-
Notifications
You must be signed in to change notification settings - Fork 1
/
nusc_model.py
289 lines (251 loc) · 13.6 KB
/
nusc_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
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
import numpy as np
import torch
import torch.nn as nn
from stl_d_lib import *
import utils
class Net(nn.Module):
def __init__(self, args):
super(Net, self).__init__()
self.args = args
self.input_dim = 77 # hybrid for both multi-lane and intersection
self.output_dim = args.nt * 2
self.feat_dim = feat_dim = 32
self.stlp_dim = stlp_dim = 6 # (vmin, vmax, dmin, dmax, d_safe)
self.lane_dim = 3
self.n_segs = args.n_segs
self.time_dim = 32
self.ego_encoder = utils.build_relu_nn(6, feat_dim, args.hiddens, activation_fn=nn.ReLU)
self.neighbor_encoder = utils.build_relu_nn(7, feat_dim, args.hiddens, activation_fn=nn.ReLU)
self.lane_encoder = utils.build_relu_nn(self.n_segs * self.lane_dim, feat_dim, args.hiddens, activation_fn=nn.ReLU)
if self.args.diffusion:
latent_dim = args.nt * 2 + self.time_dim + 1 + stlp_dim # (noise + timestep + high_level decision + stlp)
elif self.args.bc:
latent_dim = 1 + stlp_dim
elif self.args.vae:
latent_dim = args.vae_dim + 1 + stlp_dim # (noise + high_level decision + stlp)
self.traj_encoder = utils.build_relu_nn(args.nt * 2, args.vae_dim * 2, args.hiddens, activation_fn=nn.ReLU)
else:
latent_dim = 1 + stlp_dim # (stl + highlevel)
if args.use_init_hint:
latent_dim += args.nt * 2
self.policy_net = utils.build_relu_nn(latent_dim + feat_dim * 7, args.nt * 2, args.hiddens, activation_fn=nn.ReLU)
if self.args.rect_head:
rect_out_dim = args.nt * 2
extra_in_dim = 0
if self.args.diverse_loss and self.args.no_arch==False and self.args.diverse_fuse_type=="cat":
extra_in_dim += args.nt * 2
if self.args.diverse_loss:
self.merge_net = utils.build_relu_nn(args.nt*2, args.nt*2, [32, 32], activation_fn=nn.ReLU)
self.rect_net = utils.build_relu_nn(latent_dim - self.time_dim + feat_dim * 7 + extra_in_dim, rect_out_dim, args.rect_hiddens, activation_fn=nn.ReLU)
def pos_encoding(self, t, channels):
inv_freq = 1.0 / (10000 ** (torch.arange(0, channels, 2, device=t.device).float() / channels))
pos_enc_a = torch.sin(t.repeat(1, channels // 2) * inv_freq)
pos_enc_b = torch.cos(t.repeat(1, channels // 2) * inv_freq)
pos_enc = torch.cat([pos_enc_a, pos_enc_b], dim=-1)
return pos_enc
def encode_feat(self, nn_input, ext=None):
bs = nn_input["ego_traj"].shape[0]
# normalization
ego = nn_input["ego_traj"][:, 0]
ego_un = ego.unsqueeze(1)
# neighbor normalization
neis_ = nn_input["neighbors"]
neis_xyth = normalize_xyth(neis_[..., 1:4], ego_un, neis_[..., 0])
neis_input = torch.cat([neis_[..., 0:1], neis_xyth, neis_[..., 4:7]], dim=-1)
# lane normalization
tmp_di = {}
for key in ["curr", "left", "right"]:
item = normalize_xyth(nn_input["%slane_wpts"%(key)], ego_un, nn_input["%s_id"%(key)])
tmp_di[key] = item
lanes = torch.stack((tmp_di["curr"], tmp_di["left"], tmp_di["right"]), dim=1) # (N, 3, nseg, 3)
lanes_start = lanes[..., 0:1, :] # use difference encoding
lanes_diff = lanes[..., 1:, :] - lanes[..., :-1, :]
segs = lanes.shape[-2]
lanes_input = torch.cat([lanes_start, lanes_diff], dim=-2).reshape(bs, 3, segs * self.lane_dim)
ego_xyth = normalize_xyth(ego[..., :3], ego[..., :3])
ego_input = torch.cat([ego_xyth, ego[..., 3:]], dim=-1)
# encoder part
ego_feat = self.ego_encoder(ego_input) # (N, nfeat)
nei_feat = self.neighbor_encoder(neis_input) # (N, n_neis, nfeat)
nei_feat_min = torch.min(nei_feat, dim=1)[0]
nei_feat_avg = torch.mean(nei_feat, dim=1)
nei_feat_max = torch.max(nei_feat, dim=1)[0]
nei_feat = torch.cat([nei_feat_min, nei_feat_avg, nei_feat_max], dim=-1) # (N, 2 * nfeat)
# nei_feat = nei_feat_max
lanes_feat = self.lane_encoder(lanes_input) # (N, 3, nfeat)
lanes_feat = lanes_feat.reshape(bs, -1)
feature = torch.cat([ego_feat, nei_feat, lanes_feat], dim=-1) # (N, 5 * nfeat)
return feature
def forward(self, nn_input, ext=None, get_feature=False, prev_feature=None, sample=False, n_randoms=None):
bs = nn_input["ego_traj"].shape[0]
multi_check = any([self.args.diffusion, self.args.vae, self.args.bc]) and self.args.gt_data_training==False
if prev_feature is not None:
feature = prev_feature
else:
feature = self.encode_feat(nn_input)
if multi_check:
k = feature.shape[-1]
if n_randoms is None:
n_randoms = self.args.n_randoms
n_rep = n_randoms * 3
feature = feature.reshape(bs, 1, k).repeat(1, n_rep, 1).reshape(-1, k)
n = feature.shape[0] # n might be not bs
if multi_check:
stlp_dense_feat = nn_input["stlp_dense"][:,0]
else:
stlp_dense_feat = ext["gt_stlp"]
if self.args.diffusion:
time_feat = self.pos_encoding(ext["timestep"], self.time_dim)
if multi_check:
policy_input = torch.cat([feature, ext["noise"], time_feat, ext["highlevel"], stlp_dense_feat], dim=-1) # (n, args.nt * 2)
else:
n_rep = self.args.n_randoms
feature_tmp = feature.reshape(bs, 1, -1).repeat(1, n_rep, 1).reshape(bs * n_rep, -1)
highlevel_tmp = ext["highlevel"].reshape(bs, 1, -1).repeat(1, n_rep, 1).reshape(bs * n_rep, -1)
stlp_tmp = stlp_dense_feat.reshape(bs, 1, -1).repeat(1, n_rep, 1).reshape(bs * n_rep, -1)
policy_input = torch.cat([feature_tmp, ext["noise"], time_feat, highlevel_tmp, stlp_tmp], dim=-1) # (n, args.nt * 2)
elif self.args.bc:
policy_input = torch.cat([feature, ext["highlevel"], stlp_dense_feat], dim=-1)
elif self.args.vae:
if sample is not False:
latent = sample
latent_mean, latent_logstd, latent_std = None, None, None
else:
if multi_check:
code = self.traj_encoder(ext["trajopt_controls"].reshape(-1, self.args.nt * 2))
else:
n_rep = self.args.n_randoms
feature_tmp = feature.reshape(bs, 1, -1).repeat(1, n_rep, 1).reshape(bs * n_rep, -1)
highlevel_tmp = ext["highlevel"].reshape(bs, 1, -1).repeat(1, n_rep, 1).reshape(bs * n_rep, -1)
stlp_tmp = stlp_dense_feat.reshape(bs, 1, -1).repeat(1, n_rep, 1).reshape(bs * n_rep, -1)
code = self.traj_encoder(ext["gt_controls"].reshape(-1, self.args.nt * 2))
code = code[:,None,:].repeat(1, self.args.n_randoms, 1).reshape(bs * n_rep, self.args.vae_dim*2)
latent_mean = code[..., :self.args.vae_dim]
latent_logstd = code[..., self.args.vae_dim:]
latent_std = torch.exp(latent_logstd)
latent = ext["noise"] * latent_std + latent_mean
if multi_check:
policy_input = torch.cat([feature, latent, ext["highlevel"], stlp_dense_feat], dim=-1) # (n, args.nt * 2)
else:
policy_input = torch.cat([feature_tmp, latent, highlevel_tmp, stlp_tmp], dim=-1) # (n, args.nt * 2)
else:
policy_input = torch.cat([feature, nn_input["gt_high_level"], stlp_dense_feat], dim=-1)
if self.args.use_init_hint:
policy_input = torch.cat([policy_input, nn_input["params_init"].reshape(list(policy_input.shape[:-1]) + [self.args.nt * 2])], dim=-1)
raw_controls = self.policy_net(policy_input)
if self.args.diffusion:
raw_controls = raw_controls + ext["noise"]
raw_controls = raw_controls.reshape(-1, self.args.nt, 2)
if self.args.diffusion:
steer = raw_controls[..., 0] # * self.args.mul_w_max
accel = raw_controls[..., 1] # * self.args.mul_a_max
else:
steer = torch.nn.Tanh()(raw_controls[..., 0]) * self.args.mul_w_max
accel = torch.nn.Tanh()(raw_controls[..., 1]) * self.args.mul_a_max
controls = torch.stack([steer, accel], dim=-1)
if get_feature:
return controls, feature
else:
if self.args.vae:
return controls, latent_mean, latent_logstd, latent_std
else:
return controls
def rect_forward(self, feature, highlevel, stlp_dense_feat, init_controls, scores, extras=None):
n = feature.shape[0]
# print(feature.shape, highlevel.shape, stlp_dense_feat.shape, init_controls.shape)
if self.args.diverse_loss and self.args.no_arch==False:
fused_controls = self.merge_net(init_controls.reshape(-1, self.args.nt*2))
bs =int(init_controls.shape[0] / 3 / self.args.n_randoms)
fused_controls = fused_controls.reshape(bs, self.args.n_randoms, 3, self.args.nt*2)
fused_controls = fused_controls.permute(0, 2, 1, 3)
NS = self.args.n_shards
fused_controls = fused_controls.reshape(bs, 3, NS, self.args.n_randoms // NS, self.args.nt*2)
fused_controls = torch.max(fused_controls, dim=3, keepdim=True)[0]
# print("Fused",fused_controls.shape)
fused_controls = fused_controls.repeat(1, 1, 1, self.args.n_randoms // NS, 1).reshape(bs, 3, self.args.n_randoms, self.args.nt*2)
fused_controls = fused_controls.permute(0, 2, 1, 3)
fused_controls = fused_controls.reshape(init_controls.shape[0], self.args.nt, 2)
if self.args.diverse_fuse_type=="add":
fused_controls = init_controls + fused_controls
if self.args.diverse_fuse_type=="cat":
policy_input = torch.cat([feature, highlevel, stlp_dense_feat, init_controls.reshape(n, self.args.nt*2), fused_controls.reshape(n, self.args.nt*2)], dim=-1)
elif self.args.diverse_fuse_type=="add":
policy_input = torch.cat([feature, highlevel, stlp_dense_feat, fused_controls.reshape(n, self.args.nt*2)], dim=-1)
else:
raise NotImplementedError
else:
policy_input = torch.cat([feature, highlevel, stlp_dense_feat, init_controls.reshape(n, self.args.nt*2)], dim=-1)
raw_controls_aug = self.rect_net(policy_input)
raw_controls_aug = raw_controls_aug.reshape(n, self.args.nt, 2)
if self.args.interval:
init_w = init_controls[..., 0]
init_a = init_controls[..., 1]
raw_controls = torch.nn.Tanh()(raw_controls_aug)
w_mask = (raw_controls[..., 0]>=0).float()
a_mask = (raw_controls[..., 1]>=0).float()
raw_controls_w0 = raw_controls[..., 0] * (init_w-(-self.args.mul_w_max))
raw_controls_w1 = raw_controls[..., 0] * (self.args.mul_w_max-init_w)
raw_controls_a0 = raw_controls[..., 1] * (init_a-(-self.args.mul_a_max))
raw_controls_a1 = raw_controls[..., 1] * (self.args.mul_a_max-init_a)
w_merge = raw_controls_w0 * (1-w_mask) + raw_controls_w1 * w_mask
a_merge = raw_controls_a0 * (1-a_mask) + raw_controls_a1 * a_mask
raw_controls = torch.stack([w_merge, a_merge], dim=-1)
else:
raw_controls = raw_controls_aug
violated=((scores<0).float()[:,None,None])
raw_controls = init_controls + raw_controls * violated
if self.args.clip_rect:
w_merge = torch.clip(raw_controls[..., 0], -self.args.mul_w_max, self.args.mul_w_max)
a_merge = torch.clip(raw_controls[..., 1], -self.args.mul_a_max, self.args.mul_a_max)
raw_controls = torch.stack([w_merge, a_merge], dim=-1)
return raw_controls
def normalize_xyth(state, base, valid=None, no_theta=False):
assert len(state.shape) == len(base.shape) and state.shape[0]==base.shape[0]
x = state[..., 0]
y = state[..., 1]
if no_theta==False:
th = state[..., 2]
base_x = base[..., 0]
base_y = base[..., 1]
base_th = base[..., 2]
if valid is not None:
x_trans = x - base_x * valid
y_trans = y - base_y * valid
else:
x_trans = x - base_x
y_trans = y - base_y
x_rel = x_trans * torch.cos(base_th) + y_trans * torch.sin(base_th)
y_rel = -x_trans * torch.sin(base_th) + y_trans * torch.cos(base_th)
if no_theta==False:
if valid is not None:
th_rel = th - base_th * valid
else:
th_rel = th - base_th
return torch.stack([x_rel, y_rel, th_rel], dim=-1)
else:
return torch.stack([x_rel, y_rel], dim=-1)
def normalize_xyth_np(state, base, valid=None, no_theta=False):
assert len(state.shape) == len(base.shape) and state.shape[0]==base.shape[0]
x = state[..., 0]
y = state[..., 1]
if no_theta==False:
th = state[..., 2]
base_x = base[..., 0]
base_y = base[..., 1]
base_th = base[..., 2]
if valid is not None:
x_trans = x - base_x * valid
y_trans = y - base_y * valid
else:
x_trans = x - base_x
y_trans = y - base_y
x_rel = x_trans * np.cos(base_th) + y_trans * np.sin(base_th)
y_rel = -x_trans * np.sin(base_th) + y_trans * np.cos(base_th)
if no_theta==False:
if valid is not None:
th_rel = th - base_th * valid
else:
th_rel = th - base_th
return np.stack([x_rel, y_rel, th_rel], axis=-1)
else:
return np.stack([x_rel, y_rel], axis=-1)