/
a2c_agent.py
200 lines (170 loc) · 7.91 KB
/
a2c_agent.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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
"""
A2C - synchronous advantage actor critic with generalized advantage estimation
Adapted from https://arxiv.org/pdf/1602.01783.pdf, Algorithm S3
"""
import time
import numpy as np
from yarlp.agent.base_agent import Agent, add_advantage
from yarlp.model.networks import cnn
from yarlp.utils.experiment_utils import get_network
from yarlp.model.model_factories import a2c_model_factory
from yarlp.utils.env_utils import ParallelEnvs
from yarlp.utils.metric_logger import explained_variance
from yarlp.utils.schedules import PiecewiseSchedule, ConstantSchedule
from dateutil.relativedelta import relativedelta as rd
class A2CAgent(Agent):
def __init__(
self, env,
policy_network=None,
policy_network_params={
'final_dense_weights_initializer': 0.01
},
policy_learning_rate=5e-4,
value_network_params={
'final_dense_weights_initializer': 1.0
},
entropy_weight=0.01,
model_file_path=None,
adaptive_std=False, init_std=1.0, min_std=1e-6,
n_steps=5,
max_timesteps=1000000,
grad_norm_clipping=0.5,
gae_lambda=0.98,
checkpoint_freq=10000,
save_freq=50000,
policy_learning_rate_schedule=None,
*args, **kwargs):
super().__init__(env, *args, **kwargs)
assert isinstance(self._env, ParallelEnvs),\
"env must be ParallelEnvs class for A2C agent"
if policy_network is None:
policy_network = cnn
pn = get_network(policy_network, policy_network_params)
self._policy = a2c_model_factory(
env, network=pn,
policy_network_params=policy_network_params,
value_network_params=value_network_params,
learning_rate=policy_learning_rate,
has_learning_rate_schedule=True,
entropy_weight=entropy_weight,
min_std=min_std, init_std=init_std, adaptive_std=adaptive_std,
grad_norm_clipping=grad_norm_clipping,
model_file_path=model_file_path)
self.tf_object_attributes.add('_policy')
self.unserializables.add('_env')
policy_weight_sums = sum(
[np.sum(a) for a in self._policy.get_weights()])
self.logger.logger.info(
'Policy network weight sums: {}'.format(policy_weight_sums))
self.n_steps = n_steps
self.max_timesteps = max_timesteps
self.t = 0
self.checkpoint_freq = checkpoint_freq
self.save_freq = save_freq
self.num_envs = self.env.num_envs
self.env_id = self.env.env_id
self.start_seed = self.env.start_seed
self.is_atari = self.env.is_atari
if policy_learning_rate_schedule is None:
lr = ConstantSchedule(policy_learning_rate)
elif isinstance(policy_learning_rate_schedule, list):
lr = PiecewiseSchedule(
policy_learning_rate_schedule,
outside_value=policy_learning_rate_schedule[-1][-1])
self.lr_schedule = lr
def set_env(self):
self._env = ParallelEnvs(self.env_id, self.num_envs,
self.start_seed, self.is_atari)
def train(self):
self.last_saved_reward = -np.inf
num_envs = self._env.num_envs
batch_ob_shape = (
num_envs * self.n_steps, *self.env.observation_space.shape)
obs = self.env.reset()
while self.t < self.max_timesteps:
mb_obs, mb_rewards, mb_actions = [], [], []
mb_values, mb_dones = [], []
for n in range(self.n_steps):
actions = self.get_batch_actions(obs)
values = self._policy.G(
self._policy['vf'],
{self._policy['state']: obs})
mb_obs.append(obs)
mb_actions.append(actions)
mb_values.append(values)
obs, rewards, dones, _ = self.env.step(actions)
mb_dones.append(dones)
mb_rewards.append(rewards)
self.t += num_envs
mb_obs = np.asarray(mb_obs).swapaxes(1, 0).reshape(batch_ob_shape)
mb_rewards = np.asarray(mb_rewards).swapaxes(1, 0)
mb_actions = np.asarray(mb_actions).swapaxes(1, 0)
mb_values = np.asarray(mb_values).swapaxes(1, 0)
mb_dones = np.asarray(mb_dones).swapaxes(1, 0)
last_values = self._policy.G(
self._policy['vf'],
{self._policy['state']: obs})
mb_discounted_rewards = []
for n in range(self._env.num_envs):
rollout = {
"baseline_preds": mb_values[n].flatten(),
"next_baseline_pred": last_values[n].flatten(),
"dones": mb_dones[n],
"rewards": mb_rewards[n]
}
add_advantage(rollout, self._discount, Lambda=0)
mb_discounted_rewards.append(rollout["discounted_rewards"])
mb_actions = mb_actions.flatten()
mb_values = mb_values.flatten()
mb_discounted_rewards = np.asarray(mb_discounted_rewards).flatten()
mb_rewards = mb_discounted_rewards
lr = self.lr_schedule.value(self.t)
# fit the value and policy networks
policy_loss = self._policy.update(
mb_obs, mb_rewards,
mb_actions, mb_rewards - mb_values, lr)
policy_update_feed_dict = self._policy.build_update_feed_dict(
self._policy, mb_obs, mb_rewards,
mb_actions, mb_rewards - mb_values, lr)
policy_entropy = self._policy.G(
self._policy['entropy'], policy_update_feed_dict)
vf_loss = self._policy.G(
self._policy['vf_loss'], policy_update_feed_dict)
ev = explained_variance(mb_rewards, mb_values)
if self.t > 0 \
and self.t % self.checkpoint_freq == 0 \
and len(self._env.get_episode_rewards()) > 0:
# log things
episode_rewards = self._env.get_episode_rewards(
int(self.reward_len / self.num_envs))
self.logger.set_metrics_for_iter(episode_rewards)
self.logger.add_metric('timesteps_so_far', self.t)
eta = (self.max_timesteps - self.t) /\
(self.t / round(time.time() - self.logger._start_time, 6))
self.logger.add_metric(
'ETA',
str(rd(seconds=eta)))
self.logger.add_metric('policy_loss', policy_loss)
self.logger.add_metric('vf_loss', vf_loss)
self.logger.add_metric('policy_entropy', policy_entropy)
self.logger.add_metric('vf_explained_var', ev)
self.logger.add_metric('env_id', self._env_id)
self.logger.add_metric(
'last_saved_reward', self.last_saved_reward)
self.logger.add_metric(
'policy_learning_rate', self.lr_schedule.value(self.t))
self.logger.log()
# save model if necessary
if self.t % self.save_freq == 0 and self.t \
and self.logger._log_dir is not None \
and len(self._env.get_episode_rewards()) > 0:
running_reward = np.mean(
self.logger._running_reward)
if running_reward > self.last_saved_reward:
self.logger.logger.info(
'Saving best model, {} -> {}'.format(
self.last_saved_reward, running_reward))
self.save(self.logger._log_dir, 'best_agent')
self.last_saved_reward = running_reward
self.logger.logger.info('Saving model for checkpoint')
self.save(self.logger._log_dir)