From bd9333eda60a9029a198acaeacb1eca4312bd1e8 Mon Sep 17 00:00:00 2001
From: YurenHao0426 <blackhao0426@gmail.com>
Date: Mon, 4 May 2026 23:05:16 -0500
Subject: =?UTF-8?q?Initial=20release:=20GRAFT=20(KAFT)=20=E2=80=94=20NeurI?=
 =?UTF-8?q?PS=202026=20submission=20code?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Topology-factorized Jacobian-aligned feedback for deep GNNs. Includes:
- src/: GraphGrAPETrainer (KAFT) + BP / DFA / DFA-GNN / VanillaGrAPE baselines
        + multi-probe alignment estimator + dataset / sparse-mm utilities.
- experiments/: 19 runners reproducing every figure / table in the paper.
- figures/: 4 generators + the 4 PDFs cited in the report.
- paper/: NeurIPS .tex and consolidated experiments_master notes.

Smoke test: 50-epoch Cora GCN L=4 gives BP 77.3% / KAFT 79.0%.
---
 experiments/run_ablation_20seeds.py | 115 ++++++++++++++++++++++++++++++++++++
 1 file changed, 115 insertions(+)
 create mode 100644 experiments/run_ablation_20seeds.py

(limited to 'experiments/run_ablation_20seeds.py')

diff --git a/experiments/run_ablation_20seeds.py b/experiments/run_ablation_20seeds.py
new file mode 100644
index 0000000..61055ed
--- /dev/null
+++ b/experiments/run_ablation_20seeds.py
@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+"""Ablation study with 20 seeds: BP → DFA → DFA-GNN → VanillaGrAPE → GRAFT."""
+
+import torch
+import numpy as np
+import json
+import os
+from scipy import stats as scipy_stats
+from src.data import load_dataset
+from src.trainers import BPTrainer, DFATrainer, DFAGNNTrainer, VanillaGrAPETrainer, GraphGrAPETrainer
+
+device = 'cuda:0'
+SEEDS = list(range(20))
+EPOCHS = 200
+OUT_DIR = 'results/ablation_20seeds'
+
+METHODS = {
+    'BP': (BPTrainer, {}),
+    'DFA': (DFATrainer, {}),
+    'DFA-GNN': (DFAGNNTrainer, {'topo_mode': 'fixed_A'}),
+    'VanillaGrAPE': (VanillaGrAPETrainer, {
+        'diffusion_alpha': 0.5, 'diffusion_iters': 10,
+        'lr_feedback': 0.5, 'num_probes': 64, 'topo_mode': 'fixed_A'
+    }),
+    'GRAFT': (GraphGrAPETrainer, {
+        'diffusion_alpha': 0.5, 'diffusion_iters': 10,
+        'lr_feedback': 0.5, 'num_probes': 64, 'topo_mode': 'fixed_A'
+    }),
+}
+
+
+def train_one(cls, common, extra, seed):
+    torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed)
+    t = cls(**common, **extra)
+    if hasattr(t, 'align_mode'):
+        t.align_mode = 'chain_norm'
+    bv, bt = 0, 0
+    for ep in range(EPOCHS):
+        t.train_step()
+        if ep % 5 == 0:
+            v = t.evaluate('val_mask')
+            te = t.evaluate('test_mask')
+            if v > bv: bv, bt = v, te
+    del t; torch.cuda.empty_cache()
+    return bt
+
+
+def main():
+    os.makedirs(OUT_DIR, exist_ok=True)
+    per_seed_file = os.path.join(OUT_DIR, 'per_seed_data.json')
+    if os.path.exists(per_seed_file):
+        with open(per_seed_file) as f:
+            per_seed_data = json.load(f)
+    else:
+        per_seed_data = {}
+
+    results = {}
+
+    for ds_name in ['Cora', 'CiteSeer', 'PubMed']:
+        data = load_dataset(ds_name, device=device)
+        common = dict(data=data, hidden_dim=64, lr=0.01, weight_decay=5e-4,
+                      num_layers=6, residual_alpha=0.0, backbone='gcn')
+
+        for mname, (cls, extra) in METHODS.items():
+            key = f"{ds_name}_{mname}"
+            print(f"\n=== {key} (20 seeds) ===", flush=True)
+
+            if key not in per_seed_data:
+                per_seed_data[key] = {}
+
+            for seed in SEEDS:
+                sk = str(seed)
+                if sk in per_seed_data[key]:
+                    print(f"  seed {seed}: cached", flush=True)
+                    continue
+                acc = train_one(cls, common, extra, seed)
+                per_seed_data[key][sk] = acc
+                print(f"  seed {seed}: {acc*100:.1f}%", flush=True)
+
+                with open(per_seed_file, 'w') as f:
+                    json.dump(per_seed_data, f, indent=2)
+
+            accs = np.array([per_seed_data[key][str(s)] for s in SEEDS]) * 100
+            results[key] = {
+                'mean': float(accs.mean()), 'std': float(accs.std()),
+                'accs': accs.tolist(),
+            }
+            print(f"  {mname}: {accs.mean():.1f} ± {accs.std():.1f}%")
+
+        del data; torch.cuda.empty_cache()
+
+    # Paired t-tests between adjacent methods
+    print("\n=== Paired t-tests (adjacent methods) ===")
+    method_names = list(METHODS.keys())
+    for ds in ['Cora', 'CiteSeer', 'PubMed']:
+        print(f"\n{ds}:")
+        for i in range(len(method_names) - 1):
+            m1, m2 = method_names[i], method_names[i+1]
+            a1 = np.array(results[f"{ds}_{m1}"]['accs'])
+            a2 = np.array(results[f"{ds}_{m2}"]['accs'])
+            t_stat, p_val = scipy_stats.ttest_rel(a2, a1)
+            sig = '***' if p_val < 0.001 else ('**' if p_val < 0.01 else ('*' if p_val < 0.05 else 'ns'))
+            delta = a2.mean() - a1.mean()
+            results[f"{ds}_{m1}_vs_{m2}"] = {
+                'delta': float(delta), 't_stat': float(t_stat), 'p_value': float(p_val)
+            }
+            print(f"  {m1} → {m2}: Δ{delta:+.1f}% p={p_val:.4f} {sig}")
+
+    with open(os.path.join(OUT_DIR, 'results.json'), 'w') as f:
+        json.dump(results, f, indent=2)
+    print(f"\nSaved to {OUT_DIR}/results.json")
+
+
+if __name__ == '__main__':
+    main()
-- 
cgit v1.2.3