forked from ngrupen/multiagent_fairness
-
Notifications
You must be signed in to change notification settings - Fork 1
/
configs.py
197 lines (164 loc) · 8.01 KB
/
configs.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
import torch
from datetime import datetime
import sys
import os
import logging
import time
import shutil
from utils.misc import create_exp_dir
logger = logging.getLogger("logger")
#--------------------------------------
# Configs
#--------------------------------------
class BaseConfig(object):
######## general #########
seed = 612
n_threads = None
######## environment ########
env = 'med_pnp' # environment
gamma = 0.99 # discount factor
discrete = False # discrete action space
######## training #########
opt = 'adam' # optimizer
actor_hidden = [128, 128] # actor hidden units (policy)
critic_hidden = [128, 128, 128] # critic hidden units (value function)
actor_lr = 0.0001 # actor learning rate
critic_lr = 0.001 # critic learning rate
n_epochs = 42000 # number of training epochs
update_steps = 5 # number of steps between policy updates
batch_size = 512 # batch size
buffer_length = 500000 # replay buffer length
tau = 0.01 # polyak averaging factor
clip_norm = 0.5 # gradient clip norm
use_curriculum = True # curriculum learning flag
normalize = False # normalize inputs
norm_obs_var_clip = 1e-6 # threshold to clip obs variance
warmup_episodes = 333 # number of experience episodes before training begins
checkpoint_interval = 2000 # episodes between model checkpoints
num_landmarks = 0
######## testing #########
n_epochs_test = 100 # number of test epochs
######## book-keeping ########
log_interval = 50 # episodes between log updates
checkpoint_path = None # path for loading model checkpoints
scripts = ['main.py', 'configs.py']
def show(self):
attrs = [attr for attr in dir(self) if (not attr.startswith('__') and attr != "show")]
logger.info('\n'.join("%s: %s" % (item, getattr(self, item)) for item in attrs))
class Config_DDPG_Symmetric(BaseConfig):
algorithm = 'ddpg_symmetric' # algorithm name
pred_vel_start = 0.5 # curriculum start value
pred_vel_end = 1.2 # curriculum end value
decay = 5000 # number of episodes over curriculum
pred_test_vel = 0.9 # predator test speed
epsilon_start = 0.95 # epsilon start for e-greedy policy
epsilon_end = 0.05 # epsilon end for e-greedy policy
test_prey = 'cosine' # bot policy to use for prey
test_predator = 'greedy' # bot policy to use for predators
nb_pred = 3
nb_prey = 1
# inherited from other configs
use_sensor_range = True # predators have sensing range
comm_type = 'none' # predators have perfect communication
comm_range = 0.75 # communication range for perfect communication
comm_noise = 0.5 # noise in communication channel
distance_start = 1.5 # curriculum start value
distance_end = 1.5 # curriculum end value
init_range_thresh = 1.0 # percentage predators init outside sensing range
class Config_DDPG_Speed_Fair(BaseConfig):
algorithm = 'ddpg_speed_fair' # algorithm name
pred_vel_start = 1.2 # curriculum start value
pred_vel_end = 0.5 # curriculum end value
decay = 5000 # number of episodes over curriculum
pred_test_vel = 0.9 # predator test speed
epsilon_start = 0.95 # epsilon start for e-greedy policy
epsilon_end = 0.05 # epsilon end for e-greedy policy
test_prey = 'cosine' # bot policy to use for prey
test_predator = 'greedy' # bot policy to use for predators
lambda_coeff = 0.0 # strength of fairness constraint
nb_pred = 3
nb_prey = 1
# inherited from other configs
use_sensor_range = False # predators have sensing range
comm_type = 'none' # predators have perfect communication
comm_range = 0.75 # communication range for perfect communication
comm_noise = 0.5 # noise in communication channel
distance_start = 1.5 # curriculum start value
distance_end = 1.5 # curriculum end value
init_range_thresh = 1.0 # percentage predators init outside sensing range
#--------------------------------------
# Helper functions
#--------------------------------------
def preprocess(args):
config = define_configs(args)
logger = define_logger(args, config.directory)
logger.info("\n"*10)
logger.info("cmd line: python " + " ".join(sys.argv) + "\n"*2)
logger.info("Simulation configurations:\n" + "-"*30)
config.show()
logger.info("\n" * 5)
return config, logger
def define_configs(args):
if args.algorithm == 'ddpg_symmetric':
config = Config_DDPG_Symmetric()
from algorithms import ddpg_symmetric
config.multiagent_fn = ddpg_symmetric.DDPG_Runner
elif args.algorithm == 'ddpg_speed_fair':
config = Config_DDPG_Speed_Fair()
from algorithms import ddpg_speed_fair
config.multiagent_fn = ddpg_speed_fair.DDPG_Runner
else:
raise ValueError("Invalid choice of configuration")
config = read_flags(config, args)
# process seed (same seed for train every time)
# if config.mode == 'test':
# config.seed = 612
# else:
seed_tmp = time.time()
config.seed = int((seed_tmp - int(seed_tmp))*1e6) if args.seed is None else args.seed
print('random seed = {}'.format(config.seed))
# process directory
config.directory = "results/{}_{}/{}".format(config.algorithm,
config.env,
'exp' + datetime.now().strftime("_%m_%d_%Y__%H_%M_%S"))
print('directory = {}'.format(config.directory))
create_exp_dir(config.directory, scripts_to_save=config.scripts)
# process env
envs_to_norm = ['Pendulum-v0']
if config.env in envs_to_norm:
config.normalize_env = True
else:
config.normalize_env = False
particle_envs = ['simple_torus']
if config.env in particle_envs:
config.particle_env = True
else:
config.particle_env = False
comm_envs = []
if config.env in comm_envs:
config.comm_env = True
else:
config.comm_env = False
print("Using torch version: {}".format(torch.__version__))
print('{} GPUs'.format(torch.cuda.device_count()))
return config
def read_flags(config, args):
# assign flags into config
for arg in sorted(vars(args)):
key = arg
val = getattr(args, arg)
if val is not None:
setattr(config, key, val)
return config
def define_logger(args, directory):
logFormatter = logging.Formatter("%(message)s")
logger = logging.getLogger("logger")
logger.setLevel(logging.INFO)
fileHandler = logging.FileHandler("{0}/logger.log".format(directory))
fileHandler.setFormatter(logFormatter)
logger.addHandler(fileHandler)
if args.verbose:
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(logFormatter)
logger.addHandler(consoleHandler)
return logger