-
Notifications
You must be signed in to change notification settings - Fork 0
/
ddp_agent.py
377 lines (318 loc) · 18 KB
/
ddp_agent.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
import torch
import torch.nn as nn
import numpy as np
import random
from torch import FloatTensor as FT
import matplotlib.pyplot as plt
import argparse
from parsing import make_parser
from torch.autograd import Variable
import torch.nn.init as init
class ScaledDotProductAttention(nn.Module):
def __init__(self):
super(ScaledDotProductAttention, self).__init__()
def forward(self, Q, K, V, attn_mask=False):
scores = torch.matmul(Q, K.transpose(-1, -2)) / 1
if attn_mask is not None:
scores.masked_fill_(scores == 0, -1e9)
attn = nn.Softmax(dim=-1)(scores)
context = torch.matmul(attn, V)
return context, attn
class PositionWiseFeedForward(nn.Module):
def __init__(self, hidden_size, dec_size, device):
super(PositionWiseFeedForward, self).__init__()
self.fclayer = nn.Sequential(
nn.Linear(hidden_size, int(hidden_size/2)),
nn.ReLU(),
nn.Linear(int(hidden_size/2), dec_size),
).to(device)
def forward(self, x):
return self.fclayer(x)
class DGPS(nn.Module):
def __init__(self, args = make_parser(),
input_dim = 6, dof_seq = 21, dof = 2, Ksignum = 25, sig = 2,
buffer_size = 5000, batch_size = 32, gamma = 0.001,
enc_size = 6, dec_size = 6, hidden_size = 32, num_layers = 1):
super(DGPS, self).__init__()
self.args = args
self.device = self.args.device
self.hidden_size = hidden_size
self.num_layers = num_layers
self.enc_embed = nn.Sequential(
nn.Linear(enc_size, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, hidden_size),).to(self.device)
self.enc_embed2 = nn.Sequential(
nn.Linear(enc_size, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, hidden_size),).to(self.device)
#self.enc_embed = nn.Linear(Ksignum, hidden_size).to(self.device)
#self.enc_embed2 = nn.Linear(Ksignum, hidden_size).to(self.device)
self.eye = torch.eye(input_dim, device=self.device)
self.eye2 = torch.eye(args.output_dim, device=self.device)
self.eyedof = torch.eye(dof_seq - 1, device=self.device)
self.Al = torch.zeros((dof_seq-1, dof_seq-1), device = self.device)
for i in range(dof_seq-2):
self.Al[i,i+1] = 1
self.h = torch.zeros((dof_seq-1, 1), device = self.device)
self.h[-1,0] = 1
self.W = self.eye2
self.Ktheta = {}
self.Ktheta_buffer = torch.zeros((input_dim * Ksignum, args.output_dim), device = self.device)
self.Alu = self.Al
self.hu = self.h
self.re_u = 1
self.W = self.eye2
self.DOF_NN = (input_dim) * 2 ###!
self.ctheta = Variable((torch.rand((int(self.DOF_NN), Ksignum)) - 0.5).type(torch.FloatTensor).to(self.device),requires_grad=True)
self.slamb = sig*torch.ones((Ksignum), device = self.device)
self.fsignalu = torch.zeros((dof_seq-1, input_dim), device = self.device)
self.fsignalx = torch.zeros((dof_seq-1, args.output_dim), device = self.device)
self.fsignalq = torch.zeros((dof_seq-1, input_dim), device = self.device)
self.Ksignal = torch.empty((Ksignum, 1), device = self.device)
self.Ksignal_admittance = torch.empty((Ksignum, 1), device = self.device)
self.buffer = torch.empty((buffer_size, Ksignum + 6*input_dim + 9 + 6 + 0), device = self.device)
self.signalbuffer = torch.empty((buffer_size, dof_seq-1, input_dim + input_dim + args.output_dim), device = self.device) # u, q, x
self.input_dim = input_dim
self.output_dim = args.output_dim
self.Ksignum = Ksignum
self.gamma = gamma
self.sig = sig
self.violation_count = 0
self.violation_buffer = 0
self.learning_count = 0
self.batch_size = batch_size
self.buffer_size = buffer_size
self.costerror = 1
self.costerror_prev = 0
self.error_rev = 0
self.w_reuse = args.w_reuse
self.u_ma = 0.01*torch.ones((10,input_dim), device = self.device)
self.costerror_ma = 0.01*torch.ones(1000, device = self.device)
self.costerror_ma_kin = 0.01*torch.ones(1000, device = self.device)
self.crit_cul = torch.zeros((self.Ksignum, 1), device = self.device)
self.q_des_prev = torch.zeros((self.input_dim, 1), device = self.device)
self.q_des = torch.zeros((self.input_dim, 1), device = self.device)
self.q_des_init = torch.zeros((self.input_dim, 1), device = self.device)
self.make_weights(args.w_reuse)
self.cloudpoint = torch.ones((100,1), device = self.device)
self.sf_score_qq = 1 / dof_seq
self.null_ratio = 0.15
self.count_cul = 0
self.enc_self_attn = ScaledDotProductAttention()
self.pos_ffn = PositionWiseFeedForward(hidden_size, dec_size, args.device).to(self.device)
self.enc_self_attn2 = ScaledDotProductAttention()
self.pos_ffn2= PositionWiseFeedForward(hidden_size, dec_size, args.device).to(self.device)
self.fig = plt.figure() if args.visual == 1 else None
self.ax = self.fig.add_subplot(121) if args.visual == 1 else None
self.ax2 = self.fig.add_subplot(122) if args.visual == 1 else None
self.pos_encoding_table = self.get_sinusoid_encoding_table(dof_seq-1, hidden_size)
self.zero1 = torch.zeros((self.input_dim, self.input_dim), device = self.device)
self.zero2 = torch.zeros((self.output_dim, self.input_dim), device = self.device)
netlist = list(self.Kthetam.parameters()) + list(self.Kthetac.parameters()) + list(self.Kthetaj.parameters()) + \
list(self.Kthetaf.parameters()) + list(self.Kthetag.parameters()) + list(self.Kthetaj0.parameters()) + [self.ctheta]
self.optimizer = torch.optim.Adam(netlist, lr = gamma)
a_list = list(self.enc_self_attn.parameters()) + list(self.pos_ffn.parameters()) + list(
self.enc_embed.parameters())
self.optimizer2 = torch.optim.Adam(a_list, lr=5e-4)
b_list = list(self.enc_self_attn2.parameters()) + list(self.pos_ffn2.parameters()) + list(
self.enc_embed2.parameters())
self.optimizer3 = torch.optim.Adam(b_list, lr=5e-4)
def forward_attn(self, enc_inputs):
enc_embed = self.enc_embed(enc_inputs).unsqueeze(0)
enc_embed_q = enc_embed + 1 * self.pos_encoding_table.unsqueeze(0)
enc_outputs, attn = self.enc_self_attn(enc_embed_q, enc_embed_q, enc_embed)
enc_outputs = self.pos_ffn(enc_outputs.squeeze(0))
return enc_outputs, attn
def forward_attn2(self, enc_inputs):
enc_embed = self.enc_embed2(enc_inputs).unsqueeze(0)
enc_embed_q = enc_embed + 1 * self.pos_encoding_table.unsqueeze(0)
enc_outputs, attn = self.enc_self_attn2(enc_embed_q, enc_embed_q, enc_embed)
enc_outputs = self.pos_ffn2(enc_outputs.squeeze(0))
return enc_outputs, attn
def get_sinusoid_encoding_table(self,n_position, d_model):
def cal_angle(position, hid_idx):
return position / np.power(10000, 2 * (hid_idx // 2) / d_model)
def get_posi_angle_vec(position):
return [cal_angle(position, hid_j) for hid_j in range(d_model)]
sinusoid_table = np.array([get_posi_angle_vec(pos_i) for pos_i in range(n_position)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])
return FT(sinusoid_table).to(self.device)
def make_initial_qdes(self, q):
self.q_des_init = q
def make_weights(self, w_reuse):
if not w_reuse:
self.Kthetam = nn.ModuleList([nn.Linear(self.Ksignum, self.input_dim) for _ in range(self.input_dim + 1)]).to(self.device)
self.Kthetac = nn.ModuleList([nn.Linear(self.Ksignum, self.input_dim) for _ in range(self.input_dim)]).to(self.device)
self.Kthetaj = nn.ModuleList([nn.Linear(self.Ksignum, self.input_dim) for _ in range(self.output_dim)]).to(self.device)
self.Kthetaf = nn.ModuleList([nn.Linear(self.Ksignum, self.input_dim) for _ in range(self.input_dim)]).to(self.device)
self.Kthetag = nn.Linear(self.Ksignum, self.input_dim).to(self.device)
self.Kthetaj0 = nn.Linear(self.Ksignum, self.output_dim).to(self.device)
init.uniform_(self.Kthetaj0.weight, -0.1, 0.1)
init.uniform_(self.Kthetaj0.bias, -0.1, 0.1)
else:
raise
def save_buffer(self, Ksignal, qdot, q2dot, u, x, q, f, R, xi):
self.buffer[self.count_cul % self.buffer_size, :] = torch.cat([Ksignal.view(-1), qdot, q2dot, u, x, q, f, R, xi])
self.signalbuffer[self.count_cul % self.buffer_size, :, :] = torch.cat([self.fsignalu, self.fsignalq, self.fsignalx], dim = -1)
self.count_cul += 1
def generate_action(self, u, q, x, qv, xdes, v, f, sf, control_mode = 'force'):
device = self.device
with torch.no_grad():
self.fsignalu = self.Al @ self.fsignalu + self.h @ u.view(1,-1).tanh()
self.fsignalq = self.Al @ self.fsignalq + self.h @ q.view(1,-1)
self.fsignalx = self.Al @ self.fsignalx + self.h @ x.view(1,-1)
fsignal = torch.cat([self.fsignalq[-1,:],self.fsignalq[-2,:]],dim=0)
AA, BB = self.forward_attn(self.fsignalu)
lossA = torch.nn.functional.mse_loss(AA, self.fsignalq)
AA2, BB2 = self.forward_attn2(self.fsignalq)
lossB = torch.nn.functional.mse_loss(AA2, self.fsignalx)
sf_score_q_prev = self.sf_score_qq
BBB = BB.cpu().detach().numpy().squeeze(0)
BBB2 = BB2.cpu().detach().numpy().squeeze(0)
self.sf_score_qq = np.trace(BBB2)
"""
if self.count_cul % 100 == 50 and self.args.visual == 1:
self.ax.pcolor(BBB)
if self.violation_buffer <= 0:
self.ax.set_title("Normal: {:.3f}".format(np.trace(BBB)))
else:
self.ax.set_title("--External Perturbation--: {:.3f}".format(np.trace(BBB)))
if self.violation_buffer <= 0:
self.ax2.set_title("Normal: {:.3f}".format(np.trace(BBB2)))
else:
self.ax2.set_title("--External Perturbation--: {:.3f}".format(np.trace(BBB2)))
self.ax2.pcolor(BBB2)
self.fig.canvas.draw()
self.fig.canvas.flush_events()
self.fig.show()
"""
loss_ratio = (np.trace(BBB2)-8.5) / (sf_score_q_prev-8.5)
self.re_u = 10 / np.trace(BBB) if self.count_cul >= 500 and self.violation_buffer <= 0 else 0
if self.violation_buffer > 0:
sf = sf * 1
else:
sf = np.clip(sf * (1 + (loss_ratio - 1) * (-3.0)), sf * 0.9, sf * 1.1)
if self.re_u < 1 and self.violation_buffer < 0 and self.count_cul % 10 == 0:
print(self.count_cul, '{:.3f}, {:.3f}, {:.3f}, {:.3f}, {:.3f}'.format(sf, np.trace(BBB), np.trace(BBB2), lossA.cpu().detach().numpy(), lossB.cpu().detach().numpy()))
sf = np.clip(sf, 0.005, 2)
func_m = torch.zeros_like(self.zero1)
func_c = torch.zeros_like(self.zero1)
func_j = torch.zeros_like(self.zero2)
func_f = torch.zeros_like(self.zero1)
#feature = torch.exp()
self.Ksignal = torch.exp(-((fsignal.unsqueeze(1).repeat([1, self.Ksignum]) - self.ctheta).pow(2).sum(0)) * (1/(self.slamb**2))).unsqueeze(0)
func_g = self.Kthetag(self.Ksignal).T
func_j0 = self.Kthetaj0(self.Ksignal).T
for i in range(self.input_dim):
func_m[:,i] = self.Kthetam[i](self.Ksignal)
func_c[:,i] = self.Kthetac[i](self.Ksignal)
func_j[:,i] = self.Kthetaj[i](self.Ksignal)
func_f[:,i] = self.Kthetaf[i](self.Ksignal)
func_diag = torch.diag_embed(self.Kthetam[-1](self.Ksignal).abs()).squeeze(0)
L = torch.tril(func_m, diagonal = -1) + func_diag
func_m = L @ L.transpose(-2,-1)
#self.null_ratio = 0.0 * np.clip(sf - 0.0, 0, 0.25)
q_des_2prev = self.q_des_prev
self.q_des_prev = self.q_des
func_jinv = func_j.T @ torch.inverse(func_j @ func_j.T + sf * self.eye)
WW = self.eye
self.q_des = func_j.T @ WW @ torch.inverse(sf * self.eye + func_j @ WW @ func_j.T) @ \
(xdes - func_j0 - self.args.gain * (x - xdes))
qv_des = torch.tanh((self.q_des - self.q_des_prev) * 1000) - 0.5
qa_des = torch.tanh((self.q_des - 2 * self.q_des_prev + q_des_2prev) * 1000) - 0.5
if control_mode == 'impedance':
Kd = 5.0 * self.eye
Kp = 25.0 * self.eye
u_next = func_c @ qv + func_g \
- (func_j.T @ Kd @ func_j + sf * Kd * 1 * self.eye) @ (qv - qv_des) \
- (func_j.T @ Kp @ func_j + sf * Kp * 1 * self.eye) @ (q - self.q_des)
elif control_mode == 'admittance':
# Note that xdes is qdes
Kd = 5.0 * self.eye
Kp = 100.0 * self.eye
u_next = func_g - Kd @ qv - Kp @ (q - xdes) + func_f @ f
elif control_mode == 'force':
Kd = 5.0 * self.eye
Kp = 25.0 * self.eye
Kl = 15.0 * self.eye
u_next = func_m @ (qa_des - Kl @ (qv - qv_des)) + func_c @ (qv_des - Kl @ (q - self.q_des)) + \
func_g - Kd @ (qv - qv_des) - Kp @ Kl @ (q - self.q_des)
return u_next, sf, torch.trace(BB.squeeze(0)), torch.trace(BB2.squeeze(0)), lossA, func_j
def train(self, learn, control_mode):
if self.count_cul <= self.buffer_size - 1:
numlist = np.linspace(0, self.count_cul, self.count_cul+1)
else:
numlist = np.linspace(0, self.buffer_size-1, self.buffer_size)
numlist = numlist.tolist()
try:
replay_sample = random.sample(numlist, self.batch_size)
except:
replay_sample = numlist
replay_sample[-1] = (self.count_cul - 1) % self.buffer_size
with torch.no_grad():
Ksignal_ = self.buffer[replay_sample, 0:self.Ksignum]
f_ = self.buffer[replay_sample, self.Ksignum + 6 + 6 + 6 + 6 + 6:self.Ksignum + 6 + 6 + 6 + 6 + 6 + 6]
u_ = self.buffer[replay_sample, self.Ksignum + 6 + 6:self.Ksignum + 6 + 6 + 6]
x_ = self.buffer[replay_sample, self.Ksignum + 6 + 6 + 6:self.Ksignum + 6 + 6 + 6 + 6]
q_ = self.buffer[replay_sample, self.Ksignum + 6 + 6 + 6 + 6:self.Ksignum + 6 + 6 + 6 + 6 + 6]
qdot_ = self.buffer[replay_sample, self.Ksignum:self.Ksignum + 6]
q2dot_ = self.buffer[replay_sample, self.Ksignum + 6:self.Ksignum + 6 + 6]
fsignal_u = self.signalbuffer[replay_sample, :, 0:self.input_dim]
fsignal_q = self.signalbuffer[replay_sample, :, self.input_dim:self.input_dim*2]
fsignal_x = self.signalbuffer[replay_sample, :, self.input_dim*2:self.input_dim*2+self.output_dim]
func_g_ = self.Kthetag(Ksignal_).unsqueeze(2)
func_j0_ = self.Kthetaj0(Ksignal_).unsqueeze(2)
func_m_ = torch.zeros_like(self.zero1).repeat([Ksignal_.size(0), 1,1])
func_c_ = torch.zeros_like(self.zero1).repeat([Ksignal_.size(0), 1,1])
func_j_ = torch.zeros_like(self.zero2).repeat([Ksignal_.size(0), 1,1])
func_f_ = torch.zeros_like(self.zero1).repeat([Ksignal_.size(0), 1,1])
for x in range(self.input_dim):
func_m_[:, :, x] = self.Kthetam[x](Ksignal_)
func_c_[:, :, x] = self.Kthetac[x](Ksignal_)
try:
func_j_[:, :, x] = self.Kthetaj[x](Ksignal_)
except:
continue
func_f_[:, :, x] = self.Kthetaf[x](Ksignal_)
func_diag = torch.diag_embed(self.Kthetam[-1](Ksignal_).abs())
L = torch.tril(func_m_, diagonal=-1) + func_diag
func_m_ = L @ L.transpose(-2, -1)
self.optimizer.zero_grad()
costerror_ = torch.nn.functional.mse_loss(func_g_ + func_m_ @ q2dot_.unsqueeze(2) + func_c_ @ qdot_.unsqueeze(2), u_.unsqueeze(2))
costerror_kin = torch.nn.functional.mse_loss(func_j0_ + func_j_ @ q_.unsqueeze(2), x_[:, 0:self.output_dim].unsqueeze(2))
if self.count_cul % 10 == 0:
print('Loss:', '{:.3f}, {:.3f}'.format(costerror_.item(), costerror_kin.item()))
if control_mode == 'admittance':
self.re_u = 0
if learn and self.re_u < 1 and self.violation_buffer <= 0:
(costerror_ + costerror_kin).backward()
self.optimizer.step()
AA, BB = self.forward_attn(fsignal_u)
lossA = torch.nn.functional.mse_loss(AA, fsignal_q) + 10 * torch.nn.functional.mse_loss(BB.squeeze(0),self.eyedof)
AA2, BB2 = self.forward_attn2(fsignal_q)
#print(AA2.shape)
#print(fsignal_x.shape)
lossB = torch.nn.functional.mse_loss(AA2, fsignal_x) + 10 * torch.nn.functional.mse_loss(BB2.squeeze(0),self.eyedof)
self.optimizer2.zero_grad()
lossA.backward()
self.optimizer2.step()
self.optimizer3.zero_grad()
lossB.backward()
self.optimizer3.step()
if self.re_u >= 1:
self.violation_count += 1
self.violation_buffer = 25
if self.re_u < 1:
self.costerror_ma[self.count_cul % 1000] = torch.norm(costerror_)
self.costerror_ma_kin[self.count_cul % 1000] = torch.norm(costerror_kin)
if self.count_cul <= 0:
self.error_rev = 0
else:
self.error_rev = (torch.norm(costerror_)) / (torch.median(self.costerror_ma)) - 1
self.violation_buffer -= 1
if self.violation_buffer >= 0:
print('--- Robust Masking --- ', self.violation_buffer, 'counts left', self.re_u)
self.costerror_prev = torch.norm(costerror_kin)
return (func_j0_ + func_j_ @ q_.unsqueeze(2)).squeeze(2), lossA, lossB, costerror_, costerror_kin