-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
175 lines (143 loc) · 10.1 KB
/
trainer.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
import json
import logging
import os
import sys
from unityagents import UnityEnvironment
from agents import *
from models import *
logging.basicConfig(
level=logging.DEBUG,
format="[%(asctime)s] %(levelname)s [%(name)s.%(funcName)s:%(lineno)d] %(message)s",
stream=sys.stdout
)
launch_parm = namedtuple("launch_parm", ["algorithm", "times", "hparm"])
env_parm = namedtuple("env_parm", ["state_size", "action_size", "brain_name", "agents_num"])
ENVIRONMENT_BINARY = os.environ['DRLUD_P2_ENV']
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
path_prefix = "./hp_search_results/"
def train(agent, environment, n_episodes=1000, max_t=2000, store_weights_to="checkpoint.pth"):
scores = [] # list containing scores from each episode
for i_episode in range(1, n_episodes + 1):
env_info = environment.reset(train_mode=True)[agent.name]
states = env_info.vector_observations
episode_scores = []
for t in range(max_t):
actions = agent.act(states)
env_info = environment.step(actions)[agent.name]
next_states = env_info.vector_observations
rewards = env_info.rewards
dones = env_info.local_done
agent.step(states, actions, rewards, next_states, dones)
states = next_states
episode_scores.append(sum(rewards)/len(rewards))
if any(dones):
break
scores.append(sum(episode_scores))
last_100_steps_mean = np.mean(scores[-100:])
print('\rEpisode {}\tAverage Score: {:.2f}\tLast score: {:.2f}'.format(i_episode,
last_100_steps_mean,
np.mean(scores[-1])), end="")
if i_episode % 20 == 0:
print('\rEpisode {}\tAverage Score: {:.2f}\tLast score: {:.2f}'.format(i_episode,
last_100_steps_mean,
np.mean(scores[-1])))
torch.save(agent.actor_local.state_dict(),
store_weights_to.replace("eps", str(n_episodes)).replace("role", "actor"))
torch.save(agent.critic_local.state_dict(),
store_weights_to.replace("eps", str(n_episodes)).replace("role", "critic"))
return scores
def prepare_environment():
return UnityEnvironment(file_name=ENVIRONMENT_BINARY)
def infer_environment_properties(environment):
brain_name = environment.brain_names[0]
brain = environment.brains[brain_name]
action_size = brain.vector_action_space_size
env_info = environment.reset(train_mode=True)[brain_name]
num_agents = len(env_info.agents)
states = env_info.vector_observations
state_size = states.shape[1]
return brain_name, num_agents, action_size, state_size
def prepare_ddpg_agent(agent_config: ddpg_parm, env_parm: env_parm, seed):
return DDPGAgent(agent_config, env_parm, device, seed)
algorithm_factories = {
"ddpg": prepare_ddpg_agent
}
simulation_hyperparameter_reference = {
1: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, True, 1)),
2: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, False, 1)),
10: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 128, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
11: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 1024, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
12: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 4096, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
13: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 512, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
14: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 384, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
20: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 2e-4, 0, False, False, 1)),
21: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 4e-4, 0, False, False, 1)),
22: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 5e-5, 0, False, False, 1)),
23: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 8e-4, 0, False, False, 1)),
24: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-5, 0, False, False, 1)),
25: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 2e-5, 0, False, False, 1)),
30: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 2e-4, 1e-4, 0, False, False, 1)),
31: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 4e-4, 1e-4, 0, False, False, 1)),
32: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 7e-5, 1e-4, 0, False, False, 1)),
33: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 5e-5, 1e-4, 0, False, False, 1)),
34: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-5, 1e-4, 0, False, False, 1)),
35: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 5e-5, 5e-5, 0, False, False, 1)),
36: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-5, 1e-5, 0, False, False, 1)),
40: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 5e-3, False, False, 1)),
41: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 1e-2, False, False, 1)),
42: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 1e-3, False, False, 1)),
43: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 5e-4, False, False, 1)),
# Large sized memory to check for stability effect
50: launch_parm("ddpg", 1, ddpg_parm(int(1e6), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
51: launch_parm("ddpg", 1, ddpg_parm(int(1e6), 1024, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
52: launch_parm("ddpg", 1, ddpg_parm(int(1e7), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
53: launch_parm("ddpg", 1, ddpg_parm(int(1e7), 1024, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
55: launch_parm("ddpg", 1, ddpg_parm(int(1e8), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
56: launch_parm("ddpg", 1, ddpg_parm(int(1e8), 1024, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
57: launch_parm("ddpg", 1, ddpg_parm(int(1e8), 4096, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
# Check the effect of gradient clipping on the overall stability
60: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, True, 1)),
61: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, True, 1)),
62: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, True, True, 1)),
63: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, True, True, 1)),
# Check the effect of decreasing of the density of learning sessions over time
70: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 2)),
71: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, False, 2)),
72: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, True, False, 2)),
73: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, True, False, 2)),
74: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 5)),
75: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, False, 5)),
76: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, True, False, 5)),
77: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, True, False, 5)),
78: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 20)),
79: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, False, 20)),
80: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, True, False, 20)),
81: launch_parm("ddpg", 1, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, True, False, 20)),
# Check for general stability of outcomes over multiple runs with various random seed
90: launch_parm("ddpg", 25, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, False, False, 1)),
91: launch_parm("ddpg", 25, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, False, False, 1)),
92: launch_parm("ddpg", 25, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-4, 0, True, False, 1)),
93: launch_parm("ddpg", 25, ddpg_parm(int(1e5), 256, 0.99, 1e-3, 1e-4, 1e-3, 0, True, False, 1)),
}
def run_training_session(agent_factory, run_config: launch_parm, id):
env = prepare_environment()
(brain_name, num_agents, action_size, state_size) = infer_environment_properties(env)
agent_config = run_config.hparm
scores = []
for seed in range(run_config.times):
agent = agent_factory(agent_config, env_parm(state_size, action_size, brain_name, num_agents), seed)
scores.append(
train(agent, env, store_weights_to=f"{path_prefix}set{id}_weights_role_episode_eps_seed_{seed}.pth"))
env.close()
return scores
def ensure_training_run(id: int, parm: launch_parm):
if os.path.isfile(f"{path_prefix}set{id}_results.json"):
logging.info(f"Skipping configuration {id} with following parameters {parm}")
else:
logging.info(f"Running {id} with following parameters {parm}")
run_result = run_training_session(algorithm_factories["ddpg"], parm, id)
with open(f"{path_prefix}set{id}_results.json", "w") as fp:
json.dump(run_result, fp)
if __name__ == "__main__":
for parm_id in simulation_hyperparameter_reference:
ensure_training_run(parm_id, simulation_hyperparameter_reference[parm_id])