Spaces:
Runtime error
Runtime error
File size: 15,806 Bytes
7349148 |
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 |
import typing
import torch
import torch.nn.functional as F
from jaxtyping import Float
from src.loss.abstract_loss import AbstractLoss
from src.utils.math import sobol_sphere
def process_vector(
x: Float[torch.Tensor, "B D N"],
dirs: Float[torch.Tensor, "K D"],
) -> Float[torch.Tensor, "K B*N_valid"]:
"""
Project a 1-D sequence with a bank of linear directions.
Args
----
x : (B, D, N) tensor – predictions or ground truth
dirs : (K, D) tensor – unit-length projection directions
Returns
-------
proj : (K, B*N_valid) tensor of flattened projections
"""
B, D, N = x.shape
K, _ = dirs.shape
# linear projection: x (B,D,N) -> (B,N,K) -> (K,B*N)
proj = F.linear(x.transpose(1, 2).to(torch.float32), dirs.to(torch.float32))
proj = proj.permute(2, 0, 1).reshape(K, -1).to(x.dtype)
return proj
class VectorSWDLoss(AbstractLoss):
"""
1-D Sliced-Wasserstein Distance on sequences.
This loss computes the sliced Wasserstein distance between predicted and ground
truth sequences by projecting them onto random directions and computing the
Wasserstein distance in 1D. It supports reservoir sampling for adaptive direction
selection and various variance reduction techniques.
Parameters
----------
num_proj : int, default=64
Number of random projections to use per step (K).
distance : {"l1", "l2"}, default="l1"
Distance metric to use for computing the Wasserstein distance.
use_ucv : bool, default=False
Whether to use upper bounds control variates for variance reduction.
Mutually exclusive with use_lcv.
use_lcv : bool, default=False
Whether to use lower bounds control variates for variance reduction.
Mutually exclusive with use_ucv.
refresh_projections_every_n_steps : int, default=1
How often to refresh the projection directions. A value of 1 means
refresh every step, higher values reuse directions for multiple steps.
num_new_candidates : int, default=16
Number of new candidate directions to generate per step (M).
If 0, reservoir sampling is disabled. Must not exceed num_proj.
ess_alpha : float, default=0.5
Effective sample size threshold for resetting the reservoir.
When ESS drops below ess_alpha * reservoir_size, the reservoir is reset.
time_decay_tau : float or None, default=30.0
Time decay parameter for reservoir weights. If None, no time decay is applied.
Weights decay exponentially with age: exp(-age / time_decay_tau).
missing_value_method : {"random_replicate", "interpolate"},
default="random_replicate"
Method for handling sequences of different lengths:
- "random_replicate": Randomly replicate shorter sequences
- "interpolate": Use linear interpolation to match lengths
sampling_mode : {"gaussian", "qmc"}, default="qmc"
Method for generating random projection directions:
- "gaussian": Standard Gaussian sampling
- "qmc": Quasi-Monte Carlo sampling using Sobol sequences
Notes
-----
- Reservoir sampling is enabled when num_new_candidates > 0
- Reservoir size = num_proj - num_new_candidates
- When use_ucv or use_lcv is True, variance reduction is applied using
control variates based on the difference between sample and population means
- The loss automatically handles sequences of different lengths using the
specified missing_value_method
"""
def __init__(
self,
num_proj: int = 64,
distance: typing.Literal["l1", "l2"] = "l1",
use_ucv: bool = False,
use_lcv: bool = False,
refresh_projections_every_n_steps: int = 1,
num_new_candidates: int = 16,
ess_alpha: float = 0.5,
time_decay_tau: float | None = 30.0,
missing_value_method: typing.Literal[
"random_replicate", "interpolate"
] = "random_replicate",
sampling_mode: typing.Literal[
"gaussian",
"qmc",
] = "qmc",
):
super().__init__()
assert not (use_ucv and use_lcv), "use_ucv and use_lcv cannot both be True"
self.num_proj = num_proj
self.distance = distance
self.use_ucv = use_ucv
self.use_lcv = use_lcv
self.refresh_projections_every_n_steps = refresh_projections_every_n_steps
self.num_new_candidates = num_new_candidates # M
self.ess_alpha = ess_alpha
self.time_decay_tau = time_decay_tau
self.missing_value_method = missing_value_method
if num_new_candidates > 0 and self.refresh_projections_every_n_steps != 1:
# Print a warning that this is not recommended
print(
"WARNING: num_new_candidates > 0 (enabling reservoir sampling) and "
"refresh_projections_every_n_steps != 1 is not recommended"
)
assert (
num_new_candidates <= num_proj
), "`num_new_candidates` must not exceed `num_proj`"
# internal state for reservoir sampling
self.restir_enabled = self.num_new_candidates > 0
self.reservoir_size = self.num_proj - self.num_new_candidates
self.register_buffer("_reservoir_filters", torch.empty(0))
self.register_buffer("_reservoir_weights", torch.empty(0))
self.register_buffer("_reservoir_steps", torch.empty(0, dtype=torch.long))
self.register_buffer("_reservoir_keys", torch.empty(0))
self.register_buffer("_cumulative_weights", torch.tensor(0.0))
self.register_buffer("_has_reservoir", torch.tensor(False, dtype=torch.bool))
self._cached_dirs: typing.Optional[torch.Tensor] = None
self.sampling_mode = sampling_mode
self.sobol_engine = None
def _gaussian_proposals(self, k: int, d: int, device: torch.device) -> torch.Tensor:
"""Generate Gaussian random projection directions."""
w = torch.randn(k, d, device=device)
return w / (w.norm(dim=1, keepdim=True) + 1e-8) # unit length
def _qmc_proposals(self, k: int, d: int, device: torch.device) -> torch.Tensor:
"""Generate quasi-Monte Carlo projection directions using Sobol sequences."""
vecs, self.sobol_engine = sobol_sphere(k, d, device, self.sobol_engine)
return vecs.view(k, d)
def _draw_dirs(self, k: int, d: int, device: torch.device) -> torch.Tensor:
"""Draw projection directions using the specified sampling mode."""
if self.sampling_mode == "gaussian":
return self._gaussian_proposals(k, d, device)
if self.sampling_mode == "qmc":
return self._qmc_proposals(k, d, device)
raise ValueError("bad sampling_mode")
@staticmethod
def _duplicate_to_match(a: torch.Tensor, b: torch.Tensor, method: str):
"""
Make two tensors have the same length by duplicating the shorter one.
Args
----
a, b : (K, N₁) and (K, N₂) tensors
method : "random_replicate" or "interpolate"
Returns
-------
a, b : Tensors with matching second dimension
"""
if a.shape[1] == b.shape[1]:
return a, b
if a.shape[1] < b.shape[1]:
a, b = b, a # swap so that `a` is the larger
K, NA = a.shape
NB = b.shape[1]
# repeat / interpolate B until it matches A
if method == "random_replicate":
repeats = NA // NB
b = torch.cat([b] * repeats, dim=1)
if b.shape[1] < NA:
idx = torch.randint(0, NB, (NA - b.shape[1],), device=b.device)
b = torch.cat([b, b[:, idx]], dim=1)
else: # interpolate
b = F.interpolate(
b.unsqueeze(0), size=(NA,), mode="linear", align_corners=False
).squeeze(0)
return a, b
def reset(self):
"""Reset the reservoir sampling state."""
if self.restir_enabled:
self._reservoir_filters = torch.empty(0)
self._reservoir_weights = torch.empty(0)
self._cumulative_weights.data.fill_(0)
self._has_reservoir.fill_(False)
self._reservoir_steps = torch.empty(0, dtype=torch.long)
self._reservoir_keys = torch.empty(0)
def _wrs_multi(
self, filters: torch.Tensor, weights: torch.Tensor, step: int
) -> torch.Tensor:
"""
Weighted reservoir sampling that keeps exactly self.reservoir_size samples and
returns their indices inside the concatenated candidate set.
Args
----
filters : (K+M, D) tensor of candidate directions
weights : (K+M,) tensor of importance weights
step : Current training step
Returns
-------
keep_idx : Indices of kept samples
keep_w : Normalized weights of kept samples
"""
R = self.reservoir_size
device = weights.device
u = torch.rand_like(weights)
keys = u.pow(1.0 / weights.clamp_min(1e-9))
if not self._has_reservoir.item():
self._reservoir_filters = filters[:R]
self._reservoir_weights = weights[:R]
self._reservoir_keys = keys[:R]
self._reservoir_steps = torch.full(
(R,), step, dtype=torch.long, device=device
)
self._has_reservoir.fill_(True)
new_filters = filters[R:]
new_keys = keys[R:]
new_weights = weights[R:]
new_steps = torch.full(
(new_filters.size(0),), step, dtype=torch.long, device=device
)
all_filters = torch.cat([self._reservoir_filters, new_filters], 0)
all_keys = torch.cat([self._reservoir_keys, new_keys], 0)
all_weights = torch.cat([self._reservoir_weights, new_weights], 0)
all_steps = torch.cat([self._reservoir_steps, new_steps], 0)
topk_keys, topk_idx = torch.topk(all_keys, R, largest=True)
self._reservoir_filters = all_filters[topk_idx]
self._reservoir_weights = all_weights[topk_idx]
self._reservoir_keys = topk_keys
self._reservoir_steps = all_steps[topk_idx]
# indices w.r.t. current cand_dirs (old R first, then new M)
keep_idx = torch.cat(
[
torch.arange(R, device=device),
torch.arange(R, R + new_filters.size(0), device=device),
]
)[topk_idx]
keep_w = self._reservoir_weights / self._reservoir_weights.sum().clamp_min(
1e-12
)
return keep_idx, keep_w
def _apply_time_decay(self, step: int):
"""
Apply exponential time decay to stored reservoir weights.
Args
----
step : Current training step
"""
if self.time_decay_tau is None or not self._has_reservoir.item():
return
age = (step - self._reservoir_steps).to(torch.float32)
decay = torch.exp(-age / self.time_decay_tau).to(self._reservoir_weights.dtype)
self._reservoir_weights.mul_(decay)
self._reservoir_keys.mul_(decay) # preserve ordering consistency
def forward(
self,
pred: Float[torch.Tensor, "B D N"],
gt: Float[torch.Tensor, "B D N"],
step: int,
):
"""
Compute the sliced Wasserstein distance between predicted and ground truth
sequences.
Args
----
pred : (B, D, N) tensor of predicted sequences
gt : (B, D, N) tensor of ground truth sequences
step : Current training step for reservoir sampling
Returns
-------
loss : Scalar tensor containing the computed loss
"""
B, D, N = pred.shape
K = self.num_proj
M = self.num_new_candidates
R = self.reservoir_size
device = pred.device
gt = gt.detach()
self._apply_time_decay(step)
# Get candidate directions
if step % self.refresh_projections_every_n_steps == 0:
new_dirs = self._draw_dirs(
M if self.restir_enabled and self._has_reservoir.item() else K,
D,
device,
)
self._cached_dirs = new_dirs
else:
new_dirs = self._cached_dirs
if self.restir_enabled and self._has_reservoir.item():
cand_dirs = torch.cat(
[self._reservoir_filters, new_dirs], dim=0
) # [K+M, C,P,P]
else:
cand_dirs = new_dirs
# Project sequences
cand_pred = process_vector(pred, cand_dirs)
cand_gt = process_vector(gt, cand_dirs)
cand_pred, cand_gt = self._duplicate_to_match(
cand_pred, cand_gt, self.missing_value_method
)
cand_pred = cand_pred.sort(dim=1).values
cand_gt = cand_gt.sort(dim=1).values
# Select K directions (reservoir) & importance weights
if self.restir_enabled:
with torch.no_grad():
base = cand_pred - cand_gt
base = base.abs() if self.distance == "l1" else base.square()
ris_weights = base.mean(1) # (K+M)
keep_idx, keep_w = self._wrs_multi(cand_dirs, ris_weights, step)
w = keep_w
w_hat = keep_w
dirs = cand_dirs[keep_idx]
proj_pred = cand_pred[keep_idx]
proj_gt = cand_gt[keep_idx]
else:
dirs = cand_dirs
proj_pred = cand_pred
proj_gt = cand_gt
w = torch.full((dirs.shape[0],), 1.0 / K, device=device)
# Compute SWD
diff = proj_pred - proj_gt
diff = diff.abs() if self.distance == "l1" else diff.square()
per_slice = diff.mean(1) # (L,)
if self.use_ucv or self.use_lcv:
X_vecs = pred.permute(0, 2, 1).reshape(-1, D) # (B·N, D)
Y_vecs = gt.permute(0, 2, 1).reshape(-1, D) # (B·N, D)
m1 = X_vecs.mean(0) # (D,)
m2 = Y_vecs.mean(0)
diff_m = m1 - m2 # (D,)
theta = dirs # (L, D) already unit-norm
if self.use_ucv:
diff_X = X_vecs - m1
diff_Y = Y_vecs - m2
d = D
trSigX = diff_X.pow(2).mean()
trSigY = diff_Y.pow(2).mean()
G_bar = (diff_m @ diff_m) / d + (trSigX + trSigY)
delta2 = (theta @ diff_m) ** 2 # (L,)
proj_X = diff_X @ theta.t() # (B·N, L)
proj_Y = diff_Y @ theta.t()
varX = proj_X.pow(2).mean(0) # (L,)
varY = proj_Y.pow(2).mean(0)
G_hat = delta2 + varX + varY
else: # LCV
d = D
G_bar = (diff_m @ diff_m) / d
G_hat = (theta @ diff_m) ** 2
diff_hat_G_mean_G = G_hat - G_bar
hat_A = (w * per_slice).sum()
var_G = (w * diff_hat_G_mean_G.pow(2)).sum()
cov_AG = (w * (per_slice - hat_A) * diff_hat_G_mean_G).sum()
hat_alpha = cov_AG / (var_G + 1e-12)
loss = hat_A - hat_alpha * (w * diff_hat_G_mean_G).sum()
else:
loss = (w * per_slice).sum()
# Reservoir update
if self.restir_enabled and self.ess_alpha > 0:
with torch.no_grad():
ess = (w_hat.sum().square()) / (w_hat.square().sum() + 1e-12)
ess = torch.nan_to_num(ess, nan=0.0, posinf=R, neginf=0.0).item()
if ess < self.ess_alpha * R:
print(f"ESS: {ess} is less than {self.ess_alpha * R}, resetting")
self.reset()
return loss
|