-
Notifications
You must be signed in to change notification settings - Fork 113
/
finetune_sdm_yaml.py
174 lines (151 loc) · 6.8 KB
/
finetune_sdm_yaml.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
# --------------------------------------------------------
# DisCo - Disentangled Control for Referring Human Dance Generation in Real World
# Licensed under The Apache-2.0 license License [see LICENSE for details]
# Tan Wang (TAN317@e.ntu.edu.sg)
# Work done during internship at Microsoft
# --------------------------------------------------------
from utils.wutils_ldm import *
from agent import Agent_LDM, WarmupLinearLR, WarmupLinearConstantLR
import os
import torch
from utils.lib import *
from utils.dist import dist_init
from dataset.tsv_dataset import make_data_sampler, make_batch_data_sampler
torch.multiprocessing.set_sharing_strategy('file_system')
def get_loader_info(args, size_batch, dataset):
is_train = dataset.split == 'train'
if is_train:
images_per_gpu = min(
size_batch * max(1, (args.max_video_len // dataset.max_video_len)),
128)
images_per_batch = images_per_gpu * args.world_size
iter_per_ep = len(dataset) // images_per_batch
if args.epochs == -1: # try to add iters into args
assert args.ft_iters > 0
num_iters = args.ft_iters
args.epochs = (num_iters * images_per_batch) // len(dataset) + 1
else:
num_iters = iter_per_ep * args.epochs
else:
images_per_gpu = size_batch * (
args.max_video_len // dataset.max_video_len)
images_per_batch = images_per_gpu * args.world_size
iter_per_ep = None
num_iters = None
loader_info = (images_per_gpu, images_per_batch, iter_per_ep, num_iters)
return loader_info
def make_data_loader(
args, size_batch, dataset, start_iter=0, loader_info=None):
is_train = dataset.split == 'train'
collate_fn = None #dataset.collate_batch
is_distributed = args.distributed
if is_train:
shuffle = True
start_iter = start_iter
else:
shuffle = False
start_iter = 0
if loader_info is None:
loader_info = get_loader_info(args, size_batch, dataset)
images_per_gpu, images_per_batch, iter_per_ep, num_iters = loader_info
if hasattr(args, 'limited_samples'):
limited_samples = args.limited_samples // args.local_size
else:
limited_samples = -1
random_seed = args.seed
sampler = make_data_sampler(
dataset, shuffle, is_distributed, limited_samples=limited_samples,
random_seed=random_seed)
batch_sampler = make_batch_data_sampler(
sampler, images_per_gpu, num_iters, start_iter
)
data_loader = torch.utils.data.DataLoader(
dataset, num_workers=args.num_workers, batch_sampler=batch_sampler,
pin_memory=True, collate_fn=collate_fn
)
meta_info = (images_per_batch, iter_per_ep, num_iters)
return data_loader, meta_info
def main_worker(args):
"""
"""
cf = import_filename(args.cf)
Net, inner_collect_fn = cf.Net, cf.inner_collect_fn
dataset_cf = import_filename(args.dataset_cf)
BaseDataset = dataset_cf.BaseDataset
print(f"Args: {edict(sorted(vars(args).items()))}")
# init models
logger.info('Building models...')
model = Net(args)
if args.do_train:
logger.warning("Do training...")
# Prepare Dataset.
if getattr(args, 'refer_clip_preprocess', None):
train_dataset = BaseDataset(args, args.train_yaml, split='train', preprocesser=model.feature_extractor)
eval_dataset = BaseDataset(args, args.val_yaml, split='val', preprocesser=model.feature_extractor)
else:
train_dataset = BaseDataset(args, args.train_yaml, split='train')
eval_dataset = BaseDataset(args, args.val_yaml, split='val')
train_info = get_loader_info(args, args.local_train_batch_size,
train_dataset)
_, images_per_batch, args.iter_per_ep, args.num_iters = train_info
if args.eval_step <= 5.0:
args.eval_step = args.eval_step * args.iter_per_ep
if args.save_step <= 5.0:
args.save_step = args.save_step * args.iter_per_ep
args.eval_step = int(max(10, args.eval_step))
args.save_step = int(max(10, args.save_step))
from torch.optim import AdamW
optimizer = AdamW(model.parameters(), lr=args.learning_rate, betas=(0.9, 0.98), weight_decay=args.decay)
optimizer = getattr(model, 'optimizer', optimizer)
if args.constant_lr:
scheduler = WarmupLinearConstantLR(
optimizer,
max_iter=(args.num_iters // args.gradient_accumulate_steps) + 1,
warmup_ratio=getattr(args, 'warmup_ratio', 0.05))
else:
scheduler = WarmupLinearLR(
optimizer,
max_iter=(
args.num_iters//args.gradient_accumulate_steps)+1,
warmup_ratio=getattr(args, 'warmup_ratio', 0.05))
scheduler = getattr(model, 'scheduler', scheduler)
trainer = Agent_LDM(args, model, optimizer, scheduler)
trainer.setup_model_for_training()
train_dataloader, train_info = make_data_loader(
args, args.local_train_batch_size,
train_dataset, start_iter=trainer.global_step+1, loader_info=train_info)
logger.info(
f"Video Length {train_dataset.size_frame}")
logger.info(
f"Total batch size {images_per_batch}")
logger.info(
f"Total training steps {args.num_iters}")
logger.info(f"Starting train iter: {trainer.global_step+1}")
logger.info(
f"Training steps per epoch (accumulated) {args.iter_per_ep}")
logger.info(
f"Training dataloader length {len(train_dataloader)}")
logger.info(
f"Evaluation happens every {args.eval_step} steps")
logger.info(
f"Checkpoint saves every {args.save_step} steps")
eval_dataloader, eval_info = make_data_loader(
args, args.local_eval_batch_size,
eval_dataset)
trainer.train_eval_by_iter(train_loader=train_dataloader, eval_loader=eval_dataloader, inner_collect_fn=inner_collect_fn)
if args.eval_visu:
logger.warning("Do eval_visu...")
if getattr(args, 'refer_clip_preprocess', None):
eval_dataset = BaseDataset(args, args.val_yaml, split='val', preprocesser=model.feature_extractor)
else:
eval_dataset = BaseDataset(args, args.val_yaml, split='val')
eval_dataloader, eval_info = make_data_loader(
args, args.local_eval_batch_size,
eval_dataset)
trainer = Agent_LDM(args=args, model=model)
trainer.eval(eval_dataloader, inner_collect_fn=inner_collect_fn,
enc_dec_only='enc_dec_only' in args.eval_save_filename)
if __name__ == "__main__":
from utils.args import sharedArgs
parsed_args = sharedArgs.parse_args()
main_worker(parsed_args)