Restructure package: descriptive study documentation and clean layout

simulation/revision_e2e.py

@@ -0,0 +1,31 @@
+"""End-to-end (learnable encoder) vs decoder-only MAML, HIGH scenario.
+Verifies Reviewer-1 Comment-3 claim that cross-attention stays stable and
+effective when the encoder is also learned."""
+import types, numpy as np, torch
+import maml_semantic as M
+torch.manual_seed(0)
+
+def cfg(**k):
+    b = dict(d=64, U=4, H=4, tau=0.45, lam=0.1, snr_min=0.0, snr_max=20.0,
+             snr_step=2.0, inner_lr=0.01, inner_steps=5, outer_lr=1e-3,
+             meta_epochs=120, joint_epochs=120, batch=64, n_mc=80, seed=42,
+             scenario='HIGH', decoder_only=True)
+    b.update(k); return types.SimpleNamespace(**b)
+
+def run(decoder_only):
+    c = cfg(decoder_only=decoder_only); dev='cpu'; rng=np.random.default_rng(c.seed)
+    scen=M.SCENARIO_CONFIGS[c.scenario]
+    mdl=M.SemanticCommSystem(c.d,c.U,c.H,decoder_only=decoder_only).to(dev)
+    M.MAMLTrainer(mdl,c,dev,rng,scen).train()
+    res=M.evaluate_model(mdl,c,dev,rng,"rayleigh",scen)
+    snr=np.arange(0,20.0001,2);
+    def at(s): return res['ser'][int(np.argmin(np.abs(snr-s)))], res['cos'][int(np.argmin(np.abs(snr-s)))]
+    i10=int(np.argmin(np.abs(snr-10))); rho=res['rho'][i10]; mask=~np.eye(c.U,dtype=bool)
+    return at(4), at(10), at(16), float(np.abs(rho[mask]).mean())
+
+print("=== END-TO-END (learnable encoder) vs DECODER-ONLY, HIGH ===")
+for tag, do in [("decoder-only (frozen enc)", True), ("end-to-end (learnable enc)", False)]:
+    (s4,c4),(s10,c10),(s16,c16),rho = run(do)
+    print(f"  {tag:30s}  SER@4/10/16 = {s4:.3f}/{s10:.3f}/{s16:.3f}  "
+          f"cos@10={c10:.3f}  |rho_off|@10={rho:.3f}")
+print("DONE.")