-
Notifications
You must be signed in to change notification settings - Fork 6
/
train.py
311 lines (272 loc) · 11.3 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
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 argparse
import yaml
import numpy as np
import gym
from datetime import datetime
from os.path import join, dirname, abspath, exists
import sys
import os
import shutil
import logging
import collections
import time
import uuid
from pprint import pformat
import torch
try:
import GPUtil
from tensorboardX import SummaryWriter
except:
pass
from envs import registration
from envs.wrappers import StackFrame
from rl_algos import algo_class
from rl_algos.net import *
from rl_algos.base_rl_algo import ReplayBuffer
from rl_algos.sac import GaussianActor
from rl_algos.td3 import Actor, Critic #, TD3, ReplayBuffer
from rl_algos.model_based import Model
# from rl_algos.safe_td3 import SafeTD3
from rl_algos.collector import ContainerCollector, LocalCollector, ClusterCollector
def initialize_config(config_path, save_path):
# Load the config files
with open(config_path, 'r') as f:
config = yaml.load(f, Loader=yaml.FullLoader)
config["env_config"]["save_path"] = save_path
config["env_config"]["config_path"] = config_path
return config
def initialize_logging(config):
env_config = config["env_config"]
training_config = config["training_config"]
# Config logging
now = datetime.now()
string = now.strftime("%Y_%m_%d_%H_%M")
save_path = join(
env_config["save_path"],
env_config["env_id"],
training_config['algorithm'],
string,
uuid.uuid4().hex[:4]
)
print(" >>>> Saving to %s" % save_path)
if not exists(save_path):
os.makedirs(save_path)
writer = SummaryWriter(save_path)
shutil.copyfile(
env_config["config_path"],
join(save_path, "config.yaml")
)
return save_path, writer
def initialize_envs(config):
env_config = config["env_config"]
if env_config["collector"] != "local":
env_config["kwargs"]["init_sim"] = False
env = gym.make(env_config["env_id"], **env_config["kwargs"])
env = StackFrame(env, stack_frame=env_config["stack_frame"])
return env
def seed(config):
env_config = config["env_config"]
np.random.seed(env_config['seed'])
torch.manual_seed(env_config['seed'])
def get_encoder(encoder_type, args):
if encoder_type == "mlp":
encoder=MLPEncoder(**args)
elif encoder_type == 'rnn':
encoder=RNNEncoder(**args)
elif encoder_type == 'cnn':
encoder=CNNEncoder(**args)
elif encoder_type == 'transformer':
encoder=TransformerEncoder(**args)
else:
raise Exception(f"[error] Unknown encoder type {encoder_type}!")
return encoder
def initialize_policy(config, env, init_buffer=True, device=None):
training_config = config["training_config"]
state_dim = env.observation_space.shape
action_dim = np.prod(env.action_space.shape)
action_space_low = env.action_space.low
action_space_high = env.action_space.high
# find available device
if device is None:
devices = GPUtil.getAvailable(order = 'first', limit = 1, maxLoad = 0.8, maxMemory = 0.8, includeNan=False, excludeID=[], excludeUUID=[])
device = "cuda:%d" %(devices[0]) if len(devices) > 0 else "cpu"
print(" >>>> Running on device %s" %(device))
encoder_type = training_config["encoder"]
encoder_args = {
'input_dim': state_dim[-1], # np.prod(state_dim),
'num_layers': training_config['encoder_num_layers'],
'hidden_size': training_config['encoder_hidden_layer_size'],
'history_length': config["env_config"]["stack_frame"],
}
# initialize actor
input_dim = training_config['hidden_layer_size']
actor_class = GaussianActor if "SAC" in training_config["algorithm"] else Actor
actor = actor_class(
encoder=get_encoder(encoder_type, encoder_args),
head=MLP(input_dim, training_config['encoder_num_layers'], training_config['encoder_hidden_layer_size']),
action_dim=action_dim
).to(device)
actor_optim = torch.optim.Adam(
actor.parameters(),
lr=training_config['actor_lr']
)
# print("Total number of parameters: %d" %sum(p.numel() for p in actor.parameters()))
# initialize critic
input_dim += np.prod(action_dim)
critic = Critic(
encoder=get_encoder(encoder_type, encoder_args),
head=MLP(input_dim, training_config['encoder_num_layers'], training_config['encoder_hidden_layer_size']),
).to(device)
critic_optim = torch.optim.Adam(
critic.parameters(),
lr=training_config['critic_lr']
)
# initialize agents
algo = training_config["algorithm"]
if "Dyna" in algo or "SMCP" in algo or "MBPO" in algo:
model = Model(
encoder=get_encoder(encoder_type, encoder_args),
head=MLP(input_dim, training_config['encoder_num_layers'], training_config['encoder_hidden_layer_size']),
state_dim=state_dim,
deterministic=training_config['deterministic']
).to(device)
model_optim = torch.optim.Adam(
model.parameters(),
lr=training_config['model_lr']
)
policy = algo_class[algo](
model, model_optim,
actor, actor_optim,
critic, critic_optim,
action_range=[action_space_low, action_space_high],
device=device,
**training_config["policy_args"]
)
elif "Safe" in algo:
safe_critic = Critic(
encoder=get_encoder(encoder_type, encoder_args),
head=MLP(input_dim, training_config['encoder_num_layers'], training_config['encoder_hidden_layer_size']),
).to(device)
safe_critic_optim = torch.optim.Adam(
safe_critic.parameters(),
lr=training_config['critic_lr']
)
policy = algo_class[algo](
safe_critic, safe_critic_optim,
actor, actor_optim,
critic, critic_optim,
action_range=[action_space_low, action_space_high],
device=device,
**training_config["policy_args"]
)
else:
policy = algo_class[algo](
actor, actor_optim,
critic, critic_optim,
action_range=[action_space_low, action_space_high],
device=device,
**training_config["policy_args"]
)
if init_buffer:
replay_buffer = ReplayBuffer(
state_dim, action_dim, training_config['buffer_size'],
device=device,
reward_norm=False # config['training_config']["reward_norm"]
)
else:
replay_buffer = None
return policy, replay_buffer
def train(env, policy, replay_buffer, config):
env_config = config["env_config"]
training_config = config["training_config"]
save_path, writer = initialize_logging(config)
print(" >>>> initialized logging")
if env_config['collector'] == 'container':
collector = ContainerCollector(policy, env, replay_buffer, config)
elif env_config['collector'] == 'local':
collector = LocalCollector(policy, env, replay_buffer)
elif env_config['collector'] == 'cluster':
collector = ClusterCollector(policy, env, replay_buffer, config)
else:
raise ValueError
training_args = training_config["training_args"]
print(" >>>> Pre-collect experience")
collector.collect(n_steps=training_config['pre_collect'])
print(" >>>> Start training")
n_steps = 0
n_iter = 0
n_ep = 0
epinfo_buf = collections.deque(maxlen=300)
world_ep_buf = collections.defaultdict(lambda: collections.deque(maxlen=20))
t0 = time.time()
while n_steps < training_args["max_step"]:
# Linear decaying exploration noise from "start" -> "end"
if "TD3" in training_config["algorithm"]:
policy.exploration_noise = \
- (training_config["exploration_noise_start"] - training_config["exploration_noise_end"]) \
* n_steps / training_args["max_step"] + training_config["exploration_noise_start"]
steps, epinfo = collector.collect(n_steps=training_args["collect_per_step"])
n_steps += steps
n_iter += 1
n_ep += len(epinfo)
epinfo_buf.extend(epinfo)
for d in epinfo:
world = d["world"].split("/")[-1]
world_ep_buf[world].append(d)
loss_infos = []
for _ in range(training_args["update_per_step"]):
loss_info = policy.train(replay_buffer, training_args["batch_size"])
loss_infos.append(loss_info)
loss_info = {}
for k in loss_infos[0].keys():
loss_info[k] = np.mean([li[k] for li in loss_infos if li[k] is not None])
t1 = time.time()
log = {
"Episode_return": np.mean([epinfo["ep_rew"] for epinfo in epinfo_buf]),
"Episode_length": np.mean([epinfo["ep_len"] for epinfo in epinfo_buf]),
"Success": np.mean([epinfo["success"] for epinfo in epinfo_buf]),
"Time": np.mean([epinfo["ep_time"] for epinfo in epinfo_buf]),
"Collision": np.mean([epinfo["collision"] for epinfo in epinfo_buf]),
"fps": n_steps / (t1 - t0),
"n_episode": n_ep,
"Steps": n_steps
}
if "TD" in training_config["algorithm"] or "DDPG" in training_config["algorithm"]:
log.update({
"Exploration_noise": policy.exploration_noise,
})
if "SAC" in training_config["algorithm"]:
log.update({
"Alpha": policy.alpha,
})
log.update(loss_info)
print(pformat(log))
if n_iter % training_config["log_intervals"] == 0:
for k in log.keys():
writer.add_scalar('train/' + k, log[k], global_step=n_steps)
policy.save(save_path, "last_policy")
print("Logging to %s" %save_path)
for k in world_ep_buf.keys():
writer.add_scalar(k + "/Episode_return", np.mean([epinfo["ep_rew"] for epinfo in world_ep_buf[k]]), global_step=n_steps)
writer.add_scalar(k + "/Episode_length", np.mean([epinfo["ep_len"] for epinfo in world_ep_buf[k]]), global_step=n_steps)
writer.add_scalar(k + "/Success", np.mean([epinfo["success"] for epinfo in world_ep_buf[k]]), global_step=n_steps)
writer.add_scalar(k + "/Time", np.mean([epinfo["ep_time"] for epinfo in world_ep_buf[k]]), global_step=n_steps)
writer.add_scalar(k + "/Collision", np.mean([epinfo["collision"] for epinfo in world_ep_buf[k]]), global_step=n_steps)
if __name__ == "__main__":
torch.set_num_threads(8)
parser = argparse.ArgumentParser(description = 'Start training')
parser.add_argument('--config_path', dest='config_path', default="configs/e2e_default.yaml")
parser.add_argument('--device', dest='device', default=None)
logging.getLogger().setLevel("INFO")
args = parser.parse_args()
CONFIG_PATH = args.config_path
SAVE_PATH = "logging/"
print(">>>>>>>> Loading the configuration from %s" % CONFIG_PATH)
config = initialize_config(CONFIG_PATH, SAVE_PATH)
seed(config)
print(">>>>>>>> Creating the environments")
env = initialize_envs(config)
print(">>>>>>>> Initializing the policy")
policy, replay_buffer = initialize_policy(config, env, device=args.device)
print(">>>>>>>> Start training")
train(env, policy, replay_buffer, config)