Reproducibility package: UWCA semantic multiple access (TWC submission)

rev2/e5_nonlinear.py

@@ -0,0 +1,55 @@
+"""E5 — Nonlinear inter-user semantic structure (R1.7, R2.5, R3.2).
+
+Embeddings are produced by fixed random per-user nonlinear view networks
+e_u = normalize(g_u([kappa*s ; p_u])), so the inter-user dependence is
+nonlinear and NOT captured by any scalar coefficient or linear covariance.
+NONLIN-HIGH: shared s (kappa=1); NONLIN-LOW: independent s per user.
+Shows the trained UWCA decoder still exploits the shared structure while the
+scalar-parameterized genie LMMSE (mis-specified here) cannot fully.
+"""
+import numpy as np
+import torch
+
+import lib
+from lib import (UWCA, DEVICE, ViewNets, block_masks, eval_scheme, save_json,
+                 set_seed, train_multitask, SNR_GRID)
+
+rng = set_seed(42)
+d, U, H = 64, 4, 4
+masks = block_masks(U, d)
+vnets = ViewNets(d, U).to(DEVICE)
+tasks = [{"snr_db": float(s)} for s in np.arange(0, 21, 4)]
+
+out = {"snr": SNR_GRID.tolist(), "cases": {}}
+for case, (kappa, shared) in {"NONLIN-HIGH": (1.0, True),
+                              "NONLIN-LOW": (1.0, False)}.items():
+    def gen(n, kappa=kappa, shared=shared):
+        return vnets.gen(n, d, U, rng, kappa, shared)
+
+    # empirical mean pairwise cosine (the "effective" relevance)
+    E = gen(2048)
+    C = torch.einsum("nud,nvd->uv", E, E) / 2048
+    off = C[~torch.eye(U, dtype=torch.bool, device=C.device)]
+    beta_emp = float(off.mean())
+    print(f"[E5-{case}] empirical mean pairwise cosine = {beta_emp:.3f}",
+          flush=True)
+
+    # mis-specified scalar-model LMMSE uses beta_emp for every pair
+    B = np.full((U, U), beta_emp)
+    np.fill_diagonal(B, 1.0)
+
+    model = UWCA(d, U, H).to(DEVICE)
+    train_multitask(model, gen, tasks, epochs=300, tag=f"E5-{case}")
+
+    res = {"beta_emp": beta_emp}
+    for scheme in ["uwca", "ofdma", "noma", "lmmse_genie"]:
+        sers, coss = [], []
+        for snr in SNR_GRID:
+            s, c = eval_scheme(scheme, gen, {"snr_db": float(snr)}, n_mc=200,
+                               model=model, B=B, masks=masks)
+            sers.append(s); coss.append(c)
+        res[scheme] = {"ser": sers, "cos": coss}
+        print(f"[E5-{case}] {scheme}: SER@10dB={sers[5]:.3f}", flush=True)
+    out["cases"][case] = res
+
+save_json("e5_nonlinear.json", out)