-
Notifications
You must be signed in to change notification settings - Fork 14
/
trainer.py
207 lines (187 loc) · 5.95 KB
/
trainer.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
from oil.datasetup.datasets import split_dataset
import torch
from torch.utils.data import DataLoader
from torch.optim import Adam
from oil.utils.utils import LoaderTo, FixedNumpySeed, cosLr
from biases.datasets import RigidBodyDataset
from biases.dynamics_trainer import IntegratedDynamicsTrainer
from biases.models.constrained_hnn import CHNN, CHLC
from biases.models.hnn import HNN
from biases.models.lnn import LNN
from biases.models.nn import NN, DeltaNN
from biases.systems.chain_pendulum import ChainPendulum
from typing import Union, Tuple
import sys
import argparse
import numpy as np
import csv
import os
def str_to_class(classname):
return getattr(sys.modules[__name__], classname)
def make_trainer(
chunk_len: int,
angular: Union[Tuple, bool],
body,
bs: int,
dataset,
dt: float,
lr: float,
n_train: int,
n_val: int,
n_test: int,
net_cfg: dict,
network,
num_epochs: int,
regen: bool,
seed: int = 0,
device=torch.device("cuda"),
dtype=torch.float32,
trainer_config={},
):
# Create Training set and model
splits = {"train": n_train, "val": n_val, "test": n_test}
dataset = dataset(
n_systems=n_train + n_val + n_test,
regen=regen,
chunk_len=chunk_len,
body=body,
dt=dt,
integration_time=10,
angular_coords=angular,
)
# dataset=CartpoleDataset(batch_size=500,regen=regen)
with FixedNumpySeed(seed):
datasets = split_dataset(dataset, splits)
model = network(G=dataset.body.body_graph, **net_cfg).to(device=device, dtype=dtype)
# Create train and Dev(Test) dataloaders and move elems to gpu
dataloaders = {
k: LoaderTo(
DataLoader(
v, batch_size=min(bs, splits[k]), num_workers=0, shuffle=(k == "train")
),
device=device,
dtype=dtype,
)
for k, v in datasets.items()
}
dataloaders["Train"] = dataloaders["train"]
# Initialize optimizer and learning rate schedule
opt_constr = lambda params: Adam(params, lr=lr)
lr_sched = cosLr(num_epochs)
return IntegratedDynamicsTrainer(
model,
dataloaders,
opt_constr,
lr_sched,
log_args={"timeFrac": 1 / 4, "minPeriod": 0.0},
**trainer_config
)
def parse_cmdline():
parser = argparse.ArgumentParser()
parser.add_argument("--batch-size", type=int, default=200, help="Batch size")
parser.add_argument(
"--chunk-len",
type=int,
default=5,
help="Length of each chunk of training trajectory",
)
parser.add_argument(
"--exp-dir",
type=str,
default="./",
help="Directory to save files from this experiment",
)
parser.add_argument(
"--hidden_size", type=int, default=200, help="Number of hidden units"
)
parser.add_argument("--lr", type=float, default=1e-3, help="Learning rate")
parser.add_argument(
"--n-test", type=int, default=100, help="Number of test trajectories"
)
parser.add_argument(
"--n-train", type=int, default=800, help="Number of train trajectories"
)
parser.add_argument(
"--n-val", type=int, default=100, help="Number of validation trajectories"
)
parser.add_argument(
"--network",
type=str,
help="Dynamics network",
choices=["NN", "DeltaNN", "HNN", "LNN", "CHNN", "CHLC"],
)
parser.add_argument(
"--num-epochs", type=int, default=300, help="Number of training epochs"
)
parser.add_argument(
"--num-layers", type=int, default=3, help="Number of hidden layers"
)
parser.add_argument(
"--num-masses", type=int, default=1, help="Number of masses in ChainPendulum",
)
parser.add_argument(
"--regen",
action="store_true",
default=False,
help="Forcibly regenerate training data",
)
return parser.parse_args()
if __name__ == "__main__":
args = parse_cmdline()
body = ChainPendulum(args.num_masses)
euclidean_coords = args.network not in [
"NN",
"LNN",
"HNN",
] # TODO: try NN in euclidean
dof_ndim = body.d if euclidean_coords else body.D
angular_dims = body.angular_dims
net_cfg = {
"dof_ndim": dof_ndim,
"angular_dims": angular_dims,
"hidden_size": args.hidden_size,
"num_layers": args.num_layers,
"wgrad": True,
}
trainer = make_trainer(
angular=not euclidean_coords,
body=body,
bs=args.batch_size,
chunk_len=args.chunk_len,
dataset=RigidBodyDataset,
dt=0.1,
lr=args.lr,
network=str_to_class(args.network),
n_test=args.n_test,
n_train=args.n_train,
n_val=args.n_val,
net_cfg=net_cfg,
num_epochs=args.num_epochs,
regen=args.regen,
)
# Create target directory & all intermediate directories if don't exists
if not os.path.exists(args.exp_dir):
os.makedirs(args.exp_dir)
print("Directory ", args.exp_dir, " Created ")
else:
print("Directory ", args.exp_dir, " already exists")
with open(args.exp_dir + "/args.csv", "w") as csvfile:
args_dict = vars(args) # convert to dict
writer = csv.DictWriter(csvfile, fieldnames=args_dict.keys())
writer.writeheader()
writer.writerow(args_dict)
trainer.train(args.num_epochs)
print("Saving training logs")
ax = trainer.logger.scalar_frame.plot()
ax.set(yscale="log")
figure_path = args.exp_dir + "/log.png"
ax.figure.savefig(figure_path)
print("Saving test rollouts")
rollouts_path = args.exp_dir + "/test_rollouts"
rollout_errs = trainer.test_rollouts(
angular_to_euclidean=not euclidean_coords, pert_eps=1e-4
)
np.save(rollouts_path, rollout_errs.detach().cpu().numpy())
print("Saving model state_dict")
model_path = args.exp_dir + "/model.pt"
torch.save(trainer.model.to("cpu").state_dict(), model_path)