src/loss.py

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


class SIGReg(torch.nn.Module):
    def __init__(self, knots=17,device='cpu'):
        super().__init__()
        self.device = device
        t = torch.linspace(0, 3, knots, dtype=torch.float32)
        dt = 3 / (knots - 1)
        weights = torch.full((knots,), 2 * dt, dtype=torch.float32)
        weights[[0, -1]] = dt
        window = torch.exp(-t.square() / 2.0)
        self.register_buffer("t", t)
        self.register_buffer("phi", window)
        self.register_buffer("weights", weights * window)

    def forward(self, proj):
        N, V = proj.shape[:2]
        proj = proj.reshape(N, V, -1).transpose(0, 1)
        A = torch.randn(proj.size(-1), 256, device=self.device)
        A = A.div_(A.norm(p=2, dim=0))
        x_t = (proj @ A).unsqueeze(-1) * self.t
        err = (x_t.cos().mean(-3) - self.phi).square() + x_t.sin().mean(-3).square()
        statistic = (err @ self.weights) * proj.size(-2) * (256/proj.size(-2))
        return statistic.mean()
    

def gram_anchor_spatial(xs, xt, eps=1e-6):
    """
    xs, xt: (N, V, D, H, W)
    Gram anchoring over spatial tokens H*W.
    """

    N, V, Ds, H, W = xs.shape
    N, V, Dt, H, W = xt.shape
    T = H * W

    # 1. Normalize over feature dimension D
    xs = F.normalize(xs, p=2, dim=2)    # (N, V, D, H, W)
    xt = F.normalize(xt, p=2, dim=2)

    # 2. Flatten spatial tokens -> (N, V, D, T)
    xs = xs.reshape(N, V, Ds, T)
    xt = xt.reshape(N, V, Dt, T)

    # 3. Move token dim second: (N, V, T, D)
    xs = xs.permute(0, 1, 3, 2)
    xt = xt.permute(0, 1, 3, 2)

    # 4. Merge N and V -> (N*V, T, D)
    xs = xs.reshape(N * V, T, Ds)
    xt = xt.reshape(N * V, T, Dt)

    # 5. Compute Gram matrices -> (N*V, T, T)
    Gs = torch.matmul(xs, xs.transpose(-1, -2))
    Gt = torch.matmul(xt, xt.transpose(-1, -2))

    # 6. Frobenius loss
    loss = (Gs - Gt).pow(2).mean()

    return loss