-
Notifications
You must be signed in to change notification settings - Fork 24
/
train.py
181 lines (150 loc) · 5.16 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
181
# Copyright (c) Meta Platforms, Inc. All Rights Reserved
import json
import os
import random
import numpy as np
import torch
from data_loaders.dataloader import get_dataloader, load_data, TrainDataset
from model.networks import PureMLP
from runner.train_mlp import train_step
from runner.training_loop import TrainLoop
from utils import dist_util
from utils.model_util import create_model_and_diffusion
from utils.parser_util import train_args
def train_diffusion_model(args, dataloader):
print("creating model and diffusion...")
args.arch = args.arch[len("diffusion_") :]
num_gpus = torch.cuda.device_count()
args.num_workers = args.num_workers * num_gpus
model, diffusion = create_model_and_diffusion(args)
if num_gpus > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
dist_util.setup_dist()
model = torch.nn.DataParallel(model).cuda()
print(
"Total params: %.2fM"
% (sum(p.numel() for p in model.module.parameters()) / 1000000.0)
)
else:
dist_util.setup_dist(args.device)
model.to(dist_util.dev())
print(
"Total params: %.2fM"
% (sum(p.numel() for p in model.parameters()) / 1000000.0)
)
print("Training...")
TrainLoop(args, model, diffusion, dataloader).run_loop()
print("Done.")
def train_mlp_model(args, dataloader):
print("creating MLP model...")
args.arch = args.arch[len("mlp_") :]
num_gpus = torch.cuda.device_count()
args.num_workers = args.num_workers * num_gpus
model = PureMLP(
args.latent_dim,
args.input_motion_length,
args.layers,
args.sparse_dim,
args.motion_nfeat,
)
model.train()
if num_gpus > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
dist_util.setup_dist()
model = torch.nn.DataParallel(model).cuda()
print(
"Total params: %.2fM"
% (sum(p.numel() for p in model.module.parameters()) / 1000000.0)
)
else:
dist_util.setup_dist(args.device)
model.to(dist_util.dev())
print(
"Total params: %.2fM"
% (sum(p.numel() for p in model.parameters()) / 1000000.0)
)
# initialize optimizer
optimizer = torch.optim.Adam(
model.parameters(), lr=args.lr, weight_decay=args.weight_decay
)
nb_iter = 0
avg_loss = 0.0
avg_lr = 0.0
while (nb_iter + 1) < args.num_steps:
for (motion_target, motion_input) in dataloader:
loss, optimizer, current_lr = train_step(
motion_input,
motion_target,
model,
optimizer,
nb_iter,
args.num_steps,
args.lr,
args.lr / 10.0,
dist_util.dev(),
args.lr_anneal_steps,
)
avg_loss += loss
avg_lr += current_lr
if (nb_iter + 1) % args.log_interval == 0:
avg_loss = avg_loss / args.log_interval
avg_lr = avg_lr / args.log_interval
print("Iter {} Summary: ".format(nb_iter + 1))
print(f"\t lr: {avg_lr} \t Training loss: {avg_loss}")
avg_loss = 0
avg_lr = 0
if (nb_iter + 1) == args.num_steps:
break
nb_iter += 1
with open(
os.path.join(args.save_dir, "model-iter-" + str(nb_iter + 1) + ".pth"),
"wb",
) as f:
torch.save(model.state_dict(), f)
def main():
args = train_args()
torch.backends.cudnn.benchmark = False
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.save_dir is None:
raise FileNotFoundError("save_dir was not specified.")
elif os.path.exists(args.save_dir) and not args.overwrite:
raise FileExistsError("save_dir [{}] already exists.".format(args.save_dir))
elif not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
args_path = os.path.join(args.save_dir, "args.json")
with open(args_path, "w") as fw:
json.dump(vars(args), fw, indent=4, sort_keys=True)
print("creating data loader...")
motions, sparses, mean, std = load_data(
args.dataset,
args.dataset_path,
"train",
input_motion_length=args.input_motion_length,
)
dataset = TrainDataset(
args.dataset,
mean,
std,
motions,
sparses,
args.input_motion_length,
args.train_dataset_repeat_times,
args.no_normalization,
)
dataloader = get_dataloader(
dataset, "train", batch_size=args.batch_size, num_workers=args.num_workers
)
# args.lr_anneal_steps = (
# args.lr_anneal_steps // args.train_dataset_repeat_times
# ) * len(
# dataloader
# ) # the input lr_anneal_steps is by epoch, here convert it to the number of steps
model_type = args.arch.split("_")[0]
if model_type == "diffusion":
train_diffusion_model(args, dataloader)
elif model_type == "mlp":
train_mlp_model(args, dataloader)
if __name__ == "__main__":
main()