Spaces:
Runtime error
Runtime error
File size: 19,231 Bytes
cfb7702 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_efficient_distloss import flatten_eff_distloss
import pytorch_lightning as pl
from pytorch_lightning.utilities.rank_zero import rank_zero_info, rank_zero_debug
import models
from models.utils import cleanup
from models.ray_utils import get_rays
import systems
from systems.base import BaseSystem
from systems.criterions import PSNR, binary_cross_entropy
import pdb
def ranking_loss(error, penalize_ratio=0.7, extra_weights=None, type="mean"):
error, indices = torch.sort(error)
# only sum relatively small errors
s_error = torch.index_select(
error, 0, index=indices[: int(penalize_ratio * indices.shape[0])]
)
if extra_weights is not None:
weights = torch.index_select(
extra_weights, 0, index=indices[: int(penalize_ratio * indices.shape[0])]
)
s_error = s_error * weights
if type == "mean":
return torch.mean(s_error)
elif type == "sum":
return torch.sum(s_error)
@systems.register("pinhole-neus-system")
class PinholeNeuSSystem(BaseSystem):
"""
Two ways to print to console:
1. self.print: correctly handle progress bar
2. rank_zero_info: use the logging module
"""
def prepare(self):
self.criterions = {"psnr": PSNR()}
self.train_num_samples = self.config.model.train_num_rays * (
self.config.model.num_samples_per_ray
+ self.config.model.get("num_samples_per_ray_bg", 0)
)
self.train_num_rays = self.config.model.train_num_rays
self.cos = torch.nn.CosineSimilarity(dim=-1, eps=1e-6)
def forward(self, batch):
return self.model(batch["rays"])
def preprocess_data(self, batch, stage):
if "index" in batch: # validation / testing
index = batch["index"]
else:
if self.config.model.batch_image_sampling:
index = torch.randint(
0,
len(self.dataset.all_images),
size=(self.train_num_rays,),
device=self.dataset.all_images.device,
)
else:
index = torch.randint(
0,
len(self.dataset.all_images),
size=(1,),
device=self.dataset.all_images.device,
)
if stage in ["train"]:
c2w = self.dataset.all_c2w[index]
x = torch.randint(
0,
self.dataset.w,
size=(self.train_num_rays,),
device=self.dataset.all_images.device,
)
y = torch.randint(
0,
self.dataset.h,
size=(self.train_num_rays,),
device=self.dataset.all_images.device,
)
if self.dataset.directions.ndim == 3: # (H, W, 3)
directions = self.dataset.directions[y, x]
# origins = self.dataset.origins[y, x]
elif self.dataset.directions.ndim == 4: # (N, H, W, 3)
directions = self.dataset.directions[index, y, x]
# origins = self.dataset.origins[index, y, x]
rays_o, rays_d = get_rays(directions, c2w)
rgb = (
self.dataset.all_images[index, y, x]
.view(-1, self.dataset.all_images.shape[-1])
.to(self.rank)
)
normal = (
self.dataset.all_normals_world[index, y, x]
.view(-1, self.dataset.all_normals_world.shape[-1])
.to(self.rank)
)
fg_mask = self.dataset.all_fg_masks[index, y, x].view(-1).to(self.rank)
rgb_mask = self.dataset.all_rgb_masks[index, y, x].view(-1).to(self.rank)
view_weights = self.dataset.view_weights[index, y, x].view(-1).to(self.rank)
else:
c2w = self.dataset.all_c2w[index][0]
if self.dataset.directions.ndim == 3: # (H, W, 3)
directions = self.dataset.directions
# origins = self.dataset.origins
elif self.dataset.directions.ndim == 4: # (N, H, W, 3)
directions = self.dataset.directions[index][0]
# origins = self.dataset.origins[index][0]
rays_o, rays_d = get_rays(directions, c2w)
rgb = (
self.dataset.all_images[index]
.view(-1, self.dataset.all_images.shape[-1])
.to(self.rank)
)
normal = (
self.dataset.all_normals_world[index]
.view(-1, self.dataset.all_images.shape[-1])
.to(self.rank)
)
fg_mask = self.dataset.all_fg_masks[index].view(-1).to(self.rank)
rgb_mask = self.dataset.all_rgb_masks[index].view(-1).to(self.rank)
view_weights = None
cosines = self.cos(rays_d, normal)
rays = torch.cat([rays_o, F.normalize(rays_d, p=2, dim=-1)], dim=-1)
if stage in ["train"]:
if self.config.model.background_color == "white":
self.model.background_color = torch.ones(
(3,), dtype=torch.float32, device=self.rank
)
elif self.config.model.background_color == "black":
self.model.background_color = torch.zeros(
(3,), dtype=torch.float32, device=self.rank
)
elif self.config.model.background_color == "random":
self.model.background_color = torch.rand(
(3,), dtype=torch.float32, device=self.rank
)
else:
raise NotImplementedError
else:
self.model.background_color = torch.ones(
(3,), dtype=torch.float32, device=self.rank
)
if self.dataset.apply_mask:
rgb = rgb * fg_mask[..., None] + self.model.background_color * (
1 - fg_mask[..., None]
)
batch.update(
{
"rays": rays,
"rgb": rgb,
"normal": normal,
"fg_mask": fg_mask,
"rgb_mask": rgb_mask,
"cosines": cosines,
"view_weights": view_weights,
}
)
def training_step(self, batch, batch_idx):
out = self(batch)
cosines = batch["cosines"]
fg_mask = batch["fg_mask"]
rgb_mask = batch["rgb_mask"]
view_weights = batch["view_weights"]
cosines[cosines > -0.1] = 0
mask = (fg_mask > 0) & (cosines < -0.1)
rgb_mask = out["rays_valid_full"][..., 0] & (rgb_mask > 0)
grad_cosines = self.cos(batch["rays"][..., 3:], out["comp_normal"]).detach()
# grad_cosines = cosines
loss = 0.0
# update train_num_rays
if self.config.model.dynamic_ray_sampling:
train_num_rays = int(
self.train_num_rays
* (self.train_num_samples / out["num_samples_full"].sum().item())
)
self.train_num_rays = min(
int(self.train_num_rays * 0.9 + train_num_rays * 0.1),
self.config.model.max_train_num_rays,
)
erros_rgb_mse = F.mse_loss(
out["comp_rgb_full"][rgb_mask], batch["rgb"][rgb_mask], reduction="none"
)
# erros_rgb_mse = erros_rgb_mse * torch.exp(grad_cosines.abs())[:, None][rgb_mask] / torch.exp(grad_cosines.abs()[rgb_mask]).sum()
# loss_rgb_mse = ranking_loss(erros_rgb_mse.sum(dim=1), penalize_ratio=0.7, type='sum')
loss_rgb_mse = ranking_loss(
erros_rgb_mse.sum(dim=1), penalize_ratio=0.7, type="mean"
)
self.log("train/loss_rgb_mse", loss_rgb_mse, prog_bar=True, rank_zero_only=True)
loss += loss_rgb_mse * self.C(self.config.system.loss.lambda_rgb_mse)
loss_rgb_l1 = F.l1_loss(
out["comp_rgb_full"][rgb_mask], batch["rgb"][rgb_mask], reduction="none"
)
loss_rgb_l1 = ranking_loss(
loss_rgb_l1.sum(dim=1),
extra_weights=view_weights[rgb_mask],
penalize_ratio=0.8,
)
self.log("train/loss_rgb", loss_rgb_l1)
loss += loss_rgb_l1 * self.C(self.config.system.loss.lambda_rgb_l1)
normal_errors = 1 - F.cosine_similarity(
out["comp_normal"], batch["normal"], dim=1
)
# normal_errors = normal_errors * cosines.abs() / cosines.abs().sum()
normal_errors = (
normal_errors * torch.exp(cosines.abs()) / torch.exp(cosines.abs()).sum()
)
loss_normal = ranking_loss(
normal_errors[mask],
penalize_ratio=0.8,
# extra_weights=view_weights[mask],
type="sum",
)
self.log("train/loss_normal", loss_normal, prog_bar=True, rank_zero_only=True)
loss += loss_normal * self.C(self.config.system.loss.lambda_normal)
loss_eikonal = (
(torch.linalg.norm(out["sdf_grad_samples"], ord=2, dim=-1) - 1.0) ** 2
).mean()
self.log("train/loss_eikonal", loss_eikonal, prog_bar=True, rank_zero_only=True)
loss += loss_eikonal * self.C(self.config.system.loss.lambda_eikonal)
opacity = torch.clamp(out["opacity"].squeeze(-1), 1.0e-3, 1.0 - 1.0e-3)
loss_mask = binary_cross_entropy(
opacity, batch["fg_mask"].float(), reduction="none"
)
loss_mask = ranking_loss(
loss_mask, penalize_ratio=0.9, extra_weights=view_weights
)
self.log("train/loss_mask", loss_mask, prog_bar=True, rank_zero_only=True)
loss += loss_mask * (
self.C(self.config.system.loss.lambda_mask)
if self.dataset.has_mask
else 0.0
)
loss_opaque = binary_cross_entropy(opacity, opacity)
self.log("train/loss_opaque", loss_opaque)
loss += loss_opaque * self.C(self.config.system.loss.lambda_opaque)
loss_sparsity = torch.exp(
-self.config.system.loss.sparsity_scale * out["random_sdf"].abs()
).mean()
self.log(
"train/loss_sparsity", loss_sparsity, prog_bar=True, rank_zero_only=True
)
loss += loss_sparsity * self.C(self.config.system.loss.lambda_sparsity)
if self.C(self.config.system.loss.lambda_curvature) > 0:
assert (
"sdf_laplace_samples" in out
), "Need geometry.grad_type='finite_difference' to get SDF Laplace samples"
loss_curvature = out["sdf_laplace_samples"].abs().mean()
self.log("train/loss_curvature", loss_curvature)
loss += loss_curvature * self.C(self.config.system.loss.lambda_curvature)
# distortion loss proposed in MipNeRF360
# an efficient implementation from https://github.com/sunset1995/torch_efficient_distloss
if self.C(self.config.system.loss.lambda_distortion) > 0:
loss_distortion = flatten_eff_distloss(
out["weights"], out["points"], out["intervals"], out["ray_indices"]
)
self.log("train/loss_distortion", loss_distortion)
loss += loss_distortion * self.C(self.config.system.loss.lambda_distortion)
if (
self.config.model.learned_background
and self.C(self.config.system.loss.lambda_distortion_bg) > 0
):
loss_distortion_bg = flatten_eff_distloss(
out["weights_bg"],
out["points_bg"],
out["intervals_bg"],
out["ray_indices_bg"],
)
self.log("train/loss_distortion_bg", loss_distortion_bg)
loss += loss_distortion_bg * self.C(
self.config.system.loss.lambda_distortion_bg
)
if self.C(self.config.system.loss.lambda_3d_normal_smooth) > 0:
if "random_sdf_grad" not in out:
raise ValueError(
"random_sdf_grad is required for normal smooth loss, no normal is found in the output."
)
if "normal_perturb" not in out:
raise ValueError(
"normal_perturb is required for normal smooth loss, no normal_perturb is found in the output."
)
normals_3d = out["random_sdf_grad"]
normals_perturb_3d = out["normal_perturb"]
loss_3d_normal_smooth = (normals_3d - normals_perturb_3d).abs().mean()
self.log(
"train/loss_3d_normal_smooth", loss_3d_normal_smooth, prog_bar=True
)
loss += loss_3d_normal_smooth * self.C(
self.config.system.loss.lambda_3d_normal_smooth
)
losses_model_reg = self.model.regularizations(out)
for name, value in losses_model_reg.items():
self.log(f"train/loss_{name}", value)
loss_ = value * self.C(self.config.system.loss[f"lambda_{name}"])
loss += loss_
self.log("train/inv_s", out["inv_s"], prog_bar=True)
for name, value in self.config.system.loss.items():
if name.startswith("lambda"):
self.log(f"train_params/{name}", self.C(value))
self.log("train/num_rays", float(self.train_num_rays), prog_bar=True)
return {"loss": loss}
"""
# aggregate outputs from different devices (DP)
def training_step_end(self, out):
pass
"""
"""
# aggregate outputs from different iterations
def training_epoch_end(self, out):
pass
"""
def validation_step(self, batch, batch_idx):
out = self(batch)
psnr = self.criterions["psnr"](
out["comp_rgb_full"].to(batch["rgb"]), batch["rgb"]
)
W, H = self.dataset.img_wh
self.save_image_grid(
f"it{self.global_step}-{batch['index'][0].item()}.png",
[
{
"type": "rgb",
"img": batch["rgb"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
{
"type": "rgb",
"img": out["comp_rgb_full"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
]
+ (
[
{
"type": "rgb",
"img": out["comp_rgb_bg"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
{
"type": "rgb",
"img": out["comp_rgb"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
]
if self.config.model.learned_background
else []
)
+ [
{"type": "grayscale", "img": out["depth"].view(H, W), "kwargs": {}},
{
"type": "rgb",
"img": out["comp_normal"].view(H, W, 3),
"kwargs": {"data_format": "HWC", "data_range": (-1, 1)},
},
],
)
return {"psnr": psnr, "index": batch["index"]}
"""
# aggregate outputs from different devices when using DP
def validation_step_end(self, out):
pass
"""
def validation_epoch_end(self, out):
out = self.all_gather(out)
if self.trainer.is_global_zero:
out_set = {}
for step_out in out:
# DP
if step_out["index"].ndim == 1:
out_set[step_out["index"].item()] = {"psnr": step_out["psnr"]}
# DDP
else:
for oi, index in enumerate(step_out["index"]):
out_set[index[0].item()] = {"psnr": step_out["psnr"][oi]}
psnr = torch.mean(torch.stack([o["psnr"] for o in out_set.values()]))
self.log("val/psnr", psnr, prog_bar=True, rank_zero_only=True)
self.export()
def test_step(self, batch, batch_idx):
out = self(batch)
psnr = self.criterions["psnr"](
out["comp_rgb_full"].to(batch["rgb"]), batch["rgb"]
)
W, H = self.dataset.img_wh
self.save_image_grid(
f"it{self.global_step}-test/{batch['index'][0].item()}.png",
[
{
"type": "rgb",
"img": batch["rgb"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
{
"type": "rgb",
"img": out["comp_rgb_full"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
]
+ (
[
{
"type": "rgb",
"img": out["comp_rgb_bg"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
{
"type": "rgb",
"img": out["comp_rgb"].view(H, W, 3),
"kwargs": {"data_format": "HWC"},
},
]
if self.config.model.learned_background
else []
)
+ [
{"type": "grayscale", "img": out["depth"].view(H, W), "kwargs": {}},
{
"type": "rgb",
"img": out["comp_normal"].view(H, W, 3),
"kwargs": {"data_format": "HWC", "data_range": (-1, 1)},
},
],
)
return {"psnr": psnr, "index": batch["index"]}
def test_epoch_end(self, out):
"""
Synchronize devices.
Generate image sequence using test outputs.
"""
out = self.all_gather(out)
if self.trainer.is_global_zero:
out_set = {}
for step_out in out:
# DP
if step_out["index"].ndim == 1:
out_set[step_out["index"].item()] = {"psnr": step_out["psnr"]}
# DDP
else:
for oi, index in enumerate(step_out["index"]):
out_set[index[0].item()] = {"psnr": step_out["psnr"][oi]}
psnr = torch.mean(torch.stack([o["psnr"] for o in out_set.values()]))
self.log("test/psnr", psnr, prog_bar=True, rank_zero_only=True)
self.save_img_sequence(
f"it{self.global_step}-test",
f"it{self.global_step}-test",
"(\d+)\.png",
save_format="mp4",
fps=30,
)
self.export()
def export(self):
mesh = self.model.export(self.config.export)
self.save_mesh(
f"it{self.global_step}-{self.config.model.geometry.isosurface.method}{self.config.model.geometry.isosurface.resolution}.obj",
ortho_scale=self.config.export.ortho_scale,
**mesh,
)
|