Spaces:
Sleeping
Sleeping
File size: 16,601 Bytes
c3d0544 | 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 | # SPDX-FileCopyrightText: Copyright (c) 2023 - 2025 NVIDIA CORPORATION & AFFILIATES.
# SPDX-FileCopyrightText: All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from typing import List, Tuple
import torch
import torch.nn.functional as F
from torch import Tensor
# TODO enum causes segmentation faults with current torch script. Go back to enum after torch script update
"""
@enum.unique
class InterpolationType(enum.Enum):
NEAREST_NEIGHBOR = (1, 1)
LINEAR = (2, 2)
SMOOTH_STEP_1 = (3, 2)
SMOOTH_STEP_2 = (4, 2)
GAUSSIAN = (6, 5)
def __init__(self, index, stride):
self.index = index
self.stride = stride
"""
@torch.jit.script
def linear_step(x: Tensor) -> Tensor:
"""
Clips the input tensor values between 0 and 1 using a linear step.
This function constrains each element in the input tensor to be in the range [0, 1].
Values below 0 are set to 0, and values above 1 are set to 1.
Parameters
----------
x: Tensor
Input tensor to be clipped.
Returns
-------
Tensor
A tensor with values clipped between 0 and 1.
Example
-------
>>> x = torch.tensor([-0.5, 0.5, 1.5])
>>> linear_step(x)
tensor([0., 0.5, 1.])
"""
return torch.clip(x, 0, 1)
@torch.jit.script
def smooth_step_1(x: Tensor) -> Tensor:
"""
Compute the smooth step interpolation of the input tensor values.
This function applies the smooth step function: \(f(x) = 3x^2 - 2x^3\)
to each element in the input tensor, and then clips the result to be in the
range [0, 1]. It's useful for creating a smooth transition between two values.
parameters
----------
x: Tensor
Input tensor, with values expected to be in the range [0, 1] for
meaningful interpolation.
Returns
-------
Tensor
A tensor with smooth step interpolated values, clipped between 0 and 1.
"""
return torch.clip(3 * x**2 - 2 * x**3, 0, 1)
@torch.jit.script
def smooth_step_2(x: Tensor) -> Tensor:
"""
Compute the enhanced smooth step interpolation of the input tensor values.
This function applies the enhanced smooth step function:
\(f(x) = x^3 (6x^2 - 15x + 10)\) to each element in the input tensor.
The result is then clipped to be in the range [0, 1].
Parameters
----------
x: Tensor
Input tensor, with values expected to be in the range [0, 1] for meaningful
interpolation.
Returns
-------
Tensor
A tensor with enhanced smooth step interpolated values, clipped between 0 and 1.
"""
return torch.clip(x**3 * (6 * x**2 - 15 * x + 10), 0, 1)
@torch.jit.script
def nearest_neighbor_weighting(dist_vec: Tensor, dx: Tensor) -> Tensor:
"""
Compute the nearest neighbor weighting for the given distance vector.
This function returns a tensor of ones with a shape derived from the input
`dist_vec`. The resulting tensor represents weights in the context of nearest
neighbor interpolation, where the closest point has a weight of one and all other
points have a weight of zero.
Parameters:
----------
dist_vec: Tensor
A tensor representing the distances from a set of points.
The last two dimensions are expected to be spatial dimensions.
dx: Tensor
A tensor representing spacing between points.
While it's provided as an input, it doesn't influence the output for this
function since nearest neighbor weights are constant.
Returns
-------
Tensor
A tensor filled with ones and shaped according to `dist_vec`
but with the last two dimensions reduced to single dimensions.
"""
return torch.ones(dist_vec.shape[:-2] + [1] + [1], device=dist_vec.device)
@torch.jit.script
def _hyper_cube_weighting(lower_point: Tensor, upper_point: Tensor) -> Tensor:
dim = lower_point.shape[-1]
weights = []
weights = [upper_point[..., 0], lower_point[..., 0]]
for i in range(1, dim):
new_weights = []
for w in weights:
new_weights.append(w * upper_point[..., i])
new_weights.append(w * lower_point[..., i])
weights = new_weights
weights = torch.stack(weights, dim=-1)
return torch.unsqueeze(weights, dim=-1)
@torch.jit.script
def linear_weighting(dist_vec: Tensor, dx: Tensor) -> Tensor:
"""
Compute the linear weighting based on the distance vector and spacing.
Parameters
----------
dist_vec: Tensor
Distance vector for interpolation points.
dx: Tensor
Spacing between points.
Returns
-------
Tensor
Weights derived from the linear interpolation of the distance vector.
"""
normalized_dist_vec = dist_vec / dx
lower_point = normalized_dist_vec[..., 0, :]
upper_point = -normalized_dist_vec[..., -1, :]
return _hyper_cube_weighting(lower_point, upper_point)
@torch.jit.script
def smooth_step_1_weighting(dist_vec: Tensor, dx: Tensor) -> Tensor:
"""
Compute the weighting using the `smooth_step_1` function on the normalized
distance vector.
Parameters
----------
dist_vec: Tensor
Distance vector for interpolation points.
dx: Tensor
Spacing between points.
Returns
-------
Tensor
Weights derived using the `smooth_step_1` interpolation of the distance vector.
"""
normalized_dist_vec = dist_vec / dx
lower_point = smooth_step_1(normalized_dist_vec[..., 0, :])
upper_point = smooth_step_1(-normalized_dist_vec[..., -1, :])
return _hyper_cube_weighting(lower_point, upper_point)
@torch.jit.script
def smooth_step_2_weighting(dist_vec: Tensor, dx: Tensor) -> Tensor:
"""
Compute the weighting using the `smooth_step_2` function on the normalized
distance vector.
Parameters
----------
dist_vec: Tensor
Distance vector for interpolation points.
dx: Tensor
pacing between points.
Returns
-------
Tensor
Weights derived using the `smooth_step_2` interpolation of the distance vector.
"""
normalized_dist_vec = dist_vec / dx
lower_point = smooth_step_2(normalized_dist_vec[..., 0, :])
upper_point = smooth_step_2(-normalized_dist_vec[..., -1, :])
return _hyper_cube_weighting(lower_point, upper_point)
@torch.jit.script
def gaussian_weighting(dist_vec: Tensor, dx: Tensor) -> Tensor:
"""
Compute the Gaussian weighting based on the distance vector and spacing.
Parameters
----------
dist_vec: Tensor
Distance vector for interpolation points.
dx: Tensor
Spacing between points.
Returns
-------
Tensor
Gaussian weights for the provided distance vector.
"""
dim = dx.size(-1)
sharpen = 2.0
sigma = dx / sharpen
factor = 1.0 / ((2.0 * math.pi) ** (dim / 2.0) * sigma.prod())
gaussian = torch.exp(-0.5 * torch.square((dist_vec / sigma)))
gaussian = factor * gaussian.prod(dim=-1)
norm = gaussian.sum(dim=2, keepdim=True)
weights = torch.unsqueeze(gaussian / norm, dim=3)
return weights
# @torch.jit.script
def _gather_nd(params: Tensor, indices: Tensor) -> Tensor:
"""As seen here https://discuss.pytorch.org/t/how-to-do-the-tf-gather-nd-in-pytorch/6445/30"""
orig_shape = list(indices.shape)
num_samples = 1
for s in orig_shape[:-1]:
num_samples *= s
m = orig_shape[-1]
n = len(params.shape)
if m <= n:
out_shape = orig_shape[:-1] + list(params.shape)[m:]
else:
raise ValueError(
f"the last dimension of indices must less or equal to the rank of params. Got indices:{indices.shape}, params:{params.shape}. {m} > {n}"
)
indices = indices.reshape((num_samples, m)).transpose(0, 1).tolist()
output = params[indices] # (num_samples, ...)
return output.reshape(out_shape).contiguous()
@torch.jit.script
def index_values_high_mem(points: Tensor, idx: Tensor) -> Tensor:
"""
Index values from the `points` tensor using the provided indices `idx`.
Parameters
----------
points: Tensor
The source tensor from which values will be indexed.
idx: Tensor
The tensor containing indices for indexing.
Returns
-------
Tensor
Indexed values from the `points` tensor.
"""
idx = idx.unsqueeze(3).repeat_interleave(points.size(-1), dim=3)
points = points.unsqueeze(1).repeat_interleave(idx.size(1), dim=1)
out = torch.gather(points, dim=2, index=idx)
return out
# @torch.jit.script
def index_values_low_mem(points: Tensor, idx: Tensor) -> Tensor:
"""
Input:
points: (b,m,c) float32 array, known points
idx: (b,n,3) int32 array, indices to known points
Output:
out: (b,m,n,c) float32 array, interpolated point values
"""
device = points.device
idxShape = idx.shape
batch_size = idxShape[0]
num_points = idxShape[1]
K = idxShape[2]
num_features = points.shape[2]
batch_indices = torch.reshape(
torch.tile(
torch.unsqueeze(torch.arange(0, batch_size).to(device), dim=0),
(num_points * K,),
),
[-1],
) # BNK
point_indices = torch.reshape(idx, [-1]) # BNK
vertices = _gather_nd(
points, torch.stack((batch_indices, point_indices), dim=1)
) # BNKxC
vertices4d = torch.reshape(
vertices, [batch_size, num_points, K, num_features]
) # BxNxKxC
return vertices4d
@torch.jit.script
def _grid_knn_idx(
query_points: Tensor,
grid: List[Tuple[float, float, int]],
stride: int,
padding: bool = True,
) -> Tensor:
# set k
k = stride // 2
# set device
device = query_points.device
# find nearest neighbors of query points from a grid
# dx vector on grid
dx = torch.tensor([(x[1] - x[0]) / (x[2] - 1) for x in grid])
dx = dx.view(1, 1, len(grid)).to(device)
# min point on grid (this will change if we are padding the grid)
start = torch.tensor([val[0] for val in grid]).to(device)
if padding:
start = start - (k * dx)
start = start.view(1, 1, len(grid))
# this is the center nearest neighbor in the grid
center_idx = (((query_points - start) / dx) + (stride / 2.0 % 1.0)).to(torch.int64)
# index window
idx_add = (
torch.arange(-((stride - 1) // 2), stride // 2 + 1).view(1, 1, -1).to(device)
)
# find all index in window around center index
# TODO make for more general diminsions
if len(grid) == 1:
idx_row_0 = center_idx[..., 0:1] + idx_add
idx = idx_row_0.view(idx_row_0.shape[0:2] + torch.Size([int(stride)]))
elif len(grid) == 2:
dim_size_1 = grid[1][2]
if padding:
dim_size_1 += 2 * k
idx_row_0 = dim_size_1 * (center_idx[..., 0:1] + idx_add)
idx_row_0 = idx_row_0.unsqueeze(-1)
idx_row_1 = center_idx[..., 1:2] + idx_add
idx_row_1 = idx_row_1.unsqueeze(2)
idx = (idx_row_0 + idx_row_1).view(
idx_row_0.shape[0:2] + torch.Size([int(stride**2)])
)
elif len(grid) == 3:
dim_size_1 = grid[1][2]
dim_size_2 = grid[2][2]
if padding:
dim_size_1 += 2 * k
dim_size_2 += 2 * k
idx_row_0 = dim_size_2 * dim_size_1 * (center_idx[..., 0:1] + idx_add)
idx_row_0 = idx_row_0.unsqueeze(-1).unsqueeze(-1)
idx_row_1 = dim_size_2 * (center_idx[..., 1:2] + idx_add)
idx_row_1 = idx_row_1.unsqueeze(2).unsqueeze(-1)
idx_row_2 = center_idx[..., 2:3] + idx_add
idx_row_2 = idx_row_2.unsqueeze(2).unsqueeze(3)
idx = (idx_row_0 + idx_row_1 + idx_row_2).view(
idx_row_0.shape[0:2] + torch.Size([int(stride**3)])
)
else:
raise RuntimeError
return idx
# TODO currently the `tolist` operation is not supported by torch script and when fixed torch script will be used
# @torch.jit.script
def interpolation(
query_points: Tensor,
context_grid: Tensor,
grid: List[Tuple[float, float, int]],
interpolation_type: str = "smooth_step_2",
mem_speed_trade: bool = True,
) -> Tensor:
"""
Interpolate values at `query_points` based on `context_grid` using specified
interpolation methods.
Parameters
----------
query_points: Tensor
Points at which interpolation is to be performed.
context_grid: Tensor
Source grid from which values are to be interpolated.
grid: List[Tuple[float, float, int]]
Describes the grid's range and resolution.
interpolation_type: str, optional
Type of interpolation to be used, by default "smooth_step_2".
mem_speed_trade: bool, optional
Trade-off between memory usage and speed.
If True, uses low memory indexing, by default True.
Returns
-------
Tensor
Interpolated values at the `query_points`.
"""
# set stride TODO this will be replaced with InterpolationType later
if interpolation_type == "nearest_neighbor":
stride = 1
elif interpolation_type == "linear":
stride = 2
elif interpolation_type == "smooth_step_1":
stride = 2
elif interpolation_type == "smooth_step_2":
stride = 2
elif interpolation_type == "gaussian":
stride = 5
else:
raise RuntimeError(f"Interpolation type {interpolation_type} not supported")
# set device
device = query_points.device
# useful values
dims = len(grid)
nr_channels = context_grid.size(0)
dx = [((x[1] - x[0]) / (x[2] - 1)) for x in grid]
# generate mesh grid of position information [grid_dim_1, grid_dim_2, ..., 2-3]
# NOTE the mesh grid is padded by stride//2
k = stride // 2
linspace = [
torch.linspace(x[0] - k * dx_i, x[1] + k * dx_i, x[2] + 2 * k)
for x, dx_i in zip(grid, dx)
]
meshgrid = torch.meshgrid(linspace)
meshgrid = torch.stack(meshgrid, dim=-1).to(device)
# pad context grid by k to avoid cuts on corners
padding = dims * (k, k)
context_grid = F.pad(context_grid, padding)
# reshape query points, context grid and mesh grid for easier indexing
# [1, grid_dim_1*grid_dim_2*..., 2-4]
nr_grid_points = int(torch.tensor([x[2] + 2 * k for x in grid]).prod())
meshgrid = meshgrid.view(1, nr_grid_points, dims)
context_grid = torch.reshape(context_grid, [1, nr_channels, nr_grid_points])
context_grid = torch.swapaxes(context_grid, 1, 2)
query_points = query_points.unsqueeze(0)
# compute index of nearest neighbor on grid to query points
idx = _grid_knn_idx(query_points, grid, stride, padding=True)
# index mesh grid to get distance vector
if mem_speed_trade:
mesh_grid_idx = index_values_low_mem(meshgrid, idx)
else:
mesh_grid_idx = index_values_high_mem(meshgrid, idx)
dist_vec = query_points.unsqueeze(2) - mesh_grid_idx
# make tf dx vec (for interpolation function)
dx = torch.tensor(dx, dtype=torch.float32)
dx = torch.reshape(dx, [1, 1, 1, dims]).to(device)
# compute bump function
if interpolation_type == "nearest_neighbor":
weights = nearest_neighbor_weighting(dist_vec, dx)
elif interpolation_type == "linear":
weights = linear_weighting(dist_vec, dx)
elif interpolation_type == "smooth_step_1":
weights = smooth_step_1_weighting(dist_vec, dx)
elif interpolation_type == "smooth_step_2":
weights = smooth_step_2_weighting(dist_vec, dx)
elif interpolation_type == "gaussian":
weights = gaussian_weighting(dist_vec, dx)
else:
raise RuntimeError
# index context grid with index
if mem_speed_trade:
context_grid_idx = index_values_low_mem(context_grid, idx)
else:
context_grid_idx = index_values_high_mem(context_grid, idx)
# interpolate points
product = weights * context_grid_idx
interpolated_points = product.sum(dim=2)
return interpolated_points[0]
|