variants.py

"""
Model extensions for Claim 5 (Sec. 7.1-7.2):
  (1) softmax variant   -> reuses capacity.train_run(softmax=True)
  (2) value-channel / message-retrieval variant (Sec. 7.2): frozen random value map
      W_V, mix messages with raw max-over-heads scores, train MSE to neighbor message.

Both keep the key-query path identical (max-over-heads, no 1/sqrt(d_k)) and test whether
the multi-head advantage persists.
"""
import math, argparse, json
import numpy as np
import torch
import torch.nn.functional as F
from capacity import (make_embeddings, make_permutation, sample_contexts,
                      MaxHeadAttention)


def value_retrieval_run(m, d_model, D_K, h, seed, device, d_msg=None,
                        ell=16, rho=0.5, lr=1e-3, max_steps=20000, batch=64,
                        check_every=500, patience=8, pool=20000, verbose=False):
    """Sec. 7.2: value-routing message retrieval. Reports test MSE (lower=better)."""
    torch.manual_seed(seed)
    d_msg = d_msg or d_model
    X = make_embeddings(m, d_model, seed, device)
    pi = make_permutation(m, seed)
    pi_t = torch.as_tensor(pi)
    WV = (torch.randn(d_model, d_msg, generator=torch.Generator().manual_seed(seed + 5))
          / math.sqrt(d_model)).to(device)            # frozen random value map
    Y_all = X @ WV                                     # per-node messages (m, d_msg)

    def batch_tensors(ctx):
        ctx_t = torch.as_tensor(ctx)
        Xc = X[ctx_t.to(device)]                       # (n, ell, d_model)
        tgt = pi_t[ctx_t]                              # (n, ell) neighbor id
        # valid position: neighbor in-context
        inctx = (tgt.unsqueeze(2) == ctx_t.unsqueeze(1))  # (n, ell, ell)
        valid = inctx.any(dim=2)                        # (n, ell)
        ymsg = Y_all[tgt.to(device)]                    # (n, ell, d_msg) true neighbor msg
        Vc = X[ctx_t.to(device)] @ WV                   # (n, ell, d_msg) in-context messages
        return Xc, Vc, ymsg, valid.to(device)

    model = MaxHeadAttention(d_model, D_K, h, agg="max").to(device)
    opt = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=0.0)

    train_ctx = sample_contexts(pi, m, ell, rho, pool, seed + 3)
    val_ctx = sample_contexts(pi, m, ell, rho, 500, seed + 1)
    test_ctx = sample_contexts(pi, m, ell, rho, 2000, seed + 2)
    Xtr, Vtr, Ytr, valtr = batch_tensors(train_ctx)
    Xv, Vv, Yv, valv = batch_tensors(val_ctx)
    Xte, Vte, Yte, valte = batch_tensors(test_ctx)

    def mse(Xc, Vc, ymsg, valid):
        S = model.scores(Xc)                # (n, ell, ell) raw max-over-heads
        yhat = torch.einsum("nij,njd->nid", S, Vc)   # (n, ell, d_msg)
        err = ((yhat - ymsg) ** 2).sum(-1)           # (n, ell)
        return err[valid].mean()

    g = torch.Generator().manual_seed(seed + 7)
    order = torch.randperm(pool, generator=g); ptr = 0
    best_val = float("inf"); bad = 0
    for step in range(1, max_steps + 1):
        if ptr + batch > pool:
            order = torch.randperm(pool, generator=g); ptr = 0
        idx = order[ptr:ptr + batch]; ptr += batch
        loss = mse(Xtr[idx], Vtr[idx], Ytr[idx], valtr[idx])
        opt.zero_grad(); loss.backward(); opt.step()
        if step % check_every == 0:
            with torch.no_grad():
                v = mse(Xv, Vv, Yv, valv).item()
            if verbose:
                print(f"  step {step} valMSE {v:.4f}", flush=True)
            if v < best_val - 1e-4:
                best_val = v; bad = 0
            else:
                bad += 1
                if bad >= patience:
                    break
    with torch.no_grad():
        test_mse = mse(Xte, Vte, Yte, valte).item()
    return {"m": m, "d_model": d_model, "D_K": D_K, "h": h, "d_k": D_K // h,
            "seed": seed, "test_mse": test_mse, "steps": step}


if __name__ == "__main__":
    dev = "cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu")
    print("device", dev)
    # Value retrieval: m=64, compressed d_model=16 -> multi-head advantage (Fig 6 top)
    for dm in [16, 32]:
        print(f"\n=== Value retrieval m=64 d_model={dm}  (Sec 7.2, Fig 6) ===")
        for h in [1, 2, 4, 8]:
            for DK in [16, 32]:
                if DK % h or DK // h < 2:
                    continue
                r = value_retrieval_run(64, dm, DK, h, seed=0, device=dev,
                                        max_steps=8000, batch=64)
                print(f"  d_model={dm} h={h} D_K={DK} test_MSE={r['test_mse']:.4f}")