src/losses/sensor_invariance.py

from __future__ import annotations

import math

import torch
import torch.nn as nn
import torch.nn.functional as F


class SensorInvarianceLoss(nn.Module):
    """Cross-sensor pair + adversarial sensor loss.

    Returns a 3-tuple ``(total, l_pair_detached, l_adv_detached)`` so callers
    can log sub-components without extra forward passes.

    GRL anti-correlation fix: l_adv is clamped at the random-guess baseline
    ``log(num_sensors)`` before being weighted into the total.  This prevents
    the backbone from overshooting — i.e. learning to actively anti-encode
    sensor identity — while still driving the discriminator toward confusion.
    Once l_adv >= log(N) the adversarial pressure on the backbone drops to
    zero and the GRL game reaches a stable equilibrium instead of cycling.
    """

    def __init__(self, delta: float = 0.05, lambda_adv: float = 0.3):
        super().__init__()
        self.delta = float(delta)
        self.lambda_adv = float(lambda_adv)

    def forward(
        self,
        score_s1: torch.Tensor,
        score_s2: torch.Tensor,
        sensor_logits: torch.Tensor,
        sensor_labels: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        l_pair = F.relu((score_s1 - score_s2).abs() - self.delta).mean()
        l_adv = F.cross_entropy(sensor_logits, sensor_labels)

        # Clamp adversarial contribution at the random-guess CE baseline.
        # Above this ceiling the backbone is already fooling the discriminator
        # sufficiently; continued gradient reversal only creates anti-correlation
        # oscillation without improving sensor invariance.
        rand_ce = math.log(sensor_logits.size(1))
        l_adv_clamped = l_adv.clamp(max=rand_ce)

        total = l_pair + self.lambda_adv * l_adv_clamped
        return total, l_pair.detach(), l_adv.detach()