gdcn.py

1"""2GDCN：门控交叉网络（第15章 §15.5.9）34Gated Cross Layer：5    c_{l+1} = c_0 ⊙ (W^c_l c_l + b_l) ⊙ σ(W^g_l c_l) + c_l6           =  Feature Crossing      ×   Information Gate  +  残差7- W^c 学交叉关系，W^g 学逐维门控，σ 把门控压到 (0,1)8- 门控为 1 → 退化为 DCN-V2；门控学到的值可以直接看"哪一维交互被保留"9两种组合方式：10    GDCN-S（Stacked）：c_0 → GCN → c_L → DNN11    GDCN-P（Parallel）：c_0 → GCN 与 DNN 并行，输出拼接1213运行：python3 ch04_feature_interaction/gdcn.py14"""1516from __future__ import annotations1718import pathlib19import sys2021import torch22import torch.nn as nn2324sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))2526from _train_utils import (  # noqa: E40227    N_FIELDS, MLP, BaselineMLP, count_params, embedding_table, print_comparison,28    run_experiment, stack_embeddings,29)3031DIM = 432FLAT = N_FIELDS * DIM333435class GatedCrossLayer(nn.Module):36    def __init__(self, dim: int):37        super().__init__()38        self.wc = nn.Linear(dim, dim)39        self.wg = nn.Linear(dim, dim)40        self.b = nn.Parameter(torch.zeros(dim))4142    def forward(self, c0: torch.Tensor, cl: torch.Tensor):43        cross = c0 * (self.wc(cl) + self.b)44        gate = torch.sigmoid(self.wg(cl))45        return cross * gate + cl464748class GDCN(nn.Module):49    def __init__(self, mode: str = "S", dim: int = DIM, n_cross: int = 3,50                 hidden=(32,), use_gate: bool = True):51        super().__init__()52        assert mode in {"S", "P"}53        self.mode = mode54        self.dim = dim55        self.emb = embedding_table(dim)56        self.cross = nn.ModuleList([GatedCrossLayer(FLAT) for _ in range(n_cross)])57        self.use_gate = use_gate58        if mode == "S":59            self.deep = MLP(FLAT, hidden, out_dim=None, dropout=0.2)60            self.head = nn.Linear(hidden[-1], 1)61        else:62            self.deep = MLP(FLAT, hidden, out_dim=None, dropout=0.2)63            self.head = nn.Linear(FLAT + hidden[-1], 1)6465    def gated_cross(self, c0: torch.Tensor) -> torch.Tensor:66        c = c067        for layer in self.cross:68            if self.use_gate:69                c = layer(c0, c)70            else:  # 去掉门控即 DCN-V2 的交叉层71                c = c0 * (layer.wc(c) + layer.b) + c72        return c7374    def forward(self, x):75        c0 = stack_embeddings(self.emb, x).flatten(1)76        if self.mode == "S":77            return self.head(self.deep(self.gated_cross(c0))).squeeze(-1)78        return self.head(torch.cat([self.gated_cross(c0), self.deep(c0)], dim=1)).squeeze(-1)798081def main() -> None:82    print("GDCN：c_{l+1} = c0 ⊙ (W^c c_l + b) ⊙ σ(W^g c_l) + c_l（§15.5.9）\n")83    _, a_s, _ = run_experiment(lambda: GDCN(mode="S"), "GDCN-S", steps=600)84    _, a_p, _ = run_experiment(lambda: GDCN(mode="P"), "GDCN-P", steps=600)85    _, a_nogate, _ = run_experiment(86        lambda: GDCN(mode="S", use_gate=False), "GDCN-S 无门控", steps=600)87    _, a_base, _ = run_experiment(BaselineMLP, "Baseline", steps=600)8889    print_comparison([90        ("Baseline MLP", a_base, count_params(BaselineMLP())),91        ("GDCN-S（串联）", a_s, count_params(GDCN(mode="S"))),92        ("GDCN-P（并联）", a_p, count_params(GDCN(mode="P"))),93        ("GDCN-S 无门控(≈DCN-V2)", a_nogate, count_params(GDCN(mode="S", use_gate=False))),94    ])9596    # 看门控值：每一维交叉信息被保留了多少97    torch.manual_seed(0)98    model = GDCN(mode="S")99    from _train_utils import make_dataset, to_tensor100    Xtr, ytr = make_dataset(4_000, seed=1)101    from _train_utils import fit102    Xtr, ytr = to_tensor(Xtr, ytr)103    fit(model, Xtr, ytr, Xtr[:2000], ytr[:2000], steps=200, name="")104    with torch.no_grad():105        inner = Xtr[:256]106        c = stack_embeddings(model.emb, inner).flatten(1)107        gates = []108        for layer in model.cross:109            gates.append(torch.sigmoid(layer.wg(c)).mean(0))110            c = layer(c0=stack_embeddings(model.emb, inner).flatten(1), cl=c)111    print("\n各层 Gate 平均值（取 256 个样本、40 维求均值）：")112    for l, g in enumerate(gates):113        print(f"  layer {l}: mean={g.mean():.3f}  min={g.min():.3f}  max={g.max():.3f}")114    print("\nGate 不是常数而是随样本变化：同一维交叉信息在有的样本里被放大、")115    print("在另一些样本里被抑制，这就是正文所说的'实例级筛选'（§15.5.9）。")116117118if __name__ == "__main__":119    main()