src/models/mdgt/relational_pe.py

"""
Relational Positional Encoding (RPE).

For every pair of minutiae (i, j) we compute a 7-dim geometric relation
vector and project it through a small MLP to obtain a learned embedding:

    rᵢⱼ = [Δxᵢⱼ, Δyᵢⱼ, dᵢⱼ, cos(αᵢⱼ), sin(αᵢⱼ), cos(Δθᵢⱼ), sin(Δθᵢⱼ)]  ∈ ℝ⁷

    PE(i,j) = MLP_rel(rᵢⱼ)  ∈ ℝᵈ

Rationale for each component:

    (Δx, Δy)         Raw displacement — translation-invariant by construction:
                      (xⱼ+t) − (xᵢ+t) = xⱼ − xᵢ.  Lets the MLP learn any
                      spatial pattern from relative position.

    dᵢⱼ              Explicit Euclidean distance.  Although the MLP could learn
                      √(Δx²+Δy²), providing it explicitly accelerates convergence
                      — the model doesn't have to "reinvent" the concept of distance.

    (cos α, sin α)   Direction angle from minutia i to j on the plane.
                      Combined with d, this gives a complete polar-coordinate
                      representation of relative position.

    (cos Δθ, sin Δθ) Difference in ridge orientation — the most biometrically
                      discriminative feature.  Two minutiae at the same distance
                      but different Δθ are very different fingerprint evidence.
                      Using (cos, sin) handles angle periodicity correctly.
"""

import torch
import torch.nn as nn


def compute_pairwise_relations(
    x: torch.Tensor,
    y: torch.Tensor,
    theta: torch.Tensor,
) -> torch.Tensor:
    """Compute pairwise 7-dim relational features among minutiae.

    Args:
        x:     (B, N) x-coordinates.
        y:     (B, N) y-coordinates.
        theta: (B, N) orientations in **radians**.

    Returns:
        rel:   (B, N, N, 7) — [Δx, Δy, d, cos α, sin α, cos Δθ, sin Δθ]
    """
    # Δx, Δy — raw displacement (translation invariant)
    dx = x.unsqueeze(2) - x.unsqueeze(1)   # (B, N, N)
    dy = y.unsqueeze(2) - y.unsqueeze(1)

    # d — Euclidean distance (explicit, helps convergence)
    dist = torch.sqrt(dx ** 2 + dy ** 2 + 1e-8)

    # Normalize spatial features so all 7 dims are ~ [-1, 1] scale.
    # Without this, pixel-scale dx/dy/dist (0–600) drown out cos/sin (−1 to 1).
    # Normalization is per-sample: divide by the max distance in each sample.
    scale = dist.amax(dim=(1, 2), keepdim=True).clamp(min=1.0)   # (B, 1, 1)
    dx_n = dx / scale
    dy_n = dy / scale
    dist_n = dist / scale

    # α — direction angle from i to j (polar coordinate)
    alpha = torch.atan2(dy, dx)             # (B, N, N)
    cos_alpha = torch.cos(alpha)
    sin_alpha = torch.sin(alpha)

    # Δθ — orientation difference (most biometric-discriminative)
    dtheta = theta.unsqueeze(2) - theta.unsqueeze(1)
    cos_dtheta = torch.cos(dtheta)
    sin_dtheta = torch.sin(dtheta)

    rel = torch.stack([
        dx_n, dy_n, dist_n,
        cos_alpha, sin_alpha,
        cos_dtheta, sin_dtheta,
    ], dim=-1)                              # (B, N, N, 7)
    return rel


class RelationalPE(nn.Module):
    """Project pairwise geometric relations into a learned embedding space.

    MLP_rel: Linear(7, hidden) → LayerNorm → GELU → ... → Linear(hidden, out)

    The output can be used as:
      • additive attention bias  (scalar per head), or
      • multiplicative gate on attention logits, or
      • concatenated edge features.

    Parameters
    ----------
    input_dim   : raw relation dimensionality (default 7).
    hidden_dim  : MLP hidden width.
    output_dim  : final embedding size.
    num_layers  : depth of the projection MLP.
    activation  : nonlinearity name (``"gelu"`` | ``"relu"``).
    """

    def __init__(
        self,
        input_dim: int = 7,
        hidden_dim: int = 64,
        output_dim: int = 64,
        num_layers: int = 2,
        activation: str = "gelu",
    ):
        super().__init__()
        act = nn.GELU() if activation == "gelu" else nn.ReLU()

        layers: list[nn.Module] = []
        dims = [input_dim] + [hidden_dim] * (num_layers - 1) + [output_dim]
        for i in range(len(dims) - 1):
            layers.append(nn.Linear(dims[i], dims[i + 1]))
            if i < len(dims) - 2:          # no act after final projection
                layers.append(nn.LayerNorm(dims[i + 1]))
                layers.append(act)
        self.mlp = nn.Sequential(*layers)

        self._init_weights()

    # ------------------------------------------------------------------
    def _init_weights(self):
        for m in self.mlp:
            if isinstance(m, nn.Linear):
                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    # ------------------------------------------------------------------
    def forward(
        self,
        minutiae: torch.Tensor,
    ) -> torch.Tensor:
        """
        Args:
            minutiae: (B, N, 3) — columns are (x, y, θ).

        Returns:
            rpe: (B, N, N, output_dim) learned relational embeddings.
        """
        x = minutiae[..., 0]
        y = minutiae[..., 1]
        theta = minutiae[..., 2]

        rel = compute_pairwise_relations(x, y, theta)       # (B, N, N, 7)
        rpe = self.mlp(rel)                                  # (B, N, N, out)
        return rpe