Spaces:
Running
on
Zero
Running
on
Zero
File size: 29,277 Bytes
1ea89dd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 |
import numpy as np
import torch
import torch.nn as nn
from .utils import FNS, masked_mean
def log_safe(x):
"""The same as torch.log(x), but clamps the input to prevent NaNs."""
x = torch.as_tensor(x)
return torch.log(torch.min(x, torch.tensor(33e37).to(x)))
def log1p_safe(x):
"""The same as torch.log1p(x), but clamps the input to prevent NaNs."""
x = torch.as_tensor(x)
return torch.log1p(torch.min(x, torch.tensor(33e37).to(x)))
def exp_safe(x):
"""The same as torch.exp(x), but clamps the input to prevent NaNs."""
x = torch.as_tensor(x)
return torch.exp(torch.min(x, torch.tensor(87.5).to(x)))
def expm1_safe(x):
"""The same as tf.math.expm1(x), but clamps the input to prevent NaNs."""
x = torch.as_tensor(x)
return torch.expm1(torch.min(x, torch.tensor(87.5).to(x)))
def inv_softplus(y):
"""The inverse of tf.nn.softplus()."""
y = torch.as_tensor(y)
return torch.where(y > 87.5, y, torch.log(torch.expm1(y)))
def logit(y):
"""The inverse of tf.nn.sigmoid()."""
y = torch.as_tensor(y)
return -torch.log(1.0 / y - 1.0)
def affine_sigmoid(logits, lo=0, hi=1):
"""Maps reals to (lo, hi), where 0 maps to (lo+hi)/2."""
if not lo < hi:
raise ValueError("`lo` (%g) must be < `hi` (%g)" % (lo, hi))
logits = torch.as_tensor(logits)
lo = torch.as_tensor(lo)
hi = torch.as_tensor(hi)
alpha = torch.sigmoid(logits) * (hi - lo) + lo
return alpha
def inv_affine_sigmoid(probs, lo=0, hi=1):
"""The inverse of affine_sigmoid(., lo, hi)."""
if not lo < hi:
raise ValueError("`lo` (%g) must be < `hi` (%g)" % (lo, hi))
probs = torch.as_tensor(probs)
lo = torch.as_tensor(lo)
hi = torch.as_tensor(hi)
logits = logit((probs - lo) / (hi - lo))
return logits
def affine_softplus(x, lo=0, ref=1):
"""Maps real numbers to (lo, infinity), where 0 maps to ref."""
if not lo < ref:
raise ValueError("`lo` (%g) must be < `ref` (%g)" % (lo, ref))
x = torch.as_tensor(x)
lo = torch.as_tensor(lo)
ref = torch.as_tensor(ref)
shift = inv_softplus(torch.tensor(1.0))
y = (ref - lo) * torch.nn.Softplus()(x + shift) + lo
return y
def inv_affine_softplus(y, lo=0, ref=1):
"""The inverse of affine_softplus(., lo, ref)."""
if not lo < ref:
raise ValueError("`lo` (%g) must be < `ref` (%g)" % (lo, ref))
y = torch.as_tensor(y)
lo = torch.as_tensor(lo)
ref = torch.as_tensor(ref)
shift = inv_softplus(torch.tensor(1.0))
x = inv_softplus((y - lo) / (ref - lo)) - shift
return x
def students_t_nll(x, df, scale):
"""The NLL of a Generalized Student's T distribution (w/o including TFP)."""
x = torch.as_tensor(x)
df = torch.as_tensor(df)
scale = torch.as_tensor(scale)
log_partition = (
torch.log(torch.abs(scale))
+ torch.lgamma(0.5 * df)
- torch.lgamma(0.5 * df + torch.tensor(0.5))
+ torch.tensor(0.5 * np.log(np.pi))
)
return (
0.5 * ((df + 1.0) * torch.log1p((x / scale) ** 2.0 / df) + torch.log(df))
+ log_partition
)
def lossfun(x, alpha, scale, approximate=False, epsilon=1e-6):
r"""Implements the general form of the loss.
This implements the rho(x, \alpha, c) function described in "A General and
Adaptive Robust Loss Function", Jonathan T. Barron,
https://arxiv.org/abs/1701.03077.
Args:
x: The residual for which the loss is being computed. x can have any shape,
and alpha and scale will be broadcasted to match x's shape if necessary.
Must be a tensor of floats.
alpha: The shape parameter of the loss (\alpha in the paper), where more
negative values produce a loss with more robust behavior (outliers "cost"
less), and more positive values produce a loss with less robust behavior
(outliers are penalized more heavily). Alpha can be any value in
[-infinity, infinity], but the gradient of the loss with respect to alpha
is 0 at -infinity, infinity, 0, and 2. Must be a tensor of floats with the
same precision as `x`. Varying alpha allows
for smooth interpolation between a number of discrete robust losses:
alpha=-Infinity: Welsch/Leclerc Loss.
alpha=-2: Geman-McClure loss.
alpha=0: Cauchy/Lortentzian loss.
alpha=1: Charbonnier/pseudo-Huber loss.
alpha=2: L2 loss.
scale: The scale parameter of the loss. When |x| < scale, the loss is an
L2-like quadratic bowl, and when |x| > scale the loss function takes on a
different shape according to alpha. Must be a tensor of single-precision
floats.
approximate: a bool, where if True, this function returns an approximate and
faster form of the loss, as described in the appendix of the paper. This
approximation holds well everywhere except as x and alpha approach zero.
epsilon: A float that determines how inaccurate the "approximate" version of
the loss will be. Larger values are less accurate but more numerically
stable. Must be great than single-precision machine epsilon.
Returns:
The losses for each element of x, in the same shape and precision as x.
"""
assert (scale > 0).all()
if approximate:
# Compute an approximate form of the loss which is faster, but innacurate
# when x and alpha are near zero.
b = torch.abs(alpha - 2) + epsilon
d = torch.where(alpha >= 0, alpha + epsilon, alpha - epsilon)
loss = (b / d) * (torch.pow((x / scale) ** 2 / b + 1.0, 0.5 * d) - 1.0)
else:
# This will be used repeatedly.
squared_scaled_x = (x / scale) ** 2
# The loss when alpha == 2.
loss_two = 0.5 * squared_scaled_x
# The loss when alpha == 0.
loss_zero = log1p_safe(0.5 * squared_scaled_x)
# The loss when alpha == -infinity.
loss_neginf = -torch.expm1(-0.5 * squared_scaled_x)
# The loss when alpha == +infinity.
loss_posinf = expm1_safe(0.5 * squared_scaled_x)
# The loss when not in one of the above special cases.
machine_epsilon = torch.tensor(torch.finfo(alpha.dtype).eps).to(x)
# Clamp |2-alpha| to be >= machine epsilon so that it's safe to divide by.
beta_safe = torch.max(machine_epsilon, torch.abs(alpha - 2.0))
# Clamp |alpha| to be >= machine epsilon so that it's safe to divide by.
alpha_safe = torch.where(
alpha >= 0, torch.ones_like(alpha), -torch.ones_like(alpha)
) * torch.max(machine_epsilon, torch.abs(alpha))
loss_otherwise = (beta_safe / alpha_safe) * (
torch.pow(squared_scaled_x / beta_safe + 1.0, 0.5 * alpha) - 1.0
)
# Select which of the cases of the loss to return.
loss = torch.where(
alpha == -float("inf"),
loss_neginf,
torch.where(
alpha == 0,
loss_zero,
torch.where(
alpha == 2,
loss_two,
torch.where(alpha == float("inf"), loss_posinf, loss_otherwise),
),
),
)
return loss
from pkg_resources import resource_stream
def interpolate1d(x, values, tangents):
r"""Perform cubic hermite spline interpolation on a 1D spline.
The x coordinates of the spline knots are at [0 : 1 : len(values)-1].
Queries outside of the range of the spline are computed using linear
extrapolation. See https://en.wikipedia.org/wiki/Cubic_Hermite_spline
for details, where "x" corresponds to `x`, "p" corresponds to `values`, and
"m" corresponds to `tangents`.
Args:
x: A tensor of any size of single or double precision floats containing the
set of values to be used for interpolation into the spline.
values: A vector of single or double precision floats containing the value
of each knot of the spline being interpolated into. Must be the same
length as `tangents` and the same type as `x`.
tangents: A vector of single or double precision floats containing the
tangent (derivative) of each knot of the spline being interpolated into.
Must be the same length as `values` and the same type as `x`.
Returns:
The result of interpolating along the spline defined by `values`, and
`tangents`, using `x` as the query values. Will be the same length and type
as `x`.
"""
assert torch.is_tensor(x)
assert torch.is_tensor(values)
assert torch.is_tensor(tangents)
float_dtype = x.dtype
assert values.dtype == float_dtype
assert tangents.dtype == float_dtype
assert len(values.shape) == 1
assert len(tangents.shape) == 1
assert values.shape[0] == tangents.shape[0]
x_lo = torch.floor(torch.clamp(x, torch.as_tensor(0), values.shape[0] - 2)).type(
torch.int64
)
x_hi = x_lo + 1
# Compute the relative distance between each `x` and the knot below it.
t = x - x_lo.type(float_dtype)
# Compute the cubic hermite expansion of `t`.
t_sq = t**2
t_cu = t * t_sq
h01 = -2.0 * t_cu + 3.0 * t_sq
h00 = 1.0 - h01
h11 = t_cu - t_sq
h10 = h11 - t_sq + t
# Linearly extrapolate above and below the extents of the spline for all
# values.
value_before = tangents[0] * t + values[0]
value_after = tangents[-1] * (t - 1.0) + values[-1]
# Cubically interpolate between the knots below and above each query point.
neighbor_values_lo = values[x_lo]
neighbor_values_hi = values[x_hi]
neighbor_tangents_lo = tangents[x_lo]
neighbor_tangents_hi = tangents[x_hi]
value_mid = (
neighbor_values_lo * h00
+ neighbor_values_hi * h01
+ neighbor_tangents_lo * h10
+ neighbor_tangents_hi * h11
)
# Return the interpolated or extrapolated values for each query point,
# depending on whether or not the query lies within the span of the spline.
return torch.where(
t < 0.0, value_before, torch.where(t > 1.0, value_after, value_mid)
)
def partition_spline_curve(alpha):
"""Applies a curve to alpha >= 0 to compress its range before interpolation.
This is a weird hand-crafted function designed to take in alpha values and
curve them to occupy a short finite range that works well when using spline
interpolation to model the partition function Z(alpha). Because Z(alpha)
is only varied in [0, 4] and is especially interesting around alpha=2, this
curve is roughly linear in [0, 4] with a slope of ~1 at alpha=0 and alpha=4
but a slope of ~10 at alpha=2. When alpha > 4 the curve becomes logarithmic.
Some (input, output) pairs for this function are:
[(0, 0), (1, ~1.2), (2, 4), (3, ~6.8), (4, 8), (8, ~8.8), (400000, ~12)]
This function is continuously differentiable.
Args:
alpha: A numpy array or tensor (float32 or float64) with values >= 0.
Returns:
An array/tensor of curved values >= 0 with the same type as `alpha`, to be
used as input x-coordinates for spline interpolation.
"""
alpha = torch.as_tensor(alpha)
x = torch.where(
alpha < 4,
(2.25 * alpha - 4.5) / (torch.abs(alpha - 2) + 0.25) + alpha + 2,
5.0 / 18.0 * log_safe(4 * alpha - 15) + 8,
)
return x
def inv_partition_spline_curve(x):
"""The inverse of partition_spline_curve()."""
x = torch.as_tensor(x)
assert (x >= 0).all()
alpha = torch.where(
x < 8,
0.5 * x
+ torch.where(
x <= 4,
1.25 - torch.sqrt(1.5625 - x + 0.25 * x**2),
-1.25 + torch.sqrt(9.5625 - 3 * x + 0.25 * x**2),
),
3.75 + 0.25 * exp_safe(x * 3.6 - 28.8),
)
return alpha
class Distribution:
# This is only a class so that we can pre-load the partition function spline.
def __init__(self):
# Load the values, tangents, and x-coordinate scaling of a spline that
# approximates the partition function. This was produced by running
# the script in fit_partition_spline.py
with resource_stream(__name__, "resources/partition_spline.npz") as spline_file:
with np.load(spline_file, allow_pickle=False) as f:
self._spline_x_scale = torch.tensor(f["x_scale"])
self._spline_values = torch.tensor(f["values"])
self._spline_tangents = torch.tensor(f["tangents"])
def log_base_partition_function(self, alpha):
r"""Approximate the distribution's log-partition function with a 1D spline.
Because the partition function (Z(\alpha) in the paper) of the distribution
is difficult to model analytically, we approximate it with a (transformed)
cubic hermite spline: Each alpha is pushed through a nonlinearity before
being used to interpolate into a spline, which allows us to use a relatively
small spline to accurately model the log partition function over the range
of all non-negative input values.
Args:
alpha: A tensor or scalar of single or double precision floats containing
the set of alphas for which we would like an approximate log partition
function. Must be non-negative, as the partition function is undefined
when alpha < 0.
Returns:
An approximation of log(Z(alpha)) accurate to within 1e-6
"""
alpha = torch.as_tensor(alpha)
assert (alpha >= 0).all()
# Transform `alpha` to the form expected by the spline.
x = partition_spline_curve(alpha)
# Interpolate into the spline.
return interpolate1d(
x * self._spline_x_scale.to(x),
self._spline_values.to(x),
self._spline_tangents.to(x),
)
def nllfun(self, x, alpha, scale):
r"""Implements the negative log-likelihood (NLL).
Specifically, we implement -log(p(x | 0, \alpha, c) of Equation 16 in the
paper as nllfun(x, alpha, shape).
Args:
x: The residual for which the NLL is being computed. x can have any shape,
and alpha and scale will be broadcasted to match x's shape if necessary.
Must be a tensor or numpy array of floats.
alpha: The shape parameter of the NLL (\alpha in the paper), where more
negative values cause outliers to "cost" more and inliers to "cost"
less. Alpha can be any non-negative value, but the gradient of the NLL
with respect to alpha has singularities at 0 and 2 so you may want to
limit usage to (0, 2) during gradient descent. Must be a tensor or numpy
array of floats. Varying alpha in that range allows for smooth
interpolation between a Cauchy distribution (alpha = 0) and a Normal
distribution (alpha = 2) similar to a Student's T distribution.
scale: The scale parameter of the loss. When |x| < scale, the NLL is like
that of a (possibly unnormalized) normal distribution, and when |x| >
scale the NLL takes on a different shape according to alpha. Must be a
tensor or numpy array of floats.
Returns:
The NLLs for each element of x, in the same shape and precision as x.
"""
# `scale` and `alpha` must have the same type as `x`.
# x = torch.as_tensor(x)
# alpha = torch.as_tensor(alpha)
# scale = torch.as_tensor(scale)
# assert (alpha >= 0).all()
# assert (scale >= 0).all()
# float_dtype = x.dtype
# assert alpha.dtype == float_dtype
# assert scale.dtype == float_dtype
loss = lossfun(x, alpha, scale, approximate=False)
log_partition = torch.log(scale) + self.log_base_partition_function(alpha)
nll = loss + log_partition
return nll
def draw_samples(self, alpha, scale):
r"""Draw samples from the robust distribution.
This function implements Algorithm 1 the paper. This code is written to
allow
for sampling from a set of different distributions, each parametrized by its
own alpha and scale values, as opposed to the more standard approach of
drawing N samples from the same distribution. This is done by repeatedly
performing N instances of rejection sampling for each of the N distributions
until at least one proposal for each of the N distributions has been
accepted.
All samples are drawn with a zero mean, to use a non-zero mean just add each
mean to each sample.
Args:
alpha: A tensor/scalar or numpy array/scalar of floats where each element
is the shape parameter of that element's distribution.
scale: A tensor/scalar or numpy array/scalar of floats where each element
is the scale parameter of that element's distribution. Must be the same
shape as `alpha`.
Returns:
A tensor with the same shape and precision as `alpha` and `scale` where
each element is a sample drawn from the distribution specified for that
element by `alpha` and `scale`.
"""
alpha = torch.as_tensor(alpha)
scale = torch.as_tensor(scale)
assert (alpha >= 0).all()
assert (scale >= 0).all()
float_dtype = alpha.dtype
assert scale.dtype == float_dtype
cauchy = torch.distributions.cauchy.Cauchy(0.0, np.sqrt(2.0))
uniform = torch.distributions.uniform.Uniform(0, 1)
samples = torch.zeros_like(alpha)
accepted = torch.zeros(alpha.shape).type(torch.bool)
while not accepted.type(torch.uint8).all():
# Draw N samples from a Cauchy, our proposal distribution.
cauchy_sample = torch.reshape(
cauchy.sample((np.prod(alpha.shape),)), alpha.shape
)
cauchy_sample = cauchy_sample.type(alpha.dtype)
# Compute the likelihood of each sample under its target distribution.
nll = self.nllfun(
cauchy_sample,
torch.as_tensor(alpha).to(cauchy_sample),
torch.tensor(1).to(cauchy_sample),
)
# Bound the NLL. We don't use the approximate loss as it may cause
# unpredictable behavior in the context of sampling.
nll_bound = lossfun(
cauchy_sample,
torch.tensor(0.0, dtype=cauchy_sample.dtype),
torch.tensor(1.0, dtype=cauchy_sample.dtype),
approximate=False,
) + self.log_base_partition_function(alpha)
# Draw N samples from a uniform distribution, and use each uniform sample
# to decide whether or not to accept each proposal sample.
uniform_sample = torch.reshape(
uniform.sample((np.prod(alpha.shape),)), alpha.shape
)
uniform_sample = uniform_sample.type(alpha.dtype)
accept = uniform_sample <= torch.exp(nll_bound - nll)
# If a sample is accepted, replace its element in `samples` with the
# proposal sample, and set its bit in `accepted` to True.
samples = torch.where(accept, cauchy_sample, samples)
accepted = accepted | accept
# Because our distribution is a location-scale family, we sample from
# p(x | 0, \alpha, 1) and then scale each sample by `scale`.
samples *= scale
return samples
class AdaptiveLossFunction(nn.Module):
"""The adaptive loss function on a matrix.
This class behaves differently from lossfun() and
distribution.nllfun(), which are "stateless", allow the caller to specify the
shape and scale of the loss, and allow for arbitrary sized inputs. This
class only allows for rank-2 inputs for the residual `x`, and expects that
`x` is of the form [batch_index, dimension_index]. This class then
constructs free parameters (torch Parameters) that define the alpha and scale
parameters for each dimension of `x`, such that all alphas are in
(`alpha_lo`, `alpha_hi`) and all scales are in (`scale_lo`, Infinity).
The assumption is that `x` is, say, a matrix where x[i,j] corresponds to a
pixel at location j for image i, with the idea being that all pixels at
location j should be modeled with the same shape and scale parameters across
all images in the batch. If the user wants to fix alpha or scale to be a
constant,
this can be done by setting alpha_lo=alpha_hi or scale_lo=scale_init
respectively.
"""
def __init__(
self,
num_dims,
alpha_lo=0.001,
alpha_hi=1.999,
alpha_init=None,
scale_lo=1e-5,
scale_init=1.0,
):
"""Sets up the loss function.
Args:
num_dims: The number of dimensions of the input to come.
float_dtype: The floating point precision of the inputs to come.
device: The device to run on (cpu, cuda, etc).
alpha_lo: The lowest possible value for loss's alpha parameters, must be
>= 0 and a scalar. Should probably be in (0, 2).
alpha_hi: The highest possible value for loss's alpha parameters, must be
>= alpha_lo and a scalar. Should probably be in (0, 2).
alpha_init: The value that the loss's alpha parameters will be initialized
to, must be in (`alpha_lo`, `alpha_hi`), unless `alpha_lo` == `alpha_hi`
in which case this will be ignored. Defaults to (`alpha_lo` +
`alpha_hi`) / 2
scale_lo: The lowest possible value for the loss's scale parameters. Must
be > 0 and a scalar. This value may have more of an effect than you
think, as the loss is unbounded as scale approaches zero (say, at a
delta function).
scale_init: The initial value used for the loss's scale parameters. This
also defines the zero-point of the latent representation of scales, so
SGD may cause optimization to gravitate towards producing scales near
this value.
"""
super(AdaptiveLossFunction, self).__init__()
if not np.isscalar(alpha_lo):
raise ValueError(
"`alpha_lo` must be a scalar, but is of type {}".format(type(alpha_lo))
)
if not np.isscalar(alpha_hi):
raise ValueError(
"`alpha_hi` must be a scalar, but is of type {}".format(type(alpha_hi))
)
if alpha_init is not None and not np.isscalar(alpha_init):
raise ValueError(
"`alpha_init` must be None or a scalar, but is of type {}".format(
type(alpha_init)
)
)
if not alpha_lo >= 0:
raise ValueError("`alpha_lo` must be >= 0, but is {}".format(alpha_lo))
if not alpha_hi >= alpha_lo:
raise ValueError(
"`alpha_hi` = {} must be >= `alpha_lo` = {}".format(alpha_hi, alpha_lo)
)
if alpha_init is not None and alpha_lo != alpha_hi:
if not (alpha_init > alpha_lo and alpha_init < alpha_hi):
raise ValueError(
"`alpha_init` = {} must be in (`alpha_lo`, `alpha_hi`) = ({} {})".format(
alpha_init, alpha_lo, alpha_hi
)
)
if not np.isscalar(scale_lo):
raise ValueError(
"`scale_lo` must be a scalar, but is of type {}".format(type(scale_lo))
)
if not np.isscalar(scale_init):
raise ValueError(
"`scale_init` must be a scalar, but is of type {}".format(
type(scale_init)
)
)
if not scale_lo > 0:
raise ValueError("`scale_lo` must be > 0, but is {}".format(scale_lo))
if not scale_init >= scale_lo:
raise ValueError(
"`scale_init` = {} must be >= `scale_lo` = {}".format(
scale_init, scale_lo
)
)
self.num_dims = num_dims
self.distribution = Distribution()
if alpha_lo == alpha_hi:
# If the range of alphas is a single item, then we just fix `alpha` to be a constant.
fixed_alpha = torch.tensor([[alpha_lo]]).repeat(1, self.num_dims)
self.register_parameter(
"fixed_alpha", torch.nn.Parameter(fixed_alpha, requires_grad=False)
)
self.alpha = lambda: self.fixed_alpha
else:
# Otherwise we construct a "latent" alpha variable and define `alpha`
# As an affine function of a sigmoid on that latent variable, initialized
# such that `alpha` starts off as `alpha_init`.
if alpha_init is None:
alpha_init = (alpha_lo + alpha_hi) / 2.0
latent_alpha_init = inv_affine_sigmoid(alpha_init, lo=alpha_lo, hi=alpha_hi)
self.register_parameter(
"latent_alpha",
torch.nn.Parameter(
latent_alpha_init.clone()
.detach()
.view(1, 1)
.repeat(1, self.num_dims),
requires_grad=True,
),
)
self.alpha = lambda: affine_sigmoid(
self.latent_alpha, lo=alpha_lo, hi=alpha_hi
)
if scale_lo == scale_init:
# If the difference between the minimum and initial scale is zero, then
# we just fix `scale` to be a constant.
fixed_scale = torch.tensor([[scale_init]]).repeat(1, self.num_dims)
self.register_parameter(
"fixed_scale", torch.nn.Parameter(fixed_scale, requires_grad=False)
)
self.scale = lambda: self.fixed_scale
else:
# Otherwise we construct a "latent" scale variable and define `scale`
# As an affine function of a softplus on that latent variable.
self.register_parameter(
"latent_scale",
torch.nn.Parameter(torch.zeros((1, self.num_dims)), requires_grad=True),
)
self.scale = lambda: affine_softplus(
self.latent_scale, lo=scale_lo, ref=scale_init
)
def lossfun(self, x, **kwargs):
"""Computes the loss on a matrix.
Args:
x: The residual for which the loss is being computed. Must be a rank-2
tensor, where the innermost dimension is the batch index, and the
outermost dimension must be equal to self.num_dims. Must be a tensor or
numpy array of type self.float_dtype.
**kwargs: Arguments to be passed to the underlying distribution.nllfun().
Returns:
A tensor of the same type and shape as input `x`, containing the loss at
each element of `x`. These "losses" are actually negative log-likelihoods
(as produced by distribution.nllfun()) and so they are not actually
bounded from below by zero. You'll probably want to minimize their sum or
mean.
"""
assert len(x.shape) == 2
return self.distribution.nllfun(x, self.alpha(), self.scale(), **kwargs)
class RobustLoss(nn.Module):
def __init__(
self,
weight: float = 1.0,
adaptive: bool = False,
scale: float = 1.0,
alpha: float = 1.0,
input_fn: str = "linear",
output_fn: str = "linear",
num_dims: int = 3,
):
super().__init__()
self.name: str = self.__class__.__name__
self.output_fn = FNS[output_fn]
self.input_fn = FNS[input_fn]
self.weight: float = weight
if not adaptive:
self.loss = AdaptiveLossFunction(
num_dims=num_dims,
alpha_lo=alpha,
alpha_hi=alpha,
scale_lo=scale,
scale_init=scale,
)
else:
self.loss = AdaptiveLossFunction(
num_dims=num_dims, alpha_init=alpha, scale_init=scale
)
@torch.autocast(device_type="cuda", enabled=False, dtype=torch.float32)
def forward(
self,
input: torch.Tensor,
target: torch.Tensor,
mask: torch.Tensor | None = None,
**kwargs,
) -> torch.Tensor:
error = self.input_fn(input) - self.input_fn(target)
loss_map = self.loss.lossfun(error.reshape(-1, error.shape[-1]))
mean_error = masked_mean(
data=loss_map.view(*error.shape).mean(dim=-1), mask=mask, dim=(-1,)
).mean(dim=-1)
mean_error = self.output_fn(mean_error)
return mean_error
@classmethod
def build(cls, config):
obj = cls(
weight=config["weight"],
alpha=config["alpha"],
scale=config["scale"],
adaptive=config["adaptive"],
output_fn=config["output_fn"],
input_fn=config["input_fn"],
)
return obj
|