-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
188 lines (160 loc) · 7.75 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
182
183
184
185
186
187
188
import os
import sys
import argparse
import time
import datetime
import json
import random
import numpy as np
import torch
from bnp_options import *
from utils import *
from eval import *
from env.room_env import RoomEnv
from env.atari_env import AtariEnv
def get_args():
parser = argparse.ArgumentParser(
description='Trains BNP learning of options')
# Options settings
parser.add_argument('--K', type=int, default=1,
help='initial number of options before the truncation')
parser.add_argument('--tolerance', type=float, default=0.5,
help='threshold below which an option is considered unused')
parser.add_argument('--check-options-usage', type=int, default=10,
help='Number of epochs between checks of options usage to add a new option.')
parser.add_argument('--fixed-options', action="store_true", default=False,
help='prevent any option from being added')
# Networks settings
parser.add_argument('--hidden-layer-sizes-policy', nargs='*', default=[16],
help='number of hidden units per layer in policy network')
parser.add_argument('--hidden-layer-sizes-termination', nargs='*', default=[16],
help='number of hidden units per layer in termination network')
parser.add_argument('--LSTM-hidden-layer-size', type=int, default=32,
help='dimension of LSTM hidden state')
parser.add_argument('--LSTM-MLP-hidden-layer-sizes', nargs='*', default=[32, 32],
help='number of hidden units per layer in MLP after LSTM')
parser.add_argument('--action-space', type=str, default='discrete',
help='discrete or continuous action space')
# Training settings
parser.add_argument('--learning-rate', type=float, default=1e-3,
help='learning rate')
parser.add_argument('--clip', type=float, default=5.,
help='gradient clipping')
parser.add_argument('--batch-size', type=int, default=512,
help='batch_size')
parser.add_argument('--max-epochs', type=int, default=500,
help='number of training epochs')
parser.add_argument('--random-seed', type=int, default=0,
help='Used to seed random number generators')
# Relaxation settings
parser.add_argument('--relaxation-type', type=str, default='GS',
help='GS for Gumbel-Softmax (IGR not implemented)')
parser.add_argument('--temperature', type=float, default=1.,
help='Gumbel-Softmax temperature')
parser.add_argument('--temperature-ratio', type=float, default=0.999,
help='Decay rate of GS temperature')
parser.add_argument('--entropy-factor', type=float, default=5.,
help='Initial entropy factor')
parser.add_argument('--entropy-ratio', type=float, default=0.995,
help='Decay rate of entropy factor')
# Environment settings
parser.add_argument('--env-type', type=str, default='room',
help='type of environment')
parser.add_argument('--nb-rooms', type=int, default=6,
help='number of rooms in the room environment')
parser.add_argument('--nb-traj', type=int, default=1000,
help='number of trajectories in the expert dataset')
parser.add_argument('--noise-level', type=float, default=0.,
help='noise percentage in expert trajectories')
parser.add_argument('--max-steps', type=int, default=6,
help='maximum number of steps in an expert trajectory')
parser.add_argument('--demo-file', type=str, default='',
help='path to the expert trajectories file')
parser.add_argument('--atari-env-name', type=str, default='',
help='name of the atari env')
# Misc settings
parser.add_argument('--save-dir', type=str, default='',
help='directory where model and config are saved')
parser.add_argument('--results-file', type=str, default=None,
help='file where results are saved')
args = parser.parse_args()
return args
def setup_env(args):
n_rooms = args.nb_rooms
max_steps = args.max_steps
nb_traj = args.nb_traj
noise_level = args.noise_level
if args.env_type == 'room':
env = RoomEnv(rng=rng_env, n_rooms=n_rooms, max_steps=max_steps)
data = env.generate_expert_trajectories(n_traj=args.nb_traj, noise_level=args.noise_level,
max_steps=args.max_steps, action_seed=action_seed)
elif args.env_type == 'atari':
if args.atari_env_name == '':
env_name = args.demo_file.split('/')[-2]
env_name = env_name[0].upper() + env_name[1:]
else:
env_name = args.atari_env_name
env = AtariEnv(f'{env_name}-ramNoFrameskip-v4', path=args.demo_file)
data = env.get_expert_trajectories(max_steps=args.max_steps)
else:
raise AssertionError("environment not defined.")
return env, data
def split_train_test(data, rng_split, split=0.01):
n_traj = len(data[0])
perm = rng_split.permutation(n_traj)
test_indices = perm[:int(split*n_traj)]
test_data_states = data[0][test_indices]
test_data_actions = data[1][test_indices]
test_data_rewards = None if data[2] is None else data[2][test_indices]
test_data = (test_data_states, test_data_actions, test_data_rewards)
train_indices = perm[int(split*n_traj):]
train_data_states = data[0][train_indices]
train_data_actions = data[1][train_indices]
train_data_rewards = None if data[2] is None else data[2][train_indices]
train_data = (train_data_states, train_data_actions, train_data_rewards)
return train_data, test_data
if __name__ == "__main__":
args = get_args()
params = vars(args)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Creating folder for this run
os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
start_time_s = time.time()
run_ID = f"{args.env_type}_{datetime.datetime.now().strftime('%b%d_%H-%M-%S')}"
if args.save_dir == '':
run_dir = f"runs/{run_ID}"
else:
run_dir = args.save_dir
os.makedirs(run_dir, exist_ok=True)
with open(os.path.join(run_dir, "config.json"), "w") as f:
f.write(json.dumps(params, indent=4))
# This will be used to generate the seeds for other RNGs.
rng_master = np.random.RandomState(args.random_seed)
np.random.seed(args.random_seed) # there were some issue with reproducibility
random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
env_seed = rng_master.randint(100000)
action_seed = rng_master.randint(100000)
split_seed = rng_master.randint(100000)
rng_env = np.random.RandomState(env_seed)
rng_split = np.random.RandomState(split_seed)
# Environment setup
env, data = setup_env(args)
train_data, test_data = split_train_test(data, rng_split)
# Training
model = BNPOptions(train_data, env.state_dim, env.action_dim, device, rng=rng_master, **vars(args))
model.train()
model.save(os.path.join(run_dir, "checkpoint.pth"))
# Evaluation
score = compute_score(model, env, test_data, device)
print(f"Achieved a score of {score:.3f}.")
if args.results_file is not None:
with open(args.results_file, 'a') as f:
f.write(' '.join(sys.argv))
f.write('\n')
f.write(str(score))
f.write(' ')
f.write(str(model.K))
f.write(' ')
f.write(f"[{' '.join([str(epoch) for epoch in model.new_option_hist])}]")
f.write('\n')