-
Notifications
You must be signed in to change notification settings - Fork 1
/
Hierarchy.py
216 lines (172 loc) · 10.5 KB
/
Hierarchy.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
import numpy as np
import lasagne
import theano
from lasagne.regularization import regularize_network_params, l2
from super_agent import MdpAgent
from agentnet.environment import SessionPoolEnvironment
from agentnet.utils.layers import get_layer_dtype
from agentnet.learning import qlearning_n_step
from controller import Controller
from metacontroller import MetaController
class HierarchicalAgent(MdpAgent):
def __init__(self, pool, observation_shape, n_actions, n_parallel_games=1,
replay_seq_len=20, replay_batch_size=20, pool_size=None, n_steps=3, gamma=0.99,
split_into=1,): #gru0_size=128):
self.n_parallel_games = n_parallel_games
self.replay_seq_len = replay_seq_len
self.replay_batch_size = replay_batch_size
self.pool_size = pool_size
self.n_steps = n_steps
self.n_actions = n_actions
self.gamma = gamma
self.split_into = split_into
self.controller = Controller(observation_shape, n_actions)
self.metacontroller = MetaController(self.controller)#, gru0_size)
# Prepare replay pool
self.controller_env = SessionPoolEnvironment(observations=self.controller.agent.observation_layers,
actions=self.controller.resolver,
agent_memories=self.controller.agent.agent_states)
self.metacontroller_env = SessionPoolEnvironment(observations=self.metacontroller.agent.observation_layers,
actions=self.metacontroller.resolver,
agent_memories=self.metacontroller.agent.agent_states)
# get interaction sessions
observation_log, action_tensor, extrinsic_reward_log, memory_log, is_alive_tensor, _ = \
pool.interact(self.step, n_steps=self.replay_seq_len)
preceding_memory_states = list(pool.prev_memory_states)
self.reload_pool(observation_log, action_tensor, extrinsic_reward_log, is_alive_tensor,
memory_log, preceding_memory_states)
if pool_size is None:
controller_batch_env = self.controller_env
metacontroller_batch_env = self.metacontroller_env
else:
controller_batch_env = self.controller_env.sample_session_batch(self.replay_batch_size)
metacontroller_batch_env = self.metacontroller_env.sample_session_batch(self.replay_batch_size)
self.loss = self.build_loss(controller_batch_env, self.controller.agent, 50) + \
self.build_loss(metacontroller_batch_env, self.metacontroller.agent, 10)
self.eval_fun = self.build_eval_fun(metacontroller_batch_env)
weights = self.controller.weights + self.metacontroller.weights
updates = lasagne.updates.adadelta(self.loss, weights, learning_rate=0.01)
mean_session_reward = metacontroller_batch_env.rewards.sum(axis=1).mean()
train_fun = theano.function([], [self.loss, mean_session_reward], updates=updates)
super(HierarchicalAgent, self).__init__([self.controller_env, self.metacontroller_env],
pool, train_fun, pool_size, replay_seq_len)
# raise NotImplementedError
def reload_pool(self, observation_tensor, action_tensor, extrinsic_reward_tensor, is_alive_tensor,
memory_tensor, preceding_memory_states):
batch_size = observation_tensor.shape[0]
# print observation_tensor.shape
meta_obs_log, goal_log, meta_V, itrs = memory_tensor[-4:]
itr = itrs[0]
pivot = len(self.controller.agent.state_variables)
controller_preceding_states = preceding_memory_states[:pivot]
metacontroller_preceding_states = preceding_memory_states[pivot:-4]
###CONTROLLER###
# load them into experience replay environment for controller
# controller_preceding_states =!!!!!!!!!!!!!!!!!!!!!!!!!!!
ctrl_shape = (batch_size * self.split_into, self.replay_seq_len / self.split_into)
intrinsic_rewards = np.concatenate([np.zeros([meta_V.shape[0], 1]), np.diff(meta_V, axis=1)], axis=1)
# print [observation_tensor.reshape(ctrl_shape + self.controller.observation_shape[1:]),
# goal_log.reshape(ctrl_shape)][0].shape
self.controller_env.load_sessions([observation_tensor.reshape(ctrl_shape + self.controller.observation_shape[1:]),
goal_log.reshape(ctrl_shape)],
action_tensor.reshape(ctrl_shape),
intrinsic_rewards.reshape(ctrl_shape),
is_alive_tensor.reshape(ctrl_shape),
# controller_preceding_states
)
###METACONTROLLER###
# separate case for metacontroller
extrinsic_reward_sums = np.diff(
np.concatenate(
[np.zeros_like(extrinsic_reward_tensor[:, 0, None]),
extrinsic_reward_tensor.cumsum(axis=-1)[:, itr == 0]],
axis=1
)
)
self.metacontroller_env.load_sessions(meta_obs_log[:, itr == 0][:, :10],
goal_log[:, itr == 0][:, :10],
extrinsic_reward_sums[:, :10],
is_alive_tensor[:, itr == 0][:, :10],
metacontroller_preceding_states)
def update_pool(self, observation_tensor, action_tensor, extrinsic_reward_tensor, is_alive_tensor,
memory_tensor, preceding_memory_states):
batch_size = observation_tensor.shape[0]
meta_obs_log, goal_log, meta_V, itrs = memory_tensor[-4:]
itr = itrs[0]
pivot = len(self.controller.agent.state_variables)
controller_preceding_states = preceding_memory_states[:pivot]
metacontroller_preceding_states = preceding_memory_states[pivot:-4]
###CONTROLLER###
# load them into experience replay environment for controller
# controller_preceding_states =!!!!!!!!!!!!!!!!!!!!!!!!!!!
ctrl_shape = (batch_size * self.split_into, self.replay_seq_len / self.split_into)
intrinsic_rewards = np.concatenate([np.zeros([meta_V.shape[0], 1]), np.diff(meta_V, axis=1)], axis=1)
self.controller_env.append_sessions([observation_tensor.reshape(ctrl_shape + self.controller.observation_shape[1:]),
goal_log.reshape(ctrl_shape)],
action_tensor.reshape(ctrl_shape),
intrinsic_rewards.reshape(ctrl_shape),
is_alive_tensor.reshape(ctrl_shape),
controller_preceding_states,
max_pool_size=self.pool_size,
)
###METACONTROLLER###
# separate case for metacontroller
extrinsic_reward_sums = np.diff(
np.concatenate(
[np.zeros_like(extrinsic_reward_tensor[:, 0, None]),
extrinsic_reward_tensor.cumsum(axis=-1)[:, itr == 0]],
axis=1
)
)
self.metacontroller_env.append_sessions(meta_obs_log[:, itr == 0][:, :10],
goal_log[:, itr == 0][:, :10],
extrinsic_reward_sums[:, :10],
is_alive_tensor[:, itr == 0][:, :10],
metacontroller_preceding_states,
max_pool_size=self.pool_size)
def step(self, env_observation, prev_memories='zeros'):
""" returns actions and new states given observation and prev state
Prev state in default setup should be [prev window,]"""
batch_size = self.n_parallel_games
if prev_memories == 'zeros':
controller_mem = metacontroller_mem = 'zeros'
meta_inp = np.zeros((batch_size,) + tuple(self.metacontroller.observation_shape[1:]), dtype='float32')
itr = -1
# goal will be defined by "if itr ==0" clause
else:
pivot = len(self.controller.agent.state_variables)
controller_mem, metacontroller_mem = prev_memories[:pivot], prev_memories[pivot:-4]
meta_inp, goal, meta_V, itrs = prev_memories[-4:]
itr = itrs[0]
itr = (itr + 1) % self.metacontroller.period
if itr == 0:
goal, metacontroller_mem, meta_V = self.metacontroller.step(meta_inp, metacontroller_mem, batch_size)
#print env_observation.shape
action, controller_mem, meta_inp = self.controller.step(env_observation, goal, controller_mem, batch_size)
new_memories = controller_mem + metacontroller_mem + [meta_inp, goal, meta_V, [itr] * self.n_parallel_games]
return action, new_memories
def build_loss(self, env, agent, replay_seq_len):
# get agent's Qvalues obtained via experience replay
_, _, _, _, qvalues_seq = agent.get_sessions(
env,
# initial_hidden = env.preceding_agent_memories,
session_length=replay_seq_len,
batch_size=env.batch_size,
optimize_experience_replay=True,
)
scaled_reward_seq = env.rewards
elwise_mse_loss = qlearning_n_step.get_elementwise_objective(qvalues_seq,
env.actions[0],
scaled_reward_seq,
env.is_alive,
gamma_or_gammas=self.gamma,
n_steps=self.n_steps)
# compute mean over "alive" fragments
mse_loss = elwise_mse_loss.sum() / env.is_alive.sum()
# regularize network weights
reg_l2 = regularize_network_params(agent.state_variables.keys(), l2) * 10 ** -5
return mse_loss + reg_l2
def build_eval_fun(self, env):
mean_session_reward = env.rewards.sum(axis=1).mean() / self.replay_seq_len
eval_fun = theano.function([], [mean_session_reward])
return eval_fun