-
Notifications
You must be signed in to change notification settings - Fork 209
/
exp_runner.py
399 lines (322 loc) · 17.5 KB
/
exp_runner.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
import os
import time
import logging
import argparse
import numpy as np
import cv2 as cv
import trimesh
import torch
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from shutil import copyfile
from icecream import ic
from tqdm import tqdm
from pyhocon import ConfigFactory
from models.dataset import Dataset
from models.fields import RenderingNetwork, SDFNetwork, SingleVarianceNetwork, NeRF
from models.renderer import NeuSRenderer
class Runner:
def __init__(self, conf_path, mode='train', case='CASE_NAME', is_continue=False):
self.device = torch.device('cuda')
# Configuration
self.conf_path = conf_path
f = open(self.conf_path)
conf_text = f.read()
conf_text = conf_text.replace('CASE_NAME', case)
f.close()
self.conf = ConfigFactory.parse_string(conf_text)
self.conf['dataset.data_dir'] = self.conf['dataset.data_dir'].replace('CASE_NAME', case)
self.base_exp_dir = self.conf['general.base_exp_dir']
os.makedirs(self.base_exp_dir, exist_ok=True)
self.dataset = Dataset(self.conf['dataset'])
self.iter_step = 0
# Training parameters
self.end_iter = self.conf.get_int('train.end_iter')
self.save_freq = self.conf.get_int('train.save_freq')
self.report_freq = self.conf.get_int('train.report_freq')
self.val_freq = self.conf.get_int('train.val_freq')
self.val_mesh_freq = self.conf.get_int('train.val_mesh_freq')
self.batch_size = self.conf.get_int('train.batch_size')
self.validate_resolution_level = self.conf.get_int('train.validate_resolution_level')
self.learning_rate = self.conf.get_float('train.learning_rate')
self.learning_rate_alpha = self.conf.get_float('train.learning_rate_alpha')
self.use_white_bkgd = self.conf.get_bool('train.use_white_bkgd')
self.warm_up_end = self.conf.get_float('train.warm_up_end', default=0.0)
self.anneal_end = self.conf.get_float('train.anneal_end', default=0.0)
# Weights
self.igr_weight = self.conf.get_float('train.igr_weight')
self.mask_weight = self.conf.get_float('train.mask_weight')
self.is_continue = is_continue
self.mode = mode
self.model_list = []
self.writer = None
# Networks
params_to_train = []
self.nerf_outside = NeRF(**self.conf['model.nerf']).to(self.device)
self.sdf_network = SDFNetwork(**self.conf['model.sdf_network']).to(self.device)
self.deviation_network = SingleVarianceNetwork(**self.conf['model.variance_network']).to(self.device)
self.color_network = RenderingNetwork(**self.conf['model.rendering_network']).to(self.device)
params_to_train += list(self.nerf_outside.parameters())
params_to_train += list(self.sdf_network.parameters())
params_to_train += list(self.deviation_network.parameters())
params_to_train += list(self.color_network.parameters())
self.optimizer = torch.optim.Adam(params_to_train, lr=self.learning_rate)
self.renderer = NeuSRenderer(self.nerf_outside,
self.sdf_network,
self.deviation_network,
self.color_network,
**self.conf['model.neus_renderer'])
# Load checkpoint
latest_model_name = None
if is_continue:
model_list_raw = os.listdir(os.path.join(self.base_exp_dir, 'checkpoints'))
model_list = []
for model_name in model_list_raw:
if model_name[-3:] == 'pth' and int(model_name[5:-4]) <= self.end_iter:
model_list.append(model_name)
model_list.sort()
latest_model_name = model_list[-1]
if latest_model_name is not None:
logging.info('Find checkpoint: {}'.format(latest_model_name))
self.load_checkpoint(latest_model_name)
# Backup codes and configs for debug
if self.mode[:5] == 'train':
self.file_backup()
def train(self):
self.writer = SummaryWriter(log_dir=os.path.join(self.base_exp_dir, 'logs'))
self.update_learning_rate()
res_step = self.end_iter - self.iter_step
image_perm = self.get_image_perm()
for iter_i in tqdm(range(res_step)):
data = self.dataset.gen_random_rays_at(image_perm[self.iter_step % len(image_perm)], self.batch_size)
rays_o, rays_d, true_rgb, mask = data[:, :3], data[:, 3: 6], data[:, 6: 9], data[:, 9: 10]
near, far = self.dataset.near_far_from_sphere(rays_o, rays_d)
background_rgb = None
if self.use_white_bkgd:
background_rgb = torch.ones([1, 3])
if self.mask_weight > 0.0:
mask = (mask > 0.5).float()
else:
mask = torch.ones_like(mask)
mask_sum = mask.sum() + 1e-5
render_out = self.renderer.render(rays_o, rays_d, near, far,
background_rgb=background_rgb,
cos_anneal_ratio=self.get_cos_anneal_ratio())
color_fine = render_out['color_fine']
s_val = render_out['s_val']
cdf_fine = render_out['cdf_fine']
gradient_error = render_out['gradient_error']
weight_max = render_out['weight_max']
weight_sum = render_out['weight_sum']
# Loss
color_error = (color_fine - true_rgb) * mask
color_fine_loss = F.l1_loss(color_error, torch.zeros_like(color_error), reduction='sum') / mask_sum
psnr = 20.0 * torch.log10(1.0 / (((color_fine - true_rgb)**2 * mask).sum() / (mask_sum * 3.0)).sqrt())
eikonal_loss = gradient_error
mask_loss = F.binary_cross_entropy(weight_sum.clip(1e-3, 1.0 - 1e-3), mask)
loss = color_fine_loss +\
eikonal_loss * self.igr_weight +\
mask_loss * self.mask_weight
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
self.iter_step += 1
self.writer.add_scalar('Loss/loss', loss, self.iter_step)
self.writer.add_scalar('Loss/color_loss', color_fine_loss, self.iter_step)
self.writer.add_scalar('Loss/eikonal_loss', eikonal_loss, self.iter_step)
self.writer.add_scalar('Statistics/s_val', s_val.mean(), self.iter_step)
self.writer.add_scalar('Statistics/cdf', (cdf_fine[:, :1] * mask).sum() / mask_sum, self.iter_step)
self.writer.add_scalar('Statistics/weight_max', (weight_max * mask).sum() / mask_sum, self.iter_step)
self.writer.add_scalar('Statistics/psnr', psnr, self.iter_step)
if self.iter_step % self.report_freq == 0:
print(self.base_exp_dir)
print('iter:{:8>d} loss = {} lr={}'.format(self.iter_step, loss, self.optimizer.param_groups[0]['lr']))
if self.iter_step % self.save_freq == 0:
self.save_checkpoint()
if self.iter_step % self.val_freq == 0:
self.validate_image()
if self.iter_step % self.val_mesh_freq == 0:
self.validate_mesh()
self.update_learning_rate()
if self.iter_step % len(image_perm) == 0:
image_perm = self.get_image_perm()
def get_image_perm(self):
return torch.randperm(self.dataset.n_images)
def get_cos_anneal_ratio(self):
if self.anneal_end == 0.0:
return 1.0
else:
return np.min([1.0, self.iter_step / self.anneal_end])
def update_learning_rate(self):
if self.iter_step < self.warm_up_end:
learning_factor = self.iter_step / self.warm_up_end
else:
alpha = self.learning_rate_alpha
progress = (self.iter_step - self.warm_up_end) / (self.end_iter - self.warm_up_end)
learning_factor = (np.cos(np.pi * progress) + 1.0) * 0.5 * (1 - alpha) + alpha
for g in self.optimizer.param_groups:
g['lr'] = self.learning_rate * learning_factor
def file_backup(self):
dir_lis = self.conf['general.recording']
os.makedirs(os.path.join(self.base_exp_dir, 'recording'), exist_ok=True)
for dir_name in dir_lis:
cur_dir = os.path.join(self.base_exp_dir, 'recording', dir_name)
os.makedirs(cur_dir, exist_ok=True)
files = os.listdir(dir_name)
for f_name in files:
if f_name[-3:] == '.py':
copyfile(os.path.join(dir_name, f_name), os.path.join(cur_dir, f_name))
copyfile(self.conf_path, os.path.join(self.base_exp_dir, 'recording', 'config.conf'))
def load_checkpoint(self, checkpoint_name):
checkpoint = torch.load(os.path.join(self.base_exp_dir, 'checkpoints', checkpoint_name), map_location=self.device)
self.nerf_outside.load_state_dict(checkpoint['nerf'])
self.sdf_network.load_state_dict(checkpoint['sdf_network_fine'])
self.deviation_network.load_state_dict(checkpoint['variance_network_fine'])
self.color_network.load_state_dict(checkpoint['color_network_fine'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.iter_step = checkpoint['iter_step']
logging.info('End')
def save_checkpoint(self):
checkpoint = {
'nerf': self.nerf_outside.state_dict(),
'sdf_network_fine': self.sdf_network.state_dict(),
'variance_network_fine': self.deviation_network.state_dict(),
'color_network_fine': self.color_network.state_dict(),
'optimizer': self.optimizer.state_dict(),
'iter_step': self.iter_step,
}
os.makedirs(os.path.join(self.base_exp_dir, 'checkpoints'), exist_ok=True)
torch.save(checkpoint, os.path.join(self.base_exp_dir, 'checkpoints', 'ckpt_{:0>6d}.pth'.format(self.iter_step)))
def validate_image(self, idx=-1, resolution_level=-1):
if idx < 0:
idx = np.random.randint(self.dataset.n_images)
print('Validate: iter: {}, camera: {}'.format(self.iter_step, idx))
if resolution_level < 0:
resolution_level = self.validate_resolution_level
rays_o, rays_d = self.dataset.gen_rays_at(idx, resolution_level=resolution_level)
H, W, _ = rays_o.shape
rays_o = rays_o.reshape(-1, 3).split(self.batch_size)
rays_d = rays_d.reshape(-1, 3).split(self.batch_size)
out_rgb_fine = []
out_normal_fine = []
for rays_o_batch, rays_d_batch in zip(rays_o, rays_d):
near, far = self.dataset.near_far_from_sphere(rays_o_batch, rays_d_batch)
background_rgb = torch.ones([1, 3]) if self.use_white_bkgd else None
render_out = self.renderer.render(rays_o_batch,
rays_d_batch,
near,
far,
cos_anneal_ratio=self.get_cos_anneal_ratio(),
background_rgb=background_rgb)
def feasible(key): return (key in render_out) and (render_out[key] is not None)
if feasible('color_fine'):
out_rgb_fine.append(render_out['color_fine'].detach().cpu().numpy())
if feasible('gradients') and feasible('weights'):
n_samples = self.renderer.n_samples + self.renderer.n_importance
normals = render_out['gradients'] * render_out['weights'][:, :n_samples, None]
if feasible('inside_sphere'):
normals = normals * render_out['inside_sphere'][..., None]
normals = normals.sum(dim=1).detach().cpu().numpy()
out_normal_fine.append(normals)
del render_out
img_fine = None
if len(out_rgb_fine) > 0:
img_fine = (np.concatenate(out_rgb_fine, axis=0).reshape([H, W, 3, -1]) * 256).clip(0, 255)
normal_img = None
if len(out_normal_fine) > 0:
normal_img = np.concatenate(out_normal_fine, axis=0)
rot = np.linalg.inv(self.dataset.pose_all[idx, :3, :3].detach().cpu().numpy())
normal_img = (np.matmul(rot[None, :, :], normal_img[:, :, None])
.reshape([H, W, 3, -1]) * 128 + 128).clip(0, 255)
os.makedirs(os.path.join(self.base_exp_dir, 'validations_fine'), exist_ok=True)
os.makedirs(os.path.join(self.base_exp_dir, 'normals'), exist_ok=True)
for i in range(img_fine.shape[-1]):
if len(out_rgb_fine) > 0:
cv.imwrite(os.path.join(self.base_exp_dir,
'validations_fine',
'{:0>8d}_{}_{}.png'.format(self.iter_step, i, idx)),
np.concatenate([img_fine[..., i],
self.dataset.image_at(idx, resolution_level=resolution_level)]))
if len(out_normal_fine) > 0:
cv.imwrite(os.path.join(self.base_exp_dir,
'normals',
'{:0>8d}_{}_{}.png'.format(self.iter_step, i, idx)),
normal_img[..., i])
def render_novel_image(self, idx_0, idx_1, ratio, resolution_level):
"""
Interpolate view between two cameras.
"""
rays_o, rays_d = self.dataset.gen_rays_between(idx_0, idx_1, ratio, resolution_level=resolution_level)
H, W, _ = rays_o.shape
rays_o = rays_o.reshape(-1, 3).split(self.batch_size)
rays_d = rays_d.reshape(-1, 3).split(self.batch_size)
out_rgb_fine = []
for rays_o_batch, rays_d_batch in zip(rays_o, rays_d):
near, far = self.dataset.near_far_from_sphere(rays_o_batch, rays_d_batch)
background_rgb = torch.ones([1, 3]) if self.use_white_bkgd else None
render_out = self.renderer.render(rays_o_batch,
rays_d_batch,
near,
far,
cos_anneal_ratio=self.get_cos_anneal_ratio(),
background_rgb=background_rgb)
out_rgb_fine.append(render_out['color_fine'].detach().cpu().numpy())
del render_out
img_fine = (np.concatenate(out_rgb_fine, axis=0).reshape([H, W, 3]) * 256).clip(0, 255).astype(np.uint8)
return img_fine
def validate_mesh(self, world_space=False, resolution=64, threshold=0.0):
bound_min = torch.tensor(self.dataset.object_bbox_min, dtype=torch.float32)
bound_max = torch.tensor(self.dataset.object_bbox_max, dtype=torch.float32)
vertices, triangles =\
self.renderer.extract_geometry(bound_min, bound_max, resolution=resolution, threshold=threshold)
os.makedirs(os.path.join(self.base_exp_dir, 'meshes'), exist_ok=True)
if world_space:
vertices = vertices * self.dataset.scale_mats_np[0][0, 0] + self.dataset.scale_mats_np[0][:3, 3][None]
mesh = trimesh.Trimesh(vertices, triangles)
mesh.export(os.path.join(self.base_exp_dir, 'meshes', '{:0>8d}.ply'.format(self.iter_step)))
logging.info('End')
def interpolate_view(self, img_idx_0, img_idx_1):
images = []
n_frames = 60
for i in range(n_frames):
print(i)
images.append(self.render_novel_image(img_idx_0,
img_idx_1,
np.sin(((i / n_frames) - 0.5) * np.pi) * 0.5 + 0.5,
resolution_level=4))
for i in range(n_frames):
images.append(images[n_frames - i - 1])
fourcc = cv.VideoWriter_fourcc(*'mp4v')
video_dir = os.path.join(self.base_exp_dir, 'render')
os.makedirs(video_dir, exist_ok=True)
h, w, _ = images[0].shape
writer = cv.VideoWriter(os.path.join(video_dir,
'{:0>8d}_{}_{}.mp4'.format(self.iter_step, img_idx_0, img_idx_1)),
fourcc, 30, (w, h))
for image in images:
writer.write(image)
writer.release()
if __name__ == '__main__':
print('Hello Wooden')
torch.set_default_tensor_type('torch.cuda.FloatTensor')
FORMAT = "[%(filename)s:%(lineno)s - %(funcName)20s() ] %(message)s"
logging.basicConfig(level=logging.DEBUG, format=FORMAT)
parser = argparse.ArgumentParser()
parser.add_argument('--conf', type=str, default='./confs/base.conf')
parser.add_argument('--mode', type=str, default='train')
parser.add_argument('--mcube_threshold', type=float, default=0.0)
parser.add_argument('--is_continue', default=False, action="store_true")
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--case', type=str, default='')
args = parser.parse_args()
torch.cuda.set_device(args.gpu)
runner = Runner(args.conf, args.mode, args.case, args.is_continue)
if args.mode == 'train':
runner.train()
elif args.mode == 'validate_mesh':
runner.validate_mesh(world_space=True, resolution=512, threshold=args.mcube_threshold)
elif args.mode.startswith('interpolate'): # Interpolate views given two image indices
_, img_idx_0, img_idx_1 = args.mode.split('_')
img_idx_0 = int(img_idx_0)
img_idx_1 = int(img_idx_1)
runner.interpolate_view(img_idx_0, img_idx_1)