-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
178 lines (145 loc) · 6.19 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
from mpi4py import MPI
import numpy as np
import json
import os
import subprocess
import sys
import time
import platform
import argparse
from model import Model
from es import make_es
from utils import Seeder, MasterComm, WorkerComm, Logger
from config import tasks
COMM = MPI.COMM_WORLD
SIZE = COMM.Get_size()
RANK = COMM.Get_rank()
class Master(object):
def __init__(self):
task = tasks[args.game_name]
self.model = Model(task)
self.model.make_env()
num_params = self.model.num_params
print('Number of parameters: %d' % num_params)
self.logger = Logger(args)
self.pop_size = args.num_workers * args.num_trails
self.es = make_es(args, self.pop_size, num_params)
self.seeder = Seeder(args.seed)
self.comm = MasterComm(COMM, args, num_params)
def run(self):
start_time = int(time.time())
history, eval_info = [], []
i = 0
while True:
i += 1
solutions = self.es.ask()
seeds = self.get_seeds()
self.comm.distribute_solutions(seeds, solutions)
results = self.comm.gather_results()
rewards, steps = results[:, 0], results[:, 1]
self.es.tell(rewards)
best_params, best_reward, curr_reward, sigma = self.es.result()
self.model.set_model_params(np.array(best_params).round(4))
curr_time = int(time.time()) - start_time
info = (i, curr_time, np.mean(rewards), np.min(rewards),
np.max(rewards), np.std(rewards), np.mean(steps), sigma)
self.logger.write_params(self.es.get_curr_params())
self.logger.log_gen(info)
history.append(info)
self.logger.write_history(history)
if i == 1:
best_reward_eval = np.mean(rewards)
if i % args.eval_interal == 0:
curr_params = np.array(self.es.get_curr_params().round(4))
reward_eval = self.eval_batch(curr_params)
curr_params = curr_params.tolist()
improvement = reward_eval - best_reward_eval
eval_info.append([i, reward_eval, curr_params])
self.logger.write_eval(eval_info)
if len(eval_info) == 1 or reward_eval > best_reward_eval:
best_reward_eval = reward_eval
best_params_eval = curr_params
self.logger.write_best([best_params_eval, best_reward_eval])
eval_log = (i, reward_eval, improvement, best_reward_eval)
self.logger.log_eval(eval_log)
def eval_batch(self, model_params):
solutions = []
for i in range(self.pop_size):
solutions.append(np.copy(model_params))
seeds = np.arange(self.pop_size)
self.comm.distribute_solutions(
seeds, solutions, is_train=False, max_len=-1)
results = self.comm.gather_results()
rewards = results[:, 0]
return np.mean(rewards)
def get_seeds(self):
if args.antithetic:
seeds = self.seeder.next_batch(int(self.pop_size / 2))
seeds = seeds + seeds
else:
seeds = self.seeder.next_batch(self.pop_size)
return seeds
class Worker(object):
def __init__(self):
task = tasks[args.game_name]
self.model = Model(task)
self.model.make_env()
num_params = self.model.num_params
self.pop_size = args.num_workers * args.num_trails
self.comm = WorkerComm(COMM, args, num_params)
def run(self):
while True:
solutions = self.comm.receive_solution()
results = []
for solution in solutions: # multi trails per worker
worker_id, jobidx, seed, is_train, max_len, weights = solution
assert worker_id == RANK, 'worker_id=%d rank=%d' % (worker_id, RANK)
# simulate
self.model.set_model_params(weights)
rewards, timesteps = self.model.simulate(
is_train=is_train, render_mode=False,
num_ep=1, seed=seed, max_len=max_len)
fitness = np.mean(rewards)
timestep = np.mean(timesteps)
results.append([worker_id, jobidx, fitness, timestep])
self.comm.send_results(results)
def main():
if RANK == 0:
print('Master started. %d processes.' % SIZE)
Master().run()
else:
print('Worker-%d started. %d processes.' % (RANK, SIZE))
Worker().run()
def mpi_run():
if os.getenv('IN_MPI') is None:
# fork processes
env = os.environ.copy()
env.update(IN_MPI='1')
mpi_cmd = ['mpirun', '-np', str(args.num_workers + 1)]
script = [sys.executable, '-u'] + sys.argv
if platform.system() == 'Darwin': # for Mac
mpi_cmd.extend(['--hostfile', 'hostfile'])
cmd = mpi_cmd + script
print('RUN: %s' % (' '.join(cmd)))
subprocess.check_call(cmd, env=env)
sys.exit() # admin process exit
else:
main()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('game_name', type=str, help='robo_pendulum, robo_ant, robo_humanoid, etc.')
parser.add_argument('-e', '--es_name', type=str,
help='ses, pepg, openes, ga, cma.', default='cma')
parser.add_argument('--eval_interal', type=int, default=25, help='evaluate every k generations')
parser.add_argument('-n', '--num_workers', type=int, default=8)
parser.add_argument('-t', '--num_trails', type=int, help='trials per worker', default=4)
parser.add_argument('--antithetic', type=bool, default=True,
help='set to 0 to disable antithetic sampling')
parser.add_argument('-s', '--seed', type=int, default=111, help='initial seed')
parser.add_argument('--sigma_init', type=float, default=0.10, help='sigma_init')
parser.add_argument('--sigma_decay', type=float, default=0.999, help='sigma_decay')
parser.add_argument('--log_dir', type=str, default='./log/', help='log directory')
global args
args = parser.parse_args()
assert args.num_workers > 0, 'Number of workers suppose to > 0.'
mpi_run()