-
Notifications
You must be signed in to change notification settings - Fork 31
/
agents.py
147 lines (119 loc) · 5.74 KB
/
agents.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
import numpy as np
import torch.nn.functional as F
import torch.nn as nn
import torch
import torch.optim as optim
from torch.distributions.categorical import Categorical
from model import CnnActorCriticNetwork, ICMModel
class ICMAgent(object):
def __init__(
self,
input_size,
output_size,
num_env,
num_step,
gamma,
lam=0.95,
learning_rate=1e-4,
ent_coef=0.01,
clip_grad_norm=0.5,
epoch=3,
batch_size=128,
ppo_eps=0.1,
eta=0.01,
use_gae=True,
use_cuda=False,
use_noisy_net=False):
self.model = CnnActorCriticNetwork(input_size, output_size, use_noisy_net)
self.num_env = num_env
self.output_size = output_size
self.input_size = input_size
self.num_step = num_step
self.gamma = gamma
self.lam = lam
self.epoch = epoch
self.batch_size = batch_size
self.use_gae = use_gae
self.ent_coef = ent_coef
self.eta = eta
self.ppo_eps = ppo_eps
self.clip_grad_norm = clip_grad_norm
self.device = torch.device('cuda' if use_cuda else 'cpu')
self.icm = ICMModel(input_size, output_size, use_cuda)
self.optimizer = optim.Adam(list(self.model.parameters()) + list(self.icm.parameters()),
lr=learning_rate)
self.icm = self.icm.to(self.device)
self.model = self.model.to(self.device)
def get_action(self, state):
state = torch.Tensor(state).to(self.device)
state = state.float()
policy, value = self.model(state)
action_prob = F.softmax(policy, dim=-1).data.cpu().numpy()
action = self.random_choice_prob_index(action_prob)
return action, value.data.cpu().numpy().squeeze(), policy.detach()
@staticmethod
def random_choice_prob_index(p, axis=1):
r = np.expand_dims(np.random.rand(p.shape[1 - axis]), axis=axis)
return (p.cumsum(axis=axis) > r).argmax(axis=axis)
def compute_intrinsic_reward(self, state, next_state, action):
state = torch.FloatTensor(state).to(self.device)
next_state = torch.FloatTensor(next_state).to(self.device)
action = torch.LongTensor(action).to(self.device)
action_onehot = torch.FloatTensor(
len(action), self.output_size).to(
self.device)
action_onehot.zero_()
action_onehot.scatter_(1, action.view(len(action), -1), 1)
real_next_state_feature, pred_next_state_feature, pred_action = self.icm(
[state, next_state, action_onehot])
intrinsic_reward = self.eta * F.mse_loss(real_next_state_feature, pred_next_state_feature, reduction='none').mean(-1)
return intrinsic_reward.data.cpu().numpy()
def train_model(self, s_batch, next_s_batch, target_batch, y_batch, adv_batch, old_policy):
s_batch = torch.FloatTensor(s_batch).to(self.device)
next_s_batch = torch.FloatTensor(next_s_batch).to(self.device)
target_batch = torch.FloatTensor(target_batch).to(self.device)
y_batch = torch.LongTensor(y_batch).to(self.device)
adv_batch = torch.FloatTensor(adv_batch).to(self.device)
sample_range = np.arange(len(s_batch))
ce = nn.CrossEntropyLoss()
forward_mse = nn.MSELoss()
with torch.no_grad():
policy_old_list = torch.stack(old_policy).permute(1, 0, 2).contiguous().view(-1, self.output_size).to(
self.device)
m_old = Categorical(F.softmax(policy_old_list, dim=-1))
log_prob_old = m_old.log_prob(y_batch)
# ------------------------------------------------------------
for i in range(self.epoch):
np.random.shuffle(sample_range)
for j in range(int(len(s_batch) / self.batch_size)):
sample_idx = sample_range[self.batch_size * j:self.batch_size * (j + 1)]
# --------------------------------------------------------------------------------
# for Curiosity-driven
action_onehot = torch.FloatTensor(self.batch_size, self.output_size).to(self.device)
action_onehot.zero_()
action_onehot.scatter_(1, y_batch[sample_idx].view(-1, 1), 1)
real_next_state_feature, pred_next_state_feature, pred_action = self.icm(
[s_batch[sample_idx], next_s_batch[sample_idx], action_onehot])
inverse_loss = ce(
pred_action, y_batch[sample_idx])
forward_loss = forward_mse(
pred_next_state_feature, real_next_state_feature.detach())
# ---------------------------------------------------------------------------------
policy, value = self.model(s_batch[sample_idx])
m = Categorical(F.softmax(policy, dim=-1))
log_prob = m.log_prob(y_batch[sample_idx])
ratio = torch.exp(log_prob - log_prob_old[sample_idx])
surr1 = ratio * adv_batch[sample_idx]
surr2 = torch.clamp(
ratio,
1.0 - self.ppo_eps,
1.0 + self.ppo_eps) * adv_batch[sample_idx]
actor_loss = -torch.min(surr1, surr2).mean()
critic_loss = F.mse_loss(
value.sum(1), target_batch[sample_idx])
entropy = m.entropy().mean()
self.optimizer.zero_grad()
loss = (actor_loss + 0.5 * critic_loss - 0.001 * entropy) + forward_loss + inverse_loss
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
self.optimizer.step()