-
Notifications
You must be signed in to change notification settings - Fork 0
/
launch_speedup.py
399 lines (293 loc) · 12.3 KB
/
launch_speedup.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 argparse
import contextlib
import logging
import os
import sys
from tqdm import tqdm
import torch
from threestudio.utils.base import (
Updateable,
update_end_if_possible,
update_if_possible,
)
import time
from datetime import datetime, timedelta
from threestudio.utils.config import dump_config
import numpy as np
logging.getLogger("lightning").setLevel(logging.ERROR)
class ColoredFilter(logging.Filter):
"""
A logging filter to add color to certain log levels.
"""
RESET = "\033[0m"
RED = "\033[31m"
GREEN = "\033[32m"
YELLOW = "\033[33m"
BLUE = "\033[34m"
MAGENTA = "\033[35m"
CYAN = "\033[36m"
COLORS = {
"WARNING": YELLOW,
"INFO": GREEN,
"DEBUG": BLUE,
"CRITICAL": MAGENTA,
"ERROR": RED,
}
RESET = "\x1b[0m"
def __init__(self):
super().__init__()
def filter(self, record):
if record.levelname in self.COLORS:
color_start = self.COLORS[record.levelname]
record.levelname = f"{color_start}[{record.levelname}]"
record.msg = f"{record.msg}{self.RESET}"
return True
def run(rank, total_ranks, queues):
import argparse
from threestudio.utils.config import ExperimentConfig, load_config
parser = argparse.ArgumentParser()
parser.add_argument("--config", required=True, help="path to config file")
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.",
)
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()
# parse YAML config to OmegaConf
config: ExperimentConfig
config = load_config(args.config, cli_args=extras)
device = torch.device(f"cuda:{rank}")
print('Start process ', rank)
if rank == 0:
system, dm, dataiters = prepare_train(args, extras, device, config, rank=rank)
if args.test:
system.load_state_dict(torch.load(config.resume, map_location='cuda:0')['state_dict'], strict=False)
test_dataset =dm.test_dataloader()
system.eval()
global_step = config.trainer.max_steps
with torch.no_grad():
for batch in tqdm(test_dataset):
batch = to_device(batch, device)
update_if_possible(test_dataset.dataset, 0, global_step)
system.do_update_step(0, global_step)
system.test_step(batch, global_step)
system.do_update_step_end(0, global_step)
update_end_if_possible( test_dataset.dataset, 0, global_step )
system.on_test_epoch_end(global_step)
return
train_loop(system, config, device, dm, queues, config.seed)
for i in range(total_ranks - 1):
queues[0].put(None)
else:
if args.test:
return
system, dm, dataiters = prepare_train(args, extras, device, config,rank=rank)
optimizers = system.configure_optimizers()
run_worker(system, optimizers, dataiters, queues, device, config.seed)
def main():
import torch.multiprocessing as mp
torch.autograd.set_detect_anomaly(True)
mp.set_start_method('spawn', force=True)
# torch.set_num_threads(1)
queues = mp.Queue(), mp.Queue(), mp.Queue()
processes = []
num_processes = torch.cuda.device_count()
if num_processes == 1:
run(0,1,queues)
exit(0)
for rank in range(num_processes):
p = mp.Process(target=run, args=(rank, num_processes, queues))
p.start()
processes.append(p)
for p in processes:
p.join() # wait for all subprocesses to finish
def prepare_train(args, extras, device, cfg, rank=0) -> None:
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
if args.typecheck:
from jaxtyping import install_import_hook
install_import_hook("threestudio", "typeguard.typechecked")
import threestudio
from threestudio.systems.base import BaseSystem
dm = threestudio.find(cfg.data_type)(cfg.data)
dm.setup()
if rank > 0:
dataiters = dm.get_all_train_iters()
else:
dataiters = None
if 'loggers' in cfg.system:
cfg.system.pop('loggers')
system: BaseSystem = threestudio.find(cfg.system_type)(
cfg.system, device, resumed=cfg.resume is not None
)
if rank == 0:
date_str = datetime.today().strftime('%Y-%m-%d-%H:%M:%S')
cfg.trial_dir = cfg.trial_dir + f'@{date_str}'
save_dir = os.path.join(cfg.trial_dir, "save")
system.set_save_dir(save_dir)
config_dir = os.path.join(cfg.trial_dir, "configs")
os.makedirs(config_dir, exist_ok=True)
dump_config(os.path.join(config_dir, "parsed.yaml"), cfg)
system.to(device)
system.on_fit_start(device)
system.do_update_step(0, 0)
system.do_update_step_end(0, 0)
return system, dm, dataiters
def to_device(batch, device):
for k, v in batch.items():
if isinstance(v, torch.Tensor):
batch[k] = v.to(device)
elif isinstance(v, dict):
batch[k] = to_device(batch[k], device)
return batch
def run_worker(system, optimizers,dataiters, queues, device, seed_offset):
while True:
ret = queues[0].get()
if ret is None:
return
(params, step) = ret
system.set_params(params, clone=False)
res = take_step(system, optimizers, dataiters, device, step, seed_offset)
myparams = system.get_params(include_occs=False)
grads = [para.grad for para in myparams]
queues[1].put(
(grads, step)
)
if 'loss_lora' in res:
optimizers['optimizer_guidance'].zero_grad()
res['loss_lora'].backward()
optimizers['optimizer_guidance'].step()
def take_step(system, optimizers, dataiters, device, step, seed_offset):
np.random.seed(step+seed_offset)
torch.manual_seed(step+seed_offset)
torch.cuda.manual_seed(step+seed_offset)
batch = next(dataiters.get_current_iter(step))
batch = to_device(batch, device)
system.do_update_step(0, step)
res = system.training_step(batch, step)
system.do_update_step_end(0, step)
optimizers['optimizer'].zero_grad()
res['loss'].backward()
return res
def optimizer_state_clone(optimizer_from, optimizer_to):
optimizer_to.load_state_dict(optimizer_from.state_dict())
def train_loop(system, config, device, datamodule, queues, seed_offset):
train_config, speedup_config = config.trainer, config.speedup
save_dir = system.get_save_dir()
test_dataset =datamodule.test_dataloader()
val_dataset =datamodule.val_dataloader()
thresh = speedup_config.threshold
T = train_config.max_steps
P = speedup_config.P
systems = [None for _ in range(T+1)]
optimizers = [None for _ in range(T+1)]
begin_idx, end_idx = 0, P
total_iters = 0
start_time = time.time()
pbar = tqdm(total=T)
for step in range(P+1):
systems[step] = system.clone(device)
optimizers[step] = systems[step].configure_optimizers()['optimizer']
last_vis_time = 0
while begin_idx < T:
parallel_len = end_idx - begin_idx
pred_f = [None for _ in range(parallel_len)]
for i in range(parallel_len):
step = begin_idx + i
params =[p.data for p in systems[step].get_params(include_occs=True, include_binaries=True)]
queues[0].put( (params, step) )
for i in range(parallel_len):
_preds, _step = queues[1].get()
_i = _step - begin_idx
pred_f[_i] = _preds
rollout_system = systems[begin_idx]
rollout_optimizer = optimizers[begin_idx]
ind = None
errors_all = 0
for i in range(parallel_len):
step = begin_idx + i
if ind is None and rollout_system.renderer.cfg.get('grid_prune', False) and step > 0:
np.random.seed(step+seed_offset)
torch.manual_seed(step+seed_offset)
torch.cuda.manual_seed(step+seed_offset)
rollout_system.renderer.do_update_step(0, step)
rollout_system.set_grads_from_grads(pred_f[i])
rollout_optimizer.step()
rollout_optimizer.zero_grad()
if ind is None and rollout_system.renderer.cfg.get('grid_prune', False) and step > 0:
rollout_system.renderer.do_update_step_end(0, step)
# compute error
error = rollout_system.compute_error_from_system(systems[step+1])
if speedup_config.adaptivity_type == 'median':
if i == parallel_len // 2:
errors_all = error
elif speedup_config.adaptivity_type == 'mean':
errors_all += error / parallel_len
else:
raise ValueError('Adaptivity not supported')
if ind is None and (error > thresh or i == parallel_len - 1):
ind = step+1
optimizer_state_clone(optimizer_from=rollout_optimizer, optimizer_to=optimizers[step+1])
if ind is not None:
systems[step+1] = rollout_system.clone(device, systems[step+1])
thresh = thresh * speedup_config.ema_decay + (errors_all ) * (1 - speedup_config.ema_decay)
new_begin_idx = ind
new_end_idx = min(new_begin_idx + parallel_len, T)
for step in range(end_idx+1, new_end_idx+1):
systems[step] = rollout_system.clone(device, systems[step - 1 - parallel_len])
optimizers[step] = optimizers[step - 1 - parallel_len]
progress = new_begin_idx - begin_idx
begin_idx = new_begin_idx
end_idx = new_end_idx
total_iters += 1
pbar.update(progress)
elapsed = time.time() - start_time
if elapsed >= last_vis_time + train_config.val_check_interval and train_config.visualize_progress:
elapsed = time.time() - start_time
systems[begin_idx].eval()
with torch.no_grad():
systems[begin_idx].set_save_dir(save_dir)
batch = next(iter(val_dataset))
batch = to_device(batch, device=device)
systems[begin_idx].validation_step(batch, begin_idx, [str(timedelta(seconds=elapsed)).split('.')[0]] if train_config.display_time else None)
last_vis_time = elapsed
systems[begin_idx].train()
pbar.close()
elapsed = time.time() - start_time
print('Effective Iters:', total_iters)
system = systems[T]
system.set_save_dir(save_dir)
system.eval()
global_step = T
with torch.no_grad():
for batch in tqdm(test_dataset):
batch = to_device(batch, device)
update_if_possible(test_dataset.dataset, 0, global_step)
system.do_update_step(0, global_step)
system.test_step(batch, global_step, [ str(timedelta(seconds=elapsed)).split('.')[0]] if train_config.display_time else None )
system.do_update_step_end(0, global_step)
update_end_if_possible( test_dataset.dataset, 0, global_step )
system.on_test_epoch_end(global_step)
save_dict = {'epoch': 0, 'global_step': T,
'state_dict': system.state_dict()}
torch.save(save_dict, os.path.join(save_dir, 'last.ckpt'))
if __name__ == "__main__":
main()