-
Notifications
You must be signed in to change notification settings - Fork 0
/
inverse_model.py
95 lines (83 loc) · 3.12 KB
/
inverse_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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
class FWModel(nn.Module):
def __init__(self, cnf, delta=False):
super().__init__()
self.delta = delta
self.device = torch.device("cuda" if cnf.main.gpu else "cpu")
self.hidden = 256
self.linear = nn.Linear(14, self.hidden)
self.linear2 = nn.Linear(self.hidden, self.hidden)
self.head = nn.Linear(self.hidden, 7)
self.opt = optim.Adam(self.parameters(), lr=0.001)
self.to(self.device)
def forward(self, x, y):
"""predicts state x action -> next state / delta next state
x: state
y: action
"""
if x.dim() > 1:
x = torch.cat([x, y], dim=1)
else:
x = torch.cat([x, y], dim=0)
x = F.relu(self.linear(x))
x = F.relu(self.linear2(x))
return self.head(x)
def train(self, states, nstates, actions, eval=False):
actions = torch.tensor(actions).float().to(self.device)
this_state = torch.tensor(states).float().to(self.device)
next_state = torch.tensor(nstates).float().to(self.device)
if self.delta:
nstates = nstates - states
self.opt.zero_grad()
predicted_nstates = self.forward(states, actions)
loss = F.mse_loss(predicted_nstates, nstates)
if not eval:
loss.backward
self.opt.step()
return loss
class IVModel(nn.Module):
def __init__(self, cnf, depth, act=F.relu, delta=False):
super().__init__()
self.delta = delta
self.device = torch.device("cuda" if cnf.main.gpu else "cpu")
self.hidden = 256
self.act = act
self.depth = depth
self.linear = nn.Linear(cnf.icm.embedding_size * 2, self.hidden)
self.linear2 = nn.Linear(self.hidden, self.hidden)
self.linear3 = nn.Linear(self.hidden, self.hidden)
self.linear4 = nn.Linear(self.hidden, self.hidden)
self.head = nn.Linear(self.hidden, cnf.env.action_dim)
self.opt = optim.Adam(self.parameters(), lr=0.001)
self.to(self.device)
def forward(self, x, y):
if self.delta:
y = y - x
if x.dim() > 1:
x = torch.cat([x, y], dim=1)
else:
x = torch.cat([x, y], dim=0)
x = self.act(self.linear(x))
if self.depth > 1:
x = self.act(self.linear2(x))
if self.depth > 2:
x = self.act(self.linear3(x))
if self.depth > 3:
x = self.act(self.linear3(x))
return self.head(x)
def train(self, states, nstates, actions, eval=False):
if self.delta:
nstates = nstates - states
actions = torch.tensor(actions).float().to(self.device)
this_state = torch.tensor(states).float().to(self.device)
next_state = torch.tensor(nstates).float().to(self.device)
self.opt.zero_grad()
predicted_action = self.forward(this_state, next_state)
loss = F.mse_loss(predicted_action, actions)
if not eval:
loss.backward()
self.opt.step()
return loss