-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_env.py
311 lines (245 loc) · 12.1 KB
/
train_env.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
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
import copy
import logging
import random
import cv2
import hydra
import os
import time
import torch
from omegaconf import DictConfig
import logging
from logger import Logger
from models.PPOLSTMAgent import PPOLSTMAgent
from models.RuleBasedAgent import *
from models.CPCAgent import *
from models.utils.RolloutStorage import RolloutStorage
from models.utils.ManagerReplayStorage import ManagerReplayStorage
from recorder import VideoRecorder
from tocenv.env import *
from utils.logging import *
import numpy as np
from utils.svo import svo
from utils.observation import ma_obs_to_numpy
logger = logging.getLogger(os.path.basename(__file__))
class Workspace(object):
def __init__(self, cfg):
self.work_dir = os.getcwd()
print(f'Workspace: {self.work_dir}')
self.cfg = cfg
self.logger = Logger(self.work_dir,
save_tb=cfg.log_save_tb,
log_frequency=cfg.log_frequency,
agent=cfg.ra_agent.name)
self.preferences = list(eval(self.cfg.svo))
prefer = ['green', 'purple', 'blue', 'orange']
self.num_agent = len(prefer)
self.env = TOCEnv(agents=prefer,
map_size=(cfg.env.width, cfg.env.height),
episode_max_length=cfg.env.episode_length,
apple_spawn_ratio=cfg.env.apple_spawn_ratio,
)
self.device = torch.device(cfg.device)
self.env.reset()
cfg.ra_agent.obs_dim = self.env.observation_space.shape
cfg.ra_agent.action_dim = self.env.action_space.n - 1
cfg.ma_agent.obs_dim = (1, 256, 256, 3)
cfg.ma_agent.action_dim = 5
try:
cfg.ra_agent.seq_len = self.env.episode_length
except:
pass
try:
cfg.ma_agent.seq_len = self.env.episode_length
except:
pass
self.ra_agent = hydra.utils.instantiate(cfg.ra_agent)
self.ma_agent = hydra.utils.instantiate(cfg.ma_agent)
if type(self.ra_agent) in [CPCAgentGroup]:
self.ra_replay_buffer = RolloutStorage(agent_type='ac',
num_agent=self.num_agent,
num_step=cfg.env.episode_length,
batch_size=cfg.ra_agent.batch_size,
num_obs=(self.ra_agent.obs_dim[1], self.ra_agent.obs_dim[2], 3),
num_action=7,
num_rec=128)
if type(self.ma_agent) in [CPCAgentGroup]:
self.ma_replay_buffer = RolloutStorage(agent_type='ac',
num_agent=1,
num_step=cfg.env.episode_length,
batch_size=cfg.ra_agent.batch_size,
num_obs=(self.ma_agent.obs_dim[1], self.ma_agent.obs_dim[2], 3),
num_action=5,
num_rec=128)
self.writer = None
self.ra_agent.logger = self.logger
self.ma_agent.logger = self.logger
self.video_recorder = VideoRecorder(self.work_dir if cfg.save_video else None)
# self.video_recorder_blue = VideoRecorder(self.work_dir if cfg.save_video else None)
# self.video_recorder_red = VideoRecorder(self.work_dir if cfg.save_video else None)
self.step = 0
def evaluate(self):
average_episode_reward = 0
average_ma_reward = 0
average_svo_reward = np.array([0] * 4)
self.video_recorder.init(enabled=True)
for episode in range(self.cfg.num_eval_episodes):
obs, _ = self.env.reset()
episode_step = 0
done = False
episode_reward = 0
epi_ma_reward = 0
episode_svo_reward = np.array([0] * 4)
while not done:
if type(self.ra_agent) in [RuleBasedAgent, RuleBasedAgentGroup]:
obs = self.env.get_numeric_observation()
if type(self.ra_agent) is CPCAgentGroup:
action, cpc_info = self.ra_agent.act(self.ra_replay_buffer, obs, episode_step, sample=True)
else:
action = self.ra_agent.act(obs, sample=True)
obs, rewards, dones, env_info = self.env.step(action)
done = True in dones
if episode_step == self.env.episode_length:
done = True
ma_obs = self.env.render(coordination=False)
self.video_recorder.record(self.env)
ma_obs_in = np.expand_dims(ma_obs, axis=0)
if type(self.ma_agent) is CPCAgentGroup:
ma_action, ma_cpc_info = self.ma_agent.act(self.ma_replay_buffer, ma_obs_in, episode_step, sample=True)
else:
ma_action = self.ma_agent.act(ma_obs_in, sample=True)
# MA reward shaping
for i in range(self.num_agent):
episode_svo_reward[i] += svo(rewards, i, self.preferences)
if type(self.ra_agent) in [CPCAgentGroup]:
self.ra_replay_buffer.add(obs, action, rewards, dones, cpc_info)
if episode_step == 0:
ma_reward = np.zeros((1, 1))
else:
ma_reward = np.reshape(env_info['step_eaten_apple'], (1, -1))
epi_ma_reward += ma_reward[0]
if type(self.ma_agent) in [CPCAgentGroup]:
self.ma_replay_buffer.add(ma_obs_in, ma_action[0], ma_reward, dones, ma_cpc_info)
self.env.punish_agent(ma_action[0])
episode_reward += sum(rewards)
episode_step += 1
average_episode_reward += episode_reward
average_ma_reward += epi_ma_reward
average_svo_reward += episode_svo_reward
self.video_recorder.save(f'{self.step}.mp4')
if self.cfg.save_model:
self.ra_agent.save(self.step)
average_episode_reward /= self.cfg.num_eval_episodes
average_ma_reward /= self.cfg.num_eval_episodes
average_svo_reward //= self.cfg.num_eval_episodes
# Clear Buffer
self.ma_replay_buffer.after_update()
self.ra_replay_buffer.after_update()
self.logger.log('eval/episode_reward', average_episode_reward, self.step)
self.logger.log('eval/ma_apple', average_ma_reward, self.step)
for i in range(4):
self.logger.log('eval/agent{}_SVO_reward'.format(i), average_svo_reward[i], self.step)
self.logger.dump(self.step)
def run(self):
episode, episode_step, episode_reward, done = 1, 0, 0, True
start_time = time.time()
env_info = None
''' MA variables '''
prev_ma_obs = None
arr_ma_obs = []
ma_action = 0
epi_ma_reward = 0
while self.step < self.cfg.num_train_steps + 1:
if done or self.step % self.cfg.eval_frequency == 0:
if self.step > 0:
self.logger.log('train/duration', time.time() - start_time, self.step)
start_time = time.time()
self.logger.dump(self.step, save=(self.step > self.cfg.num_seed_steps))
if hasattr(self, 'ra_replay_buffer'):
self.ra_agent.train(self.ra_replay_buffer, self.logger, self.step)
if hasattr(self, 'ma_replay_buffer'):
self.ma_agent.train(self.ma_replay_buffer, self.logger, self.step)
if self.step > 0 and self.step % self.cfg.eval_frequency == 0:
self.logger.log('eval/episode', episode - 1, self.step)
self.evaluate()
start_time = time.time()
self.logger.log('train/episode_reward', episode_reward, self.step)
''' Log Environment Statistics '''
if env_info:
total_apples = 0
for item in env_info['agents']:
total_apples += item['eaten_apples']
log_statistics_to_writer(self.logger, self.step, env_info['statistics'])
log_agent_to_writer(self.logger, self.step, env_info['agents'])
self.logger.log('agent_{0}/train/episode_reward'.format(4), epi_ma_reward, self.step)
self.logger.log('agent_{0}/train/total_apple'.format(4), total_apples, self.step)
self.logger.log('train/episode', episode, self.step)
obs, env_info = self.env.reset()
episode_reward = 0
episode_step = 0
episode += 1
epi_ma_reward = 0
if type(self.ra_agent) in [RuleBasedAgent, RuleBasedAgentGroup]:
obs = self.env.get_numeric_observation()
'''RA actions'''
if self.step < self.cfg.num_seed_steps:
# Define random actions
if type(self.ra_agent) is CPCAgentGroup:
print('RA : ', obs.shape)
action, cpc_info = self.ra_agent.act(self.ra_replay_buffer, obs, episode_step, sample=True)
else:
action = self.ra_agent.act(obs, sample=True)
else:
if type(self.ra_agent) is CPCAgentGroup:
action, cpc_info = self.ra_agent.act(self.ra_replay_buffer, obs, episode_step, sample=True)
else:
action = self.ra_agent.act(obs, sample=True)
#self.env.set_apple_color_ratio(random.random())
next_obs, rewards, dones, env_info = self.env.step(action)
ma_obs = self.env.render(coordination=False)
ma_obs_in = np.expand_dims(ma_obs, axis=0)
'''MA action'''
if self.step < self.cfg.num_seed_steps:
# Define random actions
if type(self.ma_agent) is CPCAgentGroup:
ma_action, ma_cpc_info = self.ma_agent.act(self.ma_replay_buffer, ma_obs_in, episode_step, sample=True)
else:
ma_action = self.ma_agent.act(ma_obs_in, sample=True)
else:
if type(self.ma_agent) is CPCAgentGroup:
ma_action, ma_cpc_info = self.ma_agent.act(self.ma_replay_buffer, ma_obs_in, episode_step, sample=True)
else:
ma_action = self.ma_agent.act(ma_obs_in, sample=True)
if self.cfg.render:
cv2.imshow('TOCEnv', ma_obs)
cv2.waitKey(1)
done = True in dones
if episode_step >= self.env.episode_length:
done = True
episode_reward += sum(rewards)
modified_rewards = np.zeros(self.num_agent)
# Applying prosocial SVO
for i in range(self.num_agent):
modified_rewards[i] = svo(rewards, i, self.preferences)
if type(self.ra_agent) in [CPCAgentGroup]:
self.ra_replay_buffer.add(obs, action, modified_rewards, dones, cpc_info)
# If This is episode's first step, add nothing
if episode_step == 0:
ma_reward = np.zeros((1, 1))
else:
ma_reward = np.reshape(env_info['step_eaten_apple'], (1, -1))
if int(ma_action[0]) > 0:
ma_reward = ma_reward + np.array([[float(self.cfg.ma_beam_reward)]])
if type(self.ma_agent) in [CPCAgentGroup]:
#print('ma', np.sum(ma_obs_in))
self.ma_replay_buffer.add(ma_obs_in, ma_action[0], ma_reward, dones, ma_cpc_info)
self.env.punish_agent(ma_action[0])
epi_ma_reward += ma_reward
obs = next_obs
episode_step += 1
self.step += 1
@hydra.main(config_path="config", config_name="train_env")
def main(cfg: DictConfig) -> None:
workspace = Workspace(cfg)
workspace.run()
if __name__ == '__main__':
main()