Repository navigation
Expand file tree
/
Copy pathppo.py
More file actions
executable file
·118 lines (103 loc) · 4.89 KB
/
Copy pathppo.py
File metadata and controls
executable file
·118 lines (103 loc) · 4.89 KB
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
# # Unity ML Agents
# ## Proximal Policy Optimization (PPO)
# Contains an implementation of PPO as described [here](https://arxiv.org/abs/1707.06347).
from docopt import docopt
import os
from ppo.models import *
from ppo.trainer import Trainer
from unityagents import UnityEnvironment
_USAGE = '''
Usage:
ppo (<env>) [options]
Options:
--help Show this message.
--max-step=<n> Maximum number of steps to run environment [default: 5e6].
--run-path=<path> The sub-directory name for model and summary statistics [default: ppo].
--load Whether to load the model or randomly initialize [default: False].
--train Whether to train model, or only run inference [default: True].
--summary-freq=<n> Frequency at which to save training statistics [default: 10000].
--save-freq=<n> Frequency at which to save model [default: 50000].
--gamma=<n> Reward discount rate [default: 0.99].
--lambd=<n> Lambda parameter for GAE [default: 0.95].
--time-horizon=<n> How many steps to collect per agent before adding to buffer [default: 2048].
--beta=<n> Strength of entropy regularization [default: 1e-3].
--num-epoch=<n> Number of gradient descent steps per batch of experiences [default: 5].
--epsilon=<n> Acceptable threshold around ratio of old and new policy probabilities [default: 0.2].
--buffer-size=<n> How large the experience buffer should be before gradient descent [default: 2048].
--learning-rate=<rate> Model learning rate [default: 3e-4].
--hidden-units=<n> Number of units in hidden layer [default: 64].
--batch-size=<n> How many experiences per gradient descent update step [default: 64].
'''
options = docopt(_USAGE)
print(options)
# General parameters
max_steps = float(options['--max-step'])
model_path = './models/{}'.format(str(options['--run-path']))
summary_path = './summaries/{}'.format(str(options['--run-path']))
load_model = options['--load']
train_model = options['--train']
summary_freq = int(options['--summary-freq'])
save_freq = int(options['--save-freq'])
env_name = options['<env>']
# Algorithm-specific parameters for tuning
gamma = float(options['--gamma'])
lambd = float(options['--lambd'])
time_horizon = int(options['--time-horizon'])
beta = float(options['--beta'])
num_epoch = int(options['--num-epoch'])
epsilon = float(options['--epsilon'])
buffer_size = int(options['--buffer-size'])
learning_rate = float(options['--learning-rate'])
hidden_units = int(options['--hidden-units'])
batch_size = int(options['--batch-size'])
env = UnityEnvironment(file_name=env_name)
print(str(env))
brain_name = env.brain_names[0]
tf.reset_default_graph()
# Create the Tensorflow model graph
ppo_model = create_agent_model(env, lr=learning_rate,
h_size=hidden_units, epsilon=epsilon,
beta=beta)
is_continuous = (env.brains[brain_name].action_space_type == "continuous")
use_observations = (env.brains[brain_name].number_observations > 0)
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(summary_path):
os.makedirs(summary_path)
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session() as sess:
# Instantiate model parameters
if load_model:
print('Loading Model...')
ckpt = tf.train.get_checkpoint_state(model_path)
saver.restore(sess, ckpt.model_checkpoint_path)
else:
sess.run(init)
steps = sess.run(ppo_model.global_step)
summary_writer = tf.summary.FileWriter(summary_path)
info = env.reset(train_mode=train_model)[brain_name]
trainer = Trainer(ppo_model, sess, info, is_continuous, use_observations)
while steps <= max_steps or not train_model:
if env.global_done:
info = env.reset(train_mode=train_model)[brain_name]
# Decide and take an action
new_info = trainer.take_action(info, env, brain_name)
info = new_info
trainer.process_experiences(info, time_horizon, gamma, lambd)
if len(trainer.training_buffer['actions']) > buffer_size and train_model:
# Perform gradient descent with experience buffer
trainer.update_model(batch_size, num_epoch)
if steps % summary_freq == 0 and steps != 0 and train_model:
# Write training statistics to tensorboard.
trainer.write_summary(summary_writer, steps)
if steps % save_freq == 0 and steps != 0 and train_model:
# Save Tensorflow model
save_model(sess, model_path=model_path, steps=steps, saver=saver)
steps += 1
sess.run(ppo_model.increment_step)
# Final save Tensorflow model
if steps != 0 and train_model:
save_model(sess, model_path=model_path, steps=steps, saver=saver)
env.close()
export_graph(model_path, env_name)