-
Notifications
You must be signed in to change notification settings - Fork 4
/
agents.py
129 lines (115 loc) · 5.97 KB
/
agents.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
import argparse
import sys
sys.path.append('../keras-rl')
from PIL import Image
import numpy as np
import gym
from keras.models import Model
from keras.layers import Flatten, Input, Dense
from keras.optimizers import Adam
from keras.regularizers import l2
import keras.backend as K
from rl.agents.dqn import DQNAgent, DQfDAgent
from rl.policy import EpsGreedyQPolicy, LinearAnnealedPolicy, GreedyQPolicy
from rl.memory import PartitionedMemory, PrioritizedMemory
from rl.core import Processor
from rl.callbacks import TrainEpisodeLogger, ModelIntervalCheckpoint
from rl.util import load_demo_data_from_file
from record_demonstrations import demonstrate, reward_threshold_subset
parser = argparse.ArgumentParser()
parser.add_argument('--mode',choices=['train','test','demonstrate'],default='test')
parser.add_argument('--model',choices=['student','expert'],default='expert')
args = parser.parse_args()
class RocketProcessor(Processor):
def process_observation(self, observation):
return np.array(observation, dtype='float32')
def process_state_batch(self, batch):
return np.array(batch).astype('float32')
def process_reward(self, reward):
return np.sign(reward) * np.log(1 + abs(reward))
def process_demo_data(self, demo_data):
for step in demo_data:
step[0] = self.process_observation(step[0])
step[2] = self.process_reward(step[2])
return demo_data
env = gym.make("LunarLander-v2")
np.random.seed(231)
env.seed(123)
nb_actions = env.action_space.n
WINDOW_LENGTH = 2
input_shape = (WINDOW_LENGTH, 8)
# DQfD "student" model architecture
sensors = Input(shape=(input_shape))
s_dense = Flatten()(sensors)
s_dense = Dense(64, activation='relu', kernel_regularizer=l2(.0001))(s_dense)
s_dense2= Dense(128, activation='relu', kernel_regularizer=l2(.0001))(s_dense)
s_dense3 = Dense(64, activation='relu', kernel_regularizer=l2(.0001))(s_dense2)
s_actions = Dense(nb_actions, activation='linear', kernel_regularizer=l2(.0001))(s_dense3)
student_model = Model(inputs=sensors, outputs=s_actions)
# "Expert" (regular dqn) model architecture
e_dense = Flatten()(sensors)
e_dense = Dense(64, activation='relu')(e_dense)
e_dense2= Dense(128, activation='relu')(e_dense)
e_dense3 = Dense(64, activation='relu')(e_dense2)
e_actions = Dense(nb_actions, activation='linear')(e_dense3)
expert_model = Model(inputs=sensors, outputs=e_actions)
processor = RocketProcessor()
model_saves = './model_saves/'
if __name__ == "__main__":
if args.model == 'student':
# load expert data
expert_demo_data = load_demo_data_from_file(model_saves + 'demos.npy')
expert_demo_data = reward_threshold_subset(expert_demo_data,0)
print(expert_demo_data.shape)
expert_demo_data = processor.process_demo_data(expert_demo_data)
# memory
memory = PartitionedMemory(limit=500000, pre_load_data=expert_demo_data, alpha=.6, start_beta=.4, end_beta=.4, window_length=WINDOW_LENGTH)
# policy
policy = EpsGreedyQPolicy(.01)
# agent
dqfd = DQfDAgent(model=student_model, nb_actions=nb_actions, policy=policy, memory=memory,
processor=processor, enable_double_dqn=True, enable_dueling_network=True, gamma=.99, target_model_update=10000,
train_interval=1, delta_clip=1., pretraining_steps=15000, n_step=10, large_margin=.8, lam_2=1)
lr = .00025
dqfd.compile(Adam(lr), metrics=['mae'])
weights_filename = model_saves + 'student_lander15k_weights.h5f'
checkpoint_weights_filename = model_saves +'student_lander15k_weights{step}.h5f'
log_filename = model_saves + 'student_lander15k_REWARD_DATA.txt'
callbacks = [TrainEpisodeLogger(log_filename),
ModelIntervalCheckpoint(checkpoint_weights_filename, interval=1000000)
]
if args.mode == 'train':
dqfd.fit(env, callbacks=callbacks, nb_steps=4250000, verbose=0, nb_max_episode_steps=1500)
dqfd.save_weights(weights_filename, overwrite=True)
if args.mode == 'test':
dqfd.load_weights(model_saves + 'student_lander15k_weights.h5f')
dqfd.test(env, nb_episodes=12, visualize=True, verbose=2, nb_max_start_steps=30)
if args.mode == 'demonstrate':
print("DQfD cannot demonstrate.")
if args.model == 'expert':
# memory
memory = PrioritizedMemory(limit=500000, alpha=.6, start_beta=.4, end_beta=.4, steps_annealed=5000000, window_length=WINDOW_LENGTH)
# policy
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(), attr='eps', value_max=1., value_min=.02, value_test=.01,
nb_steps=1000000)
# agent
dqn = DQNAgent(model=expert_model, nb_actions=nb_actions, policy=policy, memory=memory,
processor=processor, enable_double_dqn=True, enable_dueling_network=True, gamma=.99, target_model_update=10000,
train_interval=1, delta_clip=1., nb_steps_warmup=50000)
lr = .00025
dqn.compile(Adam(lr), metrics=['mae'])
weights_filename = model_saves + 'expert_lander_weights.h5f'
checkpoint_weights_filename = model_saves +'expert_lander_weights{step}.h5f'
log_filename = model_saves + 'expert_lander_REWARD_DATA.txt'
callbacks = [TrainEpisodeLogger(log_filename),
ModelIntervalCheckpoint(checkpoint_weights_filename, interval=1000000)
]
if args.mode == 'train':
dqn.fit(env, callbacks=callbacks, nb_steps=4250000, verbose=0, nb_max_episode_steps=1500)
dqn.save_weights(weights_filename, overwrite=True)
if args.mode == 'test':
dqn.load_weights(model_saves + 'expert_lander_weights.h5f')
dqn.test(env, nb_episodes=5, visualize=True, verbose=2, nb_max_start_steps=30)
if args.mode == 'demonstrate':
dqn.load_weights(model_saves + 'expert_lander_weights.h5f')
demonstrate(dqn, env, 75000, model_saves + 'demos.npy')