-
Notifications
You must be signed in to change notification settings - Fork 3
/
optimize.py
266 lines (233 loc) · 14.3 KB
/
optimize.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
import configargparse
import datetime
import torch.optim.optimizer
from torch.utils.tensorboard import SummaryWriter
from utils.render import renderNovelView
from utils.utils import *
from models.mpi_generator import MPIGenerator
from models.feature_generator import FeatureGenerator
from datasets import find_scenedata_def
from models.losses import *
import lpips
import shutil
class OptimizePerScene():
def __init__(self, model, ckpt_filepath, neighbor_view_num, depth_sample_num, epochs, learning_rate, lr_ds_epoch_idx, loss_rgb_weight, loss_ssim_weight, loss_lpips_weight,
summary_scalars_freq, summary_images_freq, save_ckpt_freq, device):
self.feature_generator = model["feature_generator"].to(device)
self.mpi_generator = model["mpi_generator"].to(device)
self.load_ckpt(ckpt_filepath)
self.neighbor_view_num = neighbor_view_num
self.depth_sample_num = depth_sample_num
self.epochs = epochs
self.device = device
# optimizer setting
self.optimizer = torch.optim.Adam(self.mpi_generator.parameters(), lr=learning_rate, betas=(0.9, 0.999))
milestones = [int(epoch_idx) for epoch_idx in lr_ds_epoch_idx.split(':')[0].split(',')]
lr_gamma = 1 / float(lr_ds_epoch_idx.split(':')[1])
self.lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(self.optimizer, milestones, gamma=lr_gamma, last_epoch=-1)
self.loss_rgb_weight = loss_rgb_weight
self.loss_ssim_weight = loss_ssim_weight
self.loss_lpips_weight = loss_lpips_weight
self.summary_scalars_freq = summary_scalars_freq
self.summary_images_freq = summary_images_freq
self.save_ckpt_freq = save_ckpt_freq
# loss calculater definition
self.ssim_calculator = SSIM()
self.lpips_calculator = lpips.LPIPS(net="vgg").to(device)
self.lpips_calculator.requires_grad = True
def load_ckpt(self, ckpt_filepath):
state_dict = torch.load(ckpt_filepath)
self.feature_generator.load_state_dict(state_dict["feature_generator"])
self.mpi_generator.load_state_dict(state_dict["mpi_generator"])
def set_data(self, scene_data):
if self.device == torch.device("cuda"):
scene_data = dict2cuda(scene_data)
self.image_ref = scene_data["image_ref"]
self.depth_min_ref, self.depth_max_ref = scene_data["depth_min_ref"], scene_data["depth_max_ref"]
self.K_ref = scene_data["K_ref"]
self.E_ref = scene_data["E_ref"]
self.depth_ref = scene_data["depth_ref"]
self.images_tgt = scene_data["images_tgt"]
self.Ks_tgt, self.Ts_tgt_ref = scene_data["Ks_tgt"], scene_data["Ts_tgt_ref"] # [B, N, 3 ,3], [B, N, 4, 4]
self.height, self.width = self.image_ref.shape[2], self.image_ref.shape[3]
def optimize(self, scene_data, logdir, neighbor_view_indices):
logger = SummaryWriter(logdir)
self.set_data(scene_data)
with torch.no_grad():
conv1_out, block1_out, block2_out, block3_out, block4_out = self.feature_generator(self.image_ref)
self.mpi_generator.train()
for epoch_idx in range(self.epochs):
# network forward, generate mpi representations
summary_scalars_epoch = ScalarDictMerge()
summary_images_epoch = {}
for neighbor_image_idx in neighbor_view_indices:
mpi_outputs = self.mpi_generator(input_features=[conv1_out, block1_out, block2_out, block3_out, block4_out], depth_sample_num=self.depth_sample_num)
rgb_mpi_ref_dict = {
"scale_0": mpi_outputs["MPI_{}".format(0)][:, :, :3, :, :],
"scale_1": mpi_outputs["MPI_{}".format(1)][:, :, :3, :, :],
"scale_2": mpi_outputs["MPI_{}".format(2)][:, :, :3, :, :],
"scale_3": mpi_outputs["MPI_{}".format(3)][:, :, :3, :, :],
}
sigma_mpi_ref_dict = {
"scale_0": mpi_outputs["MPI_{}".format(0)][:, :, 3:, :, :],
"scale_1": mpi_outputs["MPI_{}".format(1)][:, :, 3:, :, :],
"scale_2": mpi_outputs["MPI_{}".format(2)][:, :, 3:, :, :],
"scale_3": mpi_outputs["MPI_{}".format(3)][:, :, 3:, :, :],
}
summary_scalars, summary_images = self.optimize_per_image(rgb_mpi_ref_dict, sigma_mpi_ref_dict, neighbor_image_idx)
summary_scalars_epoch.update(summary_scalars)
if neighbor_image_idx == 0:
summary_images_epoch = summary_images
if (epoch_idx+1) % self.summary_scalars_freq == 0:
save_scalars(logger, "Optimize", summary_scalars_epoch.mean(), epoch_idx+1) # scalars for random sampled tgt-view image
if (epoch_idx+1) % self.summary_images_freq == 0:
for scale in range(4):
save_images(logger, "Optimize_scale_{}".format(scale), summary_images_epoch["scale_{}".format(scale)], epoch_idx+1) # summary images for random sampled tgt-image
print("Optimize, Epoch:{}/{}, loss:{:.4f}".format(epoch_idx, self.epochs, summary_scalars_epoch.mean()["loss"]))
if (epoch_idx+1) % self.save_ckpt_freq == 0:
torch.save({
"epoch": epoch_idx,
"feature_generator": self.feature_generator.state_dict(),
"mpi_generator": self.mpi_generator.state_dict(),
"optimizer": self.optimizer.state_dict(),},
"{}/mpimodel_{:0>4}.ckpt".format(logdir, epoch_idx))
self.lr_scheduler.step()
def optimize_per_image(self, rgb_mpi_ref_dict, sigma_mpi_ref_dict, neighbor_image_idx):
with torch.no_grad():
T_ref_ref = torch.tensor([[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 1.0]],dtype=torch.float32, device=self.device).unsqueeze(0).repeat(1, 1, 1)
T_tgt_ref = self.Ts_tgt_ref[:, neighbor_image_idx, :, :]
summary_scalars, summary_images = {}, {}
loss_per_image, loss_rgb_per_image, loss_ssim_per_image, loss_lpips_per_image = 0.0, 0.0, 0.0, 0.0
for scale in range(4):
with torch.no_grad():
# rescale intrinsics for ref-view, tgt-views
K_ref_scaled = self.K_ref / (2 ** scale)
K_ref_scaled[:, 2, 2] = 1
K_tgt_scaled = self.Ks_tgt[:, neighbor_image_idx, :, :] / (2 ** scale)
K_tgt_scaled[:, 2, 2] = 1 # [B, 3, 3]
height_render, width_render = self.height // 2 ** scale, self.width // 2 ** scale
# rescale image_ref, depth_ref, images_tgt
image_ref = F.interpolate(self.image_ref, size=(height_render, width_render), mode="bilinear") # [B, 3, H//scale, W//scale]
depth_ref = F.interpolate(self.depth_ref.unsqueeze(1), size=(height_render, width_render), mode="nearest") # Not for loss, for monitor depth MAE
image_tgt = F.interpolate(self.images_tgt[:, neighbor_image_idx, :, :, :], size=(height_render, width_render), mode="bilinear") # [B, 3, H//scale, W//scale]
# render ref-view syn image
ref_rgb_syn, ref_depth_syn, ref_mask = renderNovelView(
rbg_MPI_ref=rgb_mpi_ref_dict["scale_{}".format(scale)],
sigma_MPI_ref=sigma_mpi_ref_dict["scale_{}".format(scale)],
depth_min_ref=self.depth_min_ref,
depth_max_ref=self.depth_max_ref,
depth_hypothesis_num=self.depth_sample_num,
T_tgt_ref=T_ref_ref,
K_ref=K_ref_scaled,
K_tgt=K_ref_scaled,
height_render=height_render,
width_render=width_render,
)
tgt_rgb_syn, tgt_depth_syn, tgt_mask = renderNovelView(
rbg_MPI_ref=rgb_mpi_ref_dict["scale_{}".format(scale)],
sigma_MPI_ref=sigma_mpi_ref_dict["scale_{}".format(scale)],
depth_min_ref=self.depth_min_ref,
depth_max_ref=self.depth_max_ref,
depth_hypothesis_num=self.depth_sample_num,
T_tgt_ref=T_tgt_ref,
K_ref=K_ref_scaled,
K_tgt=K_tgt_scaled,
height_render=height_render,
width_render=width_render,
)
loss_rgb = loss_fcn_rgb_L1(tgt_rgb_syn, tgt_mask, image_tgt) * self.loss_rgb_weight
loss_ssim = loss_fcn_rgb_SSIM(self.ssim_calculator, tgt_rgb_syn, tgt_mask, image_tgt) * self.loss_ssim_weight
loss_lpips = loss_fcn_rgb_lpips(self.lpips_calculator, tgt_rgb_syn, tgt_mask, image_tgt) * self.loss_lpips_weight
loss = loss_rgb + loss_ssim + loss_lpips
loss_rgb_per_image = loss_rgb_per_image + loss_rgb
loss_ssim_per_image = loss_ssim_per_image + loss_ssim
loss_lpips_per_image = loss_lpips_per_image + loss_lpips
loss_per_image = loss_per_image + loss
with torch.no_grad():
summary_images["scale_{}".format(scale)] = {
"ref_image": image_ref,
"ref_rgb_syn": ref_rgb_syn,
"tgt_rgb_syn": tgt_rgb_syn,
"ref_depth_syn": ref_depth_syn,
"ref_depth": depth_ref,
"ref_depth_diff": torch.abs(depth_ref - ref_depth_syn),
"tgt_mask": tgt_mask
}
self.optimizer.zero_grad()
loss_per_image.backward()
self.optimizer.step()
with torch.no_grad():
summary_scalars = {
"loss": loss_per_image.item(),
"loss_rgb": loss_rgb_per_image.item(),
"loss_ssim": loss_ssim_per_image.item(),
"loss_lpips": loss_lpips_per_image.item(),
# "depth_MAE": torch.mean(torch.abs(summary_images["scale_0"]["ref_depth_syn"] - summary_images["scale_0"]["ref_depth"]))
}
return summary_scalars, summary_images
if __name__ == '__main__':
parser = configargparse.ArgumentParser(description="Per Scene Optimization")
parser.add_argument('--config', is_config_file=True, help='config file path')
# data parameters
parser.add_argument("--dataset", type=str, default="levir_nvs", help="select dataset")
parser.add_argument("--dataset_dirpath", type=str, default=r"D:\Datasets\LEVIR_NVS", help="dataset directory path")
parser.add_argument("--scene_id", type=str, default="scene_005", help="scene id in LEVIR-NVS")
parser.add_argument("--sample_id", type=str, default="000", help="sample id in LEVIR-NVS per scene")
# optimization parameters
parser.add_argument("--epochs", type=int, default=500, help="optimization epoch number")
parser.add_argument("--learning_rate", type=float, default=0.001, help="learning rate")
parser.add_argument("--lr_ds_epoch_idx", type=str, default="50,100,200,300,400:2", help="epoch ids to downscale lr and the downscale rate")
parser.add_argument("--ckpt_filepath", type=str, default="./checkpoints/ASI_prior.ckpt",
help="load a specific checkpoint")
# log writer and random seed parameters
parser.add_argument("--logdir", type=str, default="./checkpoints/optimization/{}_{}", help="the directory to save optimize logs, tensorboard event log")
parser.add_argument("--summary_scalars_freq", type=int, default=1, help="save summary scalar frequency")
parser.add_argument("--summary_images_freq", type=int, default=10, help="save summary images frequency")
parser.add_argument("--save_ckpt_freq", type=int, default=1, help="save checkpoint frequency, 1 means per epoch")
parser.add_argument("--seed", type=int, default=23, metavar="S", help="random seed, ensure training can recurrence")
# model parameters
parser.add_argument("--depth_sample_num", type=int, default=32, help="depth sample number in decoder")
parser.add_argument("--feature_generator_model_type", type=str, default="resnet18", help="feature generator model type")
parser.add_argument("--neighbor_view_num", type=int, default=20, help="neighbor view number")
# loss weights
parser.add_argument("--loss_rgb_weight", type=float, default=2.0, help="loss rgb weight")
parser.add_argument("--loss_ssim_weight", type=float, default=1.0, help="loss depth weight")
parser.add_argument("--loss_lpips_weight", type=float, default=1.0, help="loss depth weight")
args = parser.parse_args()
print_args(args)
# fix random seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
SceneData = find_scenedata_def(args.dataset)
# model, decoder optimizer definition and load init network ckpt
feature_generator = FeatureGenerator(model_type=args.feature_generator_model_type, pretrained=True, device=device).to(device)
mpi_generator = MPIGenerator(feature_out_chs=feature_generator.encoder_channels).to(device)
model = {"feature_generator": feature_generator, "mpi_generator": mpi_generator}
scene_optimizer = OptimizePerScene(
model=model,
ckpt_filepath=args.ckpt_filepath,
neighbor_view_num=args.neighbor_view_num,
depth_sample_num=args.depth_sample_num,
epochs=args.epochs,
learning_rate=args.learning_rate,
lr_ds_epoch_idx=args.lr_ds_epoch_idx,
loss_rgb_weight=args.loss_rgb_weight,
loss_ssim_weight=args.loss_ssim_weight,
loss_lpips_weight=args.loss_lpips_weight,
summary_scalars_freq=args.summary_scalars_freq,
summary_images_freq=args.summary_images_freq,
save_ckpt_freq=args.save_ckpt_freq,
device=device
)
# Optimization
current_time = str(datetime.datetime.now().strftime('%Y.%m.%d_%H:%M:%S'))
print("Start time: ", current_time)
scene_data = SceneData(args.dataset_dirpath, args.scene_id, args.sample_id, neighbor_view_num=args.neighbor_view_num).loadSceneData()
neighbor_view_indices = list(range(0, 20, 2)) # training view indices
logdir = args.logdir.format(scene_data["scene_id"], scene_data["sample_id"])
# copy optimization config files
shutil.copy(args.config, os.path.join(logdir, "config.txt"))
scene_optimizer.optimize(scene_data, logdir, neighbor_view_indices)
current_time = str(datetime.datetime.now().strftime('%Y.%m.%d_%H:%M:%S'))
print("End time: ", current_time)