forked from orrivlin/Navigation-HER
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dqn_HER.py
165 lines (145 loc) · 5.92 KB
/
dqn_HER.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
"""
@author: orrivlin
"""
import torch
import numpy as np
import copy
import torch.nn.functional as F
from collections import deque
from Models import ConvNet
import random
from log_utils import logger, mean_val
from HER import HER
from copy import deepcopy as dc
class DQN_HER:
def __init__(self, env, gamma, buffer_size, ddqn):
self.env = env
[Sdim,Adim] = env.get_dims()
self.model = ConvNet(Sdim[0],Sdim[0],3,Adim).cuda()
self.target_model = copy.deepcopy(self.model).cuda()
self.her = HER()
self.gamma = gamma
self.optimizer = torch.optim.Adam(self.model.parameters(),lr=0.0001)
self.batch_size = 16
self.epsilon = 0.1
self.buffer_size = buffer_size
self.step_counter = 0
self.epsi_high = 0.9
self.epsi_low = 0.1
self.steps = 0
self.count = 0
self.decay = 2000
self.eps = self.epsi_high
self.update_target_step = 3000
self.log = logger()
self.log.add_log('tot_return')
self.log.add_log('avg_loss')
self.log.add_log('final_dist')
self.log.add_log('buffer')
self.image_mean = 0
self.image_std = 0
self.ddqn = ddqn
self.replay_buffer = deque(maxlen=buffer_size)
def run_episode(self):
self.her.reset()
obs, done = self.env.reset()
done = False
state = self.env.get_tensor(obs)
sum_r = 0
mean_loss = mean_val()
min_dist = 100000
max_t = 50
for t in range(max_t):
self.steps += 1
self.eps = self.epsi_low + (self.epsi_high-self.epsi_low) * (np.exp(-1.0 * self.steps/self.decay))
Q = self.model(self.norm(state.cuda()))
num = np.random.rand()
if (num < self.eps):
action = torch.randint(0,Q.shape[1],(1,)).type(torch.LongTensor)
else:
action = torch.argmax(Q,dim=1)
new_obs, reward, done, dist = self.env.step(obs,action.item())
new_state = self.env.get_tensor(new_obs)
sum_r = sum_r + reward
if dist < min_dist:
min_dist = dist
if (t+1) == max_t:
done = True
self.replay_buffer.append([dc(state.squeeze(0).numpy()),dc(action),dc(reward),dc(new_state.squeeze(0).numpy()),dc(done)])
self.her.keep([state.squeeze(0).numpy(),action,reward,new_state.squeeze(0).numpy(),done])
loss = self.update_model()
mean_loss.append(loss)
state = dc(new_state)
obs = dc(new_obs)
self.step_counter = self.step_counter + 1
if (self.step_counter > self.update_target_step):
self.target_model.load_state_dict(self.model.state_dict())
self.step_counter = 0
print('updated target model')
her_list = self.her.backward()
for item in her_list:
self.replay_buffer.append(item)
self.log.add_item('tot_return',sum_r)
self.log.add_item('avg_loss',mean_loss.get())
self.log.add_item('final_dist',min_dist)
def gather_data(self):
self.her.reset()
obs, done = self.env.reset()
done = False
state = self.env.get_tensor(obs)
sum_r = 0
min_dist = 100000
max_t = 50
for t in range(max_t):
self.eps = 1.0
Q = self.model(state.cuda())
num = np.random.rand()
if (num < self.eps):
action = torch.randint(0,Q.shape[1],(1,)).type(torch.LongTensor)
else:
action = torch.argmax(Q,dim=1)
new_obs, reward, done, dist = self.env.step(obs,action.item())
new_state = self.env.get_tensor(new_obs)
sum_r = sum_r + reward
if dist < min_dist:
min_dist = dist
if (t+1) == max_t:
done = True
self.replay_buffer.append([dc(state.squeeze(0).numpy()),dc(action),dc(reward),dc(new_state.squeeze(0).numpy()),dc(done)])
state = dc(new_state)
obs = dc(new_obs)
return min_dist
def calc_norm(self):
S0, A0, R1, S1, D1 = zip(*self.replay_buffer)
S0 = torch.tensor( S0, dtype=torch.float)
self.image_mean = S0.mean(dim=0).cuda()
self.image_std = S0.std(dim=0).cuda()
def norm(self,state):
return state - self.image_mean
def update_model(self):
self.optimizer.zero_grad()
num = len(self.replay_buffer)
K = np.min([num,self.batch_size])
samples = random.sample(self.replay_buffer, K)
S0, A0, R1, S1, D1 = zip(*samples)
S0 = torch.tensor( S0, dtype=torch.float)
A0 = torch.tensor( A0, dtype=torch.long).view(K, -1)
R1 = torch.tensor( R1, dtype=torch.float).view(K, -1)
S1 = torch.tensor( S1, dtype=torch.float)
D1 = torch.tensor( D1, dtype=torch.float)
S0 = self.norm(S0.cuda())
S1 = self.norm(S1.cuda())
if self.ddqn == True:
model_next_acts = self.model(S1).detach().max(dim=1)[1]
target_q = R1.squeeze().cuda() + self.gamma*self.target_model(S1).gather(1,model_next_acts.unsqueeze(1)).squeeze()*(1 - D1.cuda())
else:
target_q = R1.squeeze().cuda() + self.gamma*self.target_model(S1).max(dim=1)[0].detach()*(1 - D1.cuda())
policy_q = self.model(S0).gather(1,A0.cuda())
L = F.smooth_l1_loss(policy_q.squeeze(),target_q.squeeze())
L.backward()
self.optimizer.step()
return L.detach().item()
def run_epoch(self):
self.run_episode()
self.log.add_item('buffer',len(self.replay_buffer))
return self.log