File size: 10,863 Bytes
d4cbafd | 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 | """
FutureInteractionGraphV5 β uncertainty-aware semantic scorer + sparse RelTrajEncoder.
Builds on V4 (learnable top-N sparse graph) with an enhanced edge scorer:
V4 scorer: [mean_rel, std_rel, min_dist] β [E, 5] (geometric only)
V5 scorer: [y0_emb_i, y0_emb_j_scaled, β [E, 2*D_score + 7]
Ο_i, Ο_j,
mean_rel, std_rel, min_dist]
Where:
y0_emb_{i,j} β trajectory embedding from clean y_0_hat (not noisy y_t).
Projected from y_abs [B,K,A,T,2] which is already y_0_hat
unnormalised β no noise, same source as edge geometry.
Ο_i, Ο_j β explicit per-agent mean uncertainty, allowing the scorer to
learn "uncertain target prefers certain source."
When sigma_agent is None (pass 1 of the two-pass forward), certainty scaling
is skipped and Ο features are zeroed β scorer still runs on geometric +
semantic features, gracefully degrading to a noisier but functional signal.
"""
import torch
import torch.nn as nn
from models.graph_interaction_nba_v4 import FutureInteractionGraphV4
from models.graph_interaction_nba_v3 import RelTrajEncoder
def _heading_diff(pos_i: torch.Tensor,
pos_j: torch.Tensor) -> torch.Tensor:
"""Per-timestep relative heading angle difference.
Args:
pos_i, pos_j: [E, T, 2] absolute positions of target and source
Returns:
[E, T, 1] angle difference wrapped to [-Ο, Ο]
β positive: j is turning more CCW than i
β β 0: parallel motion; β Β±Ο: opposing motion
"""
vel_i = pos_i[:, 1:] - pos_i[:, :-1] # [E, T-1, 2]
vel_j = pos_j[:, 1:] - pos_j[:, :-1] # [E, T-1, 2]
h_i = torch.atan2(vel_i[..., 1], vel_i[..., 0]) # [E, T-1]
h_j = torch.atan2(vel_j[..., 1], vel_j[..., 0]) # [E, T-1]
diff = torch.atan2(torch.sin(h_j - h_i),
torch.cos(h_j - h_i)) # [E, T-1] wrapped
# Repeat last step so length matches T
diff = torch.cat([diff, diff[:, -1:]], dim=1) # [E, T]
return diff.unsqueeze(-1) # [E, T, 1]
class FutureInteractionGraphV5(FutureInteractionGraphV4):
"""Sparse interaction graph with uncertainty-aware semantic edge scorer.
Extra constructor kwarg (beyond V4):
y0_score_dim (int, default 32): projection dim for y_0_hat trajectory
embedding used in the scorer.
"""
def __init__(self, embed_dim: int, future_steps: int, num_agents: int,
num_heads: int = 4, dropout: float = 0.1,
num_gnn_layers: int = 2, time_dim: int = 128,
top_n_neighbors: int = 5, rel_traj_hidden: int = 32,
y0_score_dim: int = 32):
super().__init__(
embed_dim = embed_dim,
future_steps = future_steps,
num_agents = num_agents,
num_heads = num_heads,
dropout = dropout,
num_gnn_layers = num_gnn_layers,
time_dim = time_dim,
top_n_neighbors = top_n_neighbors,
rel_traj_hidden = rel_traj_hidden,
)
self.y0_score_dim = y0_score_dim
# Replace V4's encoder (in_channels=2) with one that also accepts
# the relative heading angle channel β in_channels=3.
self.rel_traj_encoder = RelTrajEncoder(
out_dim = embed_dim,
T = future_steps,
D_hidden = rel_traj_hidden,
num_heads = 4,
in_channels = 3, # rel_pos(2) + heading_diff(1)
)
# Project each agent's clean y_0_hat trajectory to scoring space.
# Input: flattened future trajectory [T*2].
self.y0_score_proj = nn.Sequential(
nn.Linear(future_steps * 2, y0_score_dim),
nn.ReLU(inplace=True),
)
# Pairwise interaction scorer.
#
# Step 1 β separate projections for i and j:
self.score_proj_i = nn.Linear(y0_score_dim, y0_score_dim)
self.score_proj_j = nn.Linear(y0_score_dim, y0_score_dim)
#
# Step 2 β explicit pairwise interaction terms:
# h_i * h_j [D_s] β element-wise similarity
# h_i - h_j [D_s] β directional asymmetry (i doing X, j doing Y)
# β cat β [2*D_s]
#
# Step 3 β head combines interaction + uncertainty + geometry:
# [2*D_s + Ο_i(1) + Ο_j(1) + mean_rel(2) + std_rel(2) + min_dist(1)]
head_in = y0_score_dim * 2 + 7
self.edge_scorer = nn.Sequential(
nn.Linear(head_in, 32),
nn.ReLU(inplace=True),
nn.Linear(32, 1),
)
# ------------------------------------------------------------------
# Forward
# ------------------------------------------------------------------
def forward(
self,
y_emb: torch.Tensor, # [B, K, A, D]
y_abs: torch.Tensor, # [B, K, A, T, 2] β y_0_hat unnorm
t_emb: torch.Tensor, # [B, D]
tau: torch.Tensor, # [B] β [0, 1]
sigma_agent: torch.Tensor = None, # [B, K, A, T] or None
) -> torch.Tensor: # [B, K, A, D]
B, K, A, D = y_emb.shape
T = y_abs.shape[3]
E0 = self._E0
N = self.top_n
# ---- Full per-mode relative trajectories [B*K*E0, T, 2] ----------
pos_bk = y_abs.reshape(B * K * A, T, 2)
edge_index_bk = self._make_batched_edge_index(B * K) # [2, B*K*E0]
pos_i_t = pos_bk[edge_index_bk[1]] # [B*K*E0, T, 2]
pos_j_t = pos_bk[edge_index_bk[0]] # [B*K*E0, T, 2]
rel_pos_t = pos_j_t - pos_i_t # [B*K*E0, T, 2]
# ---- Geometric features (all edges) ------------------------------
mean_rel = rel_pos_t.mean(dim=1) # [E, 2]
std_rel = rel_pos_t.std(dim=1) # [E, 2]
min_dist = (rel_pos_t.norm(dim=-1)
.min(dim=1).values
.unsqueeze(-1)) # [E, 1]
# ---- Semantic features from clean y_0_hat -----------------------
y0_flat = y_abs.reshape(B * K * A, T * 2) # [B*K*A, T*2]
y0_emb = self.y0_score_proj(y0_flat) # [B*K*A, D_s]
y0_emb_i = y0_emb[edge_index_bk[1]] # [E, D_s] target
y0_emb_j = y0_emb[edge_index_bk[0]] # [E, D_s] source
# ---- Uncertainty features ----------------------------------------
if sigma_agent is not None:
sigma_mean = sigma_agent.mean(dim=-1) # [B, K, A]
sigma_bka = sigma_mean.reshape(B * K * A) # [B*K*A]
sigma_i = sigma_bka[edge_index_bk[1]].unsqueeze(-1) # [E, 1]
sigma_j = sigma_bka[edge_index_bk[0]].unsqueeze(-1) # [E, 1]
tau_bka = sigma_bka
else:
# Pass 1: no uncertainty available β zero sigma features
sigma_i = torch.zeros(rel_pos_t.size(0), 1, device=y_abs.device)
sigma_j = torch.zeros_like(sigma_i)
tau_bka = (tau
.unsqueeze(1).unsqueeze(2)
.expand(-1, K, A)
.reshape(B * K * A))
# ---- Enhanced scorer --------------------------------------------
# Pairwise interaction: separate projections then explicit relation terms
h_i = self.score_proj_i(y0_emb_i) # [E, D_s]
h_j = self.score_proj_j(y0_emb_j) # [E, D_s]
interact = torch.cat([h_i * h_j, # similarity
h_i - h_j], dim=-1) # [E, 2*D_s]
score_feat = torch.cat([
interact, # pairwise relation [E, 2*D_s]
sigma_i, # target uncertainty [E, 1]
sigma_j, # source uncertainty [E, 1]
mean_rel, # mean relative position [E, 2]
std_rel, # spatial spread over time [E, 2]
min_dist, # closest approach distance [E, 1]
], dim=-1) # [E, 2*D_s+7]
scores = self.edge_scorer(score_feat).squeeze(-1) # [E]
# ---- Top-N selection per target agent ---------------------------
scores_grouped = scores.view(B * K * A, A - 1)
_, top_idx = scores_grouped.topk(N, dim=-1, sorted=False)
mask = torch.zeros(B * K * A, A - 1,
device=scores.device, dtype=torch.bool)
mask.scatter_(1, top_idx, True)
mask_flat = mask.view(-1) # [B*K*E0]
# ---- RelTrajEncoder on selected edges only ----------------------
# ---- Heading diff on selected edges only ------------------------
heading = _heading_diff(pos_i_t[mask_flat],
pos_j_t[mask_flat]) # [B*K*A*N, T, 1]
rel_pos_sparse = torch.cat(
[rel_pos_t[mask_flat], heading], dim=-1
) # [B*K*A*N, T, 3]
if sigma_agent is not None:
sigma_full = sigma_agent.reshape(B * K * A, T)
sigma_i_t = sigma_full[edge_index_bk[1][mask_flat]]
sigma_j_t = sigma_full[edge_index_bk[0][mask_flat]]
sigma_bias = sigma_i_t - sigma_j_t # [B*K*A*N, T]
else:
sigma_bias = None
edge_attr_sparse = self.rel_traj_encoder(
rel_pos_sparse, sigma_bias
) # [B*K*A*N, D]
# ---- Sparse GNN pass --------------------------------------------
edge_index_sparse = edge_index_bk[:, mask_flat]
temb_bka = (t_emb
.unsqueeze(1).unsqueeze(2)
.expand(-1, K, A, -1)
.reshape(B * K * A, D))
nodes = y_emb.reshape(B * K * A, D)
for layer in self.gnn_layers:
nodes = layer(nodes, edge_index_sparse, edge_attr_sparse,
temb_agent=temb_bka, tau=tau_bka)
# ---- Gated residual ---------------------------------------------
orig = y_emb.reshape(B * K * A, D)
gate = self.gate_proj(torch.cat([orig, nodes], dim=-1))
out = orig + gate * self.out_proj(nodes)
return out.view(B, K, A, D)
|