MoFlow/models/graph_interaction_nba_v9.py

"""
FutureInteractionGraphV9 — V8 (kinematic) + receiver-adaptive temporal readout.

Instead of fixed attention-pooling over T, the receiving node's embedding
queries the temporal edge features via cross-attention. This makes the edge
encoding adaptive: a defender cares about early timesteps (before screen),
a help defender cares about late timesteps (after rotation).

Combines V8's kinematic channels with V9's adaptive readout.
"""

import torch
import torch.nn as nn
from models.graph_interaction_nba_v8 import FutureInteractionGraphV8, _compute_kinematic_features
from models.graph_interaction_nba_v3 import _TemporalSelfAttn
from models.graph_interaction_nba_v5 import _heading_diff


class AdaptiveRelTrajEncoder(nn.Module):
    """RelTrajEncoder with cross-attention readout conditioned on receiver node.

    Architecture:
      1. Per-timestep projection: [E, T, in_ch] → [E, T, D_hidden]
      2. Temporal PE + self-attention over T
      3. Cross-attention: node_i queries temporal features → [E, D_hidden]
      4. Output projection → [E, out_dim]
    """

    def __init__(self, out_dim, T=20, D_hidden=32, num_heads=4,
                 in_channels=8, node_dim=128):
        super().__init__()
        self.input_proj = nn.Linear(in_channels, D_hidden)
        self.t_pe = nn.Embedding(T, D_hidden)
        self.attn = _TemporalSelfAttn(D_hidden, num_heads)

        # Cross-attention: node_i (query) attends to temporal features (KV)
        self.q_proj = nn.Linear(node_dim, D_hidden)
        self.k_proj = nn.Linear(D_hidden, D_hidden)
        self.v_proj = nn.Linear(D_hidden, D_hidden)
        self.cross_scale = D_hidden ** -0.5

        self.out_proj = nn.Linear(D_hidden, out_dim)

    def forward(self, rel_features, sigma_bias=None, node_i_emb=None):
        """
        Args:
            rel_features: [E, T, in_ch]
            sigma_bias:   [E, T] or None
            node_i_emb:   [E, node_dim] — receiving node's embedding
        Returns:
            [E, out_dim]
        """
        E, T, _ = rel_features.shape
        h = self.input_proj(rel_features)                          # [E, T, D_h]
        h = h + self.t_pe(torch.arange(T, device=h.device))
        h = self.attn(h)                                           # [E, T, D_h]

        if node_i_emb is not None:
            # Cross-attention: node_i queries temporal features
            q = self.q_proj(node_i_emb).unsqueeze(1)               # [E, 1, D_h]
            k = self.k_proj(h)                                     # [E, T, D_h]
            v = self.v_proj(h)                                     # [E, T, D_h]

            attn_logits = (q * k).sum(dim=-1) * self.cross_scale   # [E, T]
            if sigma_bias is not None:
                attn_logits = attn_logits + sigma_bias
            attn_w = attn_logits.softmax(dim=-1).unsqueeze(-1)     # [E, T, 1]
            pooled = (v * attn_w).sum(dim=1)                       # [E, D_h]
        else:
            # Fallback: standard pooling (for backward compatibility)
            w_logits = (h * h.mean(dim=1, keepdim=True)).sum(dim=-1, keepdim=True)
            if sigma_bias is not None:
                w_logits = w_logits + sigma_bias.unsqueeze(-1)
            w = w_logits.softmax(dim=1)
            pooled = (h * w).sum(dim=1)

        return self.out_proj(pooled)


class FutureInteractionGraphV9(FutureInteractionGraphV8):
    """V8 (kinematic) + receiver-adaptive temporal readout."""

    def __init__(self, embed_dim, future_steps, num_agents,
                 num_heads=4, dropout=0.1, num_gnn_layers=2,
                 time_dim=128, top_n_neighbors=5, rel_traj_hidden=32,
                 y0_score_dim=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, y0_score_dim=y0_score_dim)

        # Replace standard RelTrajEncoder with adaptive version
        self.rel_traj_encoder = AdaptiveRelTrajEncoder(
            out_dim=embed_dim, T=future_steps,
            D_hidden=rel_traj_hidden, num_heads=4,
            in_channels=8, node_dim=embed_dim)

    def forward(self, y_emb, y_abs, t_emb, tau, sigma_agent=None):
        B, K, A, D = y_emb.shape
        T = y_abs.shape[3]
        E0 = self._E0
        N = self.top_n

        # ---- Scoring (same as V6) ----
        y0_flat = y_abs.reshape(B * K * A, T * 2)
        y0_emb = self.y0_score_proj(y0_flat)

        if sigma_agent is not None:
            sigma_mean = sigma_agent.mean(dim=-1).reshape(B * K * A, 1)
            tau_bka = sigma_mean.squeeze(-1)
        else:
            sigma_mean = torch.zeros(B * K * A, 1, device=y_abs.device)
            tau_bka = (tau.unsqueeze(1).unsqueeze(2)
                       .expand(-1, K, A).reshape(B * K * A))

        node_feat = torch.cat([y0_emb, sigma_mean], dim=-1)
        q_bka = self.W_q(node_feat)
        k_bka = self.W_k(node_feat)

        pos_bk = y_abs.reshape(B * K * A, T, 2)
        edge_index_bk = self._make_batched_edge_index(B * K)

        pos_i_t = pos_bk[edge_index_bk[1]]
        pos_j_t = pos_bk[edge_index_bk[0]]

        # Geometric features for scoring
        rel_pos_t = pos_j_t - pos_i_t
        mean_rel = rel_pos_t.mean(dim=1)
        std_rel = rel_pos_t.std(dim=1)
        min_dist = rel_pos_t.norm(dim=-1).min(dim=1).values.unsqueeze(-1)
        heading_full = _heading_diff(pos_i_t, pos_j_t)
        heading_mean = heading_full.mean(dim=1)

        q_i = q_bka[edge_index_bk[1]]
        k_j = k_bka[edge_index_bk[0]]
        semantic_score = (q_i * k_j).sum(dim=-1) * self.scale
        geo_feat = torch.cat([mean_rel, std_rel, min_dist, heading_mean], dim=-1)
        geo_bias = self.geo_mlp(geo_feat).squeeze(-1)
        scores = semantic_score + geo_bias

        # ---- Top-N selection ----
        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)

        # ---- Kinematic features + adaptive readout (NEW) ----
        kinematic_sparse = _compute_kinematic_features(
            pos_i_t[mask_flat], pos_j_t[mask_flat])

        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
        else:
            sigma_bias = None

        # Get receiver node embeddings for cross-attention
        node_i_emb = y_emb.reshape(B * K * A, D)[edge_index_bk[1][mask_flat]]

        edge_attr_sparse = self.rel_traj_encoder(
            kinematic_sparse, sigma_bias, node_i_emb=node_i_emb)

        # ---- GNN + gated residual ----
        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)

        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)