Files
2026-05-15 10:27:24 +02:00

78 lines
2.9 KiB
Python

"""pairwise_bridge — PairwiseAdditiveBridge.
For N input streams, compute the shifted-multiply ((1+a)(1+b) - 1) fusion for
every pair, then mix them with learnable per-pair scalar weights.
Motivation: the basic additive form (1+a)(1+b)...(1+N) - 1 helps for 2 streams
but regresses for 3+ because the triple-and-higher cross-terms (abc, abcd, …)
have explosive variance. This bridge keeps only the 2-way interactions —
O(N²) pair-experts rather than O(2^N) cross-terms — and lets the model learn
which pairs matter.
For 2 streams this reduces to a single weighted (1+a)(1+b)-1 → equivalent (up
to the scaling factor) to FusionBridge(additive=True).
"""
from __future__ import annotations
from itertools import combinations
import torch
import torch.nn as nn
from v4.classes.accessory.se_block import SEBlock
class PairwiseAdditiveBridge(nn.Module):
"""Sum of per-pair shifted-multiplies.
Parameters
----------
input_dims : ordered list of input embedding dims
fusion_dim : projection / output dimension
use_ln : LayerNorm after each per-stream projection (default: True)
use_se : SE gate on the fused vector
se_reduction : SE bottleneck factor
"""
def __init__(
self,
input_dims: list[int],
fusion_dim: int = 256,
use_ln: bool = True,
use_se: bool = True,
se_reduction: int = 16,
):
super().__init__()
self.out_dim = fusion_dim
self.W = nn.ModuleList([nn.Linear(d, fusion_dim) for d in input_dims])
self.ln = nn.ModuleList(
[nn.LayerNorm(fusion_dim) if use_ln else nn.Identity()
for _ in input_dims]
)
n_pairs = max(1, len(input_dims) * (len(input_dims) - 1) // 2)
# initialise to uniform mixing so each pair contributes equally at start
self.pair_weights = nn.Parameter(torch.full((n_pairs,), 1.0 / n_pairs))
self.se = SEBlock(fusion_dim, reduction=se_reduction, residual=True) if use_se else None
def forward(self, embeddings: list[torch.Tensor]) -> torch.Tensor:
assert len(embeddings) == len(self.W), (
f"PairwiseAdditiveBridge expects {len(self.W)} inputs, got {len(embeddings)}"
)
projected = [self.ln[i](self.W[i](e)) for i, e in enumerate(embeddings)]
if len(projected) == 1:
h = projected[0]
else:
pairs = list(combinations(range(len(projected)), 2))
h = 0
for idx, (i, j) in enumerate(pairs):
pair_fusion = (1.0 + projected[i]) * (1.0 + projected[j]) - 1.0
h = h + self.pair_weights[idx] * pair_fusion
if self.se is not None:
h, _ = self.se(h)
return h
def set_phase(self, phase: str) -> None:
enabled = phase not in ("tower_warmup", "cd_warmup")
for p in self.parameters():
p.requires_grad_(enabled)