-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathmontezuma_demo_experiment.py
177 lines (151 loc) · 6.66 KB
/
montezuma_demo_experiment.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
# compile cython modules
import os
os.system('python experience_replay_setup.py build_ext --inplace')
# load dependencies
import tensorflow as tf
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0, "Not enough GPU hardware devices available"
tf.config.experimental.set_memory_growth(physical_devices[0], True)
from tensorflow.keras.models import load_model
from tensorflow.keras.optimizers import RMSprop, Adam
from tensorflow.keras import initializers
import gym
import numpy as np
from deep_q_agents import EpsAnnDQNAgent
from deep_q_networks import DeepQNetwork
from experience_replay import PrioritizedExperienceReplay
from atari_preprocessing import atari_montezuma_processor, ProcessedAtariEnv
from openai_baseline_wrappers import make_atari, wrap_deepmind
from load_data import LoadAtariHeadData
#create environment
frame_processor = atari_montezuma_processor
game_id = 'MontezumaRevengeNoFrameskip-v4'
game_name = 'montezuma_revenge'
env = make_atari(game_id)
env = wrap_deepmind(env)
env = ProcessedAtariEnv(env, frame_processor, reward_processor = lambda x: np.sign(x) * np.log(1 + np.abs(x)))
# additional env specific parameters
frame_shape = env.reset().shape
frame_skip = 4
num_stacked_frames = 4
num_actions = env.action_space.n
# replay parameters
batch_size = 32
max_frame_num = 2**20
prioritized_replay = True
prio_coeff = 0.4
is_schedule = [0.6, 1.0, 1500000]
replay_epsilon = 0.001
expert_epsilon = 1.0
memory_restore_path = None
# network training parameters
dueling = True
double_q = True
lr_schedule = [[0.0000625, 0.0000625, 10000000]]
optimizer = Adam
discount_factor = 0.99
n_step = 10
one_step_weight = 1.0/3.0
n_step_weight = 1.0/3.0
expert_weight = 1.0/3.0
l2_weight = 0.00001
large_margin_coeff = 0.8
model_restore_path = None
# network architecture
conv_layers = {'filters': [32, 64, 64, 1024],
'kernel_sizes': [8, 4, 3, 7],
'strides': [4, 2, 1, 1],
'paddings': ['valid' for _ in range(4)],
'activations': ['relu' for _ in range(4)],
'initializers': [initializers.VarianceScaling(scale = 2.0) for _ in range(4)],
'names': ['conv_%i'%(i) for i in range(1,5)]}
dense_layers = None
# exploration parameters
eps_schedule = [[0.25, 0.1, 250000],
[0.1, 0.01, 5000000],
[0.01, 0.001, 5000000]]
# training session parameters
target_interval = 10000
warmup_steps = 50000
pretrain_steps = 500000
learning_interval = 4
num_steps = 12000000
num_episodes = 10000
max_steps_per_episode = 18000
output_freq = 1000
save_freq = 500
store_memory = True
save_path = "experiments/montezuma_standard_experiment/"
# create replay memory
memory = PrioritizedExperienceReplay(frame_shape = frame_shape,
max_frame_num = max_frame_num,
num_stacked_frames = num_stacked_frames,
batch_size = batch_size,
prio_coeff = prio_coeff,
is_schedule = is_schedule,
epsilon = replay_epsilon,
restore_path = memory_restore_path)
# expert memory
data_loader = LoadAtariHeadData(game_name = game_name, frame_processor = frame_processor)
expert_memory = data_loader.demonstrations_to_per(max_frame_num = max_frame_num,
num_stacked_frames = num_stacked_frames,
frame_shape = frame_shape,
batch_size = batch_size,
prio_coeff = prio_coeff,
is_schedule = is_schedule,
epsilon = expert_epsilon,
recompute_demonstrations = True,
only_highscore = False,
frame_skip = frame_skip)
# create policy network
policy_network = DeepQNetwork(in_shape = (num_stacked_frames, *frame_shape),
conv_layers = conv_layers,
dense_layers = dense_layers,
num_actions = num_actions,
optimizer = optimizer,
lr_schedule = lr_schedule,
dueling = dueling,
one_step_weight = one_step_weight,
n_step_weight = n_step_weight,
expert_weight = expert_weight)
if model_restore_path is not None:
policy_network.model.load_weights(model_restore_path, by_name = True)
# create target network
target_network = DeepQNetwork(in_shape = (num_stacked_frames, *frame_shape),
conv_layers = conv_layers,
dense_layers = dense_layers,
num_actions = num_actions,
optimizer = optimizer,
lr_schedule = lr_schedule,
dueling = dueling,
one_step_weight = one_step_weight,
n_step_weight = n_step_weight,
expert_weight = expert_weight)
if model_restore_path is not None:
target_network.model.load_weights(model_restore_path, by_name = True)
# create agent
agent = EpsAnnDQNAgent(env = env,
memory = memory,
policy_network = policy_network,
target_network = target_network,
num_actions = num_actions,
frame_shape = frame_shape,
discount_factor = discount_factor,
save_path = save_path,
eps_schedule = eps_schedule,
double_q = double_q,
n_step = n_step,
expert_memory = expert_memory,
prioritized_replay = prioritized_replay)
agent.policy_network.model.save(save_path + "/trained_models/initial_model.h5")
# train the agent
agent.train(num_episodes = num_episodes,
num_steps = num_steps,
max_steps_per_episode = max_steps_per_episode,
warmup_steps = warmup_steps,
pretrain_steps = pretrain_steps,
target_interval = target_interval,
learning_interval = learning_interval,
output_freq = output_freq,
save_freq = save_freq,
store_memory = store_memory)