-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathmodel_adapt.py
178 lines (148 loc) · 6.64 KB
/
model_adapt.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
import sys
sys.path.insert(0, "composite")
import copy
import torch
import numpy as np
from typing import Optional
from models import Discriminator, RunningMeanStd, DiagonalPopArt
from models import ACModel as ACModelBase
class ACModel(torch.nn.Module):
def __init__(self, state_dim: int, act_dim: int, goal_dim: int=0, value_dim: int=1,
normalize_value: bool=True,
init_mu:Optional[torch.Tensor or float]=None, init_sigma:Optional[torch.Tensor or float]=None,
meta_goal_dim: int=0
):
super().__init__()
self.state_dim = state_dim
self.goal_dim_actor = goal_dim
self.goal_dim_critic = goal_dim + meta_goal_dim
self.actor = ACModelBase.Actor(state_dim, act_dim, self.goal_dim_actor, init_mu=init_mu, init_sigma=init_sigma)
self.critic = ACModelBase.Critic(state_dim, self.goal_dim_critic, value_dim)
self.actor_ob_normalizer = RunningMeanStd(state_dim, clamp=5.0)
self.critic_ob_normalizer = self.actor_ob_normalizer
self.ob_normalizer = [self.actor_ob_normalizer]
if isinstance(self.critic_ob_normalizer, torch.nn.ModuleList):
self.ob_normalizer.extend(self.critic_ob_normalizer)
if normalize_value:
self.value_normalizer = DiagonalPopArt(value_dim,
self.critic.mlp[-1].weight, self.critic.mlp[-1].bias)
else:
self.value_normalizer = None
def observe(self, obs, norm=True):
if self.goal_dim_critic > 0:
s = obs[:, :-self.goal_dim_critic]
g = obs[:, -self.goal_dim_critic:]
else:
s = obs
g = None
s = s.view(*s.shape[:-1], -1, self.state_dim)
return [normalizer(s) for normalizer in self.ob_normalizer] if norm else s, g
def eval_(self, s, seq_end_frame, g, unnorm):
v = self.critic(s[-1], seq_end_frame, g)
if unnorm and self.value_normalizer is not None:
v = self.value_normalizer(v, unnorm=True)
return v
def act(self, obs, seq_end_frame, stochastic=None, unnorm=False):
if stochastic is None:
stochastic = self.training
s, g = self.observe(obs)
pi = self.actor(s, seq_end_frame, None if g is None else g[:, :self.goal_dim_actor])
if stochastic:
a = pi.sample()
lp = pi.log_prob(a)
# if g is not None:
# g = g[...,:self.goal_dim_critic]
return a, self.eval_(s, seq_end_frame, g, unnorm), lp
else:
return pi.mean
def evaluate(self, obs, seq_end_frame, unnorm=False):
# no meta_goal passed with obs
# self.goal_dim, self.goal_dim_critic = self.goal_dim_critic, self.goal_dim
s, g = self.observe(obs)
# self.goal_dim, self.goal_dim_critic = self.goal_dim_critic, self.goal_dim
# if g is not None:
# g = g[...,:self.goal_dim_critic]
return self.eval_(s, seq_end_frame, g, unnorm)
def forward(self, obs, seq_end_frame, unnorm=False):
s, g = self.observe(obs)
pi = self.actor(s, seq_end_frame, None if g is None else g[:, :self.goal_dim_actor])
return pi, self.eval_(s, seq_end_frame, g, unnorm)
class MapCNN(torch.nn.Module):
def __init__(self, dropout=0):
super().__init__()
self.cnn = torch.nn.Sequential(
torch.nn.Conv2d(1, 32, 5, 5),
torch.nn.ReLU(inplace=True),
torch.nn.Conv2d(32, 64, 3),
torch.nn.ReLU(inplace=True),
torch.nn.Conv2d(64, 128, 3),
torch.nn.ReLU(inplace=True),
torch.nn.Dropout(dropout),
torch.nn.Conv2d(128, 256, 3),
torch.nn.Flatten()
)
def forward(self, m):
return self.cnn(m)
class AdaptNet(torch.nn.Module):
def __init__(self, meta_model, g_dim=0):
super().__init__()
actor_ob_normalizer = copy.deepcopy(meta_model.actor_ob_normalizer)
for normalizer in meta_model.ob_normalizer:
normalizer.reset_counter()
meta_model.ob_normalizer.insert(0, actor_ob_normalizer)
meta_policy = meta_model.actor
for n, p in meta_policy.named_parameters():
p.requires_grad = False
input_size = meta_policy.rnn.input_size
hidden_size = meta_policy.rnn.hidden_size
num_layers = meta_policy.rnn.num_layers
batch_first = meta_policy.rnn.batch_first
self.meta_policy = meta_policy
self.rnn = meta_policy.rnn.__class__(input_size, hidden_size, num_layers, batch_first=batch_first)
self.embed = torch.nn.Linear(meta_policy.mlp[0].in_features+g_dim, meta_policy.mlp[0].in_features)
ia_layer = lambda in_dim, out_dim: torch.nn.Linear(in_dim, out_dim)
self.ia = torch.nn.ModuleList([ia_layer(op.in_features, op.out_features)
if isinstance(op, torch.nn.Linear) else torch.nn.Identity() for op in meta_policy.mlp])
for e in self.ia:
if isinstance(e, torch.nn.Identity): continue
if isinstance(e, torch.nn.Sequential): e = e[-1]
for p in e.parameters():
torch.nn.init.zeros_(p)
for p, p_ in zip(self.rnn.parameters(), meta_policy.rnn.parameters()):
p.data.copy_(p_.data)
for p in self.embed.parameters():
torch.nn.init.zeros_(p)
self.g = None
def forward(self, s, seq_end_frame, g=None):
s, s_ = s[0], s[1]
n_inst = s.size(0)
if n_inst > self.meta_policy.all_inst.size(0):
self.meta_policy.all_inst = torch.arange(n_inst,
dtype=seq_end_frame.dtype, device=seq_end_frame.device)
ind = (self.meta_policy.all_inst[:n_inst], torch.clip(seq_end_frame, max=s.size(1)-1))
s_, _ = self.rnn(s_)
s_ = s_[ind]
s, _ = self.meta_policy.rnn(s)
s = s[ind]
if g is not None:
s = torch.cat((s, g), -1)
if self.g is None:
s_ = torch.cat((s_, g), -1)
else:
s_ = torch.cat((s_, g, self.g), -1)
elif self.g is not None:
s_ = torch.cat((s_, self.g), -1)
if isinstance(self.embed, torch.nn.ModuleList):
s_ = [embed(s_) for embed in self.embed]
else:
s_ = self.embed(s_)
s = s + s_
for j, op in enumerate(self.meta_policy.mlp):
embed = self.ia[j]
if isinstance(embed, torch.nn.Identity):
s = op(s)
else:
s = op(s) + embed(s)
mu = self.meta_policy.mu(s)
sigma = torch.exp(self.meta_policy.log_sigma(s)) + 1e-8
return torch.distributions.Normal(mu, sigma)