81 lines
2.9 KiB
Python
81 lines
2.9 KiB
Python
"""gated_bridge — GatedAdditiveBridge.
|
||
|
||
Per-sample, per-stream learned gates determine how much each stream contributes
|
||
to the fused embedding. Each gate is a sigmoid scalar produced by an MLP over
|
||
the raw input embeddings, so the gate is conditioned on the actual content of
|
||
all streams — when a stream's signal is weak for a particular sample, its gate
|
||
can attenuate toward 0; when it's informative, gate goes toward 1.
|
||
|
||
h = Σ_i g_i(x) · LN_i(W_i z_i)
|
||
g_i(x) = σ(MLP_i([z_1, z_2, …, z_N]))
|
||
|
||
No symmetry-breaking between streams — every tower is treated identically; the
|
||
gate network decides per-sample which to amplify. Gates use sigmoid (not
|
||
softmax) so they can be independently small or large; the model isn't forced
|
||
into a "pick one" distribution.
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import torch
|
||
import torch.nn as nn
|
||
|
||
from v4.classes.accessory.se_block import SEBlock
|
||
|
||
|
||
class GatedAdditiveBridge(nn.Module):
|
||
"""Per-sample sigmoid-gated additive fusion.
|
||
|
||
Parameters
|
||
----------
|
||
input_dims : ordered list of input embedding dims
|
||
fusion_dim : projection / output dimension
|
||
gate_hidden : hidden width of the gating MLP (default: fusion_dim)
|
||
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,
|
||
gate_hidden: int = 128,
|
||
use_ln: bool = True,
|
||
use_se: bool = True,
|
||
se_reduction: int = 16,
|
||
):
|
||
super().__init__()
|
||
self.out_dim = fusion_dim
|
||
self.n = len(input_dims)
|
||
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]
|
||
)
|
||
self.gate = nn.Sequential(
|
||
nn.Linear(sum(input_dims), gate_hidden),
|
||
nn.ReLU(inplace=True),
|
||
nn.Linear(gate_hidden, self.n),
|
||
nn.Sigmoid(),
|
||
)
|
||
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) == self.n, (
|
||
f"GatedAdditiveBridge expects {self.n} inputs, got {len(embeddings)}"
|
||
)
|
||
projected = [self.ln[i](self.W[i](e)) for i, e in enumerate(embeddings)]
|
||
gate_in = torch.cat(embeddings, dim=-1)
|
||
gates = self.gate(gate_in) # (B, N) ∈ (0, 1)
|
||
h = 0
|
||
for i in range(self.n):
|
||
h = h + gates[..., i:i+1] * projected[i]
|
||
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)
|