-
Notifications
You must be signed in to change notification settings - Fork 0
/
MPI_ES.py
167 lines (114 loc) · 4.06 KB
/
MPI_ES.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
import pickle
import os
import time
import gym
import cma
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
from mpi4py import MPI
from World_Model import World_Model
from Env_Runner import Env_Runner
comm = MPI.COMM_WORLD
rank = comm.Get_rank()
device = torch.device('cpu')
# hard coded
hidden_size = 256
actions = 3
latent_size = 32
dirname = os.path.dirname(os.path.abspath(__file__))
save_path_controller = dirname +"\\"+"es_controller"
# es algorithm
eps=250
pop_size=16
num_rollouts=18
def master():
es = cma.CMAEvolutionStrategy((((hidden_size+latent_size)*actions)+actions)*[0], 0.125,{'popsize': pop_size})
num_slaves = comm.Get_size()-1
f = open("log.txt", "w")
f.close()
f = open("log_pop_performance.csv", "w")
f.close()
os.mkdir(save_path_controller)
for ep in range(eps):
solutions = es.ask()
sol = []
fitness = []
slave_jobs = pop_size//num_slaves
start_time = time.time()
for slave in range(1, num_slaves+1):
packet = solutions[(slave-1)*slave_jobs:((slave-1)*slave_jobs)+slave_jobs]
packet = ("fitness", packet)
comm.send(packet, dest=slave)
for slave in range(1, num_slaves+1):
slave_result = comm.recv(source=slave)
for i in range(len(slave_result)):
sol.append(slave_result[i][0])
fitness.append(slave_result[i][1])
es.tell(sol,fitness)
end_time = time.time()
best_reward = es.result[1]
curr_reward = es.result[2]
fitness = - np.array(fitness)
fit_max = np.amax(fitness)
fit_mean = np.mean(fitness)
fit_min = np.amin(fitness)
f = open("log_pop_performance.csv", "a+")
#min, avg, max
f.write(f"{fit_min},{fit_mean},{fit_max}\n")
f.close()
f = open("log.txt", "a+")
f.write("***********************\n")
f.write(f"ep: {ep} finished | time: {end_time - start_time}\n")
f.write(f"best reward: {best_reward}\n")
f.write(f"best reward batch: {curr_reward}\n")
f.close()
weights = es.result[0]
f = open(save_path_controller + f"\\network_{ep}.pt","wb")
pickle.dump(weights,f)
f.close()
for slave in range(1, num_slaves+1):
packet = ("done", None)
comm.send(packet, dest=slave)
def slave():
env = gym.make("CarRacing-v0")
while True:
msg, solutions = comm.recv(source=0)
if msg == "done":
return
fitness = worker(solutions, env)
packet = []
for i in range(len(solutions)):
packet.append([solutions[i],fitness[i]])
comm.send(packet, dest=0)
def worker(solutions, env):
fitness_solutions = []
if not isinstance(solutions, list):
solutions = [solutions]
for weights in solutions:
wm = World_Model(dirname + "\\vae.pt",
dirname + "\\mdn_rnn.pt",
actions,
device)
w = weights[0:actions*(hidden_size+latent_size)]
b = weights[actions*(hidden_size+latent_size)::]
w = nn.Parameter(torch.tensor(np.reshape(w,(actions,hidden_size+latent_size))).type('torch.FloatTensor').to(device))
b = nn.Parameter(torch.tensor(b).type('torch.FloatTensor').to(device))
wm.set_controller(w,b)
fitness = []
for i in range(num_rollouts):
runner = Env_Runner(device)
wm.reset_rnn()
_, _, rewards = runner.run(env, wm, img_resize=(64,64))
# append negative return, because ES will try to minimize it
fitness.append(-np.sum(np.array(rewards)))
env.close()
fitness_solutions.append(np.mean(np.array(fitness)))
return fitness_solutions
if __name__ == "__main__":
if rank == 0:
master()
else:
slave()