-
-
Notifications
You must be signed in to change notification settings - Fork 482
/
train.py
180 lines (148 loc) · 6.52 KB
/
train.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
import os, sys
from opt import get_opts
import torch
from collections import defaultdict
from torch.utils.data import DataLoader
from datasets import dataset_dict
# models
from models.nerf import Embedding, NeRF
from models.rendering import render_rays
# optimizer, scheduler, visualization
from utils import *
# losses
from losses import loss_dict
# metrics
from metrics import *
# pytorch-lightning
from pytorch_lightning.callbacks import ModelCheckpoint
from pytorch_lightning import LightningModule, Trainer
from pytorch_lightning.logging import TestTubeLogger
class NeRFSystem(LightningModule):
def __init__(self, hparams):
super(NeRFSystem, self).__init__()
self.hparams = hparams
self.loss = loss_dict[hparams.loss_type]()
self.embedding_xyz = Embedding(3, 10) # 10 is the default number
self.embedding_dir = Embedding(3, 4) # 4 is the default number
self.embeddings = [self.embedding_xyz, self.embedding_dir]
self.nerf_coarse = NeRF()
self.models = [self.nerf_coarse]
if hparams.N_importance > 0:
self.nerf_fine = NeRF()
self.models += [self.nerf_fine]
def decode_batch(self, batch):
rays = batch['rays'] # (B, 8)
rgbs = batch['rgbs'] # (B, 3)
return rays, rgbs
def forward(self, rays):
"""Do batched inference on rays using chunk."""
B = rays.shape[0]
results = defaultdict(list)
for i in range(0, B, self.hparams.chunk):
rendered_ray_chunks = \
render_rays(self.models,
self.embeddings,
rays[i:i+self.hparams.chunk],
self.hparams.N_samples,
self.hparams.use_disp,
self.hparams.perturb,
self.hparams.noise_std,
self.hparams.N_importance,
self.hparams.chunk, # chunk size is effective in val mode
self.train_dataset.white_back)
for k, v in rendered_ray_chunks.items():
results[k] += [v]
for k, v in results.items():
results[k] = torch.cat(v, 0)
return results
def prepare_data(self):
dataset = dataset_dict[self.hparams.dataset_name]
kwargs = {'root_dir': self.hparams.root_dir,
'img_wh': tuple(self.hparams.img_wh)}
if self.hparams.dataset_name == 'llff':
kwargs['spheric_poses'] = self.hparams.spheric_poses
kwargs['val_num'] = self.hparams.num_gpus
self.train_dataset = dataset(split='train', **kwargs)
self.val_dataset = dataset(split='val', **kwargs)
def configure_optimizers(self):
self.optimizer = get_optimizer(self.hparams, self.models)
scheduler = get_scheduler(self.hparams, self.optimizer)
return [self.optimizer], [scheduler]
def train_dataloader(self):
return DataLoader(self.train_dataset,
shuffle=True,
num_workers=4,
batch_size=self.hparams.batch_size,
pin_memory=True)
def val_dataloader(self):
return DataLoader(self.val_dataset,
shuffle=False,
num_workers=4,
batch_size=1, # validate one image (H*W rays) at a time
pin_memory=True)
def training_step(self, batch, batch_nb):
log = {'lr': get_learning_rate(self.optimizer)}
rays, rgbs = self.decode_batch(batch)
results = self(rays)
log['train/loss'] = loss = self.loss(results, rgbs)
typ = 'fine' if 'rgb_fine' in results else 'coarse'
with torch.no_grad():
psnr_ = psnr(results[f'rgb_{typ}'], rgbs)
log['train/psnr'] = psnr_
return {'loss': loss,
'progress_bar': {'train_psnr': psnr_},
'log': log
}
def validation_step(self, batch, batch_nb):
rays, rgbs = self.decode_batch(batch)
rays = rays.squeeze() # (H*W, 3)
rgbs = rgbs.squeeze() # (H*W, 3)
results = self(rays)
log = {'val_loss': self.loss(results, rgbs)}
typ = 'fine' if 'rgb_fine' in results else 'coarse'
if batch_nb == 0:
W, H = self.hparams.img_wh
img = results[f'rgb_{typ}'].view(H, W, 3).cpu()
img = img.permute(2, 0, 1) # (3, H, W)
img_gt = rgbs.view(H, W, 3).permute(2, 0, 1).cpu() # (3, H, W)
depth = visualize_depth(results[f'depth_{typ}'].view(H, W)) # (3, H, W)
stack = torch.stack([img_gt, img, depth]) # (3, 3, H, W)
self.logger.experiment.add_images('val/GT_pred_depth',
stack, self.global_step)
log['val_psnr'] = psnr(results[f'rgb_{typ}'], rgbs)
return log
def validation_epoch_end(self, outputs):
mean_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
mean_psnr = torch.stack([x['val_psnr'] for x in outputs]).mean()
return {'progress_bar': {'val_loss': mean_loss,
'val_psnr': mean_psnr},
'log': {'val/loss': mean_loss,
'val/psnr': mean_psnr}
}
if __name__ == '__main__':
hparams = get_opts()
system = NeRFSystem(hparams)
checkpoint_callback = ModelCheckpoint(filepath=os.path.join(f'ckpts/{hparams.exp_name}',
'{epoch:d}'),
monitor='val/loss',
mode='min',
save_top_k=5,)
logger = TestTubeLogger(
save_dir="logs",
name=hparams.exp_name,
debug=False,
create_git_tag=False
)
trainer = Trainer(max_epochs=hparams.num_epochs,
checkpoint_callback=checkpoint_callback,
resume_from_checkpoint=hparams.ckpt_path,
logger=logger,
early_stop_callback=None,
weights_summary=None,
progress_bar_refresh_rate=1,
gpus=hparams.num_gpus,
distributed_backend='ddp' if hparams.num_gpus>1 else None,
num_sanity_val_steps=1,
benchmark=True,
profiler=hparams.num_gpus==1)
trainer.fit(system)