-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_remote.py
158 lines (133 loc) · 7.26 KB
/
run_remote.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
import numpy as np
import torch
import time
from time import sleep
from collections import defaultdict
import itertools
import random
from logger import log, init_logger
from env.Flatland import Flatland, FlatlandWrapper
from env.rewards.FakeRewardShaper import FakeRewardShaper
from env.DeadlockChecker import DeadlockChecker
from env.GreedyChecker import GreedyChecker
from agent.judge.Judge import ConstWindowSizeGenerator, LinearOnAgentNumberSizeGenerator
from env.observations import SimpleObservation
from agent.PPO.PPOController import PPOController
from agent.judge.Judge import Judge
from flatland.evaluators.client import FlatlandRemoteClient
from flatland.utils.rendertools import RenderTool, AgentRenderVariant
from flatland.evaluators.client import TimeoutException
from flatland.envs.agent_utils import RailAgentStatus
RANDOM_SEED = 23
torch.manual_seed(RANDOM_SEED)
torch.cuda.manual_seed_all(RANDOM_SEED)
np.random.seed(RANDOM_SEED)
random.seed(RANDOM_SEED)
init_logger("logdir", "tmp", use_wandb=False)
def evaluate_remote():
remote_client = FlatlandRemoteClient()
my_observation_builder = SimpleObservation(max_depth=3, neighbours_depth=3,
timetable=Judge(LinearOnAgentNumberSizeGenerator(0.03, 5), lr=0,
batch_size=0, optimization_epochs=0, device=torch.device("cpu")),
deadlock_checker=DeadlockChecker(), greedy_checker=GreedyChecker(), parallel=False, eval=True)
params = torch.load("generated/params.torch")
params.neighbours_depth=my_observation_builder.neighbours_depth
controller = PPOController(params, torch.device("cpu"))
controller.load_controller("generated/controller.torch")
my_observation_builder.timetable.load_judge("generated/judge.torch")
render = False
sum_reward, sum_percent_done = 0., 0.
for evaluation_number in itertools.count():
time_start = time.time()
observation, info = remote_client.env_create(obs_builder_object=my_observation_builder)
if not observation:
break
local_env = FlatlandWrapper(remote_client.env, FakeRewardShaper())
local_env.n_agents = len(local_env.agents)
log().check_time()
if render:
env_renderer = RenderTool(
local_env.env,
agent_render_variant=AgentRenderVariant.ONE_STEP_BEHIND,
show_debug=True,
screen_height=600,
screen_width=800
)
env_creation_time = time.time() - time_start
print("Evaluation Number : {}".format(evaluation_number))
time_taken_by_controller = []
time_taken_per_step = []
steps = 0
done = defaultdict(lambda: False)
while True:
try:
if render:
env_renderer.render_env(show=True, show_observations=False, show_predictions=False)
time_start = time.time()
action_dict = dict()
handles_to_ask = list()
observation = {k: torch.tensor(v, dtype=torch.float) for k, v in observation.items() if v is not None}
for i in range(local_env.n_agents):
if not done[i]:
if local_env.obs_builder.greedy_checker.greedy_position(i):
action_dict[i] = 0
elif i in observation:
handles_to_ask.append(i)
for handle in handles_to_ask:
for opp_handle in local_env.obs_builder.encountered[handle]:
if opp_handle != -1 and opp_handle not in observation:
observation[opp_handle] = torch.tensor(local_env.obs_builder._get_internal(opp_handle), dtype=torch.float)
time_taken_per_step.append(time.time() - time_start)
time_start = time.time()
controller_actions = controller.fast_select_actions(handles_to_ask, observation,
local_env.obs_builder.encountered, train=True)
action_dict.update(controller_actions)
action_dict = {k: local_env.transform_action(k, v) for k, v in action_dict.items()}
action_dict = {handle: action for handle, action in action_dict.items() if action != -1}
time_taken = time.time() - time_start
time_taken_by_controller.append(time_taken)
time_start = time.time()
observation, all_rewards, done, info = remote_client.env_step(action_dict)
num_done = sum([1 for agent in local_env.agents if agent.status == RailAgentStatus.DONE_REMOVED])
num_started = sum([1 for handle in range(len(local_env.agents)) if local_env.obs_builder.timetable.is_ready(handle)])
finished_handles = [handle for handle in range(len(local_env.agents))
if local_env.obs_builder.timetable.ready_to_depart[handle] == 2]
reward = torch.sum(local_env._max_episode_steps - local_env.obs_builder.timetable.end_time[finished_handles])
reward /= len(local_env.agents) * local_env._max_episode_steps
percent_done = float(num_done) / len(local_env.agents)
deadlocked = int(sum(local_env.obs_builder.deadlock_checker._is_deadlocked) + 0.5)
steps += 1
time_taken = time.time() - time_start
time_taken_per_step.append(time_taken)
if done['__all__']:
print("Done agents {}/{}".format(num_done, len(local_env.agents)))
print("Started agents {}/{}".format(num_started, len(local_env.agents)))
print("Deadlocked agents {}/{}".format(deadlocked, len(local_env.agents)))
print("Reward: {} Percent done: {}".format(reward, percent_done))
sum_reward += reward
sum_percent_done += percent_done
print("Total reward: {} Avg percent done: {}".format(sum_reward, sum_percent_done / (evaluation_number + 1)))
if render:
env_renderer.close_window()
break
except TimeoutException as err:
print("Timeout! Will skip this episode and go to the next.", err)
break
np_time_taken_by_controller = np.array(time_taken_by_controller)
np_time_taken_per_step = np.array(time_taken_per_step)
print("="*100)
print("="*100)
print("Evaluation Number : ", evaluation_number)
print("Current Env Path : ", remote_client.current_env_path)
print("Env Creation Time : ", env_creation_time)
print("Number of Steps : {}/{}".format(steps, local_env._max_episode_steps))
print("Mean/Std/Sum of Time taken by Controller : ", np_time_taken_by_controller.mean(), np_time_taken_by_controller.std(), np_time_taken_by_controller.sum())
print("Mean/Std/Sum of Time per Step : ", np_time_taken_per_step.mean(), np_time_taken_per_step.std(), np_time_taken_per_step.sum())
log().print_time_metrics()
log().zero_time_metrics()
print("="*100)
print("\n\n")
print("Evaluation of all environments complete...")
print(remote_client.submit())
if __name__ == "__main__":
evaluate_remote()