forked from ashawkey/torch-ngp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
renderer.py
574 lines (447 loc) · 25.1 KB
/
renderer.py
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
import math
import trimesh
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import raymarching
from .utils import custom_meshgrid
def sample_pdf(bins, weights, n_samples, det=False):
# This implementation is from NeRF
# bins: [B, T], old_z_vals
# weights: [B, T - 1], bin weights.
# return: [B, n_samples], new_z_vals
# Get pdf
weights = weights + 1e-5 # prevent nans
pdf = weights / torch.sum(weights, -1, keepdim=True)
cdf = torch.cumsum(pdf, -1)
cdf = torch.cat([torch.zeros_like(cdf[..., :1]), cdf], -1)
# Take uniform samples
if det:
u = torch.linspace(0. + 0.5 / n_samples, 1. - 0.5 / n_samples, steps=n_samples).to(weights.device)
u = u.expand(list(cdf.shape[:-1]) + [n_samples])
else:
u = torch.rand(list(cdf.shape[:-1]) + [n_samples]).to(weights.device)
# Invert CDF
u = u.contiguous()
inds = torch.searchsorted(cdf, u, right=True)
below = torch.max(torch.zeros_like(inds - 1), inds - 1)
above = torch.min((cdf.shape[-1] - 1) * torch.ones_like(inds), inds)
inds_g = torch.stack([below, above], -1) # (B, n_samples, 2)
matched_shape = [inds_g.shape[0], inds_g.shape[1], cdf.shape[-1]]
cdf_g = torch.gather(cdf.unsqueeze(1).expand(matched_shape), 2, inds_g)
bins_g = torch.gather(bins.unsqueeze(1).expand(matched_shape), 2, inds_g)
denom = (cdf_g[..., 1] - cdf_g[..., 0])
denom = torch.where(denom < 1e-5, torch.ones_like(denom), denom)
t = (u - cdf_g[..., 0]) / denom
samples = bins_g[..., 0] + t * (bins_g[..., 1] - bins_g[..., 0])
return samples
def plot_pointcloud(pc, color=None):
# pc: [N, 3]
# color: [N, 3/4]
print('[visualize points]', pc.shape, pc.dtype, pc.min(0), pc.max(0))
pc = trimesh.PointCloud(pc, color)
# axis
axes = trimesh.creation.axis(axis_length=4)
# sphere
sphere = trimesh.creation.icosphere(radius=1)
trimesh.Scene([pc, axes, sphere]).show()
class NeRFRenderer(nn.Module):
def __init__(self,
bound=1,
cuda_ray=False,
density_scale=1, # scale up deltas (or sigmas), to make the density grid more sharp. larger value than 1 usually improves performance.
min_near=0.2,
density_thresh=0.01,
bg_radius=-1,
):
super().__init__()
self.bound = bound
self.cascade = 1 + math.ceil(math.log2(bound))
self.grid_size = 128
self.density_scale = density_scale
self.min_near = min_near
self.density_thresh = density_thresh
self.bg_radius = bg_radius # radius of the background sphere.
# prepare aabb with a 6D tensor (xmin, ymin, zmin, xmax, ymax, zmax)
# NOTE: aabb (can be rectangular) is only used to generate points, we still rely on bound (always cubic) to calculate density grid and hashing.
aabb_train = torch.FloatTensor([-bound, -bound, -bound, bound, bound, bound])
aabb_infer = aabb_train.clone()
self.register_buffer('aabb_train', aabb_train)
self.register_buffer('aabb_infer', aabb_infer)
# extra state for cuda raymarching
self.cuda_ray = cuda_ray
if cuda_ray:
# density grid
density_grid = torch.zeros([self.cascade, self.grid_size ** 3]) # [CAS, H * H * H]
density_bitfield = torch.zeros(self.cascade * self.grid_size ** 3 // 8, dtype=torch.uint8) # [CAS * H * H * H // 8]
self.register_buffer('density_grid', density_grid)
self.register_buffer('density_bitfield', density_bitfield)
self.mean_density = 0
self.iter_density = 0
# step counter
step_counter = torch.zeros(16, 2, dtype=torch.int32) # 16 is hardcoded for averaging...
self.register_buffer('step_counter', step_counter)
self.mean_count = 0
self.local_step = 0
def forward(self, x, d):
raise NotImplementedError()
# separated density and color query (can accelerate non-cuda-ray mode.)
def density(self, x):
raise NotImplementedError()
def color(self, x, d, mask=None, **kwargs):
raise NotImplementedError()
def reset_extra_state(self):
if not self.cuda_ray:
return
# density grid
self.density_grid.zero_()
self.mean_density = 0
self.iter_density = 0
# step counter
self.step_counter.zero_()
self.mean_count = 0
self.local_step = 0
def run(self, rays_o, rays_d, num_steps=128, upsample_steps=128, bg_color=None, perturb=False, **kwargs):
# rays_o, rays_d: [B, N, 3], assumes B == 1
# bg_color: [3] in range [0, 1]
# return: image: [B, N, 3], depth: [B, N]
prefix = rays_o.shape[:-1]
rays_o = rays_o.contiguous().view(-1, 3)
rays_d = rays_d.contiguous().view(-1, 3)
N = rays_o.shape[0] # N = B * N, in fact
device = rays_o.device
# choose aabb
aabb = self.aabb_train if self.training else self.aabb_infer
# sample steps
nears, fars = raymarching.near_far_from_aabb(rays_o, rays_d, aabb, self.min_near)
nears.unsqueeze_(-1)
fars.unsqueeze_(-1)
#print(f'nears = {nears.min().item()} ~ {nears.max().item()}, fars = {fars.min().item()} ~ {fars.max().item()}')
z_vals = torch.linspace(0.0, 1.0, num_steps, device=device).unsqueeze(0) # [1, T]
z_vals = z_vals.expand((N, num_steps)) # [N, T]
z_vals = nears + (fars - nears) * z_vals # [N, T], in [nears, fars]
# perturb z_vals
sample_dist = (fars - nears) / num_steps
if perturb:
z_vals = z_vals + (torch.rand(z_vals.shape, device=device) - 0.5) * sample_dist
#z_vals = z_vals.clamp(nears, fars) # avoid out of bounds xyzs.
# generate xyzs
xyzs = rays_o.unsqueeze(-2) + rays_d.unsqueeze(-2) * z_vals.unsqueeze(-1) # [N, 1, 3] * [N, T, 1] -> [N, T, 3]
xyzs = torch.min(torch.max(xyzs, aabb[:3]), aabb[3:]) # a manual clip.
#plot_pointcloud(xyzs.reshape(-1, 3).detach().cpu().numpy())
# query SDF and RGB
density_outputs = self.density(xyzs.reshape(-1, 3))
#sigmas = density_outputs['sigma'].view(N, num_steps) # [N, T]
for k, v in density_outputs.items():
density_outputs[k] = v.view(N, num_steps, -1)
# upsample z_vals (nerf-like)
if upsample_steps > 0:
with torch.no_grad():
deltas = z_vals[..., 1:] - z_vals[..., :-1] # [N, T-1]
deltas = torch.cat([deltas, sample_dist * torch.ones_like(deltas[..., :1])], dim=-1)
alphas = 1 - torch.exp(-deltas * self.density_scale * density_outputs['sigma'].squeeze(-1)) # [N, T]
alphas_shifted = torch.cat([torch.ones_like(alphas[..., :1]), 1 - alphas + 1e-15], dim=-1) # [N, T+1]
weights = alphas * torch.cumprod(alphas_shifted, dim=-1)[..., :-1] # [N, T]
# sample new z_vals
z_vals_mid = (z_vals[..., :-1] + 0.5 * deltas[..., :-1]) # [N, T-1]
new_z_vals = sample_pdf(z_vals_mid, weights[:, 1:-1], upsample_steps, det=not self.training).detach() # [N, t]
new_xyzs = rays_o.unsqueeze(-2) + rays_d.unsqueeze(-2) * new_z_vals.unsqueeze(-1) # [N, 1, 3] * [N, t, 1] -> [N, t, 3]
new_xyzs = torch.min(torch.max(new_xyzs, aabb[:3]), aabb[3:]) # a manual clip.
# only forward new points to save computation
new_density_outputs = self.density(new_xyzs.reshape(-1, 3))
#new_sigmas = new_density_outputs['sigma'].view(N, upsample_steps) # [N, t]
for k, v in new_density_outputs.items():
new_density_outputs[k] = v.view(N, upsample_steps, -1)
# re-order
z_vals = torch.cat([z_vals, new_z_vals], dim=1) # [N, T+t]
z_vals, z_index = torch.sort(z_vals, dim=1)
xyzs = torch.cat([xyzs, new_xyzs], dim=1) # [N, T+t, 3]
xyzs = torch.gather(xyzs, dim=1, index=z_index.unsqueeze(-1).expand_as(xyzs))
for k in density_outputs:
tmp_output = torch.cat([density_outputs[k], new_density_outputs[k]], dim=1)
density_outputs[k] = torch.gather(tmp_output, dim=1, index=z_index.unsqueeze(-1).expand_as(tmp_output))
deltas = z_vals[..., 1:] - z_vals[..., :-1] # [N, T+t-1]
deltas = torch.cat([deltas, sample_dist * torch.ones_like(deltas[..., :1])], dim=-1)
alphas = 1 - torch.exp(-deltas * self.density_scale * density_outputs['sigma'].squeeze(-1)) # [N, T+t]
alphas_shifted = torch.cat([torch.ones_like(alphas[..., :1]), 1 - alphas + 1e-15], dim=-1) # [N, T+t+1]
weights = alphas * torch.cumprod(alphas_shifted, dim=-1)[..., :-1] # [N, T+t]
dirs = rays_d.view(-1, 1, 3).expand_as(xyzs)
for k, v in density_outputs.items():
density_outputs[k] = v.view(-1, v.shape[-1])
mask = weights > 1e-4 # hard coded
rgbs = self.color(xyzs.reshape(-1, 3), dirs.reshape(-1, 3), mask=mask.reshape(-1), **density_outputs)
rgbs = rgbs.view(N, -1, 3) # [N, T+t, 3]
#print(xyzs.shape, 'valid_rgb:', mask.sum().item())
# calculate weight_sum (mask)
weights_sum = weights.sum(dim=-1) # [N]
# calculate depth
ori_z_vals = ((z_vals - nears) / (fars - nears)).clamp(0, 1)
depth = torch.sum(weights * ori_z_vals, dim=-1)
# calculate color
image = torch.sum(weights.unsqueeze(-1) * rgbs, dim=-2) # [N, 3], in [0, 1]
# mix background color
if self.bg_radius > 0:
# use the bg model to calculate bg_color
sph = raymarching.sph_from_ray(rays_o, rays_d, self.bg_radius) # [N, 2] in [-1, 1]
bg_color = self.background(sph, rays_d.reshape(-1, 3)) # [N, 3]
elif bg_color is None:
bg_color = 1
image = image + (1 - weights_sum).unsqueeze(-1) * bg_color
image = image.view(*prefix, 3)
depth = depth.view(*prefix)
# tmp: reg loss in mip-nerf 360
# z_vals_shifted = torch.cat([z_vals[..., 1:], sample_dist * torch.ones_like(z_vals[..., :1])], dim=-1)
# mid_zs = (z_vals + z_vals_shifted) / 2 # [N, T]
# loss_dist = (torch.abs(mid_zs.unsqueeze(1) - mid_zs.unsqueeze(2)) * (weights.unsqueeze(1) * weights.unsqueeze(2))).sum() + 1/3 * ((z_vals_shifted - z_vals_shifted) * (weights ** 2)).sum()
return {
'depth': depth,
'image': image,
'weights_sum': weights_sum,
}
def run_cuda(self, rays_o, rays_d, dt_gamma=0, bg_color=None, perturb=False, force_all_rays=False, max_steps=1024, T_thresh=1e-4, **kwargs):
# rays_o, rays_d: [B, N, 3], assumes B == 1
# return: image: [B, N, 3], depth: [B, N]
prefix = rays_o.shape[:-1]
rays_o = rays_o.contiguous().view(-1, 3)
rays_d = rays_d.contiguous().view(-1, 3)
N = rays_o.shape[0] # N = B * N, in fact
device = rays_o.device
# pre-calculate near far
nears, fars = raymarching.near_far_from_aabb(rays_o, rays_d, self.aabb_train if self.training else self.aabb_infer, self.min_near)
# mix background color
if self.bg_radius > 0:
# use the bg model to calculate bg_color
sph = raymarching.sph_from_ray(rays_o, rays_d, self.bg_radius) # [N, 2] in [-1, 1]
bg_color = self.background(sph, rays_d) # [N, 3]
elif bg_color is None:
bg_color = 1
results = {}
if self.training:
# setup counter
counter = self.step_counter[self.local_step % 16]
counter.zero_() # set to 0
self.local_step += 1
xyzs, dirs, deltas, rays = raymarching.march_rays_train(rays_o, rays_d, self.bound, self.density_bitfield, self.cascade, self.grid_size, nears, fars, counter, self.mean_count, perturb, 128, force_all_rays, dt_gamma, max_steps)
#plot_pointcloud(xyzs.reshape(-1, 3).detach().cpu().numpy())
sigmas, rgbs = self(xyzs, dirs)
# density_outputs = self.density(xyzs) # [M,], use a dict since it may include extra things, like geo_feat for rgb.
# sigmas = density_outputs['sigma']
# rgbs = self.color(xyzs, dirs, **density_outputs)
sigmas = self.density_scale * sigmas
#print(f'valid RGB query ratio: {mask.sum().item() / mask.shape[0]} (total = {mask.sum().item()})')
# special case for CCNeRF's residual learning
if len(sigmas.shape) == 2:
K = sigmas.shape[0]
depths = []
images = []
for k in range(K):
weights_sum, depth, image = raymarching.composite_rays_train(sigmas[k], rgbs[k], deltas, rays, T_thresh)
image = image + (1 - weights_sum).unsqueeze(-1) * bg_color
depth = torch.clamp(depth - nears, min=0) / (fars - nears)
images.append(image.view(*prefix, 3))
depths.append(depth.view(*prefix))
depth = torch.stack(depths, axis=0) # [K, B, N]
image = torch.stack(images, axis=0) # [K, B, N, 3]
else:
weights_sum, depth, image = raymarching.composite_rays_train(sigmas, rgbs, deltas, rays, T_thresh)
image = image + (1 - weights_sum).unsqueeze(-1) * bg_color
depth = torch.clamp(depth - nears, min=0) / (fars - nears)
image = image.view(*prefix, 3)
depth = depth.view(*prefix)
results['weights_sum'] = weights_sum
else:
# allocate outputs
# if use autocast, must init as half so it won't be autocasted and lose reference.
#dtype = torch.half if torch.is_autocast_enabled() else torch.float32
# output should always be float32! only network inference uses half.
dtype = torch.float32
weights_sum = torch.zeros(N, dtype=dtype, device=device)
depth = torch.zeros(N, dtype=dtype, device=device)
image = torch.zeros(N, 3, dtype=dtype, device=device)
n_alive = N
rays_alive = torch.arange(n_alive, dtype=torch.int32, device=device) # [N]
rays_t = nears.clone() # [N]
step = 0
while step < max_steps:
# count alive rays
n_alive = rays_alive.shape[0]
# exit loop
if n_alive <= 0:
break
# decide compact_steps
n_step = max(min(N // n_alive, 8), 1)
xyzs, dirs, deltas = raymarching.march_rays(n_alive, n_step, rays_alive, rays_t, rays_o, rays_d, self.bound, self.density_bitfield, self.cascade, self.grid_size, nears, fars, 128, perturb if step == 0 else False, dt_gamma, max_steps)
sigmas, rgbs = self(xyzs, dirs)
# density_outputs = self.density(xyzs) # [M,], use a dict since it may include extra things, like geo_feat for rgb.
# sigmas = density_outputs['sigma']
# rgbs = self.color(xyzs, dirs, **density_outputs)
sigmas = self.density_scale * sigmas
raymarching.composite_rays(n_alive, n_step, rays_alive, rays_t, sigmas, rgbs, deltas, weights_sum, depth, image, T_thresh)
rays_alive = rays_alive[rays_alive >= 0]
#print(f'step = {step}, n_step = {n_step}, n_alive = {n_alive}, xyzs: {xyzs.shape}')
step += n_step
image = image + (1 - weights_sum).unsqueeze(-1) * bg_color
depth = torch.clamp(depth - nears, min=0) / (fars - nears)
image = image.view(*prefix, 3)
depth = depth.view(*prefix)
results['depth'] = depth
results['image'] = image
return results
@torch.no_grad()
def mark_untrained_grid(self, poses, intrinsic, S=64):
# poses: [B, 4, 4]
# intrinsic: [3, 3]
if not self.cuda_ray:
return
if isinstance(poses, np.ndarray):
poses = torch.from_numpy(poses)
B = poses.shape[0]
fx, fy, cx, cy = intrinsic
X = torch.arange(self.grid_size, dtype=torch.int32, device=self.density_bitfield.device).split(S)
Y = torch.arange(self.grid_size, dtype=torch.int32, device=self.density_bitfield.device).split(S)
Z = torch.arange(self.grid_size, dtype=torch.int32, device=self.density_bitfield.device).split(S)
count = torch.zeros_like(self.density_grid)
poses = poses.to(count.device)
# 5-level loop, forgive me...
for xs in X:
for ys in Y:
for zs in Z:
# construct points
xx, yy, zz = custom_meshgrid(xs, ys, zs)
coords = torch.cat([xx.reshape(-1, 1), yy.reshape(-1, 1), zz.reshape(-1, 1)], dim=-1) # [N, 3], in [0, 128)
indices = raymarching.morton3D(coords).long() # [N]
world_xyzs = (2 * coords.float() / (self.grid_size - 1) - 1).unsqueeze(0) # [1, N, 3] in [-1, 1]
# cascading
for cas in range(self.cascade):
bound = min(2 ** cas, self.bound)
half_grid_size = bound / self.grid_size
# scale to current cascade's resolution
cas_world_xyzs = world_xyzs * (bound - half_grid_size)
# split batch to avoid OOM
head = 0
while head < B:
tail = min(head + S, B)
# world2cam transform (poses is c2w, so we need to transpose it. Another transpose is needed for batched matmul, so the final form is without transpose.)
cam_xyzs = cas_world_xyzs - poses[head:tail, :3, 3].unsqueeze(1)
cam_xyzs = cam_xyzs @ poses[head:tail, :3, :3] # [S, N, 3]
# query if point is covered by any camera
mask_z = cam_xyzs[:, :, 2] > 0 # [S, N]
mask_x = torch.abs(cam_xyzs[:, :, 0]) < cx / fx * cam_xyzs[:, :, 2] + half_grid_size * 2
mask_y = torch.abs(cam_xyzs[:, :, 1]) < cy / fy * cam_xyzs[:, :, 2] + half_grid_size * 2
mask = (mask_z & mask_x & mask_y).sum(0).reshape(-1) # [N]
# update count
count[cas, indices] += mask
head += S
# mark untrained grid as -1
self.density_grid[count == 0] = -1
print(f'[mark untrained grid] {(count == 0).sum()} from {self.grid_size ** 3 * self.cascade}')
@torch.no_grad()
def update_extra_state(self, decay=0.95, S=128):
# call before each epoch to update extra states.
if not self.cuda_ray:
return
### update density grid
tmp_grid = - torch.ones_like(self.density_grid)
# full update.
if self.iter_density < 16:
#if True:
X = torch.arange(self.grid_size, dtype=torch.int32, device=self.density_bitfield.device).split(S)
Y = torch.arange(self.grid_size, dtype=torch.int32, device=self.density_bitfield.device).split(S)
Z = torch.arange(self.grid_size, dtype=torch.int32, device=self.density_bitfield.device).split(S)
for xs in X:
for ys in Y:
for zs in Z:
# construct points
xx, yy, zz = custom_meshgrid(xs, ys, zs)
coords = torch.cat([xx.reshape(-1, 1), yy.reshape(-1, 1), zz.reshape(-1, 1)], dim=-1) # [N, 3], in [0, 128)
indices = raymarching.morton3D(coords).long() # [N]
xyzs = 2 * coords.float() / (self.grid_size - 1) - 1 # [N, 3] in [-1, 1]
# cascading
for cas in range(self.cascade):
bound = min(2 ** cas, self.bound)
half_grid_size = bound / self.grid_size
# scale to current cascade's resolution
cas_xyzs = xyzs * (bound - half_grid_size)
# add noise in [-hgs, hgs]
cas_xyzs += (torch.rand_like(cas_xyzs) * 2 - 1) * half_grid_size
# query density
sigmas = self.density(cas_xyzs)['sigma'].reshape(-1).detach()
sigmas *= self.density_scale
# assign
tmp_grid[cas, indices] = sigmas
# partial update (half the computation)
# TODO: why no need of maxpool ?
else:
N = self.grid_size ** 3 // 4 # H * H * H / 4
for cas in range(self.cascade):
# random sample some positions
coords = torch.randint(0, self.grid_size, (N, 3), device=self.density_bitfield.device) # [N, 3], in [0, 128)
indices = raymarching.morton3D(coords).long() # [N]
# random sample occupied positions
occ_indices = torch.nonzero(self.density_grid[cas] > 0).squeeze(-1) # [Nz]
rand_mask = torch.randint(0, occ_indices.shape[0], [N], dtype=torch.long, device=self.density_bitfield.device)
occ_indices = occ_indices[rand_mask] # [Nz] --> [N], allow for duplication
occ_coords = raymarching.morton3D_invert(occ_indices) # [N, 3]
# concat
indices = torch.cat([indices, occ_indices], dim=0)
coords = torch.cat([coords, occ_coords], dim=0)
# same below
xyzs = 2 * coords.float() / (self.grid_size - 1) - 1 # [N, 3] in [-1, 1]
bound = min(2 ** cas, self.bound)
half_grid_size = bound / self.grid_size
# scale to current cascade's resolution
cas_xyzs = xyzs * (bound - half_grid_size)
# add noise in [-hgs, hgs]
cas_xyzs += (torch.rand_like(cas_xyzs) * 2 - 1) * half_grid_size
# query density
sigmas = self.density(cas_xyzs)['sigma'].reshape(-1).detach()
sigmas *= self.density_scale
# assign
tmp_grid[cas, indices] = sigmas
## max-pool on tmp_grid for less aggressive culling [No significant improvement...]
# invalid_mask = tmp_grid < 0
# tmp_grid = F.max_pool3d(tmp_grid.view(self.cascade, 1, self.grid_size, self.grid_size, self.grid_size), kernel_size=3, stride=1, padding=1).view(self.cascade, -1)
# tmp_grid[invalid_mask] = -1
# ema update
valid_mask = (self.density_grid >= 0) & (tmp_grid >= 0)
self.density_grid[valid_mask] = torch.maximum(self.density_grid[valid_mask] * decay, tmp_grid[valid_mask])
self.mean_density = torch.mean(self.density_grid.clamp(min=0)).item() # -1 regions are viewed as 0 density.
#self.mean_density = torch.mean(self.density_grid[self.density_grid > 0]).item() # do not count -1 regions
self.iter_density += 1
# convert to bitfield
density_thresh = min(self.mean_density, self.density_thresh)
self.density_bitfield = raymarching.packbits(self.density_grid, density_thresh, self.density_bitfield)
### update step counter
total_step = min(16, self.local_step)
if total_step > 0:
self.mean_count = int(self.step_counter[:total_step, 0].sum().item() / total_step)
self.local_step = 0
#print(f'[density grid] min={self.density_grid.min().item():.4f}, max={self.density_grid.max().item():.4f}, mean={self.mean_density:.4f}, occ_rate={(self.density_grid > 0.01).sum() / (128**3 * self.cascade):.3f} | [step counter] mean={self.mean_count}')
def render(self, rays_o, rays_d, staged=False, max_ray_batch=4096, **kwargs):
# rays_o, rays_d: [B, N, 3], assumes B == 1
# return: pred_rgb: [B, N, 3]
if self.cuda_ray:
_run = self.run_cuda
else:
_run = self.run
B, N = rays_o.shape[:2]
device = rays_o.device
# never stage when cuda_ray
if staged and not self.cuda_ray:
depth = torch.empty((B, N), device=device)
image = torch.empty((B, N, 3), device=device)
for b in range(B):
head = 0
while head < N:
tail = min(head + max_ray_batch, N)
results_ = _run(rays_o[b:b+1, head:tail], rays_d[b:b+1, head:tail], **kwargs)
depth[b:b+1, head:tail] = results_['depth']
image[b:b+1, head:tail] = results_['image']
head += max_ray_batch
results = {}
results['depth'] = depth
results['image'] = image
else:
results = _run(rays_o, rays_d, **kwargs)
return results