-
Notifications
You must be signed in to change notification settings - Fork 9
/
train_model.py
244 lines (196 loc) · 7.99 KB
/
train_model.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
# Copyright (c) Meta, Inc. and its affiliates.
# Copyright (c) Stanford University
import argparse
import os
import sys
import time
import torch
import torch.optim as optim
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.data import DataLoader
from simple_transformer_with_state import TF_RNN_Past_State
from training_data_loader import TrainSubDataset
from learning_utils import set_seed, loss_q_only_2axis, loss_constr_multi, loss_jerk
sys.path.append("../")
torch.set_printoptions(threshold=10_000)
parser = argparse.ArgumentParser(description='Transformer Training for IMU')
parser.add_argument('--batch_size', type=int, default=128,
help='batch size (default: 128)')
parser.add_argument('--cuda', action='store_true',
help='use CUDA (default: False)')
parser.add_argument('--rnn_dropout', type=float, default=0.0,
help='dropout applied to layers (default: 0.0)')
parser.add_argument('--in_dropout', type=float, default=0.0,
help='dropout applied to IMU input (default: 0.0)')
parser.add_argument('--clip', type=float, default=5.0,
help='gradient clip, -1 means no clip (default: 5.0)')
parser.add_argument('--epochs', type=int, default=10,
help='upper epoch limit (default: 10)')
parser.add_argument('--seq_len', type=int, default=40,
help='sequence window length for input (default: 40)')
parser.add_argument('--log-interval', type=int, default=100,
help='report interval (default: 100')
parser.add_argument('--lr', type=float, default=4e-4,
help='initial learning rate (default: 4e-4)')
parser.add_argument('--optim', type=str, default='Adam',
help='optimizer to use (default: Adam)')
parser.add_argument('--weight_decay', type=float, default='1e-5',
help='for AdamW')
parser.add_argument('--rnn_nhid', type=int, default=512,
help='hidden size of rnn (default: 512)')
parser.add_argument('--tf_nhid', type=int, default=1024,
help='hidden size of transformer')
parser.add_argument('--tf_in_dim', type=int, default=256,
help='input dimension of transformer')
parser.add_argument('--n_heads', type=int, default=8,
help='num of heads for transformer')
parser.add_argument('--tf_layers', type=int, default=4,
help='num of layers for transformer')
parser.add_argument('--seed', type=int, default=1111,
help='random seed (default: 1111)')
parser.add_argument('--save_path', type=str, default='output/model-tmp',
help='model save path')
parser.add_argument('--cosine_lr', action='store_true',
help='use cosine learning rate (default: False)')
parser.add_argument('--warm_start', type=str, default=None,
help='')
parser.add_argument('--double', action='store_true',
help='use double precision instead of single')
parser.add_argument('--past_dropout', type=float, default=0.8,
help='input dropout for past state in transformer')
parser.add_argument('--with_acc_sum', action='store_true',
help='')
parser.add_argument('--n_sbps', type=int, default=5,
help='')
parser.add_argument('--noise_input_hist', type=float, default=0.1,
help='')
parser.add_argument('--data_version_tag', type=str, default=None,
help='')
args = parser.parse_args()
batch_size = args.batch_size
seq_length = args.seq_len
epochs = args.epochs
n_sbps = args.n_sbps
with_acc_sum = args.with_acc_sum
d_tag = args.data_version_tag
noise_input_hist = args.noise_input_hist
if args.double:
torch.set_default_dtype(torch.float64)
set_seed(args.seed)
if torch.cuda.is_available():
if not args.cuda:
print("WARNING: You have a CUDA device, so you should probably run with --cuda")
print(args)
print("Preparing data...")
input_channels = 6 * (9 + 3)
output_channels = 18 * 6 + 3 + (n_sbps * 4)
model = TF_RNN_Past_State(
input_channels, output_channels,
rnn_hid_size=args.rnn_nhid,
tf_hid_size=args.tf_nhid, tf_in_dim=args.tf_in_dim,
n_heads=args.n_heads, tf_layers=args.tf_layers,
dropout=args.rnn_dropout, in_dropout=args.in_dropout,
past_state_dropout=args.past_dropout,
with_rnn=True,
with_acc_sum=with_acc_sum
)
if args.warm_start is not None:
model.load_state_dict(torch.load(args.warm_start + ".pt"))
# TODO: better also to load Adam state
if args.cuda:
model.cuda()
lr = args.lr
if args.optim == "AdamW":
optimizer = getattr(optim, args.optim)(model.parameters(), lr=lr, weight_decay=args.weight_decay)
else:
optimizer = getattr(optim, args.optim)(model.parameters(), lr=lr)
if args.cosine_lr:
lr_s = CosineAnnealingLR(optimizer=optimizer, T_max=args.epochs + 850) # 850 probably doesn't matter
else:
lr_s = None
def train(epoch):
# torch.autograd.set_detect_anomaly(True)
model.train()
data = TrainSubDataset(
seq_length=seq_length,
imu_combine_path="data/imu_train_" + d_tag + ".npy",
s_combine_path="data/s_train_" + d_tag + ".npy",
info_path="data/info_train_" + d_tag + ".npy",
with_acc_sum=with_acc_sum,
)
num_samples = len(data)
loader = DataLoader(data, shuffle=True, pin_memory=True,
batch_size=batch_size,
num_workers=1)
batch_idx = 1
total_loss = 0
i = 0
for (x_imu, x_s, y) in loader:
i += x_imu.size()[0]
start = time.time()
loss_func = loss_q_only_2axis
loss_func_c = loss_constr_multi
if args.double:
x_imu = x_imu.double()
x_s = x_s.double()
y = y.double()
if args.cuda:
x_imu = x_imu.cuda()
x_s = x_s.cuda()
y = y.cuda()
# TODO: not sure what's the best value for this parameter
noise_s = (torch.rand(x_s.size()) - 0.5) * (noise_input_hist * 2)
if args.cuda:
noise_s = noise_s.cuda()
y_pred = model(x_imu, x_s + noise_s)
loss_j = loss_jerk(y_pred[:, :, :-3-(n_sbps * 4)])
y_pred = y_pred.reshape(-1, y_pred.size()[-1])
y = y.reshape(-1, y.size()[-1])
loss_q = loss_func(y[:, :-(n_sbps * 4)], y_pred[:, :-(n_sbps * 4)])
loss_c = loss_func_c(y[:, -(n_sbps * 4):], y_pred[:, -(n_sbps * 4):])
loss = loss_c + loss_q
if loss_j is not None:
loss += loss_j
total_loss += loss.item()
optimizer.zero_grad()
loss.backward()
total_norm = None
if args.clip > 0:
total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip)
optimizer.step()
batch_idx += 1
if args.cosine_lr:
lr_s.step()
cur_lr = lr_s.get_last_lr()[0]
else:
cur_lr = lr
end = time.time()
if batch_idx % args.log_interval == 0:
cur_loss = total_loss / args.log_interval
processed = min(i, num_samples)
# for logging
print("total norm", total_norm)
print('Train Epoch: {:2d} [{:6d}/{:6d} ({:.0f}%)]\tLearning rate: {:.7f}\tLoss: {:.6f}\tEp Time: {:.4f}'
.format(epoch, processed, num_samples, 100. * processed / num_samples, cur_lr, cur_loss, end - start),
flush=True)
total_loss = 0
def save(m, ep_num):
if ep_num == 1 or ep_num % 10 == 0:
save_filename = os.path.join(args.save_path, "it" + str(ep_num) + ".pt")
torch.save(m.state_dict(), save_filename)
print('Saved as %s' % save_filename)
torch.save(m.state_dict(), args.save_path + ".pt")
def evaluate(ep_num):
model.eval()
print("Saving...")
save(model, ep_num)
return
try:
os.makedirs(args.save_path)
except FileExistsError:
print("warning: path existed")
except OSError:
exit()
for ep in range(1, epochs + 1):
evaluate(ep)
train(ep)