-
Notifications
You must be signed in to change notification settings - Fork 1
/
multistep_dynamics.py
261 lines (203 loc) · 7.7 KB
/
multistep_dynamics.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
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
# Finite Trajectory Similarity Code
"""Transition dynamics for the agent in original MDP."""
from typing import List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
import multistep_utils as utils
import numpy as np
from torch.nn import BatchNorm1d
import time
class MultiStepDynamicsModel(nn.Module):
def __init__(self,
obs_dim,
action_dim,
xu_enc_hidden_dim,
x_dec_hidden_dim,
rec_latent_dim,
rec_num_layers,
clip_grad_norm=0.2,
xu_enc_hidden_depth=2,
x_dec_hidden_depth=2,
rec_type='LSTM'
):
super().__init__()
self.obs_dim = obs_dim
self.action_dim = action_dim
self.clip_grad_norm = clip_grad_norm
# Manually freeze the goal locations
self.freeze_dims = None
self.rec_type = rec_type
self.rec_num_layers = rec_num_layers
self.rec_latent_dim = rec_latent_dim
self.xu_enc = utils.mlp(
obs_dim+action_dim, xu_enc_hidden_dim, rec_latent_dim, xu_enc_hidden_depth)
self.x_dec = utils.mlp(
rec_latent_dim, x_dec_hidden_dim, obs_dim, x_dec_hidden_depth)
self.apply(utils.weight_init) # Don't apply this to the recurrent unit.
mods = [self.xu_enc, self.x_dec]
if rec_num_layers > 0:
if rec_type == 'LSTM':
self.rec = nn.LSTM(
rec_latent_dim, rec_latent_dim, num_layers=rec_num_layers)
elif rec_type == 'GRU':
self.rec = nn.GRU(
rec_latent_dim, rec_latent_dim, num_layers=rec_num_layers)
else:
assert False
mods.append(self.rec)
params = utils.get_params(mods)
def __getstate__(self):
d = self.__dict__
return d
def __setstate__(self, d):
self.__dict__ = d
self.rec.flatten_parameters()
def init_hidden_state(self, init_x):
assert init_x.dim() == 2
n_batch = init_x.size(0)
if self.rec_type == 'LSTM':
h = torch.zeros(
self.rec_num_layers, n_batch, self.rec_latent_dim, device=init_x.device)
c = torch.zeros_like(h)
h = (h, c)
elif self.rec_type == 'GRU':
h = torch.zeros(
self.rec_num_layers, n_batch, self.rec_latent_dim, device=init_x.device)
else:
assert False
return h
def unroll(self, x, us, detach_xt=False):
assert x.dim() == 2
assert us.dim() == 3
n_batch = x.size(0)
assert us.size(1) == n_batch
if self.rec_num_layers > 0:
h = self.init_hidden_state(x)
pred_xs = []
xt = x
for t in range(us.size(0)):
ut = us[t]
if detach_xt:
xt = xt.detach()
xut = torch.cat((xt, ut), dim=1)
xu_emb = self.xu_enc(xut).unsqueeze(0)
if t==0:
xu_emb_1 = xu_emb.squeeze(0)
if self.rec_num_layers > 0:
xtp1_emb, h = self.rec(xu_emb, h)
else:
xtp1_emb = xu_emb
xtp1 = xt + self.x_dec(xtp1_emb.squeeze(0))
pred_xs.append(xtp1)
xt = xtp1
pred_xs = torch.stack(pred_xs)
h_states, _ = h
return pred_xs, h_states[-1]
def forward(self, x, us):
return self.unroll(x, us)
class MultiStepPixelDynamicsModel(nn.Module):
def __init__(self,
obs_shape,
action_shape,
obs_hidden_dim,
xu_enc_hidden_dim,
x_dec_hidden_dim,
rec_latent_dim,
rec_num_layers,
clip_grad_norm=0.2,
xu_enc_hidden_depth=2,
x_dec_hidden_depth=2,
rec_type='LSTM'
):
super().__init__()
self.obs_shape = obs_shape
self.action_shape = action_shape
self.obs_hidden_dim = obs_hidden_dim
self.clip_grad_norm = clip_grad_norm
# Manually freeze the goal locations
self.freeze_dims = None
self.rec_type = rec_type
self.rec_num_layers = rec_num_layers
self.rec_latent_dim = rec_latent_dim
self.obs_enc = utils.conv_mlp_encoder(
obs_shape, obs_hidden_dim, 2)
self.obs_enc_fc = nn.Linear(32 * 80 * 80, obs_hidden_dim)
self.obs_enc_ln = nn.LayerNorm(obs_hidden_dim)
self.xu_enc = utils.mlp(
obs_hidden_dim+action_shape[0], xu_enc_hidden_dim, rec_latent_dim, xu_enc_hidden_depth)
self.x_dec = utils.mlp(
rec_latent_dim, x_dec_hidden_dim, obs_hidden_dim, x_dec_hidden_depth)
self.obs_dec_fc = nn.Linear(obs_hidden_dim, 32 * 9 * 9)
self.obs_dec = utils.conv_mlp_decoder(
obs_shape, obs_hidden_dim, 2)
self.latent_fc = utils.mlp(
rec_latent_dim+action_shape[0], xu_enc_hidden_dim, rec_latent_dim, xu_enc_hidden_depth)
self.apply(utils.weight_init) # Don't apply this to the recurrent unit.
mods = [self.obs_enc, self.obs_enc_fc, self.obs_enc_ln, self.xu_enc, self.x_dec, self.obs_dec_fc, self.obs_dec, self.latent_fc]
if rec_num_layers > 0:
if rec_type == 'LSTM':
self.rec = nn.LSTM(
rec_latent_dim, rec_latent_dim, num_layers=rec_num_layers)
elif rec_type == 'GRU':
self.rec = nn.GRU(
rec_latent_dim, rec_latent_dim, num_layers=rec_num_layers)
else:
assert False
mods.append(self.rec)
params = utils.get_params(mods)
def __getstate__(self):
d = self.__dict__
return d
def __setstate__(self, d):
self.__dict__ = d
self.rec.flatten_parameters()
def init_hidden_state(self, init_x):
n_batch = init_x.size(0)
if self.rec_type == 'LSTM':
h = torch.zeros(
self.rec_num_layers, n_batch, self.rec_latent_dim, device=init_x.device)
c = torch.zeros_like(h)
h = (h, c)
elif self.rec_type == 'GRU':
h = torch.zeros(
self.rec_num_layers, n_batch, self.rec_latent_dim, device=init_x.device)
else:
assert False
return h
def unroll(self, x, us, detach_xt=False):
n_batch = x.size(0)
if self.rec_num_layers > 0:
h = self.init_hidden_state(x)
pred_xs = []
xt = x
for t in range(us.size(0)):
ut = us[t]
if detach_xt:
xt = xt.detach()
xt = self.obs_enc(xt / 255)
xt = xt.view(xt.size(0), -1)
xt = self.obs_enc_fc(xt)
xt = torch.relu(xt)
xt = self.obs_enc_ln(xt)
xut = torch.cat((xt, ut), dim=1)
xu_emb = self.xu_enc(xut).unsqueeze(0)
if t==0:
xu_emb_1 = xu_emb.squeeze(0)
if self.rec_num_layers > 0:
h_tm1 = h
xtp1_emb, h = self.rec(xu_emb, h_tm1)
else:
xtp1_emb = xu_emb
xtp1 = xt + self.x_dec(xtp1_emb.squeeze(0))
xtp1_dec = self.obs_dec_fc(xtp1)
xtp1_dec = self.obs_dec(xtp1_dec.view(-1, 32, 9, 9))
pred_xs.append(xtp1_dec)
xt = xtp1_dec
pred_xs = torch.stack(pred_xs)
h_tm1_states, _ = h_tm1
h_states, _ = h
hpred_t = self.latent_fc(torch.cat((h_tm1_states[-2],ut),dim=1))
return pred_xs, h_states[-1], hpred_t
def forward(self, x, us):
return self.unroll(x, us)