid_ids_ft [C,Tg], got {tuple(grid_ids_ft.shape)}")
if int(family_onsets_ft.dim()) != 2 or int(family_onset_count_ft.dim()) != 2:
raise RuntimeError(
f"expected family onset tensors [C,Tg], got {tuple(family_onsets_ft.shape)} / {tuple(family_onset_count_ft.shape)}"
)
if float(self.grid_rate_hz) > 0.0:
decimated = downsample_grid_payload(
grid_ft=grid_ft,
grid_ids_ft=grid_ids_ft,
family_onsets_ft=family_onsets_ft,
family_onset_count_ft=family_onset_count_ft,
grid_num_frames=int(grid_num_frames),
duration_sec=float(payload["duration_sec"]),
target_rate_hz=float(self.grid_rate_hz),
)
grid_ft = decimated["grid_ft"]
grid_ids_ft = decimated["grid_ids_ft"]
family_onsets_ft = decimated["family_onsets_ft"]
family_onset_count_ft = decimated["family_onset_count_ft"]
grid_num_frames = int(decimated["grid_num_frames"])
grid_frame_rate = float(decimated["grid_frame_rate"])
grid_times_sec_t = uniform_grid_times(
int(grid_num_frames),
float(payload["duration_sec"]),
)
if int(grid_ft.shape[-1]) != int(grid_num_frames):
raise RuntimeError(
f"grid_num_frames mismatch: grid_ft has {grid_ft.shape[-1]} frames vs {grid_num_frames}"
)
if int(grid_ids_ft.shape[-1]) != int(grid_num_frames):
raise RuntimeError(
f"grid_ids_ft mismatch: grid_ids_ft has {grid_ids_ft.shape[-1]} frames vs {grid_num_frames}"
)
if int(grid_times_sec_t.shape[0]) != int(grid_num_frames):
raise RuntimeError(
f"grid_times_sec_t mismatch: grid_times_sec_t has {grid_times_sec_t.shape[0]} frames vs {grid_num_frames}"
)
if int(token_times_sec_t.shape[0]) != int(target_num_frames):
raise RuntimeError(
f"token_times_sec_t mismatch: token_times_sec_t has {token_times_sec_t.shape[0]} frames vs {target_num_frames}"
)
class_names = tuple(str(x) for x in list(payload.get("class_names") or []))
class_id_vocab_sizes = tuple(int(x) for x in list(payload.get("class_id_vocab_sizes") or []))
feature_row_names = tuple(str(x) for x in list(payload.get("feature_row_names") or []))
return DiffusionExample(
example_path=example_path,
source_id=str(payload.get("source_id") or row.get("source_id") or ""),
source_manifest_index=int(payload.get("source_manifest_index", row.get("source_manifest_index", -1))),
source_pt_rel=str(payload.get("source_pt_rel") or row.get("source_pt_rel") or ""),
source_row_in_shard=int(payload.get("source_row_in_shard", row.get("source_row_in_shard", -1))),
beat_index=int(payload.get("beat_index", row.get("beat_index", 0))),
beat_index_end=int(payload.get("beat_index_end", row.get("beat_index_end", 0))),
split=str(payload.get("split") or row.get("split") or self.split),
kit_name=str(payload.get("kit_name") or row.get("kit_name") or ""),
source_row_id=str(payload.get("source_row_id") or row.get("source_row_id") or ""),
conditioning_mode=str(payload.get("conditioning_mode") or "midi_family_state_onset_ids"),
class_names=class_names,
class_id_vocab_sizes=class_id_vocab_sizes,
feature_row_names=feature_row_names,
grid_ft=grid_ft,
grid_ids_ft=grid_ids_ft,
family_onsets_ft=family_onsets_ft,
family_onset_count_ft=family_onset_count_ft,
grid_times_sec_t=grid_times_sec_t,
token_times_sec_t=token_times_sec_t,
beat_boundaries_sec_rel=beat_boundaries_sec_rel,
grid_num_frames=grid_num_frames,
grid_frame_rate=grid_frame_rate,
bpm=bpm,
duration_sec=duration_sec,
target_sum_td=target_sum_td,
target_full_sum_td=raw_target_sum_td,
target_dim=int(target_dim),
target_full_dim=int(raw_target_sum_td.shape[-1]),
target_layout=(
PRESNAP_TARGET_LAYOUT
if self.target_layout == PRESNAP_TARGET_LAYOUT
else str(payload.get("target_layout") or "framewise_sum")
),
pca_basis_path=(
""
if self.target_layout == PRESNAP_TARGET_LAYOUT
else str(payload.get("pca_basis_path") or "")
),
target_num_frames=target_num_frames,
source_codes_ct=source_codes_ct,
target_sum_pool_d=raw_target_sum_pool_d,
)
def collate_diffusion_examples(items: Sequence[DiffusionExample]) -> dict[str, Any]:
if not items:
raise ValueError("expected at least one DiffusionExample")
batch_size = int(len(items))
grid_len = int(max(item.grid_num_frames for item in items))
grid_dim = int(items[0].grid_ft.shape[0])
class_dim = int(items[0].grid_ids_ft.shape[0])
target_len = int(max(item.target_num_frames for item in items))
codebooks = int(items[0].source_codes_ct.shape[0])
beat_boundary_len = int(max(item.beat_boundaries_sec_rel.shape[0] for item in items))
target_dim = int(items[0].target_dim)
target_full_dim = int(items[0].target_full_dim)
if any(int(item.target_dim) != int(target_dim) for item in items):
raise ValueError("all items in a batch must share target_dim")
if any(int(item.target_full_dim) != int(target_full_dim) for item in items):
raise ValueError("all items in a batch must share target_full_dim")
target_layout = str(items[0].target_layout)
if any(str(item.target_layout) != str(target_layout) for item in items):
raise ValueError("all items in a batch must share target_layout")
pca_basis_path = str(items[0].pca_basis_path)
if any(str(item.pca_basis_path) != str(pca_basis_path) for item in items):
raise ValueError("all items in a batch must share pca_basis_path")
grid_bft = torch.zeros((batch_size, grid_dim, grid_len), dtype=torch.float32)
grid_ids_bct = torch.full((batch_size, class_dim, grid_len), -1, dtype=torch.long)
family_onsets_bft = torch.zeros((batch_size, class_dim, grid_len), dtype=torch.bool)
family_onset_count_bft = torch.zeros((batch_size, class_dim, grid_len), dtype=torch.uint8)
grid_valid_mask_bt = torch.zeros((batch_size, grid_len), dtype=torch.bool)
grid_times_sec_bt = torch.zeros((batch_size, grid_len), dtype=torch.float32)
token_times_sec_bt = torch.zeros((batch_size, target_len), dtype=torch.float32)
beat_boundaries_sec_bk = torch.zeros((batch_size, beat_boundary_len), dtype=torch.float32)
beat_boundaries_valid_mask_bk = torch.zeros((batch_size, beat_boundary_len), dtype=torch.bool)
target_btd = torch.zeros((batch_size, target_len, target_dim), dtype=torch.float32)
target_sum_btd = torch.zeros((batch_size, target_len, target_full_dim), dtype=torch.float32)
target_valid_mask_bt = torch.zeros((batch_size, target_len), dtype=torch.bool)
source_codes_bct = torch.full((batch_size, codebooks, target_len), -1, dtype=torch.int16)
target_sum_pool_bd = torch.zeros((batch_size, target_full_dim), dtype=torch.float32)
bpm_b = torch.zeros((batch_size,), dtype=torch.float32)
duration_sec_b = torch.zeros((batch_size,), dtype=torch.float32)
grid_frame_rate_b = torch.zeros((batch_size,), dtype=torch.float32)
for row_idx, item in enumerate(items):
gf = int(item.grid_num_frames)
tf = int(item.target_num_frames)
bk = int(item.beat_boundaries_sec_rel.shape[0])
grid_bft[int(row_idx), :, : int(gf)] = item.grid_ft[:, : int(gf)]
grid_ids_bct[int(row_idx), :, : int(gf)] = item.grid_ids_ft[:, : int(gf)]
family_onsets_bft[int(row_idx), :, : int(gf)] = item.family_onsets_ft[:, : int(gf)]
family_onset_count_bft[int(row_idx), :, : int(gf)] = item.family_onset_count_ft[:, : int(gf)]
grid_valid_mask_bt[int(row_idx), : int(gf)] = True
grid_times_sec_bt[int(row_idx), : int(gf)] = item.grid_times_sec_t[: int(gf)]
token_times_sec_bt[int(row_idx), : int(tf)] = item.token_times_sec_t[: int(tf)]
beat_boundaries_sec_bk[int(row_idx), : int(bk)] = item.beat_boundaries_sec_rel[: int(bk)]
beat_boundaries_valid_mask_bk[int(row_idx), : int(bk)] = True
target_btd[int(row_idx), : int(tf), :] = item.target_sum_td[: int(tf)]
target_sum_btd[int(row_idx), : int(tf), :] = item.target_full_sum_td[: int(tf)]
target_valid_mask_bt[int(row_idx), : int(tf)] = True
source_codes_bct[int(row_idx), :, : int(tf)] = item.source_codes_ct[:, : int(tf)]
target_sum_pool_bd[int(row_idx), :] = item.target_sum_pool_d
bpm_b[int(row_idx)] = float(item.bpm)
duration_sec_b[int(row_idx)] = float(item.duration_sec)
grid_frame_rate_b[int(row_idx)] = float(item.grid_frame_rate)
batch = {
"conditioning_mode": str(items[0].conditioning_mode),
"class_names": list(items[0].class_names),
"class_id_vocab_sizes": [int(x) for x in list(items[0].class_id_vocab_sizes)],
"feature_row_names": list(items[0].feature_row_names),
"grid": grid_bft.contiguous(),
"grid_ids": grid_ids_bct.contiguous(),
"family_onsets_bft": family_onsets_bft.contiguous(),
"family_onset_count_bft": family_onset_count_bft.contiguous(),
"grid_valid_mask": grid_valid_mask_bt.contiguous(),
"grid_times_sec": grid_times_sec_bt.contiguous(),
"token_times_sec": token_times_sec_bt.contiguous(),
"beat_boundaries_sec": beat_boundaries_sec_bk.contiguous(),
"beat_boundaries_valid_mask": beat_boundaries_valid_mask_bk.contiguous(),
"bpm": bpm_b.contiguous(),
"duration_sec": duration_sec_b.contiguous(),
"grid_frame_rate_b": grid_frame_rate_b.contiguous(),
"grid_num_frames_b": torch.tensor([int(item.grid_num_frames) for item in items], dtype=torch.long),
"target_layout": str(target_layout),
"target_dim": int(target_dim),
"target_full_dim": int(target_full_dim),
"target_btd": target_btd.contiguous(),
"target_sum_btd": target_sum_btd.contiguous(),
"target_valid_mask_bt": target_valid_mask_bt.contiguous(),
"source_codes_bct": source_codes_bct.contiguous(),
"target_sum_pool_bd": target_sum_pool_bd.contiguous(),
"target_pca_basis_path": str(pca_basis_path),
"target_num_frames_b": torch.tensor([int(item.target_num_frames) for item in items], dtype=torch.long),
"source_manifest_index_b": torch.tensor([int(item.source_manifest_index) for item in items], dtype=torch.long),
"source_row_in_shard_b": torch.tensor([int(item.source_row_in_shard) for item in items], dtype=torch.long),
"beat_index_b": torch.tensor([int(item.beat_index) for item in items], dtype=torch.long),
"beat_index_end_b": torch.tensor([int(item.beat_index_end) for item in items], dtype=torch.long),
"source_id": [str(item.source_id) for item in items],
"source_pt_rel": [str(item.source_pt_rel) for item in items],
"split": [str(item.split) for item in items],
"kit_name": [str(item.kit_name) for item in items],
"source_row_id": [str(item.source_row_id) for item in items],
"example_path": [str(item.example_path) for item in items],
}
return batch
@torch.no_grad()
def estimate_target_normalization(
dataloader: DataLoader,
*,
device: str | torch.device = "cpu",
eps: float = 1.0e-6,
) -> tuple[torch.Tensor, torch.Tensor]:
resolved_device = torch.device(device)
target_sum_d: torch.Tensor | None = None
target_sq_sum_d: torch.Tensor | None = None
total_frames = 0
for batch in dataloader:
target_btd = torch.as_tensor(batch["target_btd"], dtype=torch.float32, device=resolved_device)
target_valid_mask_bt = torch.as_tensor(
batch["target_valid_mask_bt"],
dtype=torch.bool,
device=resolved_device,
)
if int(target_btd.dim()) != 3:
raise ValueError(f"expected target_btd [B,T,D], got {tuple(target_btd.shape)}")
if tuple(target_valid_mask_bt.shape) != tuple(target_btd.shape[:2]):
raise ValueError(
"target_valid_mask_bt must match target_btd batch/time dimensions, got "
f"{tuple(target_valid_mask_bt.shape)} vs {tuple(target_btd.shape[:2])}"
)
if target_sum_d is None:
target_sum_d = torch.zeros((int(target_btd.shape[-1]),), dtype=torch.float32, device=resolved_device)
target_sq_sum_d = torch.zeros_like(target_sum_d)
valid_btd = target_btd * target_valid_mask_bt.unsqueeze(-1).to(dtype=target_btd.dtype)
target_sum_d += valid_btd.sum(dim=(0, 1))
target_sq_sum_d += valid_btd.square().sum(dim=(0, 1))
total_frames += int(target_valid_mask_bt.sum().item())
if target_sum_d is None or target_sq_sum_d is None or int(total_frames) <= 0:
raise RuntimeError("could not estimate target normalization from an empty dataloader")
mean_d = target_sum_d / float(total_frames)
var_d = (target_sq_sum_d / float(total_frames)) - mean_d.square()
std_d = var_d.clamp_min(float(eps)).sqrt()
return mean_d.contiguous(), std_d.contiguous()
@torch.no_grad()
def estimate_target_range(
dataloader: DataLoader,
*,
device: str | torch.device = "cpu",
) -> tuple[torch.Tensor, torch.Tensor]:
"""Per-dimension minimum and maximum over the valid target frames."""
resolved_device = torch.device(device)
target_min_d: torch.Tensor | None = None
target_max_d: torch.Tensor | None = None
for batch in dataloader:
target_btd = torch.as_tensor(batch["target_btd"], dtype=torch.float32, device=resolved_device)
valid_td = target_btd[torch.as_tensor(batch["target_valid_mask_bt"], dtype=torch.bool, device=resolved_device)]
if int(valid_td.shape[0]) == 0:
continue
batch_min_d, batch_max_d = valid_td.amin(dim=0), valid_td.amax(dim=0)
target_min_d = batch_min_d if target_min_d is None else torch.minimum(target_min_d, batch_min_d)
target_max_d = batch_max_d if target_max_d is None else torch.maximum(target_max_d, batch_max_d)
if target_min_d is None or target_max_d is None:
raise RuntimeError("could not measure the target range from an empty dataloader")
return target_min_d.contiguous(), target_max_d.contiguous()
def build_diffusion_dataloader(
cache_root: str | Path,
*,
split: str = "train",
batch_size: int = 8,
shuffle: bool = False,
num_workers: int = 0,
max_items: int = 0,
conditioning_mode: str = "seconds",
pin_memory: bool = False,
persistent_workers: bool = False,
multiprocessing_context: str | None = None,
grid_rate_hz: float = 0.0,
grid_overlay_root: str | Path | None = None,
) -> DataLoader:
dataset = DiffusionConditioningDataset(
cache_root,
split=split,
max_items=max_items,
conditioning_mode=conditioning_mode,
grid_rate_hz=float(grid_rate_hz),
grid_overlay_root=grid_overlay_root,
)
loader_kwargs: dict[str, Any] = {}
if int(num_workers) > 0:
loader_kwargs["persistent_workers"] = bool(persistent_workers)
if multiprocessing_context is not None and str(multiprocessing_context).strip():
loader_kwargs["multiprocessing_context"] = str(multiprocessing_context).strip()
return DataLoader(
dataset,
batch_size=int(batch_size),
shuffle=bool(shuffle),
num_workers=int(num_workers),
pin_memory=bool(pin_memory),
collate_fn=collate_diffusion_examples,
**loader_kwargs,
)
|