-
Notifications
You must be signed in to change notification settings - Fork 0
/
system.py
333 lines (286 loc) · 14.5 KB
/
system.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
import pytorch_lightning as pl
import torch.nn.functional as F
import torch
import numpy as np
from utils import cal_n_samples, N_to_reso, visualize_depth_numpy
from loss import vector_diffs, L1_VM, TVloss
from metrics import rgb_ssim, rgb_lpips, mse2psnr
from dataLoader.ray_utils import get_rays
from models.renderer import NeRFRenderer
from logger import logger
class TensoRF(pl.LightningModule):
def __init__(self, config):
super().__init__()
self.batch_size = config["batch_size"]
self.rank = config["device"]
self.reso_cur = config["reso_cur"]
self.reso_mask = self.reso_cur
self.n_samples = config["n_samples"]
if config["lr_decay_iters"] > 0:
self.lr_factor = config["lr_decay_target_ratio"]**(1/config["lr_decay_iters"])
else:
config["lr_decay_iters"] = config["n_iters"]
self.lr_factor = config["lr_decay_target_ratio"]**(1/config["n_iters"])
self.config = config
self.weights_dict = {
'Ortho_reg_weight': config["Ortho_weight"],
'L1_reg_weight': config["L1_weight_inital"],
'TV_weight_density': config["TV_weight_density"],
'TV_weight_app': config["TV_weight_app"]
}
self.upsamp_list = config["upsamp_list"]
self.update_AlphaMask_list = config["update_AlphaMask_list"]
self.N_voxel_list = (
torch.round(
torch.exp(
torch.linspace(
np.log(config["N_voxel_init"]),
np.log(config["N_voxel_final"]),
len(self.upsamp_list)+1
)
)
).long()
).tolist()[1:] #linear in logrithmic space
density_field_config = {
'aabb': config["aabb"],
'grid_size': config["reso_cur"],
'n_comp': config["n_lamb_sigma"],
'value_offset': config["density_shift"],
'activation': config["fea2denseAct"]
}
radiance_field_config = {
'aabb': config["aabb"],
'grid_size': config["reso_cur"],
'n_comp': config["n_lamb_sh"],
'value_offset': None,
'activation': {
'MLP': config["shadingMode"],
'pos_pe': config["pos_pe"],
'view_pe': config["view_pe"],
'fea_pe': config["fea_pe"],
'featureC': config["featureC"]
},
'dim_4d': config["data_dim_color"],
}
occupancy_grid_config = {
'aabb': config["OccupancyGrid_aabb"],
'grid_size': config["OccupancyGrid_grid_size"],
'threshold': config["alpha_mask_thre"],
}
renderer_config = {
'near_far': config["near_far"],
'white_bg': config["white_bg"],
'ndc_ray': config['ndc_ray'],
'step_ratio': config["step_ratio"],
'distance_scale': config["distance_scale"],
'n_samples': config["n_samples"],
'ray_march_weight_thres': config["rm_weight_mask_thre"],
'density_field_config': density_field_config,
'radiance_field_config': radiance_field_config,
'occupancy_grid_config': occupancy_grid_config,
'device': self.rank
}
self.model = eval(config["model_name"])(**renderer_config)
self.validation_step_outputs = []
self.test_step_outputs = []
def preprocess_data(self, batch, stage):
assert stage in ['train', 'val', 'test', 'predict']
if 'index' in batch: # validation / test / predict
index = batch['index'].cpu()
if stage == 'train':
if (
not hasattr(self, "rays")
and not hasattr(self, "rgbs")
):
self.rays = self.dataset.all_rays
self.rgbs = self.dataset.all_rgbs
if self.global_step == self.update_AlphaMask_list[1]:
self.rays, self.rgbs = self.model.filter_rays(
self.rays,
self.rgbs,
bbox_only=self.global_step == 0,
)
index = torch.randint(
0,
len(self.rays),
size=(self.batch_size,),
device=self.dataset.all_rgbs.device,
)
rays = self.rays[index].to(self.rank)
rgbs = self.rgbs[index].to(self.rank)
elif stage in ['val', 'test', 'predict']:
rays = self.dataset.all_rays[index].view(-1, 6).to(self.rank)
rgbs = self.dataset.all_rgbs[index].view(-1, 3).to(self.rank)
batch.update({
'rays': rays,
'rgbs': rgbs,
})
def forward(self, batch):
return self.model(batch["rays"])
# Refer to https://pytorch-lightning.readthedocs.io/en/1.7.2/common/lightning_module.html#hooks
# for hook order.
def on_train_epoch_start(self) -> None:
self.dataset = self.trainer.train_dataloader.dataset
def on_validation_epoch_start(self) -> None:
self.dataset = self.trainer.val_dataloaders.dataset
def on_validation_end(self) -> None:
self.dataset = self.trainer.train_dataloader.dataset
def on_test_epoch_start(self) -> None:
self.dataset = self.trainer.test_dataloaders.dataset
def on_train_batch_start(self, batch, batch_idx) -> None:
self.preprocess_data(batch, 'train')
self.update_step(self.global_step)
def on_validation_batch_start(self, batch, batch_idx) -> None:
self.preprocess_data(batch, 'val')
def on_test_batch_start(self, batch, batch_idx) -> None:
self.preprocess_data(batch, 'test')
def training_step(self, batch, batch_idx):
out = self(batch)
loss = torch.mean((out["rgb_map"] - batch["rgbs"]) ** 2)
# loss + regularization
total_loss = loss
self.log('train/psnr', mse2psnr(loss), prog_bar=True, on_step=True)
if self.weights_dict["Ortho_reg_weight"] > 0:
loss_reg = vector_diffs(self.model.DensityField.lines) + vector_diffs(self.model.RadianceField.lines)
total_loss += self.weights_dict["Ortho_reg_weight"]*loss_reg
self.log('train/reg', loss_reg, on_step=True)
# logger.add_scalar('train/reg', loss_reg.detach().item(), self.global_step)
if self.weights_dict["L1_reg_weight"] > 0:
loss_reg_L1 = L1_VM(self.model.DensityField.planes, self.model.DensityField.lines)
total_loss += self.weights_dict["L1_reg_weight"]*loss_reg_L1
self.log('train/reg_l1', loss_reg_L1, on_step=True)
# logger.add_scalar('train/reg_l1', loss_reg_L1.detach().item(), self.global_step)
if self.weights_dict["TV_weight_density"]>0:
TV_weight_density = self.lr_factor * self.weights_dict["TV_weight_density"]
loss_tv = TV_weight_density * TVloss(self.model.DensityField.planes)
total_loss = total_loss + loss_tv
self.log('train/reg_tv_density', loss_tv, on_step=True)
# logger.add_scalar('train/reg_tv_density', loss_tv.detach().item(), self.global_step)
if self.weights_dict["TV_weight_app"]>0:
TV_weight_app = self.weights_dict["TV_weight_app"] * self.lr_factor
loss_tv = TV_weight_app * TVloss(self.model.RadianceField.planes)
total_loss = total_loss + loss_tv
self.log('train/reg_tv_app', loss_tv, on_step=True)
# logger.add_scalar('train/reg_tv_app', loss_tv.detach().item(), self.global_step)
return {
'loss': total_loss
}
def update_step(self, global_step):
# Update the occupancy grid
if global_step in self.update_AlphaMask_list:
if self.reso_cur[0] * self.reso_cur[1] * self.reso_cur[2]<256**3:# update volume resolution
self.reso_mask = self.reso_cur
new_aabb = self.model.OccupancyGrid.update_alpha_volume(self.model.DensityField,
self.model.feature2density,
tuple(self.reso_mask),
self.model.step_size)
if global_step == self.update_AlphaMask_list[0]:
self.model.DensityField.shrink(new_aabb, self.model.OccupancyGrid)
self.model.RadianceField.shrink(new_aabb,self.model.OccupancyGrid)
self.model.update_stepsize()
L1_reg_weight = self.config["L1_weight_rest"]
logger.info_print(f"L1_reg_weight reset: {L1_reg_weight}")
# Upsample
if global_step in self.upsamp_list:
n_voxels = self.N_voxel_list.pop(0)
self.reso_cur = N_to_reso(n_voxels, self.model.DensityField.aabb)
self.model.DensityField.upsample(self.reso_cur)
self.model.RadianceField.upsample(self.reso_cur)
self.model.update_stepsize()
if self.config["lr_upsample_reset"]:
logger.info_print("Reset lr to initial")
lr_scale = 1 #0.1 ** (global_step / args.n_iters)
else:
lr_scale = self.config["lr_decay_target_ratio"] ** (global_step / self.config["n_iters"])
grad_vars = self.model.get_opt_params(self.config["lr_init"] * lr_scale, self.config["lr_basis"]*lr_scale)
self.trainer.optimizers = [torch.optim.Adam(grad_vars, betas=(0.9, 0.99))]
def test_step(self, batch, batch_idx):
out = self(batch)
W, H = self.dataset.img_wh
gt_rgb = batch["rgbs"].view(H, W, 3)
rgb_map = torch.clamp(out["rgb_map"], min=0.0, max=1.0).reshape(H, W, 3).cpu()
depth_map = out["depth_map"].reshape(H, W).cpu()
acc_map = torch.clamp(out["acc_map"], min=0.0, max=1.0).reshape(H, W).cpu()
depth_map, _ = visualize_depth_numpy(depth_map.numpy(), self.dataset.near_far)
# All metrics in float, not tensor
psnr = mse2psnr(torch.mean(( rgb_map.to(gt_rgb) - gt_rgb ) ** 2)).item()
ssim = rgb_ssim(rgb_map.cpu(), gt_rgb.cpu(), 1)
l_a = rgb_lpips(gt_rgb.cpu().numpy(), rgb_map.numpy(), 'alex', self.rank)
l_v = rgb_lpips(gt_rgb.cpu().numpy(), rgb_map.numpy(), 'vgg', self.rank)
logger.write_image(f'it{self.global_step}-test/rgb_{batch["index"][0].item():03d}.png', rgb_map.numpy())
logger.write_image(f'it{self.global_step}-test/depth_{batch["index"][0].item():03d}.png', depth_map)
logger.write_image(f'it{self.global_step}-test/alpha_{batch["index"][0].item():03d}.png', acc_map.numpy())
result = {
'psnr': psnr,
'ssim': ssim,
'l_alex': l_a,
'l_vgg': l_v,
'index': batch['index']
}
self.test_step_outputs.append(result)
def on_test_epoch_end(self):
out = self.test_step_outputs
if self.trainer.is_global_zero:
psnr = np.mean([o['psnr'] for o in out])
ssim = np.mean([o['ssim'] for o in out])
l_alex = np.mean([o['l_alex'] for o in out])
l_vgg = np.mean([o['l_vgg'] for o in out])
logger.info_print(f"test psnr={psnr}")
logger.info_print(f"test ssim={ssim}")
logger.info_print(f"test l_alex={l_alex}")
logger.info_print(f"test l_vgg={l_vgg}")
self.test_step_outputs.clear()
def validation_step(self, batch, batch_idx):
out = self(batch)
W, H = self.dataset.img_wh
gt_rgb = batch["rgbs"].view(H, W, 3)
rgb_map = torch.clamp(out["rgb_map"], min=0.0, max=1.0).reshape(H, W, 3).cpu()
depth_map = out["depth_map"].reshape(H, W).cpu()
acc_map = torch.clamp(out["acc_map"], min=0.0, max=1.0).reshape(H, W).cpu()
depth_map, _ = visualize_depth_numpy(depth_map.numpy(), self.dataset.near_far)
psnr = mse2psnr(torch.mean(( rgb_map.to(gt_rgb) - gt_rgb ) ** 2)).item()
ssim = rgb_ssim(rgb_map.cpu(), gt_rgb.cpu(), 1)
l_a = rgb_lpips(gt_rgb.cpu().numpy(), rgb_map.numpy(), 'alex', self.rank)
l_v = rgb_lpips(gt_rgb.cpu().numpy(), rgb_map.numpy(), 'vgg', self.rank)
logger.write_image(f'it{self.global_step}-val/rgb_{batch["index"][0].item():03d}.png', rgb_map.numpy())
logger.write_image(f'it{self.global_step}-val/depth_{batch["index"][0].item():03d}.png', depth_map)
logger.write_image(f'it{self.global_step}-val/alpha_{batch["index"][0].item():03d}.png', acc_map.numpy())
result = {
'psnr': psnr,
'ssim': ssim,
'l_alex': l_a,
'l_vgg': l_v,
'index': batch['index']
}
self.validation_step_outputs.append(result)
def on_validation_epoch_end(self):
out = self.validation_step_outputs
if self.trainer.is_global_zero:
psnr = np.mean([o['psnr'] for o in out])
ssim = np.mean([o['ssim'] for o in out])
l_alex = np.mean([o['l_alex'] for o in out])
l_vgg = np.mean([o['l_vgg'] for o in out])
logger.info_print(f"val psnr={psnr}")
logger.info_print(f"val ssim={ssim}")
logger.info_print(f"val l_alex={l_alex}")
logger.info_print(f"val l_vgg={l_vgg}")
self.validation_step_outputs.clear()
def configure_optimizers(self):
grad_vars = self.model.get_opt_params(self.config["lr_init"], self.config["lr_basis"])
optimizer = torch.optim.Adam(grad_vars, betas=(0.9, 0.99))
return optimizer
def on_save_checkpoint(self, ckpt):
ckpt['aabb'] = self.model.DensityField.aabb
ckpt['reso_cur'] = self.model.DensityField.grid_size
ckpt['n_samples'] = self.model.n_samples
ckpt['alpha_volume'] = np.packbits(self.model.OccupancyGrid.alpha_volume.reshape(-1).bool().cpu().numpy())
ckpt['OccupancyGrid_aabb'] = self.model.OccupancyGrid.aabb
ckpt['OccupancyGrid_grid_size'] = self.model.OccupancyGrid.grid_size
def on_load_checkpoint(self, ckpt):
alpha_volume = ckpt['alpha_volume']
length = torch.prod(self.model.OccupancyGrid.grid_size, dtype=torch.int)
self.model.OccupancyGrid.alpha_volume = torch.from_numpy(
np.unpackbits(alpha_volume)[:length].reshape(self.model.OccupancyGrid.grid_size.tolist()),
).to(self.rank, torch.float)
self.model.OccupancyGrid.alpha_volume = self.model.OccupancyGrid.alpha_volume.view(1, 1, *self.model.OccupancyGrid.alpha_volume.shape[-3:])
print(type(self.model.OccupancyGrid.alpha_volume.dtype))