-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_GBRv0.py
197 lines (167 loc) · 7.58 KB
/
train_GBRv0.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
import os
import numpy as np
import cv2
import json
from toolz.itertoolz import interleave
import torch.nn as nn
from stable_baselines3 import DQN, PPO
from stable_baselines3.common.vec_env import SubprocVecEnv
from stable_baselines3.common.monitor import Monitor
from stable_baselines3.common.evaluation import evaluate_policy
from stable_baselines3.common.env_util import make_vec_env
from stable_baselines3.common.callbacks import ProgressBarCallback, BaseCallback, CallbackList
from stable_baselines3.common.env_checker import check_env
from envs import GBR_v0
run_id = "gbrv0_dqn_moore_10"
evaluate = False
config = {
"rl_alg": "DQN",
"policy_type": "MlpPolicy",
"n_workers": 32, # number of parallel processes to use
"total_timesteps": 3000000, # total number of steps
"train_log_eps_freq": 1,
"train_render_eps_freq": 50,
"n_eval_eps": 100,
"train_run_dir": f"./logs/run_{run_id}/train/",
"eval_run_dir": f"./logs/run_{run_id}/eval/",
"plot_dir": f"./logs/run_{run_id}/train/plots/",
"model_dir": f"./logs/run_{run_id}/models/"
}
class TrainCallback(BaseCallback):
def __init__(self, model, n_envs,
log_freq: int, render_freq: int,
render_dir: str, model_log_dir: str,
verbose: int = 0):
super(TrainCallback, self).__init__(verbose)
self.n_envs = n_envs
self.model = model
self.envs = model.get_env()
self.train_log_eps_freq = log_freq
self.render_freq = render_freq
self.dir = render_dir
self.best_mean_reward = -np.inf
self.save_path = os.path.join(model_log_dir, "best_model")
def _on_step(self) -> bool:
eps = self.envs.get_attr('eps', indices=[-1])[-1]
stp = self.envs.get_attr('stp', indices=[-1])[-1]
if eps % self.train_log_eps_freq == 0:
if stp == 1:
returns = list(interleave(self.envs.get_attr('returns', indices=[i for i in range(self.n_envs)])))
if len(returns) > 0:
scores = list(interleave(self.envs.get_attr('performances', indices=[i for i in range(self.n_envs)])))
eps_lengths = list(interleave(self.envs.get_attr('eps_lengths', indices=[i for i in range(self.n_envs)])))
# log rewards and scores
perfLog = {
"scores": scores,
"returns": returns,
"eps_lengths": eps_lengths
}
json_obj = json.dumps(perfLog, indent=4)
with open(f"{self.dir}plots/performance_log.json", "w") as outfile:
outfile.write(json_obj)
# Mean training reward over the last 100 episodes
mean_reward = np.mean(returns[-100:])
# New best model, you could save the agent here
if mean_reward > self.best_mean_reward:
self.best_mean_reward = mean_reward
# save best model
if self.verbose >= 1:
print(f"Eps {eps} Best Avg Rwd {self.best_mean_reward:.3f} | Saving new best model to {self.save_path}")
self.model.save(self.save_path)
if eps % self.render_freq == 0:
# log renders
frame = self.envs.get_images()[-1]
bgr_array = cv2.cvtColor(frame, cv2.COLOR_RGBA2BGR)
cv2.imwrite(f"{self.dir}{eps}-{stp}.png", bgr_array)
return True
if __name__ == '__main__':
os.makedirs(config["plot_dir"], exist_ok=True)
os.makedirs(config["model_dir"], exist_ok=True)
os.makedirs(config["eval_run_dir"], exist_ok=True)
os.makedirs(config["train_run_dir"], exist_ok=True)
# multiprocess training
env = make_vec_env(lambda: GBR_v0(local=True, screen_size=500),
n_envs=config["n_workers"],
vec_env_cls=SubprocVecEnv)
# env = Monitor(env, config["train_run_dir"])
# check_env(env) # check if the env follows the gym interface
if config["rl_alg"] == 'DQN':
model = DQN(policy = config["policy_type"],
env = env,
gamma = 0.999,
learning_rate = 0.000809342,
batch_size = 128,
buffer_size = 50000,
train_freq = 8,
gradient_steps = 4,
exploration_fraction = 0.03640738,
exploration_final_eps = 0.00104112,
target_update_interval = 1,
learning_starts = 0,
policy_kwargs = dict(net_arch=[64]),
device = 'cpu', # 'cuda' 'cpu'
verbose=0) # verbose=2 for debugging
if config["rl_alg"] == 'DQN-CNN':
model = DQN(policy = "CnnPolicy",
env = env,
gamma = 0.999,
learning_rate = 0.000621429983014515,
batch_size = 256,
buffer_size = 10000,
train_freq = 256,
gradient_steps = 256,
exploration_fraction = 0.127963173155836,
exploration_final_eps = 0.0375556799144331,
target_update_interval = 5000,
learning_starts = 5000,
policy_kwargs = dict(net_arch=[256, 256, 256]),
device = 'cpu', # 'cuda' 'cpu'
verbose=0) # verbose=2 for debugging
elif config["rl_alg"] == 'PPO':
model = PPO(policy = config["policy_type"],
env = env,
learning_rate = 1.26e-05,
n_steps = 2048,
batch_size = 256,
n_epochs = 20,
gamma = 0.99,
gae_lambda = 0.99,
clip_range = 0.3,
ent_coef = 2.52e-07,
vf_coef = 2.69e-06,
max_grad_norm = 0.9,
policy_kwargs = dict(
net_arch=dict(pi=[256, 256], vf=[256, 256]),
activation_fn=nn.ReLU,
ortho_init=True,
),
device = 'cpu', # 'cuda' 'cpu'
verbose=0) # verbose=2 for debugging
if not evaluate:
# callbacks
trainLog = TrainCallback(model,
config["n_workers"],
config["train_log_eps_freq"],
config["train_render_eps_freq"],
config["train_run_dir"],
config['model_dir'],
verbose=1)
progressBar = ProgressBarCallback()
# Training
model.learn(
total_timesteps=config["total_timesteps"], # // config["n_workers"]
callback=CallbackList([
trainLog,
progressBar
])
)
env.close()
# Evaluate the trained agent
if evaluate:
trained_eval_env = Monitor(GBR_v0(local=True, screen_size=500))
# del and reload trained model
del model
model = PPO.load(f"{config['model_dir']}best_model", env=trained_eval_env, print_system_info=False)
mean_reward, std_reward = evaluate_policy(model, trained_eval_env, n_eval_episodes=config["n_eval_eps"])
print(f'Trained agent | Mean reward: {mean_reward} +/- {std_reward:.2f}')
trained_eval_env.close()