-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.py
130 lines (115 loc) · 6.28 KB
/
run.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
import numpy as np
import matplotlib.pyplot as plt
from collections import namedtuple, deque
import gym
from agent.Hdqn import Hdqn
#from utils.plotting import plot_episode_stats, plot_visited_states
#from utils import plotting
from meta_controller import meta_controller
from object_detection import object_detection
import cv2
import copy
plt.style.use('ggplot')
class Coach:
def __init__(self):
self.env = gym.make('MontezumaRevenge-v4')
self.env_actions = self.env.unwrapped.get_action_meanings()
self.agent = Hdqn()
self.goal = ''
self.goal_mask = []
self.meta = meta_controller()
self.history = deque([], maxlen = 5)
self.num_episodes = 5000
self.anneal_factor = (1.0-0.1)/self.num_episodes
self.goal_idx = {'ladder1':0,'ladder2':1,'ladder3':2,'key':3 ,'door2':4}
self.ActorExperience = namedtuple("ActorExperience", ["state", "goal", "action", "reward", "next_state"])
self.stats = {'episode_rewards': np.zeros(self.num_episodes) , 'episode_length' : np.zeros(self.num_episodes), 'goal_selected': np.zeros(5), 'goal_success': np.zeros(5)}
self.anneal_threshold = 0.8
self.ale_lives = 6
self.object_detection = object_detection()
self.initial_p = 1
self.final_p = 0.1
self.schedule_timesteps = 50000
self.time_steps = 0
def learn_subgoal(self):
goal_mask = self.object_detection.to_grayscale(self.object_detection.downsample(self.goal_mask))
action = self.agent.select_move(list(self.history)[1:5], goal_mask, self.goal_idx[self.goal])
print("GOAL", self.goal, str((self.meta.getCurrentState() , self.env_actions[action])) + "; ")
next_frame , external_reward, done, info = self.env.step(action)
print "Done", done, "Info : ", info['ale.lives'], "ale_lives", self.ale_lives
if info['ale.lives'] < self.ale_lives:
self.ale_lives = info['ale.lives']
print "Agent Died!!!! . Lives left : ", self.ale_lives
self.meta.update_state('start')
self.stats['goal_selected'][self.goal_idx[self.goal]] += 1
self.goal, self.goal_mask = self.meta.getSubgoal()
# cv2.imshow('image', next_frame)
# cv2.waitKey(10)
# cv2.imshow('image', self.goal_mask)
# cv2.waitKey(1)
next_frame_preprocessed = self.object_detection.preprocess(next_frame)
self.history.append(next_frame_preprocessed)
if external_reward > 0:
print "extrinsic_reward for goal", self.goal, " reward:", external_reward
# print self.goal_mask.shape , next_frame.shape
intrinsic_reward = self.agent.criticize(self.goal_mask, next_frame)
print("Intrinsic Reward", intrinsic_reward)
goal_reached = (intrinsic_reward > 0)
if goal_reached:
print "Goal reached!! ", self.goal
self.meta.update_state(self.goal)
if len(self.history) == 5:
exp = self.ActorExperience(copy.deepcopy(list(self.history)[0:4]), goal_mask, action, intrinsic_reward, copy.deepcopy(list(self.history)[1:5]))
self.agent.store(exp)
self.agent.update()
return external_reward, goal_reached, done
# def annealing_schedule(self):
# fraction = min(float(self.time_steps) / self.schedule_timesteps, 1.0)
# return self.initial_p + fraction * (self.final_p - self.initial_p)
def learn_global(self):
print "Annealing factor: " + str(self.anneal_factor)
for num_episode in range(self.num_episodes):
self.history.clear()
self.ale_lives = 6
total_external_reward = 0
episode_length = 0
print "\n\n### EPISODE " + str(num_episode) + "###"
self.env.reset()
self.meta = meta_controller()
done = False
while not done:
frame = self.env.render(mode='rgb_array')
frame = self.object_detection.preprocess(frame)
self.history.append(frame)
self.goal, self.goal_mask = self.meta.getSubgoal()
self.stats['goal_selected'][self.goal_idx[self.goal]] += 1
print "\nNew Goal: " + str(self.goal) + "\nState-Actions: "
goal_reached = False
while not done and not goal_reached:
self.time_steps += 1
external_reward, goal_reached, done = self.learn_subgoal()
total_external_reward += external_reward
episode_length += 1
if goal_reached:
self.stats['goal_success'][self.goal_idx[self.goal]] += 1
#Annealing
self.stats['episode_rewards'][num_episode] = total_external_reward
self.stats['episode_length'][num_episode] = episode_length
print "annealing_schedule" , self.annealing_schedule()
for goal in self.goal_idx.keys():
#self.agent.actor_epsilon[self.goal_idx[goal]] = self.annealing_schedule()
if self.stats['goal_selected'][self.goal_idx[goal]] > 0:
print("Success Rate", self.stats['goal_success'][self.goal_idx[goal]], self.stats['goal_selected'][self.goal_idx[goal]], goal)
avg_success_rate = self.stats['goal_success'][self.goal_idx[goal]] / self.stats['goal_selected'][self.goal_idx[goal]]
if avg_success_rate < self.anneal_threshold or self.stats['goal_selected'][self.goal_idx[goal]] < 100:
self.agent.actor_epsilon[self.goal_idx[goal]] -= self.anneal_factor
# self.agent.actor_epsilon[self.goal_idx[goal]] -= self.anneal_factor
# self.agent.actor_epsilon[self.goal_idx[goal]] = max(0.1, self.agent.actor_epsilon[self.goal_idx[goal]])
else:
self.agent.actor_epsilon[self.goal_idx[goal]] = 0.1
print "actor_epsilon " + str(goal) + ": " + str(self.agent.actor_epsilon[self.goal_idx[goal]])
def main():
coach = Coach()
coach.learn_global()
if __name__ == "__main__":
main()