-
Notifications
You must be signed in to change notification settings - Fork 1
/
option_model.py
151 lines (120 loc) · 5.8 KB
/
option_model.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
import torch
import torch.nn as nn
import numpy as np
from rl_algorithm.dqn.config import embedding_size, discount, device
from rl_algorithm.common.base_model import BaseModel
def init_params(m):
classname = m.__class__.__name__
if classname.find("Linear") != -1:
m.weight.data.normal_(0, 1)
m.weight.data *= 1 / torch.sqrt(m.weight.data.pow(2).sum(1, keepdim=True))
if m.bias is not None:
m.bias.data.fill_(0)
def softmax(a, ww):
c = max(np.max(a), 1)
exp_a = np.exp(np.array(a)/c * ww)
sum_exp_a = np.sum(exp_a)
y = exp_a / sum_exp_a
y = np.clip(y, 0.05, 0.95)
return y
class OptionQ(BaseModel):
def __init__(self, obs_space, num_options, is_init):
super().__init__(obs_space)
self.actor = nn.Sequential(
nn.Linear(embedding_size, embedding_size),
nn.Tanh(),
nn.Linear(embedding_size, num_options)
)
self.terminations = nn.Linear(embedding_size, num_options)
if is_init:
self.apply(init_params)
def preprocess_obs(self, obs):
x = obs.image.transpose(1, 3).transpose(2, 3)
x = x.reshape(x.shape[0], -1)
return self.image_conv(x)
def forward(self, obs):
processed_obs = self.preprocess_obs(obs)
x = self.actor(processed_obs)
return x
def get_td_error(online_net, target_net, collected_experience, algorithm):
obs = collected_experience["obs"]
new_obs = collected_experience["new_obs"]
actions = collected_experience["actions"]
rewards = collected_experience["rewards"]
dones = collected_experience["dones"]
indices = np.arange(len(collected_experience["actions"]))
if algorithm == 'dqn':
q_target= target_net(new_obs)
q_policy = online_net(obs)
if algorithm == 'drqn':
hidden_state_target = target_net.init_hidden_state(training=True)
q_target, _ = target_net(new_obs, hidden_state_target)
hidden_state = online_net.init_hidden_state(training=True)
q_policy, _ = online_net(obs, hidden_state)
max_actions = q_target.max(dim=1)[0]
Q_target = torch.tensor(rewards, device=device) + discount * max_actions * torch.tensor(dones, device=device)
Q_policy = q_policy[indices, actions]
return abs(Q_target - Q_policy)
@classmethod
def get_option_td_error(cls, self, collected_experience):
obs = collected_experience["obs"]
new_obs = collected_experience["new_obs"]
options = collected_experience["options"]
done_masks = collected_experience["dones"]
rewards = np.array(list(map(float, collected_experience["rewards"])))
rewards_i = np.array(list(map(float, collected_experience["rewards_i"])))
indices = np.arange(len(obs))
rewards_i = rewards_i * self.rnd_scale
rewards = np.add(rewards, rewards_i)
next_Q_target = self.option_target_network(new_obs)
next_termination_probs = self.option_policy_network.get_terminations(new_obs).detach()
next_options_term_prob = torch.clip(next_termination_probs[indices, options], 0.1, 0.9)
torch_rewards = torch.tensor(np.array(rewards), device=device)
gt = torch_rewards + torch.tensor(np.array(done_masks), device=device) * discount * \
((1 - next_options_term_prob) * next_Q_target[indices, options] + next_options_term_prob * next_Q_target.max(dim=-1)[0])
Q_policy = self.option_policy_network(obs)
# compute loss
td_err = (Q_policy[indices, options] - gt.detach()).pow(2).mul(0.5).mean()
return td_err
@classmethod
def get_termination_loss_batch(cls, self, collected_experience):
obs = collected_experience["obs"]
options = collected_experience["options"]
done_masks = collected_experience["dones"]
done_masks = torch.tensor(done_masks, device=device)
indices = np.arange(len(obs))
option_term_prob = self.option_policy_network.get_terminations(obs)[indices, options]
Q_target = self.option_target_network(obs)
error_before_reg = Q_target[indices, options].detach() - Q_target.max(dim=1)[0].detach()
reg = - 1 * error_before_reg.mean()
termination_error = Q_target[indices, options].detach() - Q_target.max(dim=1)[0].detach() + reg
termination_loss = option_term_prob * termination_error * done_masks
return termination_loss.mean(), termination_error.mean()
def predict_option_termination(self, obs, current_option):
processed_obs = self.preprocess_obs(obs)
termination = self.terminations(processed_obs)[:, current_option].sigmoid()
sigmoid_termonations = [self.terminations(processed_obs)[:, o].sigmoid().item() for o in range(4)]
termination = torch.clip(termination, 0.1, 0.9)
sigmoid_termonations = np.clip([self.terminations(processed_obs)[:, o].sigmoid().item() for o in range(4)], 0.1, 0.9).tolist()
return termination, sigmoid_termonations
def get_terminations(self, obs):
processed_obs = self.preprocess_obs(obs)
return self.terminations(processed_obs).sigmoid()
def choice(self, options,probs):
x = np.random.rand()
cum = 0
for i,p in enumerate(probs):
cum += p
if x < cum:
break
return options[i]
def select_option(self, obs, exploration_options, ww):
processed_obs = self.preprocess_obs(obs)
Q = self.actor(processed_obs)
exploration_ratio = Q.tolist()[0]
exploration_ratio = softmax(exploration_ratio, ww)
# next_option = np.random.choice(
# len(exploration_options), 1, p=exploration_ratio
# )[0]
next_option = self.choice([0, 1, 2, 3], exploration_ratio)
return next_option