-
Notifications
You must be signed in to change notification settings - Fork 0
/
reflow_train_ddp.py
273 lines (231 loc) · 10.3 KB
/
reflow_train_ddp.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
from absl import app
from absl import flags
from ml_collections.config_flags import config_flags
from pathlib import Path
from torch.utils import tensorboard
import torch
from torch.utils.data import DataLoader
from tqdm.auto import trange
from torchvision.utils import make_grid, save_image
import numpy as np
from loguru import logger
from accelerate import Accelerator
from diffusers.optimization import get_scheduler
from reflow.utils import get_sampling_fn, set_seed, decode_latents, get_loss_fn
from reflow.utils import ExponentialMovingAverage
from reflow.sde_lib import RectifiedFlow
from reflow.data.reflow_with_text import DataPairsWithText, get_reflow_dataset
from reflow.data.utils import get_image_transforms, LMDB_ndarray
from reflow.utils import create_models, to_device, cycle
def main(argv):
config, workdir = FLAGS.config, FLAGS.workdir
workdir = Path(workdir)
set_seed(config.seed)
# Create directories for experimental logs
sample_dir = workdir/"samples"
sample_dir.mkdir(parents=True, exist_ok=True)
tb_dir = workdir/"tensorboard"
tb_dir.mkdir(exist_ok=True)
writer = tensorboard.SummaryWriter(str(tb_dir))
accelerator = Accelerator(
gradient_accumulation_steps=config.training.gradient_accumulation_steps,
mixed_precision=config.training.mixed_precision,
)
if accelerator.is_main_process:
logger.add(str(workdir / 'exp.log'))
logger.info(f'\n{config}')
logger.info(f'comment : {FLAGS.comment}')
tokenizer, text_encoder, vae, score_model = create_models(config)
weight_dtype = torch.float32
if config.training.mixed_precision == "fp16":
weight_dtype = torch.float16
vae.to(accelerator.device, dtype=weight_dtype)
text_encoder.to(accelerator.device, dtype=weight_dtype)
# Initialize the optimizer
if config.optim.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
)
optimizer_cls = bnb.optim.AdamW8bit
else:
optimizer_cls = torch.optim.AdamW
optimizer = optimizer_cls(
score_model.parameters(),
lr=config.optim.lr,
betas=config.optim.betas,
weight_decay=config.optim.weight_decay,
eps=config.optim.eps,
)
lr_scheduler = get_scheduler(
config.optim.lr_scheduler,
optimizer=optimizer,
num_warmup_steps=config.optim.warmup *
config.training.gradient_accumulation_steps,
num_training_steps=config.training.num_steps *
config.training.gradient_accumulation_steps,
)
train_ds = get_reflow_dataset(
data_root=config.data.train_root,
tokenizer=tokenizer,
src_type='lmdb',
train=True,
random_flip=config.data.random_flip,
)
eval_ds = get_reflow_dataset(
data_root=config.data.eval_root,
tokenizer=tokenizer,
src_type='lmdb',
)
train_dl = DataLoader(
train_ds,
batch_size=config.training.batch_size,
shuffle=True,
num_workers=config.data.dl_workers,
drop_last=True,
)
eval_dl = DataLoader(
eval_ds,
batch_size=config.training.batch_size,
shuffle=False,
num_workers=config.data.dl_workers,
drop_last=True,
)
score_model, optimizer, train_dl, lr_scheduler = accelerator.prepare(
score_model, optimizer, train_dl, lr_scheduler
)
train_iter = cycle(train_dl)
eval_iter = cycle(eval_dl)
# accelerator.register_for_checkpointing(lr_scheduler)
initial_step = 0
ckpt_path = config.training.ckpt_path
if ckpt_path is not None:
# state = restore_checkpoint(ckpt_path, state, accelerator.device)
initial_step = int(ckpt_path.split(
'/')[-1].split('_')[-1][1:]) # checkpoint_s{xxx}
accelerator.load_state(f'{ckpt_path}')
ema = ExponentialMovingAverage(
score_model.parameters(), decay=config.ema.decay)
state = dict(model=score_model, ema=ema, step=initial_step)
checkpoint_dir = workdir/'checkpoints'
checkpoint_dir.mkdir(exist_ok=True)
if accelerator.is_main_process:
accelerator.init_trackers("tmp", config=vars(config))
sde = RectifiedFlow(
init_type=config.sampling.init_type,
noise_scale=config.sampling.init_noise_scale,
reflow_flag=True,
reflow_t_schedule=config.reflow.reflow_t_schedule,
reflow_loss=config.reflow.reflow_loss,
use_ode_sampler=config.sampling.use_ode_sampler,
sample_N=config.sampling.sample_N,
codec=vae,
device=accelerator.device,
)
# Building sampling functions
sampling_shape = (config.training.batch_size, config.data.num_channels,
config.data.image_size, config.data.image_size)
sampling_fn = get_sampling_fn(
config, sde, sampling_shape, eps=1e-3) # 使用 euler 或 rk45
# optimize_fn = optimization_manager(config)
# reduce_mean = config.training.reduce_mean
# train_step_fn = get_step_fn(sde, train=True, optimize_fn=optimize_fn,
# reduce_mean=reduce_mean, )
# eval_step_fn = get_step_fn(sde, train=False, optimize_fn=optimize_fn,
# reduce_mean=reduce_mean,)
reduce_mean = config.training.reduce_mean
train_loss_fn = get_loss_fn(sde, train=True, reduce_mean=reduce_mean,)
eval_loss_fn = get_loss_fn(sde, train=False, reduce_mean=reduce_mean,)
num_train_steps = config.training.num_steps
if accelerator.is_main_process:
logger.info(f'REFLOW T SCHEDULE: {config.reflow.reflow_t_schedule}')
logger.info(f'LOSS: {config.reflow.reflow_loss}')
logger.info(f"Starting reflow training loop at step {initial_step}.")
def prepare_step_fn_input(batch):
z0 = batch.pop('noise')
z1 = batch.pop('latent')
encoder_hidden_states = text_encoder(**batch)[0]
return {
'z0': z0,
'z1': z1,
'encoder_hidden_states': encoder_hidden_states,
}
pbar = trange(initial_step, num_train_steps, desc='Steps',
disable=not accelerator.is_local_main_process)
for step in pbar:
train_loss = 0.0
for _ in range(config.training.gradient_accumulation_steps):
batch = next(train_iter)
if config.training.randz0 == 'random':
# 1-reflow , random noise for same target
batch['noise'] = torch.randn_like(batch['noise'])
with accelerator.accumulate(score_model):
loss = train_loss_fn(state, prepare_step_fn_input(batch))
avg_loss = accelerator.gather(
loss.repeat(config.training.batch_size)).mean()
train_loss += avg_loss
accelerator.backward(loss)
if accelerator.sync_gradients:
accelerator.clip_grad_norm_(
score_model.parameters(), config.optim.grad_clip)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
train_loss = train_loss / config.training.gradient_accumulation_steps
if accelerator.sync_gradients:
state['step'] += 1
state['ema'].update(score_model.parameters())
pbar.set_postfix({'train_loss': train_loss.item()})
if accelerator.is_main_process:
if step % config.training.log_freq == 0:
logger.info(
f'step {step} | training_loss {train_loss.item():.5f}')
writer.add_scalar("training_loss", train_loss, step)
if step % config.training.eval_freq == 0:
eval_batch = to_device(next(eval_iter), accelerator.device)
eval_loss = eval_loss_fn(
state, prepare_step_fn_input(eval_batch))
logger.info(f'step {step} | eval_loss {eval_loss.item():.5f}')
writer.add_scalar("eval_loss", eval_loss, step)
if step != initial_step and step % config.training.snapshot_freq == 0 or step == num_train_steps-1:
# Save the checkpoint.
accelerator.save_state(
str(checkpoint_dir / f'checkpoint_s{step}'))
# save_checkpoint(str(checkpoint_dir / f'checkpoint_s{step}.pth'), state)
score_model_to_save = accelerator.unwrap_model(score_model)
ema.copy_to(score_model_to_save.parameters())
torch.save(score_model_to_save.state_dict(), str(
checkpoint_dir / f'score_model_s{step}.pth'))
if step != initial_step and step % config.training.sampling_freq == 0 or step == num_train_steps-1:
# Generate and save samples
ema.store(score_model.parameters())
ema.copy_to(score_model.parameters())
eval_batch = to_device(next(eval_iter), accelerator.device)
eval_step_fn_input = prepare_step_fn_input(eval_batch)
z0 = eval_step_fn_input.pop('z0')
z1 = eval_step_fn_input.pop('z1')
sample, n = sampling_fn(
score_model,
z=None if config.sampling.randz0 == 'random' else z0,
condition=eval_step_fn_input,
)
ema.restore(score_model.parameters())
images = decode_latents(vae, sample)
nrow = int(np.sqrt(sample.shape[0]))
image_grid = make_grid(images, nrow, padding=2)
save_image(image_grid, str(sample_dir / f'sample_s{step}.png'))
pbar.close()
if __name__ == "__main__":
FLAGS = flags.FLAGS
config_flags.DEFINE_config_file(
"config", None, "Training configuration.", lock_config=True)
flags.DEFINE_string("workdir", None, "Work directory.")
# flags.DEFINE_enum("mode", None, ["train", "eval", "reflow"], "Running mode")
flags.DEFINE_string("eval_folder", "eval",
"The folder name for storing evaluation results")
flags.DEFINE_string("comment", None, "complementary info of exp")
# flags.mark_flags_as_required(["workdir", "config", "mode"])
flags.mark_flags_as_required(["workdir", "config"])
app.run(main)