-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
289 lines (241 loc) · 10.3 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
import argparse
import os
import random
import sys
import time
import yaml
import numpy as np
import torch
from baselines import logger
from dcpg.algos import *
from dcpg.envs import make_envs
from dcpg.models import *
from dcpg.sample_utils import sample_episodes
from dcpg.storages import RolloutStorage
from test import evaluate
def main(config):
# Fix random seed
random.seed(config["seed"])
np.random.seed(config["seed"])
torch.manual_seed(config["seed"])
torch.cuda.manual_seed_all(config["seed"])
# CUDA setting
torch.set_num_threads(1)
cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if cuda else "cpu")
# Create directories
os.makedirs(config["log_dir"], exist_ok=True)
if not config["debug"]:
os.makedirs(config["output_dir"], exist_ok=True)
os.makedirs(config["save_dir"], exist_ok=True)
# Create logger
log_file = "-{}-{}-s{}".format(
config["env_name"], config["exp_name"], config["seed"]
)
if config["debug"]:
log_file += "-debug"
logger.configure(
dir=config["log_dir"], format_strs=["csv", "stdout"], log_suffix=log_file
)
print("\nLog File:", log_file)
# Create environments
envs = make_envs(
num_envs=config["num_processes"],
env_name=config["env_name"],
num_levels=config["num_levels"],
start_level=config["start_level"],
distribution_mode=config["distribution_mode"],
normalize_reward=config["normalize_reward"],
device=device,
)
obs_space = envs.observation_space
action_space = envs.action_space
# Create actor-critic
actor_critic_class = getattr(sys.modules[__name__], config["actor_critic_class"])
actor_critic_params = config["actor_critic_params"]
actor_critic = actor_critic_class(
obs_space.shape, action_space.n, **actor_critic_params
)
actor_critic.to(device)
print("\nActor-Critic Network:", actor_critic)
# Create rollout storage
rollouts = RolloutStorage(
config["num_steps"], config["num_processes"], obs_space.shape, action_space
)
rollouts.to(device)
# Create agent
agent_class = getattr(sys.modules[__name__], config["agent_class"])
agent_params = config["agent_params"]
agent = agent_class(actor_critic, **agent_params, device=device)
# Initialize environments
obs = envs.reset()
*_, infos = envs.step_wait()
levels = torch.LongTensor([info["level_seed"] for info in infos])
rollouts.obs[0].copy_(obs)
rollouts.levels[0].copy_(levels)
# Train actor-critic
num_env_steps_epoch = config["num_steps"] * config["num_processes"]
num_updates = int(config["num_env_steps"]) // num_env_steps_epoch
elapsed_time = 0
for j in range(num_updates):
# Start training
start = time.time()
# Set actor-critic to train mode
actor_critic.train()
# Sample episode
sample_episodes(envs, rollouts, actor_critic)
# Compute return
with torch.no_grad():
next_critic_outputs = actor_critic.forward_critic(rollouts.obs[-1])
next_value = next_critic_outputs["value"]
rollouts.compute_returns(next_value, config["gamma"], config["gae_lambda"])
rollouts.compute_advantages()
# Update actor-critic
train_statistics = agent.update(rollouts)
# Reset rollout storage
rollouts.after_update()
# End training
end = time.time()
elapsed_time += end - start
# Statistics
if j % config["log_interval"] == 0:
# Train statistics
total_num_steps = (j + 1) * config["num_processes"] * config["num_steps"]
time_per_epoch = elapsed_time / (j + 1)
print(
"\nUpdate {}, step {}, time per epoch {:.2f} \n".format(
j, total_num_steps, time_per_epoch
)
)
logger.logkv("train/total_num_steps", total_num_steps)
logger.logkv("train/time_per_epoch", time_per_epoch)
for key, val in train_statistics.items():
logger.logkv("train/{}".format(key), val)
# Fetch reward normalizing variables
norm_infos = envs.normalization_infos()
# Evaluate actor-critic on train environments
train_eval_statistics, train_value_statistics = evaluate(
config, actor_critic, device, test_envs=False, norm_infos=norm_infos
)
train_episode_rewards = train_eval_statistics["episode_rewards"]
train_episode_steps = train_eval_statistics["episode_steps"]
print(
"Last {} training episodes: \n"
"mean/med/std reward {:.2f}/{:.2f}/{:.2f}, "
"mean/med/std step {:.2f}/{:.2f}/{:.2f} \n".format(
len(train_episode_rewards),
np.mean(train_episode_rewards),
np.median(train_episode_rewards),
np.std(train_episode_rewards),
np.mean(train_episode_steps),
np.median(train_episode_steps),
np.std(train_episode_steps),
)
)
logger.logkv("train/mean_episode_reward", np.mean(train_episode_rewards))
logger.logkv("train/med_episode_reward", np.median(train_episode_rewards))
logger.logkv("train/std_episode_reward", np.std(train_episode_rewards))
logger.logkv("train/mean_episode_step", np.mean(train_episode_steps))
logger.logkv("train/med_episode_step", np.median(train_episode_steps))
logger.logkv("train/std_episode_step", np.std(train_episode_steps))
for key, val in train_value_statistics.items():
logger.logkv("train/{}".format(key), val)
# Evaluate actor-critic on test environments
test_eval_statistics, *_ = evaluate(
config, actor_critic, device, test_envs=True
)
test_episode_rewards = test_eval_statistics["episode_rewards"]
test_episode_steps = test_eval_statistics["episode_steps"]
print(
"Last {} test episodes: \n"
"mean/med/std reward {:.2f}/{:.2f}/{:.2f}, "
"mean/med/std step {:.2f}/{:.2f}/{:.2f} \n".format(
len(test_episode_rewards),
np.mean(test_episode_rewards),
np.median(test_episode_rewards),
np.std(test_episode_rewards),
np.mean(test_episode_steps),
np.median(test_episode_steps),
np.std(test_episode_steps),
)
)
logger.logkv("test/mean_episode_reward", np.mean(test_episode_rewards))
logger.logkv("test/med_episode_reward", np.median(test_episode_rewards))
logger.logkv("test/std_episode_reward", np.std(test_episode_rewards))
logger.logkv("test/mean_episode_step", np.mean(test_episode_steps))
logger.logkv("test/med_episode_step", np.median(test_episode_steps))
logger.logkv("test/std_episode_step", np.std(test_episode_steps))
logger.dumpkvs()
if j == num_updates - 1 and not config["debug"]:
print("\nFinal evaluation \n")
# Evaluate actor-critic on train environments
train_eval_statistics, *_ = evaluate(
config, actor_critic, device, test_envs=False
)
train_episode_rewards = train_eval_statistics["episode_rewards"]
train_episode_steps = train_eval_statistics["episode_steps"]
print(
"Last {} train episodes: \n"
"mean/med/std reward {:.2f}/{:.2f}/{:.2f}, "
"mean/med/std step {:.2f}/{:.2f}/{:.2f} \n".format(
len(train_episode_rewards),
np.mean(train_episode_rewards),
np.median(train_episode_rewards),
np.std(train_episode_rewards),
np.mean(train_episode_steps),
np.median(train_episode_steps),
np.std(train_episode_steps),
)
)
# Save train scores
np.save(
os.path.join(config["output_dir"], "scores-train{}.npy".format(log_file)),
np.array(train_episode_rewards),
)
# Evaluate actor-critic on test environments
test_eval_statistics, *_ = evaluate(
config, actor_critic, device, test_envs=True
)
test_episode_rewards = test_eval_statistics["episode_rewards"]
test_episode_steps = test_eval_statistics["episode_steps"]
print(
"Last {} test episodes: \n"
"mean/med/std reward {:.2f}/{:.2f}/{:.2f}, "
"mean/med/std step {:.2f}/{:.2f}/{:.2f} \n".format(
len(test_episode_rewards),
np.mean(test_episode_rewards),
np.median(test_episode_rewards),
np.std(test_episode_rewards),
np.mean(test_episode_steps),
np.median(test_episode_steps),
np.std(test_episode_steps),
)
)
# Save test scores
np.save(
os.path.join(config["output_dir"], "scores-test{}.npy".format(log_file)),
np.array(test_episode_rewards),
)
# Save checkpoint
torch.save(
actor_critic.state_dict(),
os.path.join(config["save_dir"], "agent{}.pt".format(log_file)),
)
if __name__ == "__main__":
# Argument
parser = argparse.ArgumentParser()
parser.add_argument("--exp_name", type=str, required=True)
parser.add_argument("--env_name", type=str, required=True)
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--debug", action="store_true")
args = parser.parse_args()
# Load config
config_file = open("configs/{}.yaml".format(args.exp_name), "r")
config = yaml.load(config_file, Loader=yaml.FullLoader)
# Update config
config["exp_name"] = args.exp_name
config["env_name"] = args.env_name
config["seed"] = args.seed
config["debug"] = args.debug
# Run main
main(config)