File size: 11,953 Bytes
c1a7f73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
""" Adapted from NVLabs/CatK """

# Not a contribution
# Changes made by NVIDIA CORPORATION & AFFILIATES enabling <CAT-K> or otherwise documented as
# NVIDIA-proprietary are not a contribution and subject to the following terms and conditions:
# SPDX-FileCopyrightText: Copyright (c) <year> NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NvidiaProprietary
#
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
# property and proprietary rights in and to this material, related
# documentation and any modifications thereto. Any use, reproduction,
# disclosure or distribution of this material and related documentation
# without an express license agreement from NVIDIA CORPORATION or
# its affiliates is strictly prohibited.

from typing import Optional, Tuple

import torch
from omegaconf import DictConfig
from torch import Tensor
from torch.distributions import Categorical, Independent, MixtureSameFamily, Normal


@torch.no_grad()
def cal_polygon_contour(
    pos: Tensor,  # [n_agent, n_step, n_target, 2]
    head: Tensor,  # [n_agent, n_step, n_target]
    width_length: Tensor,  # [n_agent, 1, 1, 2]
) -> Tensor:  # [n_agent, n_step, n_target, 4, 2]
    x, y = pos[..., 0], pos[..., 1]  # [n_agent, n_step, n_target]
    width, length = width_length[..., 0], width_length[..., 1]  # [n_agent, 1 ,1]

    half_cos = 0.5 * head.cos()  # [n_agent, n_step, n_target]
    half_sin = 0.5 * head.sin()  # [n_agent, n_step, n_target]
    length_cos = length * half_cos  # [n_agent, n_step, n_target]
    length_sin = length * half_sin  # [n_agent, n_step, n_target]
    width_cos = width * half_cos  # [n_agent, n_step, n_target]
    width_sin = width * half_sin  # [n_agent, n_step, n_target]

    left_front_x = x + length_cos - width_sin
    left_front_y = y + length_sin + width_cos
    left_front = torch.stack((left_front_x, left_front_y), dim=-1)

    right_front_x = x + length_cos + width_sin
    right_front_y = y + length_sin - width_cos
    right_front = torch.stack((right_front_x, right_front_y), dim=-1)

    right_back_x = x - length_cos + width_sin
    right_back_y = y - length_sin - width_cos
    right_back = torch.stack((right_back_x, right_back_y), dim=-1)

    left_back_x = x - length_cos - width_sin
    left_back_y = y - length_sin + width_cos
    left_back = torch.stack((left_back_x, left_back_y), dim=-1)

    polygon_contour = torch.stack(
        (left_front, right_front, right_back, left_back), dim=-2
    )

    return polygon_contour


def transform_to_global(
    pos_local: Tensor,  # [n_agent, n_step, 2]
    head_local: Optional[Tensor],  # [n_agent, n_step]
    pos_now: Tensor,  # [n_agent, 2]
    head_now: Tensor,  # [n_agent]
) -> Tuple[Tensor, Optional[Tensor]]:
    cos, sin = head_now.cos(), head_now.sin()
    rot_mat = torch.zeros((head_now.shape[0], 2, 2), device=head_now.device)
    rot_mat[:, 0, 0] = cos
    rot_mat[:, 0, 1] = sin
    rot_mat[:, 1, 0] = -sin
    rot_mat[:, 1, 1] = cos

    pos_global = torch.bmm(pos_local, rot_mat)  # [n_agent, n_step, 2]*[n_agent, 2, 2]
    pos_global = pos_global + pos_now.unsqueeze(1)
    if head_local is None:
        head_global = None
    else:
        head_global = head_local + head_now.unsqueeze(1)
    return pos_global, head_global


def transform_to_local(
    pos_global: Tensor,  # [n_agent, n_step, 2]
    head_global: Optional[Tensor],  # [n_agent, n_step]
    pos_now: Tensor,  # [n_agent, 2]
    head_now: Tensor,  # [n_agent]
) -> Tuple[Tensor, Optional[Tensor]]:
    cos, sin = head_now.cos(), head_now.sin()
    rot_mat = torch.zeros((head_now.shape[0], 2, 2), device=head_now.device)
    rot_mat[:, 0, 0] = cos
    rot_mat[:, 0, 1] = -sin
    rot_mat[:, 1, 0] = sin
    rot_mat[:, 1, 1] = cos

    pos_local = pos_global - pos_now.unsqueeze(1)
    pos_local = torch.bmm(pos_local, rot_mat)  # [n_agent, n_step, 2]*[n_agent, 2, 2]
    if head_global is None:
        head_local = None
    else:
        head_local = head_global - head_now.unsqueeze(1)
    return pos_local, head_local


def sample_next_token_traj(
    token_traj: Tensor,  # [n_agent, n_token, 4, 2]
    token_traj_all: Tensor,  # [n_agent, n_token, 6, 4, 2]
    sampling_scheme: DictConfig,
    # ! for most-likely sampling
    next_token_logits: Tensor,  # [n_agent, n_token], with grad
    # ! for nearest-pos sampling, sampling near to GT
    pos_now: Tensor,  # [n_agent, 2]
    head_now: Tensor,  # [n_agent]
    pos_next_gt: Tensor,  # [n_agent, 2]
    head_next_gt: Tensor,  # [n_agent]
    valid_next_gt: Tensor,  # [n_agent]
    token_agent_shape: Tensor,  # [n_agent, 2]
) -> Tuple[Tensor, Tensor]:
    """
    Returns:
        next_token_traj_all: [n_agent, 6, 4, 2], local coord
        next_token_idx: [n_agent], without grad
    """
    range_a = torch.arange(next_token_logits.shape[0])
    next_token_logits = next_token_logits.detach()

    if (
        sampling_scheme.criterium == "topk_prob"
        or sampling_scheme.criterium == "topk_prob_sampled_with_dist"
    ):
        topk_logits, topk_indices = torch.topk(
            next_token_logits, sampling_scheme.num_k, dim=-1, sorted=False
        )
        if sampling_scheme.criterium == "topk_prob_sampled_with_dist":
            #! gt_contour: [n_agent, 4, 2] in global coord
            gt_contour = cal_polygon_contour(
                pos_next_gt, head_next_gt, token_agent_shape
            )
            gt_contour = gt_contour.unsqueeze(1)  # [n_agent, 1, 4, 2]
            token_world_sample = token_traj[range_a.unsqueeze(1), topk_indices]
            token_world_sample = transform_to_global(
                pos_local=token_world_sample.flatten(1, 2),
                head_local=None,
                pos_now=pos_now,  # [n_agent, 2]
                head_now=head_now,  # [n_agent]
            )[0].view(*token_world_sample.shape)

            # dist: [n_agent, n_token]
            dist = torch.norm(token_world_sample - gt_contour, dim=-1).mean(-1)
            topk_logits = topk_logits.masked_fill(
                valid_next_gt.unsqueeze(1), 0.0
            ) - 1.0 * dist.masked_fill(~valid_next_gt.unsqueeze(1), 0.0)
    elif sampling_scheme.criterium == "topk_dist_sampled_with_prob":
        #! gt_contour: [n_agent, 4, 2] in global coord
        gt_contour = cal_polygon_contour(pos_next_gt, head_next_gt, token_agent_shape)
        gt_contour = gt_contour.unsqueeze(1)  # [n_agent, 1, 4, 2]
        token_world_sample = transform_to_global(
            pos_local=token_traj.flatten(1, 2),  # [n_agent, n_token*4, 2]
            head_local=None,
            pos_now=pos_now,  # [n_agent, 2]
            head_now=head_now,  # [n_agent]
        )[0].view(*token_traj.shape)

        _invalid = ~valid_next_gt
        # dist: [n_agent, n_token]
        dist = torch.norm(token_world_sample - gt_contour, dim=-1).mean(-1)
        _logits = -1.0 * dist.masked_fill(_invalid.unsqueeze(1), 0.0)

        if _invalid.any():
            _logits[_invalid] = next_token_logits[_invalid]
        _, topk_indices = torch.topk(
            _logits, sampling_scheme.num_k, dim=-1, sorted=False
        )  # [n_agent, K]
        topk_logits = next_token_logits[range_a.unsqueeze(1), topk_indices]

    else:
        raise ValueError(f"Invalid criterium: {sampling_scheme.criterium}")

    # topk_logits, topk_indices: [n_agent, K]
    topk_logits = topk_logits / sampling_scheme.temp
    samples = Categorical(logits=topk_logits).sample()  # [n_agent] in K
    next_token_idx = topk_indices[range_a, samples]
    next_token_traj_all = token_traj_all[range_a, next_token_idx]

    return next_token_idx, next_token_traj_all


def sample_next_gmm_traj(
    token_traj: Tensor,  # [n_agent, n_token, 4, 2]
    token_traj_all: Tensor,  # [n_agent, n_token, 6, 4, 2]
    sampling_scheme: DictConfig,
    # ! for most-likely sampling
    ego_mask: Tensor,  # [n_agent], bool, ego_mask.sum()==n_batch
    ego_next_logits: Tensor,  # [n_batch, n_k_ego_gmm]
    ego_next_poses: Tensor,  # [n_batch, n_k_ego_gmm, 3]
    ego_next_cov: Tensor,  # [2], one for pos, one for heading.
    # ! for nearest-pos sampling, sampling near to GT
    pos_now: Tensor,  # [n_agent, 2]
    head_now: Tensor,  # [n_agent]
    pos_next_gt: Tensor,  # [n_agent, 2]
    head_next_gt: Tensor,  # [n_agent]
    valid_next_gt: Tensor,  # [n_agent]
    token_agent_shape: Tensor,  # [n_agent, 2]
    next_token_idx: Tensor,  # [n_agent]
) -> Tuple[Tensor, Tensor]:
    """
    Returns:
        next_token_traj_all: [n_agent, 6, 4, 2], local coord
        next_token_idx: [n_agent], without grad
    """
    n_agent = token_traj.shape[0]
    n_batch = ego_next_logits.shape[0]
    next_token_traj_all = token_traj_all[torch.arange(n_agent), next_token_idx]

    # ! sample only the ego-vehicle
    assert (
        sampling_scheme.criterium == "topk_prob"
        or sampling_scheme.criterium == "topk_prob_sampled_with_dist"
    )
    topk_logits, topk_indices = torch.topk(
        ego_next_logits, sampling_scheme.num_k, dim=-1, sorted=False
    )  # [n_agent, k], [n_agent, k]
    ego_pose_topk = ego_next_poses[
        torch.arange(n_batch).unsqueeze(1), topk_indices
    ]  # [n_batch, k, 3]

    if sampling_scheme.criterium == "topk_prob_sampled_with_dist":
        # udpate topk_logits
        gt_contour = cal_polygon_contour(
            pos_next_gt[ego_mask],
            head_next_gt[ego_mask],
            token_agent_shape[ego_mask],
        )  # [n_batch, 4, 2] in global coord
        gt_contour = gt_contour.unsqueeze(1)  # [n_batch, 1, 4, 2]

        ego_pos_global, ego_head_global = transform_to_global(
            pos_local=ego_pose_topk[:, :, :2],  # [n_batch, k, 2]
            head_local=ego_pose_topk[:, :, -1],  # [n_batch, k]
            pos_now=pos_now[ego_mask],  # [n_batch, 2]
            head_now=head_now[ego_mask],  # [n_batch]
        )
        ego_contour = cal_polygon_contour(
            ego_pos_global,  # [n_batch, k, 2]
            ego_head_global,  # [n_batch, k]
            token_agent_shape[ego_mask].unsqueeze(1),
        )  # [n_batch, k, 4, 2] in global coord

        dist = torch.norm(ego_contour - gt_contour, dim=-1).mean(-1)  # [n_batch, k]
        topk_logits = topk_logits.masked_fill(
            valid_next_gt[ego_mask].unsqueeze(1), 0.0
        ) - 1.0 * dist.masked_fill(~valid_next_gt[ego_mask].unsqueeze(1), 0.0)

    topk_logits = topk_logits / sampling_scheme.temp_mode  # [n_batch, k]
    ego_pose_topk = torch.cat(
        [
            ego_pose_topk[..., :2],
            ego_pose_topk[..., [-1]].cos(),
            ego_pose_topk[..., [-1]].sin(),
        ],
        dim=-1,
    )
    cov = (
        (ego_next_cov * sampling_scheme.temp_cov)
        .repeat_interleave(2)[None, None, :]
        .expand(*ego_pose_topk.shape)
    )  # [n_batch, k, 4]
    gmm = MixtureSameFamily(
        Categorical(logits=topk_logits), Independent(Normal(ego_pose_topk, cov), 1)
    )
    ego_sample = gmm.sample()  # [n_batch, 4]

    ego_contour_local = cal_polygon_contour(
        ego_sample[:, :2],  # [n_batch, 2]
        torch.arctan2(ego_sample[:, -1], ego_sample[:, -2]),  # [n_batch]
        token_agent_shape[ego_mask],  # [n_batch, 2]
    )  # [n_batch, 4, 2] in local coord

    ego_token_local = token_traj[ego_mask]  # [n_batch, n_token, 4, 2]

    dist = torch.norm(ego_contour_local.unsqueeze(1) - ego_token_local, dim=-1).mean(
        -1
    )  # [n_batch, n_token]
    next_token_idx[ego_mask] = dist.argmin(-1)

    ego_contour_local  # [n_batch, 4, 2] in local coord
    ego_countour_start = next_token_traj_all[ego_mask][:, 0]  # [n_batch, 4, 2]
    n_step = next_token_traj_all.shape[1]
    diff = (ego_contour_local - ego_countour_start) / (n_step - 1)
    ego_token_interp = [ego_countour_start + diff * i for i in range(n_step)]
    # [n_batch, 6, 4, 2]
    next_token_traj_all[ego_mask] = torch.stack(ego_token_interp, dim=1)

    return next_token_idx, next_token_traj_all