-
Notifications
You must be signed in to change notification settings - Fork 15
/
runner.py
294 lines (248 loc) · 16 KB
/
runner.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
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
import numpy as np
import ray
import torch
from alg_parameters import *
from episodic_buffer import EpisodicBuffer
from mapf_gym import MAPFEnv
from model import Model
from od_mstar3 import od_mstar
from od_mstar3.col_set_addition import OutOfTimeError, NoSolutionError
from util import one_step, update_perf, reset_env,set_global_seeds
@ray.remote(num_cpus=1, num_gpus=SetupParameters.NUM_GPU / (TrainingParameters.N_ENVS + 1))
class Runner(object):
"""sub-process used to collect experience"""
def __init__(self, env_id):
"""initialize model0 and environment"""
self.ID = env_id
set_global_seeds(env_id*123)
self.num_agent = EnvParameters.N_AGENTS
self.imitation_num_agent = EnvParameters.N_AGENTS
self.one_episode_perf = {'num_step': 0, 'episode_reward': 0, 'invalid': 0, 'block': 0, 'num_leave_goal': 0,
'wrong_blocking': 0, 'num_collide': 0, 'reward_count': 0, 'ex_reward': 0,
'in_reward': 0}
self.env = MAPFEnv(num_agents=self.num_agent)
self.imitation_env = MAPFEnv(num_agents=self.imitation_num_agent)
self.local_device = torch.device('cuda') if SetupParameters.USE_GPU_LOCAL else torch.device('cpu')
self.local_model = Model(env_id, self.local_device)
self.hidden_state = (
torch.zeros((self.num_agent, NetParameters.NET_SIZE // 2)).to(self.local_device),
torch.zeros((self.num_agent, NetParameters.NET_SIZE // 2)).to(self.local_device))
self.message = torch.zeros((1, self.num_agent, NetParameters.NET_SIZE)).to(self.local_device)
self.done, self.valid_actions, self.obs, self.vector, self.train_valid = reset_env(self.env, self.num_agent)
self.episodic_buffer = EpisodicBuffer(0, self.num_agent)
new_xy = self.env.get_positions()
self.episodic_buffer.batch_add(new_xy)
self.imitation_episodic_buffer = EpisodicBuffer(0, self.imitation_num_agent)
def run(self, weights, total_steps):
"""run multiple steps and collect data for reinforcement learning"""
with torch.no_grad():
mb_obs, mb_vector, mb_rewards_in, mb_rewards_ex, mb_rewards_all, mb_values_in, mb_values_ex, \
mb_values_all, mb_done, mb_ps, mb_actions = [], [], [], [], [], [], [], [], [], [], []
mb_hidden_state = []
mb_message = []
mb_train_valid, mb_blocking = [], []
performance_dict = {'per_r': [], 'per_in_r': [], 'per_ex_r': [], 'per_valid_rate': [],
'per_episode_len': [], 'per_block': [],
'per_leave_goal': [], 'per_final_goals': [], 'per_half_goals': [], 'per_block_acc': [],
'per_max_goals': [], 'per_num_collide': [], 'rewarded_rate': []}
self.local_model.set_weights(weights)
for _ in range(TrainingParameters.N_STEPS):
mb_obs.append(self.obs)
mb_vector.append(self.vector)
mb_hidden_state.append(
[self.hidden_state[0].cpu().detach().numpy(), self.hidden_state[1].cpu().detach().numpy()])
mb_message.append(self.message)
actions, ps, values_in, values_ex, values_all, pre_block, self.hidden_state, num_invalid, self.message = \
self.local_model.step(self.obs, self.vector, self.valid_actions, self.hidden_state,
self.episodic_buffer.no_reward, self.message, self.num_agent)
self.one_episode_perf['invalid'] += num_invalid
mb_values_in.append(values_in)
mb_values_ex.append(values_ex)
mb_values_all.append(values_all)
mb_train_valid.append(self.train_valid)
mb_ps.append(ps)
mb_done.append(self.done)
rewards, self.valid_actions, self.obs, self.vector, self.train_valid, self.done, blockings, \
num_on_goals, self.one_episode_perf, max_on_goals, action_status, modify_actions, on_goal \
= one_step(self.env, self.one_episode_perf, actions, pre_block, self.local_model, values_all,
self.hidden_state, ps, self.episodic_buffer.no_reward, self.message, self.episodic_buffer,
self.num_agent)
new_xy = self.env.get_positions()
processed_rewards, be_rewarded, intrinsic_rewards, min_dist = self.episodic_buffer.if_reward(new_xy,
rewards,
self.done,
on_goal)
self.one_episode_perf['reward_count'] += be_rewarded
self.vector[:, :, 3] = rewards
self.vector[:, :, 4] = intrinsic_rewards
self.vector[:, :, 5] = min_dist
mb_actions.append(modify_actions)
for i in range(self.num_agent):
if action_status[i] == -3:
mb_train_valid[-1][i][int(modify_actions[i])] = 0
mb_rewards_all.append(processed_rewards)
mb_rewards_in.append(intrinsic_rewards)
mb_rewards_ex.append(rewards)
mb_blocking.append(blockings)
self.one_episode_perf['episode_reward'] += np.sum(processed_rewards)
self.one_episode_perf['ex_reward'] += np.sum(rewards)
self.one_episode_perf['in_reward'] += np.sum(intrinsic_rewards)
if self.one_episode_perf['num_step'] == EnvParameters.EPISODE_LEN // 2:
performance_dict['per_half_goals'].append(num_on_goals)
if self.done:
performance_dict = update_perf(self.one_episode_perf, performance_dict, num_on_goals, max_on_goals,
self.num_agent)
self.one_episode_perf = {'num_step': 0, 'episode_reward': 0, 'invalid': 0, 'block': 0,
'num_leave_goal': 0, 'wrong_blocking': 0, 'num_collide': 0,
'reward_count': 0, 'ex_reward': 0, 'in_reward': 0}
self.num_agent = EnvParameters.N_AGENTS
self.done, self.valid_actions, self.obs, self.vector, self.train_valid = reset_env(self.env,
self.num_agent)
self.done = True
self.hidden_state = (
torch.zeros((self.num_agent, NetParameters.NET_SIZE // 2)).to(self.local_device),
torch.zeros((self.num_agent, NetParameters.NET_SIZE // 2)).to(self.local_device))
self.message = torch.zeros((1, self.num_agent, NetParameters.NET_SIZE)).to(self.local_device)
self.episodic_buffer.reset(total_steps, self.num_agent)
new_xy = self.env.get_positions()
self.episodic_buffer.batch_add(new_xy)
mb_obs = np.concatenate(mb_obs, axis=0)
mb_vector = np.concatenate(mb_vector, axis=0)
mb_rewards_in = np.concatenate(mb_rewards_in, axis=0)
mb_rewards_ex = np.concatenate(mb_rewards_ex, axis=0)
mb_rewards_all = np.concatenate(mb_rewards_all, axis=0)
mb_values_in = np.squeeze(np.concatenate(mb_values_in, axis=0), axis=-1)
mb_values_ex = np.squeeze(np.concatenate(mb_values_ex, axis=0), axis=-1)
mb_values_all = np.squeeze(np.concatenate(mb_values_all, axis=0), axis=-1)
mb_actions = np.asarray(mb_actions, dtype=np.int64)
mb_ps = np.stack(mb_ps)
mb_done = np.asarray(mb_done, dtype=np.bool_)
mb_hidden_state = np.stack(mb_hidden_state)
mb_message = np.concatenate(mb_message, axis=0)
mb_train_valid = np.stack(mb_train_valid)
mb_blocking = np.concatenate(mb_blocking, axis=0)
last_values_in, last_values_ex, last_values_all = np.squeeze(
self.local_model.value(self.obs, self.vector, self.hidden_state, self.episodic_buffer.no_reward,
self.message))
# calculate advantages
mb_advs_in = np.zeros_like(mb_rewards_in)
mb_advs_ex = np.zeros_like(mb_rewards_ex)
mb_advs_all = np.zeros_like(mb_rewards_all)
last_gaelam_in = last_gaelam_ex = last_gaelam_all = 0
for t in reversed(range(TrainingParameters.N_STEPS)):
if t == TrainingParameters.N_STEPS - 1:
next_nonterminal = 1.0 - self.done
next_values_in = last_values_in
next_values_ex = last_values_ex
next_values_all = last_values_all
else:
next_nonterminal = 1.0 - mb_done[t + 1]
next_values_in = mb_values_in[t + 1]
next_values_ex = mb_values_ex[t + 1]
next_values_all = mb_values_all[t + 1]
delta_in = np.subtract(np.add(mb_rewards_in[t], TrainingParameters.GAMMA * next_nonterminal *
next_values_in), mb_values_in[t])
delta_ex = np.subtract(np.add(mb_rewards_ex[t], TrainingParameters.GAMMA * next_nonterminal *
next_values_ex), mb_values_ex[t])
delta_all = np.subtract(np.add(mb_rewards_all[t], TrainingParameters.GAMMA * next_nonterminal *
next_values_all), mb_values_all[t])
mb_advs_in[t] = last_gaelam_in = np.add(delta_in,
TrainingParameters.GAMMA * TrainingParameters.LAM
* next_nonterminal * last_gaelam_in)
mb_advs_ex[t] = last_gaelam_ex = np.add(delta_ex,
TrainingParameters.GAMMA * TrainingParameters.LAM
* next_nonterminal * last_gaelam_ex)
mb_advs_all[t] = last_gaelam_all = np.add(delta_all,
TrainingParameters.GAMMA * TrainingParameters.LAM
* next_nonterminal * last_gaelam_all)
mb_returns_in = np.add(mb_advs_in, mb_values_in)
mb_returns_ex = np.add(mb_advs_ex, mb_values_ex)
mb_returns_all = np.add(mb_advs_all, mb_values_all)
return mb_obs, mb_vector, mb_returns_in, mb_returns_ex, mb_returns_all, mb_values_in, mb_values_ex, \
mb_values_all, mb_actions, mb_ps, mb_hidden_state, mb_train_valid, mb_blocking, mb_message, \
len(performance_dict['per_r']), performance_dict
def imitation(self, weights, total_steps):
"""run multiple steps and collect corresponding data for imitation learning"""
with torch.no_grad():
self.local_model.set_weights(weights)
mb_obs, mb_vector, mb_hidden_state, mb_actions = [], [], [], []
mb_message = []
step = 0
episode = 0
self.imitation_num_agent = EnvParameters.N_AGENTS
while step <= TrainingParameters.N_STEPS:
self.imitation_env._reset(num_agents=self.imitation_num_agent)
self.imitation_episodic_buffer.reset(total_steps, self.imitation_num_agent)
new_xy = self.imitation_env.get_positions()
self.imitation_episodic_buffer.batch_add(new_xy)
world = self.imitation_env.get_obstacle_map()
start_positions = tuple(self.imitation_env.get_positions())
goals = tuple(self.imitation_env.get_goals())
try:
obs = None
mstar_path = od_mstar.find_path(world, start_positions, goals, inflation=2, time_limit=5)
obs, vector, actions, hidden_state, message = self.parse_path(mstar_path)
except OutOfTimeError:
print("timeout")
except NoSolutionError:
print("nosol????", start_positions)
if obs is not None: # no error
mb_obs.append(obs)
mb_vector.append(vector)
mb_actions.append(actions)
mb_hidden_state.append(hidden_state)
mb_message.append(message)
step += np.shape(vector)[0]
episode += 1
mb_obs = np.concatenate(mb_obs, axis=0)
mb_vector = np.concatenate(mb_vector, axis=0)
mb_actions = np.concatenate(mb_actions, axis=0)
mb_hidden_state = np.concatenate(mb_hidden_state, axis=0)
mb_message = np.concatenate(mb_message, axis=0)
return mb_obs, mb_vector, mb_actions, mb_hidden_state, mb_message, episode, step
def parse_path(self, path):
"""take the path generated from M* and create the corresponding inputs and actions"""
mb_obs, mb_vector, mb_actions, mb_hidden_state = [], [], [], []
mb_message = []
hidden_state = (
torch.zeros((self.imitation_num_agent, NetParameters.NET_SIZE // 2)).to(self.local_device),
torch.zeros((self.imitation_num_agent, NetParameters.NET_SIZE // 2)).to(self.local_device))
obs = np.zeros((1, self.imitation_num_agent, NetParameters.NUM_CHANNEL, EnvParameters.FOV_SIZE, EnvParameters.FOV_SIZE),
dtype=np.float32)
vector = np.zeros((1, self.imitation_num_agent, NetParameters.VECTOR_LEN), dtype=np.float32)
message = torch.zeros((1, self.imitation_num_agent, NetParameters.NET_SIZE)).to(self.local_device)
for i in range(self.imitation_num_agent):
s = self.imitation_env.observe(i + 1)
obs[:, i, :, :, :] = s[0]
vector[:, i, : 3] = s[1]
for t in range(len(path[:-1])):
mb_obs.append(obs)
mb_vector.append(vector)
mb_hidden_state.append([hidden_state[0].cpu().detach().numpy(), hidden_state[1].cpu().detach().numpy()])
mb_message.append(message)
hidden_state, message = self.local_model.generate_state(obs, vector, hidden_state, message)
actions = np.zeros(self.imitation_num_agent)
for i in range(self.imitation_num_agent):
pos = path[t][i]
new_pos = path[t + 1][i] # guaranteed to be in bounds by loop guard
direction = (new_pos[0] - pos[0], new_pos[1] - pos[1])
actions[i] = self.imitation_env.world.get_action(direction)
mb_actions.append(actions)
obs, vector, rewards, done, _, on_goal, _, valid_actions, _, _, _, _, _, _, _ = \
self.imitation_env.joint_step(actions, 0, model='imitation', pre_value=None, input_state=None,
ps=None, no_reward=None, message=None, episodic_buffer=None)
vector[:, :, -1] = actions
new_xy = self.imitation_env.get_positions()
_, _, intrinsic_reward, min_dist = self.imitation_episodic_buffer.if_reward(new_xy, rewards, done, on_goal)
vector[:, :, 3] = rewards
vector[:, :, 4] = intrinsic_reward
vector[:, :, 5] = min_dist
if not all(valid_actions): # M* can not generate collisions
print('invalid action')
return None, None, None, None
mb_obs = np.concatenate(mb_obs, axis=0)
mb_message = np.concatenate(mb_message, axis=0)
mb_vector = np.concatenate(mb_vector, axis=0)
mb_actions = np.asarray(mb_actions, dtype=np.int64)
mb_hidden_state = np.stack(mb_hidden_state)
return mb_obs, mb_vector, mb_actions, mb_hidden_state, mb_message