/
train_against.py
252 lines (208 loc) · 8.97 KB
/
train_against.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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
"""
Train a model to against existing benchmark
"""
import argparse
import time
import os
import logging as log
import math
import numpy as np
import magent
from magent.builtin.rule_model import RandomActor
def generate_map(env, map_size, handles):
width = height = map_size
init_num = map_size * map_size * 0.04
gap = 3
leftID, rightID = 0, 1
# add left square of agents
n = init_num
side = int(math.sqrt(n)) * 2
pos = []
for x in range(width//2 - gap - side, width//2 - gap - side + side, 2):
for y in range((height - side)//2, (height - side)//2 + side, 2):
pos.append([x, y, 0])
env.add_agents(handles[leftID], method="custom", pos=pos)
# add right square of agents
n = init_num
side = int(math.sqrt(n)) * 2
pos = []
for x in range(width//2 + gap, width//2 + gap + side, 2):
for y in range((height - side)//2, (height - side)//2 + side, 2):
pos.append([x, y, 0])
env.add_agents(handles[rightID], method="custom", pos=pos)
def play_a_round(env, map_size, handles, models, print_every, eps, step_batch_size=None, train=True,
train_id=1, render=False):
"""play a round of game"""
env.reset()
generate_map(env, map_size, handles)
step_ct = 0
done = False
n = len(handles)
obs = [[] for _ in range(n)]
ids = [[] for _ in range(n)]
acts = [[] for _ in range(n)]
nums = [env.get_num(handle) for handle in handles]
total_reward = [0 for _ in range(n)]
n_transition = 0
pos_reward_num = 0
total_loss, value = 0, 0
print("===== sample =====")
print("eps %s number %s" % (eps, nums))
start_time = time.time()
while not done:
# take actions for every model
for i in range(n):
obs[i] = env.get_observation(handles[i])
ids[i] = env.get_agent_id(handles[i])
# let models infer action in parallel (non-blocking)
models[i].infer_action(obs[i], ids[i], 'e_greedy', eps[i], block=False)
for i in range(n):
acts[i] = models[i].fetch_action() # fetch actions (blocking)
env.set_action(handles[i], acts[i])
# simulate one step
done = env.step()
# sample
step_reward = []
for i in range(n):
rewards = env.get_reward(handles[i])
if train and i == train_id:
alives = env.get_alive(handles[train_id])
# store samples in replay buffer (non-blocking)
models[train_id].sample_step(rewards, alives, block=False)
pos_reward_num += len(rewards[rewards > 0])
s = sum(rewards)
step_reward.append(s)
total_reward[i] += s
# render
if render:
env.render()
# stat info
nums = [env.get_num(handle) for handle in handles]
n_transition += nums[train_id]
# clear dead agents
env.clear_dead()
# check return message of previous called non-blocking function sample_step()
if train:
models[train_id].check_done()
if step_ct % print_every == 0:
print("step %3d, nums: %s reward: %s, total_reward: %s, pos_rewards %d" %
(step_ct, nums, np.around(step_reward, 2), np.around(total_reward, 2),
pos_reward_num))
step_ct += 1
if step_ct > args.n_step:
break
if step_batch_size and n_transition > step_batch_size and train:
total_loss, value = models[train_id].train(500)
n_transition = 0
sample_time = time.time() - start_time
print("steps: %d, total time: %.2f, step average %.2f" % (step_ct, sample_time, sample_time / step_ct))
# train
if train:
print("===== train =====")
start_time = time.time()
total_loss, value = models[train_id].train(500)
train_time = time.time() - start_time
print("train_time %.2f" % train_time)
return magent.round(total_loss), nums, magent.round(total_reward), magent.round(value)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--save_every", type=int, default=5)
parser.add_argument("--render_every", type=int, default=10)
parser.add_argument("--n_round", type=int, default=600)
parser.add_argument("--n_step", type=int, default=550)
parser.add_argument("--render", action="store_true")
parser.add_argument("--load_from", type=int)
parser.add_argument("--train", action="store_true")
parser.add_argument("--map_size", type=int, default=125)
parser.add_argument("--greedy", action="store_true")
parser.add_argument("--name", type=str, default="against")
parser.add_argument("--eval", action="store_true")
parser.add_argument("--opponent", type=int, default=0)
parser.add_argument('--alg', default='dqn', choices=['dqn', 'drqn', 'a2c'])
args = parser.parse_args()
# download opponent model
magent.utility.check_model('against')
# set logger
magent.utility.init_logger(args.name)
# init the game
env = magent.GridWorld("battle", map_size=args.map_size)
env.set_render_dir("build/render")
# two groups of agents
handles = env.get_handles()
# sample eval observation set
if args.eval:
print("sample eval set...")
env.reset()
generate_map(env, args.map_size, handles)
eval_obs = magent.utility.sample_observation(env, handles, n_obs=2048, step=500)
else:
eval_obs = [None, None]
# init models
names = [args.name + "-a", "battle"]
batch_size = 512
unroll_step = 16
train_freq = 5
models = []
# load opponent
if args.opponent >= 0:
from magent.builtin.tf_model import DeepQNetwork
models.append(magent.ProcessingModel(env, handles[1], names[1], 20000, 0, DeepQNetwork))
models[0].load("data/battle_model", args.opponent)
else:
models.append(magent.ProcessingModel(env, handles[1], names[1], 20000, 0, RandomActor))
# load our model
if args.alg == 'dqn':
from magent.builtin.tf_model import DeepQNetwork
models.append(magent.ProcessingModel(env, handles[0], names[0], 20001, 1000, DeepQNetwork,
batch_size=batch_size,
learning_rate=3e-4,
memory_size=2 ** 20, train_freq=train_freq, eval_obs=eval_obs[0]))
step_batch_size = None
elif args.alg == 'drqn':
from magent.builtin.tf_model import DeepRecurrentQNetwork
models.append(magent.ProcessingModel(env, handles[0], names[0], 20001, 1000, DeepRecurrentQNetwork,
batch_size=batch_size/unroll_step, unroll_step=unroll_step,
learning_rate=3e-4,
memory_size=4 * 625, train_freq=train_freq, eval_obs=eval_obs[0]))
step_batch_size = None
elif args.alg == 'a2c':
from magent.builtin.mx_model import AdvantageActorCritic
step_batch_size = 10 * args.map_size * args.map_size * 0.04
models.append(magent.ProcessingModel(env, handles[0], names[0], 20001, 1000, AdvantageActorCritic,
learning_rate=1e-3))
# load if
savedir = 'save_model'
if args.load_from is not None:
start_from = args.load_from
print("load ... %d" % start_from)
models[0].load(savedir, start_from)
else:
start_from = 0
# print debug info
print(args)
print("view_size", env.get_view_space(handles[0]))
print("feature_size", env.get_feature_space(handles[0]))
# play
start = time.time()
for k in range(start_from, start_from + args.n_round):
tic = time.time()
start = 1 if args.opponent != -1 else 0.1
train_eps = magent.utility.piecewise_decay(k, [0, 100, 250], [start, 0.1, 0.05]) if not args.greedy else 0
opponent_eps = train_eps if k < 100 else 0.05 # can use curriculum learning in first 100 steps
loss, num, reward, value = play_a_round(env, args.map_size, handles, models,
eps=[opponent_eps, train_eps], step_batch_size=step_batch_size,
train=args.train,
print_every=50,
render=args.render or (k+1) % args.render_every == 0) # for e-greedy
log.info("round %d\t loss: %s\t num: %s\t reward: %s\t value: %s" % (k, loss, num, reward, value))
print("round time %.2f total time %.2f\n" % (time.time() - tic, time.time() - start))
# save models
if (k + 1) % args.save_every == 0 and args.train:
print("save model... ")
if not os.path.exists(savedir):
os.mkdir(savedir)
for model in models:
model.save(savedir, k)
# close model processing
for model in models:
model.quit()