forked from threestudio-project/threestudio
-
Notifications
You must be signed in to change notification settings - Fork 1
/
run.py
249 lines (212 loc) · 8.26 KB
/
run.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
import argparse
import contextlib
import json
import logging
import os
import random
import sys
import time
import pytorch_lightning as pl
import torch
from pytorch_lightning import Trainer
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
from pytorch_lightning.loggers import CSVLogger, TensorBoardLogger
from pytorch_lightning.utilities.rank_zero import rank_zero_only
import launch
import threestudio
from threestudio.systems.base import BaseSystem
from threestudio.utils.callbacks import (
CodeSnapshotCallback,
ConfigSnapshotCallback,
CustomProgressBar,
ProgressCallback,
)
from threestudio.utils.config import ExperimentConfig, load_config
from threestudio.utils.misc import get_rank
from threestudio.utils.typing import Optional
def inference(cfg, logger, devices, seed):
# set a different seed for each device
pl.seed_everything(seed, workers=True)
# Image Data Infomation from Config
dm = threestudio.find(cfg.data_type)(cfg.data)
# 3D Generation Model System from Config
system: BaseSystem = threestudio.find(cfg.system_type) (
cfg.system, resumed=cfg.resume is not None, cfg_full=cfg
)
system.set_save_dir(os.path.join(cfg.trial_dir, "save"))
if args.gradio:
fh = logging.FileHandler(os.path.join(cfg.trial_dir, "logs"))
fh.setLevel(logging.INFO)
if args.verbose:
fh.setLevel(logging.DEBUG)
fh.setFormatter(logging.Formatter("[%(levelname)s] %(message)s"))
logger.addHandler(fh)
callbacks = []
if args.train:
callbacks += [
ModelCheckpoint(
dirpath=os.path.join(cfg.trial_dir, "ckpts"), **cfg.checkpoint
),
LearningRateMonitor(logging_interval="step"),
CodeSnapshotCallback(
os.path.join(cfg.trial_dir, "code"), use_version=False
),
ConfigSnapshotCallback(
args.config,
cfg,
os.path.join(cfg.trial_dir, "configs"),
use_version=False,
),
]
if args.gradio:
callbacks += [
ProgressCallback(save_path=os.path.join(cfg.trial_dir, "progress"))
]
else:
callbacks += [CustomProgressBar(refresh_rate=1)]
def write_to_text(file, lines):
with open(file, "w") as f:
for line in lines:
f.write(line + "\n")
loggers = []
if args.train:
# make tensorboard logging dir to suppress warning
rank_zero_only(
lambda: os.makedirs(os.path.join(cfg.trial_dir, "tb_logs"), exist_ok=True)
)()
loggers += [
TensorBoardLogger(cfg.trial_dir, name="tb_logs"),
CSVLogger(cfg.trial_dir, name="csv_logs"),
] + system.get_loggers()
rank_zero_only(
lambda: write_to_text(
os.path.join(cfg.trial_dir, "cmd.txt"),
["python " + " ".join(sys.argv), str(args)],
)
)()
trainer = Trainer(
callbacks=callbacks,
logger=loggers,
inference_mode=False,
accelerator="gpu",
devices=devices,
**cfg.trainer,
)
def set_system_status(system: BaseSystem, ckpt_path: Optional[str]):
if ckpt_path is None:
return
ckpt = torch.load(ckpt_path, map_location="cpu")
system.set_resume_status(ckpt["epoch"], ckpt["global_step"])
if args.train:
trainer.fit(system, datamodule=dm, ckpt_path=cfg.resume)
trainer.test(system, datamodule=dm)
if args.gradio:
# also export assets if in gradio mode
trainer.predict(system, datamodule=dm)
elif args.validate:
# manually set epoch and global_step as they cannot be automatically resumed
set_system_status(system, cfg.resume)
trainer.validate(system, datamodule=dm, ckpt_path=cfg.resume)
elif args.test:
# manually set epoch and global_step as they cannot be automatically resumed
set_system_status(system, cfg.resume)
trainer.test(system, datamodule=dm, ckpt_path=cfg.resume)
elif args.export:
set_system_status(system, cfg.resume)
trainer.predict(system, datamodule=dm, ckpt_path=cfg.resume)
def main(args, extras) -> None:
# set CUDA_VISIBLE_DEVICES if needed, then import pytorch-lightning
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
env_gpus_str = os.environ.get("CUDA_VISIBLE_DEVICES", None)
env_gpus = list(env_gpus_str.split(",")) if env_gpus_str else []
# selected_gpus = [0]
devices = -1
if len(env_gpus) > 0:
# CUDA_VISIBLE_DEVICES was set already, e.g. within SLURM srun or higher-level script.
n_gpus = len(env_gpus)
else:
selected_gpus = list(args.gpu.split(","))
n_gpus = len(selected_gpus)
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
if args.typecheck:
from jaxtyping import install_import_hook
install_import_hook("threestudio", "typeguard.typechecked")
logger = logging.getLogger("pytorch_lightning")
if args.verbose:
logger.setLevel(logging.DEBUG)
for handler in logger.handlers:
if handler.stream == sys.stderr: # type: ignore
if not args.gradio:
handler.setFormatter(logging.Formatter("%(levelname)s %(message)s"))
handler.addFilter(launch.ColoredFilter())
else:
handler.setFormatter(logging.Formatter("[%(levelname)s] %(message)s"))
launch.load_custom_modules()
# parse YAML config to OmegaConf
cfg: ExperimentConfig
with open(args.images_json, "r") as file:
images = json.load(file)
for image in images:
image_name = image["image_path"].split("/")[-1].split(".")[0].replace(" ", "_")
cfg = load_config(args.config, cli_args=extras, n_gpus=n_gpus)
cfg.data.image_path = image["image_path"]
update_values = [
# ("data.image_path", image["image_path"]),
("tag", image_name),
]
cfg.tag = cfg.tag.replace("hamburger_rgba.png", image_name)
cfg.trial_name = cfg.tag
cfg.trial_dir = os.path.join(cfg.exp_dir, cfg.trial_name)
# cfg.__post_init__()
# cfg.update_values(update_values)
# args.config.data.image_path = image["image_path"]
# cfg.data.image_path = image["image_path"]
# args.config.tag = args.config.tag.replace("hamburger_rgba.png", image_name)
for i in range(4):
random_seed = random.randint(0, 100000)
random_seed = random_seed + get_rank()
cfg.seed = random_seed
inference(cfg, logger, devices, random_seed)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--config",
default="configs/zero123.yaml",
help="path to config file (configs/zero123.yaml)",
)
parser.add_argument(
"--gpu",
default="0",
help="GPU(s) to be used. 0 means use the 1st available GPU. "
"1,2 means use the 2nd and 3rd available GPU. "
"If CUDA_VISIBLE_DEVICES is set before calling `launch.py`, "
"this argument is ignored and all available GPUs are always used.",
)
parser.add_argument(
"--images_json",
default="./load/data/images.json",
help="path to json file (./load/data/images.json)",
)
group = parser.add_mutually_exclusive_group(required=True)
group.add_argument("--train", action="store_true")
group.add_argument("--validate", action="store_true")
group.add_argument("--test", action="store_true")
group.add_argument("--export", action="store_true")
parser.add_argument(
"--gradio", action="store_true", help="if true, run in gradio mode"
)
parser.add_argument(
"--verbose", action="store_true", help="if true, set logging level to DEBUG"
)
parser.add_argument(
"--typecheck",
action="store_true",
help="whether to enable dynamic type checking",
)
args, extras = parser.parse_known_args()
if args.gradio:
# FIXME: no effect, stdout is not captured
with contextlib.redirect_stdout(sys.stderr):
main(args, extras)
else:
main(args, extras)