-
Notifications
You must be signed in to change notification settings - Fork 0
/
roa_nn.py
380 lines (341 loc) · 15.2 KB
/
roa_nn.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
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
import torch
import torch.nn as nn
import roa_utils as utils
import cgw_sim_choi as choi
def lc_forward(fc_list, x):
num_layers = len(fc_list)
for i, hidden in enumerate(range(num_layers - 1)):
x = nn.ReLU()(fc_list[i](x))
v = fc_list[-1](x)
return v
class Net(nn.Module):
def __init__(self, args):
super(Net, self).__init__()
self.args = args
self.linear_list = utils.create_fcs(args.N_REF, 1, args.net_hiddens)
def embedding(self, ref):
if self.args.X_DIM < 5:
new_ref = torch.stack([
ref[:, 0], # mu factor
ref[:, 1], # wref
], dim=-1)
else:
new_ref = torch.stack([
ref[:, 0], # mu factor
(ref[:, 1] - 17.5) / 12.5, # vref
ref[:, 2], # wref
], dim=-1)
return new_ref
def forward(self, ref):
level = lc_forward(self.linear_list, self.embedding(ref))
if self.args.net_exp:
level = torch.exp(level)
return level
class CLF(nn.Module):
def __init__(self, lqr_P, args):
super(CLF, self).__init__()
self.args = args
self.X_DIM = X_DIM = args.X_DIM
self.N_REF = N_REF = args.N_REF
if self.args.clf_mode == "ellipsoid":
if self.args.multi:
if self.args.clf_e2e:
self.linear_list = utils.create_fcs(X_DIM + N_REF, 1, args.clf_hiddens)
else:
self.linear_list = utils.create_fcs(N_REF, X_DIM * X_DIM, args.clf_hiddens)
if self.args.clf_ell_scale:
self.linear_list2 = utils.create_fcs(N_REF, X_DIM, args.clf_hiddens)
elif self.args.clf_ell_extra:
self.linear_list2 = utils.create_fcs(X_DIM + N_REF, 1, args.clf_hiddens)
else:
if self.args.clf_init_type == "eye":
P_init = torch.eye(X_DIM) / 2
elif self.args.clf_init_type == "lqr":
P_init = lqr_P / 2
elif self.args.clf_init_type == "rand":
P_init = torch.rand(X_DIM, X_DIM) / 2
else:
raise NotImplementedError
self.P = torch.nn.Parameter(P_init)
elif self.args.clf_mode == "diag":
if self.args.multi:
self.linear_list = utils.create_fcs(N_REF, X_DIM, args.clf_hiddens)
else:
self.scales = torch.nn.Parameter(torch.ones(X_DIM, ))
elif self.args.clf_mode == "nn":
if self.args.multi:
self.linear_list = utils.create_fcs(X_DIM + N_REF, 1, args.clf_hiddens)
else:
self.linear_list = utils.create_fcs(X_DIM, 1, args.clf_hiddens)
else:
raise NotImplementedError
if self.args.normalize:
assert self.X_DIM <=5
# in_means (xd, y, theta, F)
# (xd, y, xd_ref, y_ref) -> 1
self.in_means = torch.cat([args.in_means[:2], args.in_means[:2]], dim=-1)
self.in_stds = torch.cat([args.in_stds[:2], args.in_stds[:2]], dim=-1)
self.out_means = torch.zeros_like(args.in_means[:1])
self.out_stds =torch.ones_like(args.in_means[:1])
def update_device(self):
if self.args.normalize:
self.in_means = self.in_means.cuda() # .to(device)
self.in_stds = self.in_stds.cuda() # .to(device)
self.out_means = self.out_means.cuda() # .to(device)
self.out_stds = self.out_stds.cuda() # .to(device)
def embedding(self, ref):
if self.X_DIM>5:
if hasattr(self.args, "single_clf") and self.args.single_clf:
new_ref = ref * 0.0
else:
new_ref = torch.stack([
ref[:, 0], # mu factor
(ref[:, 1] - 17.5) / 12.5, # vref
ref[:, 2], # wref
], dim=-1)
else:
new_ref = torch.stack([
ref[:, 0], # xdot
ref[:, 1], # y
], dim=-1)
return new_ref
def forward(self, x):
if len(x.shape)==3:
reshape = True
N, T, K = x.shape
x = x.reshape([N*T, K])
else:
reshape = False
if self.args.normalize:
x = (x - self.in_means) / self.in_stds
origin_x = x[:, : self.X_DIM + self.N_REF]
ref = x[:, self.X_DIM: self.X_DIM + self.N_REF]
x = x[:, :self.X_DIM]
batch_size = x.shape[0]
if self.args.clf_mode == "ellipsoid":
if self.args.multi:
if self.args.clf_e2e:
v = lc_forward(self.linear_list, origin_x) # (B, 1)
v = v * v
else:
P = lc_forward(self.linear_list, self.embedding(ref)).reshape(batch_size, self.X_DIM, self.X_DIM)
if self.args.clf_ell_scale:
scale = lc_forward(self.linear_list2, self.embedding(ref)).reshape(batch_size, self.X_DIM, 1)
scale = torch.exp(scale)
Px = (scale * torch.bmm(P, torch.unsqueeze(x, dim=-1))).squeeze(-1) # (B, N, 1)
elif self.args.clf_ell_extra:
Px = torch.bmm(P, torch.unsqueeze(x, dim=-1)).squeeze(-1) # (B, N, 1)
extra = lc_forward(self.linear_list2, torch.cat([x, self.embedding(ref)], dim=-1))
if self.args.clf_ell_extra_tanh:
extra = nn.Tanh()(extra)
if self.args.clf_ell_extra_xnorm:
extra = extra * torch.norm(x, dim=-1, keepdim=True)
else:
if self.args.X_DIM<5:
Px = torch.bmm(P, torch.unsqueeze(x-ref, dim=-1)).squeeze(-1) # (B, N, 1)
if self.args.clf_ell_eye:
extra = torch.norm(x-ref, dim=-1, keepdim=True)
else:
Px = torch.bmm(P, torch.unsqueeze(x, dim=-1)).squeeze(-1) # (B, N, 1)
Px_norm = torch.norm(Px, dim=-1, keepdim=True) # (B, 1)
v = Px_norm * Px_norm / 2 # (B, 1)
if self.args.clf_ell_extra:
v = v + extra
elif self.args.clf_ell_eye:
v = v + extra * self.args.clf_eye_scale
else:
PP = self.P + self.P.t()
v = torch.bmm(torch.unsqueeze(x @ PP, dim=1), x.unsqueeze(dim=-1)).squeeze(dim=-1) / 2
elif self.args.clf_mode == "diag":
if self.args.multi:
if self.args.X_DIM < 5:
v = torch.norm(x-ref, dim=-1, keepdim=True)
v = v ** 2
else:
scale = lc_forward(self.linear_list, self.embedding(ref)).reshape(batch_size, self.X_DIM)
if self.args.clf_no_exp_scale == False:
scale = torch.exp(scale)
v = torch.norm(scale * x, dim=-1, keepdim=True)
v = v ** 2
else:
v_norm = torch.norm(x * self.scales, dim=-1, keepdim=True)
v = v_norm ** 2
else:
assert self.args.clf_mode == "nn"
if self.args.multi:
v = lc_forward(self.linear_list, origin_x)
else:
v = lc_forward(self.linear_list, x)
v = v ** 2
if reshape:
v = v.reshape(N, T, 1)
return v
class Actor(nn.Module):
def __init__(self, lqr_K, args):
super(Actor, self).__init__()
self.args = args
self.X_DIM = X_DIM = args.X_DIM
self.N_REF = N_REF = args.N_REF
self.U_DIM = U_DIM = args.U_DIM
if self.args.actor_mode == "mat" or self.args.actor_mode == "mat_nn":
if self.args.multi:
if self.args.actor_e2e:
self.linear_list = utils.create_fcs(X_DIM + N_REF, U_DIM, args.actor_hiddens)
else:
self.linear_list = utils.create_fcs(N_REF, U_DIM * X_DIM, args.actor_hiddens)
else:
if self.args.actor_init_type == "lqr":
K_init = lqr_K.float()
elif self.args.actor_init_type == "rand":
K_init = torch.random(U_DIM, X_DIM)
else:
raise NotImplementedError
self.K = torch.nn.Parameter(K_init)
if self.args.actor_mode == "nn" or self.args.actor_mode == "mat_nn":
if self.args.multi:
if self.args.actor_e2e == False:
self.linear_list2 = utils.create_fcs(X_DIM + N_REF, U_DIM, args.actor_hiddens)
else:
self.linear_list = utils.create_fcs(X_DIM + N_REF, U_DIM, args.actor_hiddens)
else:
self.linear_list = utils.create_fcs(X_DIM, U_DIM, args.actor_hiddens)
if self.args.normalize:
assert self.X_DIM <= 5
# in_means (xd, y, theta, F)
# out_means (x, xd, y)
# (xd, y, xd_ref, y_ref) -> (theta, F)
self.in_means = torch.cat([args.in_means[:2], args.in_means[:2]], dim=-1)
self.in_stds = torch.cat([args.in_stds[:2], args.in_stds[:2]], dim=-1)
self.out_means = args.out_means[1:]
self.out_stds = args.out_stds[1:]
# self.out_means = torch.zeros_like(args.out_means[1:])
# self.out_stds = torch.tensor([1.0, 2000])
# self.out_stds = torch.tensor([0.01, 1])
def update_device(self):
if self.args.normalize:
self.in_means = self.in_means.cuda() # .to(device)
self.in_stds = self.in_stds.cuda() # .to(device)
self.out_means = self.out_means.cuda() # .to(device)
self.out_stds = self.out_stds.cuda() # .to(device)
def embedding(self, ref):
if self.X_DIM > 5:
new_ref = torch.stack([
ref[:, 0], # mu factor
(ref[:, 1] - 17.5) / 12.5, # vref
ref[:, 2], # wref
], dim=-1)
else:
new_ref = torch.stack([
ref[:, 0], # xdot
ref[:, 1], # y
], dim=-1)
return new_ref
def forward(self, x, return_K=False):
batch_size = x.shape[0]
if self.args.normalize:
x = (x - self.in_means) / self.in_stds
origin_x = x[:, :self.X_DIM + self.N_REF]
ref = x[:, self.X_DIM: self.X_DIM + self.N_REF]
x = x[:, :self.X_DIM]
if self.args.multi and self.args.actor_e2e:
u_e2e = lc_forward(self.linear_list, origin_x)
u_e2e = torch.stack([
torch.clamp(u_e2e[:, 0], -self.args.tanh_w_gain, self.args.tanh_w_gain),
torch.clamp(u_e2e[:, 1], -self.args.tanh_a_gain, self.args.tanh_a_gain)
], dim=-1)
u_all = u_e2e
if self.args.actor_mode == "mat" or self.args.actor_mode == "mat_nn":
if self.args.multi:
if self.args.actor_e2e == False:
K_mat = lc_forward(self.linear_list, self.embedding(ref)).reshape(batch_size, self.U_DIM, self.X_DIM)
u_mat = torch.bmm(-K_mat, x.unsqueeze(-1)).squeeze(-1)
else:
raise NotImplementedError
else:
u_mat = torch.mm(x, -self.K.T)
if self.args.actor_clip_u:
u_mat = torch.stack([
torch.clamp(u_mat[:, 0], -self.args.tanh_w_gain, self.args.tanh_w_gain),
torch.clamp(u_mat[:, 1], -self.args.tanh_a_gain, self.args.tanh_a_gain)
], dim=-1)
if self.args.actor_mode == "nn" or self.args.actor_mode == "mat_nn":
if self.args.multi:
if self.args.actor_e2e == False:
u_nn = lc_forward(self.linear_list2, origin_x)
else:
u_nn = lc_forward(self.linear_list, origin_x)
else:
u_nn = lc_forward(self.linear_list, x)
u_nn = nn.Tanh()(u_nn)
u_nn = torch.stack([
u_nn[:, 0] * self.args.tanh_w_gain,
u_nn[:, 1] * self.args.tanh_a_gain], dim=-1)
if self.args.actor_mode == "mat":
u_all = u_mat
elif self.args.actor_mode == "nn":
u_all = u_nn
elif self.args.actor_mode == "mat_nn":
ratio = self.args.actor_nn_res_ratio
u_all = u_mat * (1-ratio) + u_nn * ratio
if self.args.normalize:
if self.args.u_norm==False:
u_all = u_all * self.out_stds + self.out_means
if return_K:
return u_all, K_mat
else:
return u_all
class CGW_Actor(nn.Module):
def __init__(self, args):
super(CGW_Actor, self).__init__()
self.args = args
self.X_DIM = X_DIM = args.X_DIM
self.N_REF = N_REF = args.N_REF
self.U_DIM = U_DIM = args.U_DIM
self.linear_list = utils.create_fcs(X_DIM + N_REF, U_DIM, args.actor_hiddens)
def forward(self, x):
u = lc_forward(self.linear_list, x)
u_nn = nn.Tanh()(u) * self.args.u_gain
prev_th = self.args.params.th1d
self.args.params.th1d = x[:, -1]
u_qp = choi.two_link_io_control_clf_qp(x[:, :-1],
self.args.params, use_torch=True, qp_bound=self.args.qp_bound)
self.args.params.th1d = prev_th
u = u_nn * self.args.nn_ratio + u_qp * (1 - self.args.nn_ratio)
return u
class CGW_CLF(nn.Module):
def __init__(self, args):
super(CGW_CLF, self).__init__()
self.args = args
self.X_DIM = X_DIM = args.X_DIM
self.N_REF = N_REF = args.N_REF
self.U_DIM = U_DIM = args.U_DIM
self.gait_data = args.gait_data
self.gait_th = args.gait_th.unsqueeze(0) # (1, n_th)
self.linear_list = utils.create_fcs(X_DIM + N_REF + X_DIM, U_DIM, args.clf_hiddens)
# TODO (get ref)
def get_ref(self, x, num_ths):
th_tar = x[:, -1].unsqueeze(1) # (n_samples, 1)
th_idx = torch.argmin(torch.abs(th_tar - self.gait_th), dim=1) # (n_samples, )
ref_list = []
assert x.shape[0] % num_ths == 0
n_batch = x.shape[0] // num_ths
for i in range(num_ths):
bi0 = i * n_batch
bi1 = (i + 1) * n_batch
gait_data = self.gait_data[th_idx[bi0]].unsqueeze(0) # (1, gait_len, 4)
gait_idx = torch.argmin(torch.abs(gait_data[:, :, 0]-x[bi0:bi1, 0].unsqueeze(1)), dim=1) # (n_batch, )
ref = gait_data[0, gait_idx]
ref_list.append(ref)
ref_list = torch.cat(ref_list, dim=0)
# TODO(check here)
return ref_list
def forward(self, x, num_ths, ref=None):
if ref is None:
ref = self.get_ref(x, num_ths)
# for i in range(ref.shape[0]):
# print("%04d %.3f %.3f %.3f %.3f"%(i, ref[i, 0], ref[i, 1], ref[i, 2], ref[i, 3]))
x_err = x[:, :-1] - ref
new_x = torch.cat([x, x_err], dim=-1) # (n_samples, 9)
# TODO use a Q^TQ format?
v = lc_forward(self.linear_list, new_x)
return v**2