diff options
Diffstat (limited to 'experiments')
| -rw-r--r-- | experiments/run_ablation_20seeds.py | 6 | ||||
| -rw-r--r-- | experiments/run_bp_kaft_depth.py (renamed from experiments/run_bp_graft_depth.py) | 12 | ||||
| -rw-r--r-- | experiments/run_combo_20seeds.py | 72 | ||||
| -rw-r--r-- | experiments/run_cora_perturb.py | 4 | ||||
| -rw-r--r-- | experiments/run_cs_full.py | 14 | ||||
| -rw-r--r-- | experiments/run_dblp_depth.py | 8 | ||||
| -rw-r--r-- | experiments/run_dblp_depth_scaling.py | 14 | ||||
| -rw-r--r-- | experiments/run_depth_extras.py | 6 | ||||
| -rw-r--r-- | experiments/run_dfagnn_depth.py | 4 | ||||
| -rw-r--r-- | experiments/run_grad_reach_20seeds.py | 22 | ||||
| -rw-r--r-- | experiments/run_hero_extras.py | 12 | ||||
| -rw-r--r-- | experiments/run_realworld_hero_L20.py | 18 | ||||
| -rw-r--r-- | experiments/run_resgcn_20seeds.py | 20 | ||||
| -rw-r--r-- | experiments/run_shallow_depth.py | 20 | ||||
| -rw-r--r-- | experiments/run_wikics_paper_setup.py | 14 |
15 files changed, 122 insertions, 124 deletions
diff --git a/experiments/run_ablation_20seeds.py b/experiments/run_ablation_20seeds.py index 61055ed..d2bd434 100644 --- a/experiments/run_ablation_20seeds.py +++ b/experiments/run_ablation_20seeds.py @@ -1,5 +1,5 @@ #!/usr/bin/env python3 -"""Ablation study with 20 seeds: BP → DFA → DFA-GNN → VanillaGrAPE → GRAFT.""" +"""Ablation study with 20 seeds: BP → DFA → DFA-GNN → VanillaGrAPE → KAFT.""" import torch import numpy as np @@ -7,7 +7,7 @@ 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 +from src.trainers import BPTrainer, DFATrainer, DFAGNNTrainer, VanillaGrAPETrainer, KAFTTrainer device = 'cuda:0' SEEDS = list(range(20)) @@ -22,7 +22,7 @@ METHODS = { 'diffusion_alpha': 0.5, 'diffusion_iters': 10, 'lr_feedback': 0.5, 'num_probes': 64, 'topo_mode': 'fixed_A' }), - 'GRAFT': (GraphGrAPETrainer, { + 'KAFT': (KAFTTrainer, { 'diffusion_alpha': 0.5, 'diffusion_iters': 10, 'lr_feedback': 0.5, 'num_probes': 64, 'topo_mode': 'fixed_A' }), diff --git a/experiments/run_bp_graft_depth.py b/experiments/run_bp_kaft_depth.py index 1e8dd76..d3119e1 100644 --- a/experiments/run_bp_graft_depth.py +++ b/experiments/run_bp_kaft_depth.py @@ -1,10 +1,10 @@ #!/usr/bin/env python3 -"""H9: BP + GRAFT depth sweep on Cora/CiteSeer/PubMed. +"""H9: BP + KAFT depth sweep on Cora/CiteSeer/PubMed. E1 already did DBLP L={8,12,16,20,24,32}. This fills the gap for Cora/CiteSeer/PubMed at L={8,10,12,16,20} so we can plot Figure 4(a)-style depth curves on 4 datasets. -BP + GRAFT only (GRAFT+ResGCN not needed for this figure — that's stacking table). +BP + KAFT only (KAFT+ResGCN not needed for this figure — that's stacking table). """ import torch @@ -12,20 +12,20 @@ import numpy as np import json import os from src.data import load_dataset -from src.trainers import BPTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, KAFTTrainer device = 'cuda:0' SEEDS = list(range(20)) EPOCHS = 200 DEPTHS = [8, 10, 12, 16, 20] -OUT_DIR = 'results/bp_graft_depth_20seeds' +OUT_DIR = 'results/bp_kaft_depth_20seeds' grape_extra = dict(diffusion_alpha=0.5, diffusion_iters=10, lr_feedback=0.5, num_probes=64, topo_mode='fixed_A') METHODS = { 'BP': (BPTrainer, {}), - 'GRAFT': (GraphGrAPETrainer, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), } @@ -90,7 +90,7 @@ def main(): del data; torch.cuda.empty_cache() # Summary - print(f"\n{'=' * 70}\nBP/GRAFT depth sweep summary\n{'=' * 70}") + print(f"\n{'=' * 70}\nBP/KAFT depth sweep summary\n{'=' * 70}") results = {} for ds in datasets_cfg: print(f"\n{ds}:") diff --git a/experiments/run_combo_20seeds.py b/experiments/run_combo_20seeds.py index 1598964..acc8846 100644 --- a/experiments/run_combo_20seeds.py +++ b/experiments/run_combo_20seeds.py @@ -1,7 +1,7 @@ #!/usr/bin/env python3 -"""Task 2ceadaa7: GRAFT + Forward Tricks combo experiments (20 seeds). +"""Task 2ceadaa7: KAFT + Forward Tricks combo experiments (20 seeds). -Combos: GRAFT+ResGCN, GRAFT+DropEdge, GRAFT+PairNorm, GRAFT+JKNet +Combos: KAFT+ResGCN, KAFT+DropEdge, KAFT+PairNorm, KAFT+JKNet Each compared to: BP, forward_trick_only, GRAFT_only, combo """ @@ -12,7 +12,7 @@ import json import os from scipy import stats as scipy_stats from src.data import load_dataset, spmm, build_normalized_adj -from src.trainers import BPTrainer, GraphGrAPETrainer, _FeedbackTrainerBase +from src.trainers import BPTrainer, KAFTTrainer, _FeedbackTrainerBase from run_deep_baselines import ResGCNTrainer, JKNetTrainer from run_dropedge import BPDropEdgeTrainer from run_pairnorm_baseline import BPPairNormTrainer, pairnorm @@ -28,10 +28,10 @@ grape_extra = dict(diffusion_alpha=0.5, diffusion_iters=10, # ═══════════════════════════════════════════════════════════════════════════ -# GRAFT + ResGCN combo (fixed version) +# KAFT + ResGCN combo (fixed version) # ═══════════════════════════════════════════════════════════════════════════ -class GRAFTResGCN(GraphGrAPETrainer): - """GRAFT backward + ResGCN forward (skip connections).""" +class GRAFTResGCN(KAFTTrainer): + """KAFT backward + ResGCN forward (skip connections).""" def forward(self): X = self.data['X'] @@ -57,10 +57,10 @@ class GRAFTResGCN(GraphGrAPETrainer): # ═══════════════════════════════════════════════════════════════════════════ -# GRAFT + DropEdge combo +# KAFT + DropEdge combo # ═══════════════════════════════════════════════════════════════════════════ -class GRAFTDropEdge(GraphGrAPETrainer): - """GRAFT backward + DropEdge forward (random edge dropping).""" +class GRAFTDropEdge(KAFTTrainer): + """KAFT backward + DropEdge forward (random edge dropping).""" def __init__(self, *args, drop_rate=0.5, **kwargs): super().__init__(*args, **kwargs) @@ -81,9 +81,9 @@ class GRAFTDropEdge(GraphGrAPETrainer): ).coalesce() def forward(self): - # DropEdge only in forward pass, GRAFT backward uses original A_hat + # DropEdge only in forward pass, KAFT backward uses original A_hat self.data['A_hat'] = self._drop_edges() - result = super().forward() # uses GraphGrAPETrainer.forward() + result = super().forward() # uses KAFTTrainer.forward() self.data['A_hat'] = self._A_hat_orig return result @@ -93,10 +93,10 @@ class GRAFTDropEdge(GraphGrAPETrainer): # ═══════════════════════════════════════════════════════════════════════════ -# GRAFT + PairNorm combo +# KAFT + PairNorm combo # ═══════════════════════════════════════════════════════════════════════════ -class GRAFTPairNorm(GraphGrAPETrainer): - """GRAFT backward + PairNorm forward (center & scale normalization).""" +class GRAFTPairNorm(KAFTTrainer): + """KAFT backward + PairNorm forward (center & scale normalization).""" def __init__(self, *args, pn_scale=1.0, **kwargs): super().__init__(*args, **kwargs) @@ -138,13 +138,13 @@ class GRAFTPairNorm(GraphGrAPETrainer): # ═══════════════════════════════════════════════════════════════════════════ -# GRAFT + JKNet combo +# KAFT + JKNet combo # ═══════════════════════════════════════════════════════════════════════════ -class GRAFTJKNet(GraphGrAPETrainer): - """GRAFT backward + JKNet forward (jumping knowledge max-pool). +class GRAFTJKNet(KAFTTrainer): + """KAFT backward + JKNet forward (jumping knowledge max-pool). Note: JKNet changes the output architecture. We max-pool hidden layers - and project to num_classes. GRAFT backward operates on hidden layers + and project to num_classes. KAFT backward operates on hidden layers as usual; the JK projection is treated as the output layer. """ @@ -187,7 +187,7 @@ class GRAFTJKNet(GraphGrAPETrainer): def _update_weights(self, inter, E0, deltas): """Override to handle JK projection separately.""" - # Update hidden layers using GRAFT feedback as usual + # Update hidden layers using KAFT feedback as usual X = self.data['X'] Hs = inter['Hs'] H0 = inter['H0'] @@ -265,7 +265,7 @@ def main(): per_seed_data = {} # Reuse existing per-seed data from other experiments - # BP, ResGCN, GRAFT from resgcn_20seeds + # BP, ResGCN, KAFT from resgcn_20seeds try: with open('results/resgcn_20seeds/per_seed_data.json') as f: resgcn_cache = json.load(f) @@ -292,11 +292,11 @@ def main(): 'DropEdge': (BPDropEdgeTrainer, {'drop_rate': 0.5}), 'PairNorm': (BPPairNormTrainer, {'pn_scale': 1.0}), 'JKNet': (JKNetTrainer, {}), - 'GRAFT': (GraphGrAPETrainer, grape_extra), - 'GRAFT+ResGCN': (GRAFTResGCN, grape_extra), - 'GRAFT+DropEdge': (GRAFTDropEdge, {**grape_extra, 'drop_rate': 0.5}), - 'GRAFT+PairNorm': (GRAFTPairNorm, {**grape_extra, 'pn_scale': 1.0}), - 'GRAFT+JKNet': (GRAFTJKNet, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), + 'KAFT+ResGCN': (GRAFTResGCN, grape_extra), + 'KAFT+DropEdge': (GRAFTDropEdge, {**grape_extra, 'drop_rate': 0.5}), + 'KAFT+PairNorm': (GRAFTPairNorm, {**grape_extra, 'pn_scale': 1.0}), + 'KAFT+JKNet': (GRAFTJKNet, grape_extra), } datasets_cfg = { @@ -330,7 +330,7 @@ def main(): cached = resgcn_cache[f"{ds_name}_BP"].get(sk) elif mname == 'ResGCN' and f"{ds_name}_ResGCN" in resgcn_cache: cached = resgcn_cache[f"{ds_name}_ResGCN"].get(sk) - elif mname == 'GRAFT' and f"{ds_name}_GRAFT" in resgcn_cache: + elif mname == 'KAFT' and f"{ds_name}_GRAFT" in resgcn_cache: cached = resgcn_cache[f"{ds_name}_GRAFT"].get(sk) elif mname == 'DropEdge': de_key = f"{ds_name}_gcn_L6" @@ -369,14 +369,14 @@ def main(): for ds in ['Cora', 'CiteSeer', 'DBLP']: print(f"\n--- {ds} GCN L=6 lr=0.01 ---") - print(f"{'Method':<18} {'Mean±Std':>12} {'vs GRAFT':>18} {'vs FwdTrick':>18}") + print(f"{'Method':<18} {'Mean±Std':>12} {'vs KAFT':>18} {'vs FwdTrick':>18}") print("-" * 70) - # Get GRAFT accs for comparison + # Get KAFT accs for comparison gr_accs = np.array([per_seed_data[f"{ds}_GRAFT"][str(s)] for s in SEEDS]) * 100 for mname in ['BP', 'ResGCN', 'DropEdge', 'PairNorm', 'JKNet', - 'GRAFT', 'GRAFT+ResGCN', 'GRAFT+DropEdge', 'GRAFT+PairNorm', 'GRAFT+JKNet']: + 'KAFT', 'KAFT+ResGCN', 'KAFT+DropEdge', 'KAFT+PairNorm', 'KAFT+JKNet']: key = f"{ds}_{mname}" if key not in per_seed_data or len(per_seed_data[key]) < 20: print(f" {mname:<16} MISSING ({len(per_seed_data.get(key, {}))} seeds)") @@ -387,8 +387,8 @@ def main(): results[key] = {'mean': float(m), 'std': float(s), 'accs': accs.tolist()} - # Paired t-test vs GRAFT - if mname != 'GRAFT': + # Paired t-test vs KAFT + if mname != 'KAFT': t_stat, p_val = scipy_stats.ttest_rel(accs, gr_accs) delta = m - gr_accs.mean() sig = '***' if p_val < 0.001 else ('**' if p_val < 0.01 else ('*' if p_val < 0.05 else 'ns')) @@ -402,8 +402,8 @@ def main(): # Paired t-test vs forward trick only fwd_map = { - 'GRAFT+ResGCN': 'ResGCN', 'GRAFT+DropEdge': 'DropEdge', - 'GRAFT+PairNorm': 'PairNorm', 'GRAFT+JKNet': 'JKNet' + 'KAFT+ResGCN': 'ResGCN', 'KAFT+DropEdge': 'DropEdge', + 'KAFT+PairNorm': 'PairNorm', 'KAFT+JKNet': 'JKNet' } if mname in fwd_map: fwd_key = f"{ds}_{fwd_map[mname]}" @@ -431,8 +431,8 @@ def main(): for ds in ['Cora', 'CiteSeer', 'DBLP']: print(f"\n{ds}:") gr_m = results.get(f"{ds}_GRAFT", {}).get('mean', 0) - for combo, fwd in [('GRAFT+ResGCN', 'ResGCN'), ('GRAFT+DropEdge', 'DropEdge'), - ('GRAFT+PairNorm', 'PairNorm'), ('GRAFT+JKNet', 'JKNet')]: + for combo, fwd in [('KAFT+ResGCN', 'ResGCN'), ('KAFT+DropEdge', 'DropEdge'), + ('KAFT+PairNorm', 'PairNorm'), ('KAFT+JKNet', 'JKNet')]: ck = f"{ds}_{combo}" fk = f"{ds}_{fwd}" if ck in results and fk in results: @@ -442,7 +442,7 @@ def main(): vs_fw = results[ck].get(f'vs_{fwd}', {}) better_than_both = c_m > gr_m and c_m > f_m marker = "✓ ADDITIVE" if better_than_both else "✗ not additive" - print(f" {combo}: {c_m:.1f} | GRAFT={gr_m:.1f} | {fwd}={f_m:.1f} → {marker}") + print(f" {combo}: {c_m:.1f} | KAFT={gr_m:.1f} | {fwd}={f_m:.1f} → {marker}") # Save with open(os.path.join(OUT_DIR, 'results.json'), 'w') as f: diff --git a/experiments/run_cora_perturb.py b/experiments/run_cora_perturb.py index 1dabc95..1124b97 100644 --- a/experiments/run_cora_perturb.py +++ b/experiments/run_cora_perturb.py @@ -13,7 +13,7 @@ import numpy as np import json import os from src.data import load_dataset, build_normalized_adj, build_row_normalized_adj -from src.trainers import BPTrainer, DFATrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, DFATrainer, KAFTTrainer device = 'cuda:0' SEEDS = [0, 1, 2, 3, 4] @@ -111,7 +111,7 @@ def main(): common = dict(data=data, hidden_dim=64, lr=0.01, weight_decay=5e-4, num_layers=L, residual_alpha=0.0, backbone='gcn') bp_accs.append(train_one(BPTrainer, common, {}, seed)) - gr_accs.append(train_one(GraphGrAPETrainer, common, grape_extra, seed)) + gr_accs.append(train_one(KAFTTrainer, common, grape_extra, seed)) bp, gr = np.mean(bp_accs)*100, np.mean(gr_accs)*100 delta = gr - bp diff --git a/experiments/run_cs_full.py b/experiments/run_cs_full.py index d66fc59..d170e13 100644 --- a/experiments/run_cs_full.py +++ b/experiments/run_cs_full.py @@ -1,6 +1,6 @@ #!/usr/bin/env python3 """H19 CitationFull-CiteSeer (4.2K, deg 2.5, 6-class) — same regime as Planetoid CiteSeer. -Quick BP + GRAFT depth sweep to confirm/extend the 'GRAFT wins on real sparse citation' story.""" +Quick BP + KAFT depth sweep to confirm/extend the 'KAFT wins on real sparse citation' story.""" import torch, sys, numpy as np, time import torch.nn as nn, torch.nn.functional as F @@ -9,7 +9,7 @@ from torch_geometric.nn import GCNConv from torch_geometric.utils import add_self_loops, degree sys.path.insert(0, '/home/yurenh2/graph-grape') -from src.trainers import GraphGrAPETrainer +from src.trainers import KAFTTrainer DATA_ROOT = '/home/yurenh2/graph-grape/data/CFull' device = torch.device('cuda:0') @@ -90,7 +90,7 @@ def bp_one(L, seed, d, train_mask, val_mask, test_mask, epochs=200, lr=5e-3, hid return bt -def graft_one(L, seed, d, A_hat, A_row, A_row_T, train_mask, val_mask, test_mask, +def kaft_one(L, seed, d, A_hat, A_row, A_row_T, train_mask, val_mask, test_mask, epochs=200, lr=5e-3, hidden=128): torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed) data = { @@ -99,7 +99,7 @@ def graft_one(L, seed, d, A_hat, A_row, A_row_T, train_mask, val_mask, test_mask 'num_features': d.x.shape[1], 'num_classes': int(d.y.max())+1, 'num_nodes': d.num_nodes, 'traces': {}, } - trainer = GraphGrAPETrainer( + trainer = KAFTTrainer( data=data, hidden_dim=hidden, lr=lr, weight_decay=0.0, lr_feedback=0.5, num_probes=64, topo_mode='fixed_A', max_topo_power=3, diffusion_alpha=0.5, diffusion_iters=10, @@ -133,14 +133,14 @@ def main(): t0 = time.time() bp = bp_one(L, s, d, tm, vm, tem) t1 = time.time() - gf = graft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem) + gf = kaft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem) t2 = time.time() bp_accs.append(bp); gf_accs.append(gf) - print(f' L={L} s={s}: BP={bp:.4f}({t1-t0:.0f}s) GRAFT={gf:.4f}({t2-t1:.0f}s)', flush=True) + print(f' L={L} s={s}: BP={bp:.4f}({t1-t0:.0f}s) KAFT={gf:.4f}({t2-t1:.0f}s)', flush=True) bp_m, bp_sd = np.mean(bp_accs), np.std(bp_accs) gf_m, gf_sd = np.mean(gf_accs), np.std(gf_accs) bp_res[L] = (bp_m, bp_sd); gf_res[L] = (gf_m, gf_sd) - print(f'>>> L={L}: BP {bp_m:.4f}±{bp_sd:.4f} GRAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True) + print(f'>>> L={L}: BP {bp_m:.4f}±{bp_sd:.4f} KAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True) if __name__ == '__main__': diff --git a/experiments/run_dblp_depth.py b/experiments/run_dblp_depth.py index d63b94a..f91440a 100644 --- a/experiments/run_dblp_depth.py +++ b/experiments/run_dblp_depth.py @@ -11,8 +11,7 @@ import os import time from torch_geometric.datasets import CitationFull from src.data import build_normalized_adj, build_row_normalized_adj, spmm, precompute_traces -from src.trainers import BPTrainer, DFATrainer, GraphGrAPETrainer -from benchmark_efficient import GraphGrAPEEfficient +from src.trainers import BPTrainer, DFATrainer, KAFTTrainer device = 'cuda:0' SEEDS = [0, 1, 2, 3, 4] @@ -108,7 +107,7 @@ def main(): row = {} for mname, cls, extra in [('BP', BPTrainer, {}), ('DFA', DFATrainer, dict(diffusion_alpha=0.5, diffusion_iters=10)), - ('GrAPE', GraphGrAPETrainer, grape_extra)]: + ('GrAPE', KAFTTrainer, grape_extra)]: accs = [train_one(cls, common, extra, s) for s in SEEDS] row[mname] = {'mean': float(np.mean(accs)), 'std': float(np.std(accs))} results[key] = row @@ -123,7 +122,6 @@ def main(): common = dict(data=dblp, hidden_dim=64, lr=0.01, weight_decay=5e-4, num_layers=L, residual_alpha=0.0, backbone=bb) bp_ms = time_method(BPTrainer, common, {}) - eff_ms = time_method(GraphGrAPEEfficient, common, dict(lr_feedback=0.5, num_probes=64, max_topo_power=3, diff_alpha=0.5, align_every=10)) key = f"DBLP_eff|{bb}|L={L}" @@ -146,7 +144,7 @@ def main(): key = f"{ds_name}|{bb}|L={L}|lr=0.01" row = {} for mname, cls, extra in [('BP', BPTrainer, {}), - ('GrAPE', GraphGrAPETrainer, grape_extra)]: + ('GrAPE', KAFTTrainer, grape_extra)]: accs = [train_one(cls, common, extra, s) for s in SEEDS] row[mname] = {'mean': float(np.mean(accs)), 'std': float(np.std(accs))} results[key] = row diff --git a/experiments/run_dblp_depth_scaling.py b/experiments/run_dblp_depth_scaling.py index 4c0bc11..86a5af1 100644 --- a/experiments/run_dblp_depth_scaling.py +++ b/experiments/run_dblp_depth_scaling.py @@ -1,16 +1,16 @@ #!/usr/bin/env python3 """E1: DBLP depth scaling — upgrade depth_stress 3-seed to 20 seeds on DBLP, extend to L={8,12,16,20,24,32}. Goal: confirm (or falsify) the preliminary -finding that GRAFT > ResGCN at L=16 (3-seed: 69.9 vs 63.7) and scales to L=32. +finding that KAFT > ResGCN at L=16 (3-seed: 69.9 vs 63.7) and scales to L=32. -BP vs ResGCN vs GRAFT vs GRAFT+ResGCN, GCN backbone, lr=0.01, 200 epochs.""" +BP vs ResGCN vs KAFT vs KAFT+ResGCN, GCN backbone, lr=0.01, 200 epochs.""" import torch import numpy as np import json import os from scipy import stats as scipy_stats -from src.trainers import BPTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, KAFTTrainer from run_deep_baselines import ResGCNTrainer from run_combo_20seeds import GRAFTResGCN from run_dblp_depth import load_dblp @@ -27,8 +27,8 @@ grape_extra = dict(diffusion_alpha=0.5, diffusion_iters=10, METHODS = { 'BP': (BPTrainer, {}), 'ResGCN': (ResGCNTrainer, {}), - 'GRAFT': (GraphGrAPETrainer, grape_extra), - 'GRAFT+ResGCN': (GRAFTResGCN, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), + 'KAFT+ResGCN': (GRAFTResGCN, grape_extra), } @@ -100,11 +100,11 @@ def main(): 'per_seed': vals.tolist()} print(f" {mname:<15} {vals.mean():5.1f} ± {vals.std():4.1f}") - # GRAFT vs ResGCN (paired) + # KAFT vs ResGCN (paired) g_accs = np.array([per_seed_data[f"DBLP_L{L}_GRAFT"][str(s)] for s in SEEDS]) * 100 r_accs = np.array([per_seed_data[f"DBLP_L{L}_ResGCN"][str(s)] for s in SEEDS]) * 100 t_gr, p_gr = scipy_stats.ttest_rel(g_accs, r_accs) - print(f" GRAFT vs ResGCN: Δ={g_accs.mean() - r_accs.mean():+.1f}, p={p_gr:.4f}") + print(f" KAFT vs ResGCN: Δ={g_accs.mean() - r_accs.mean():+.1f}, p={p_gr:.4f}") with open(os.path.join(OUT_DIR, 'results.json'), 'w') as f: json.dump(results, f, indent=2) diff --git a/experiments/run_depth_extras.py b/experiments/run_depth_extras.py index 66a7d45..e3b860b 100644 --- a/experiments/run_depth_extras.py +++ b/experiments/run_depth_extras.py @@ -1,6 +1,6 @@ #!/usr/bin/env python3 """H11: Fill depth sweep at L=14 and L=18 to densify Fig 4(a). -3 methods (BP / DFA-GNN / GRAFT) × 4 datasets × 2 depths × 20 seeds = 480 runs. +3 methods (BP / DFA-GNN / KAFT) × 4 datasets × 2 depths × 20 seeds = 480 runs. """ import torch @@ -8,7 +8,7 @@ import numpy as np import json import os from src.data import load_dataset -from src.trainers import BPTrainer, DFAGNNTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, DFAGNNTrainer, KAFTTrainer from run_dblp_depth import load_dblp device = 'cuda:0' @@ -24,7 +24,7 @@ dfagnn_extra = dict(diffusion_alpha=0.5, diffusion_iters=10, max_topo_power=3) METHODS = { 'BP': (BPTrainer, {}), 'DFA-GNN': (DFAGNNTrainer, dfagnn_extra), - 'GRAFT': (GraphGrAPETrainer, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), } diff --git a/experiments/run_dfagnn_depth.py b/experiments/run_dfagnn_depth.py index ed6e6c3..1a36485 100644 --- a/experiments/run_dfagnn_depth.py +++ b/experiments/run_dfagnn_depth.py @@ -2,9 +2,9 @@ """H7: DFA-GNN depth sweep for Figure 4(a)-style plot. Runs DFA-GNN at L ∈ {4, 8, 10, 12, 16, 20} × {Cora, CiteSeer, PubMed, DBLP} × 20 seeds. -L=6 data already exists from prior experiments; L=2/3 skipped (CiteSeer L=2 GRAFT soft spot). +L=6 data already exists from prior experiments; L=2/3 skipped (CiteSeer L=2 KAFT soft spot). -Combined with existing BP and GRAFT depth data, produces 3-method depth curves for Figure 4(a). +Combined with existing BP and KAFT depth data, produces 3-method depth curves for Figure 4(a). """ import torch diff --git a/experiments/run_grad_reach_20seeds.py b/experiments/run_grad_reach_20seeds.py index b5ad53b..b9ac8b9 100644 --- a/experiments/run_grad_reach_20seeds.py +++ b/experiments/run_grad_reach_20seeds.py @@ -12,7 +12,7 @@ import json import os from scipy import stats as scipy_stats from src.data import load_dataset, spmm -from src.trainers import BPTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, KAFTTrainer device = 'cuda:0' ALL_SEEDS = list(range(20)) @@ -31,7 +31,7 @@ def measure_one(data, L, backbone, seed): torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed) bp = BPTrainer(**common) torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed) - gr = GraphGrAPETrainer(**common, **grape_extra) + gr = KAFTTrainer(**common, **grape_extra) gr.align_mode = 'chain_norm' for _ in range(EPOCHS): @@ -46,10 +46,10 @@ def measure_one(data, L, backbone, seed): loss.backward() bp_norms = [bp.weights[l].grad.norm().item() for l in range(L)] - # GRAFT feedback norms + # KAFT feedback norms Z_gr, inter = gr.forward() E0, E_bar = gr._output_error(Z_gr) - graft_norms = [] + kaft_norms = [] for l in range(L - 1): power = min(L - l, gr.max_topo_power) topo_E = E_bar @@ -57,13 +57,13 @@ def measure_one(data, L, backbone, seed): topo_E = spmm(A, topo_E) fb = topo_E @ gr.Rs[l] relu_gate = (inter['Zs'][l].detach() > 0).float() - graft_norms.append((relu_gate * fb).norm().item()) + kaft_norms.append((relu_gate * fb).norm().item()) bp_acc = bp.evaluate('test_mask') gr_acc = gr.evaluate('test_mask') del bp, gr; torch.cuda.empty_cache() - return bp_norms, graft_norms, bp_acc, gr_acc + return bp_norms, kaft_norms, bp_acc, gr_acc def main(): @@ -101,10 +101,10 @@ def main(): bn, gn, ba, ga = measure_one(data, L, backbone, seed) per_seed_data[key][seed_key] = { - 'bp_norms': bn, 'graft_norms': gn, + 'bp_norms': bn, 'kaft_norms': gn, 'bp_acc': ba, 'gr_acc': ga } - print(f" seed {seed}: BP {ba*100:.1f}% GRAFT {ga*100:.1f}%", flush=True) + print(f" seed {seed}: BP {ba*100:.1f}% KAFT {ga*100:.1f}%", flush=True) # Save incrementally with open(old_per_seed_file, 'w') as f: @@ -121,7 +121,7 @@ def main(): t_stat, p_val = scipy_stats.ttest_rel(gr_accs, bp_accs) avg_bp_norms = np.mean([sd[str(s)]['bp_norms'] for s in ALL_SEEDS], axis=0) - avg_gr_norms = np.mean([sd[str(s)]['graft_norms'] for s in ALL_SEEDS], axis=0) + avg_gr_norms = np.mean([sd[str(s)]['kaft_norms'] for s in ALL_SEEDS], axis=0) results[key] = { 'bp_acc_mean': float(bp_accs.mean()), @@ -142,10 +142,10 @@ def main(): sig = '***' if p_val < 0.001 else ('**' if p_val < 0.01 else ('*' if p_val < 0.05 else 'ns')) print(f"\n {key}:") print(f" BP: {bp_accs.mean():.1f} ± {bp_accs.std():.1f}%") - print(f" GRAFT: {gr_accs.mean():.1f} ± {gr_accs.std():.1f}%") + print(f" KAFT: {gr_accs.mean():.1f} ± {gr_accs.std():.1f}%") print(f" Δ: {(gr_accs-bp_accs).mean():+.1f} ± {(gr_accs-bp_accs).std():.1f}% t={t_stat:.2f} p={p_val:.4f} {sig}") print(f" BP norm L0: {avg_bp_norms[0]:.6f}") - print(f" GRAFT norm L0: {avg_gr_norms[0]:.4f}") + print(f" KAFT norm L0: {avg_gr_norms[0]:.4f}") with open(os.path.join(OUT_DIR, 'results.json'), 'w') as f: json.dump(results, f, indent=2) diff --git a/experiments/run_hero_extras.py b/experiments/run_hero_extras.py index c6fed4d..ced76b8 100644 --- a/experiments/run_hero_extras.py +++ b/experiments/run_hero_extras.py @@ -1,8 +1,8 @@ #!/usr/bin/env python3 """E0a+E0c+E0e: hero-table coverage expansion. Adds 3 datasets (PubMed stack, -Coauthor-Physics, Coauthor-CS) × 5 methods (BP, DFA, DFA-GNN, GRAFT, -GRAFT+ResGCN) × 20 seeds, all GCN L=6. Goal: 6-row hero table of homophilous -citation/coauthor graphs where GRAFT or GRAFT+ResGCN is best per row.""" +Coauthor-Physics, Coauthor-CS) × 5 methods (BP, DFA, DFA-GNN, KAFT, +KAFT+ResGCN) × 20 seeds, all GCN L=6. Goal: 6-row hero table of homophilous +citation/coauthor graphs where KAFT or KAFT+ResGCN is best per row.""" import torch import numpy as np @@ -10,7 +10,7 @@ import json import os from scipy import stats as scipy_stats from src.data import load_dataset, spmm -from src.trainers import BPTrainer, DFATrainer, DFAGNNTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, DFATrainer, DFAGNNTrainer, KAFTTrainer from run_deep_baselines import ResGCNTrainer from run_combo_20seeds import GRAFTResGCN from run_large_graph_scout import load_and_check @@ -54,8 +54,8 @@ METHODS = { 'BP': (BPTrainer, {}), 'DFA': (DFATrainer, dfagnn_extra), 'DFA-GNN': (DFAGNNTrainer, dfagnn_extra), - 'GRAFT': (GraphGrAPETrainer, grape_extra), - 'GRAFT+ResGCN': (GRAFTResGCN, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), + 'KAFT+ResGCN': (GRAFTResGCN, grape_extra), } diff --git a/experiments/run_realworld_hero_L20.py b/experiments/run_realworld_hero_L20.py index 93a6c91..3b352f1 100644 --- a/experiments/run_realworld_hero_L20.py +++ b/experiments/run_realworld_hero_L20.py @@ -1,5 +1,5 @@ #!/usr/bin/env python3 -"""H33: 20-seed extension of L=20 hero on 4 real-world datasets × {BP, DFA, DFA-GNN, GRAFT}. +"""H33: 20-seed extension of L=20 hero on 4 real-world datasets × {BP, DFA, DFA-GNN, KAFT}. Paper setup (5%/class, hidden=64, lr=0.01, no scheduler, 200 epochs, GCN backbone, no dropout/BN/res). Tightens DBLP std (0.121 at 10-seed bimodal) for paper-grade stats. @@ -17,7 +17,7 @@ from torch_geometric.nn import GCNConv from torch_geometric.utils import add_self_loops, degree sys.path.insert(0, '/home/yurenh2/graph-grape') -from src.trainers import GraphGrAPETrainer +from src.trainers import KAFTTrainer device = torch.device('cuda:2') @@ -92,7 +92,7 @@ def bp_one(L, seed, d, tm, vm, tem, epochs=200, lr=0.01, hidden=64): return bt -def graft_one(L, seed, d, A_hat, A_row, A_row_T, tm, vm, tem, +def kaft_one(L, seed, d, A_hat, A_row, A_row_T, tm, vm, tem, epochs=200, lr=0.01, hidden=64): torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed) data = { @@ -101,7 +101,7 @@ def graft_one(L, seed, d, A_hat, A_row, A_row_T, tm, vm, tem, 'num_features': d.x.shape[1], 'num_classes': int(d.y.max())+1, 'num_nodes': d.num_nodes, 'traces': {}, } - trainer = GraphGrAPETrainer( + trainer = KAFTTrainer( data=data, hidden_dim=hidden, lr=lr, weight_decay=5e-4, lr_feedback=0.5, num_probes=64, topo_mode='fixed_A', max_topo_power=3, diffusion_alpha=0.5, diffusion_iters=10, @@ -149,21 +149,21 @@ def main(): t0 = time.time() bp = bp_one(L, s, d, tm, vm, tem) t1 = time.time() - gf = graft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem) + gf = kaft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem) t2 = time.time() bp_a.append(bp); gf_a.append(gf) - print(f' s={s} L={L}: BP={bp:.4f}({t1-t0:.0f}s) GRAFT={gf:.4f}({t2-t1:.0f}s)', flush=True) + print(f' s={s} L={L}: BP={bp:.4f}({t1-t0:.0f}s) KAFT={gf:.4f}({t2-t1:.0f}s)', flush=True) bp_m, bp_sd = float(np.mean(bp_a)), float(np.std(bp_a)) gf_m, gf_sd = float(np.mean(gf_a)), float(np.std(gf_a)) - out[name] = dict(seeds=seeds, BP=bp_a, GRAFT=gf_a, BP_mean=bp_m, BP_std=bp_sd, + out[name] = dict(seeds=seeds, BP=bp_a, KAFT=gf_a, BP_mean=bp_m, BP_std=bp_sd, GRAFT_mean=gf_m, GRAFT_std=gf_sd) - print(f' >>> {name} L=20 (seeds {s_lo}-{s_hi-1}): BP {bp_m:.4f}±{bp_sd:.4f} GRAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True) + print(f' >>> {name} L=20 (seeds {s_lo}-{s_hi-1}): BP {bp_m:.4f}±{bp_sd:.4f} KAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True) del d, A_hat, A_row, A_row_T torch.cuda.empty_cache() print('\n=== SUMMARY (this run) ===', flush=True) for k, v in out.items(): - print(f' {k}: BP {v["BP_mean"]:.4f}±{v["BP_std"]:.4f} GRAFT {v["GRAFT_mean"]:.4f}±{v["GRAFT_std"]:.4f}', flush=True) + print(f' {k}: BP {v["BP_mean"]:.4f}±{v["BP_std"]:.4f} KAFT {v["GRAFT_mean"]:.4f}±{v["GRAFT_std"]:.4f}', flush=True) if __name__ == '__main__': diff --git a/experiments/run_resgcn_20seeds.py b/experiments/run_resgcn_20seeds.py index 995a568..b56bb04 100644 --- a/experiments/run_resgcn_20seeds.py +++ b/experiments/run_resgcn_20seeds.py @@ -1,5 +1,5 @@ #!/usr/bin/env python3 -"""Task 7016bd94 Part 1: ResGCN vs GRAFT, 20 seeds, paired t-tests.""" +"""Task 7016bd94 Part 1: ResGCN vs KAFT, 20 seeds, paired t-tests.""" import torch import numpy as np @@ -7,7 +7,7 @@ import json import os from scipy import stats as scipy_stats from src.data import load_dataset -from src.trainers import BPTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, KAFTTrainer from run_deep_baselines import ResGCNTrainer from run_dblp_depth import load_dblp @@ -48,7 +48,7 @@ def main(): METHODS = { 'BP': (BPTrainer, {}), 'ResGCN': (ResGCNTrainer, {}), - 'GRAFT': (GraphGrAPETrainer, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), } datasets_cfg = { @@ -92,9 +92,9 @@ def main(): del data; torch.cuda.empty_cache() - # Paired t-tests: GRAFT vs ResGCN + # Paired t-tests: KAFT vs ResGCN print("\n" + "=" * 70) - print("Paired t-tests: GRAFT vs ResGCN (20 seeds)") + print("Paired t-tests: KAFT vs ResGCN (20 seeds)") print("-" * 70) for ds in ['Cora', 'CiteSeer', 'DBLP']: @@ -102,7 +102,7 @@ def main(): res_accs = np.array(results[f"{ds}_ResGCN"]['accs']) gr_accs = np.array(results[f"{ds}_GRAFT"]['accs']) - # GRAFT vs ResGCN + # KAFT vs ResGCN t_stat, p_val = scipy_stats.ttest_rel(gr_accs, res_accs) delta = gr_accs.mean() - res_accs.mean() sig = '***' if p_val < 0.001 else ('**' if p_val < 0.01 else ('*' if p_val < 0.05 else 'ns')) @@ -112,7 +112,7 @@ def main(): 'p_value': float(p_val), 'significant': bool(p_val < 0.05), } - # GRAFT vs BP + # KAFT vs BP t2, p2 = scipy_stats.ttest_rel(gr_accs, bp_accs) d2 = gr_accs.mean() - bp_accs.mean() sig2 = '***' if p2 < 0.001 else ('**' if p2 < 0.01 else ('*' if p2 < 0.05 else 'ns')) @@ -134,9 +134,9 @@ def main(): print(f"\n{ds}:") print(f" BP: {bp_accs.mean():.1f} ± {bp_accs.std():.1f}") print(f" ResGCN: {res_accs.mean():.1f} ± {res_accs.std():.1f}") - print(f" GRAFT: {gr_accs.mean():.1f} ± {gr_accs.std():.1f}") - print(f" GRAFT vs ResGCN: Δ{delta:+.1f}% p={p_val:.6f} {sig}") - print(f" GRAFT vs BP: Δ{d2:+.1f}% p={p2:.6f} {sig2}") + print(f" KAFT: {gr_accs.mean():.1f} ± {gr_accs.std():.1f}") + print(f" KAFT vs ResGCN: Δ{delta:+.1f}% p={p_val:.6f} {sig}") + print(f" KAFT vs BP: Δ{d2:+.1f}% p={p2:.6f} {sig2}") with open(os.path.join(OUT_DIR, 'results.json'), 'w') as f: json.dump(results, f, indent=2) diff --git a/experiments/run_shallow_depth.py b/experiments/run_shallow_depth.py index 68c9138..5f1cc26 100644 --- a/experiments/run_shallow_depth.py +++ b/experiments/run_shallow_depth.py @@ -1,8 +1,8 @@ #!/usr/bin/env python3 """E2: Shallow depth (L=2,3,4) on 4 datasets. Last exploratory avenue after -E1 (deep scaling) and E0-extras (more datasets) both failed to extend GRAFT's -regime. If GRAFT still wins at L=2/3 (standard GNN depth), we can counter -the reviewer attack 'L=5,6 nobody uses'. If GRAFT matches BP only at L=5,6, +E1 (deep scaling) and E0-extras (more datasets) both failed to extend KAFT's +regime. If KAFT still wins at L=2/3 (standard GNN depth), we can counter +the reviewer attack 'L=5,6 nobody uses'. If KAFT matches BP only at L=5,6, paper stays at current scope and we ship.""" import torch @@ -11,7 +11,7 @@ import json import os from scipy import stats as scipy_stats from src.data import load_dataset -from src.trainers import BPTrainer, GraphGrAPETrainer +from src.trainers import BPTrainer, KAFTTrainer from run_deep_baselines import ResGCNTrainer from run_combo_20seeds import GRAFTResGCN from run_dblp_depth import load_dblp @@ -27,8 +27,8 @@ grape_extra = dict(diffusion_alpha=0.5, diffusion_iters=10, METHODS = { 'BP': (BPTrainer, {}), - 'GRAFT': (GraphGrAPETrainer, grape_extra), - 'GRAFT+ResGCN': (GRAFTResGCN, grape_extra), + 'KAFT': (KAFTTrainer, grape_extra), + 'KAFT+ResGCN': (GRAFTResGCN, grape_extra), } @@ -108,10 +108,10 @@ def main(): t, p = scipy_stats.ttest_rel(gr_accs, bp_accs) delta = gr_accs.mean() - bp_accs.mean() print(f" {ds} L={L}: BP {bp_accs.mean():5.1f}±{bp_accs.std():4.1f} " - f"GRAFT {gr_accs.mean():5.1f}±{gr_accs.std():4.1f} " - f"GRAFT+ResGCN {stk_accs.mean():5.1f}±{stk_accs.std():4.1f} " - f"Δ(GRAFT-BP)={delta:+.1f}, p={p:.4f}") - for mname, accs in [('BP', bp_accs), ('GRAFT', gr_accs), ('GRAFT+ResGCN', stk_accs)]: + f"KAFT {gr_accs.mean():5.1f}±{gr_accs.std():4.1f} " + f"KAFT+ResGCN {stk_accs.mean():5.1f}±{stk_accs.std():4.1f} " + f"Δ(KAFT-BP)={delta:+.1f}, p={p:.4f}") + for mname, accs in [('BP', bp_accs), ('KAFT', gr_accs), ('KAFT+ResGCN', stk_accs)]: key = f"{ds}_L{L}_{mname}" results[key] = {'mean': float(accs.mean()), 'std': float(accs.std()), 'per_seed': accs.tolist()} diff --git a/experiments/run_wikics_paper_setup.py b/experiments/run_wikics_paper_setup.py index a2bf879..c9525e6 100644 --- a/experiments/run_wikics_paper_setup.py +++ b/experiments/run_wikics_paper_setup.py @@ -1,7 +1,7 @@ #!/usr/bin/env python3 """H15 WikiCS paper-setup depth sweep — Wikipedia academic articles. ~11.7K nodes, avg deg ~4.1, 10-class, undirected. Sparse + few-class -fits GRAFT's regime profile. Test BP vs GRAFT at L ∈ {3,5,10,14,20} × 5 seeds. +fits KAFT's regime profile. Test BP vs KAFT at L ∈ {3,5,10,14,20} × 5 seeds. """ import sys, time import numpy as np @@ -13,7 +13,7 @@ from torch_geometric.nn import GCNConv from torch_geometric.utils import add_self_loops, degree, to_undirected sys.path.insert(0, '/home/yurenh2/graph-grape') -from src.trainers import GraphGrAPETrainer +from src.trainers import KAFTTrainer device = torch.device('cuda:0') # CUDA_VISIBLE_DEVICES=2 maps cuda:0 → physical GPU 2 @@ -88,7 +88,7 @@ def bp_one(L, seed, d, tm, vm, tem, epochs=200, lr=0.01, hidden=64): return bt -def graft_one(L, seed, d, A_hat, A_row, A_row_T, tm, vm, tem, +def kaft_one(L, seed, d, A_hat, A_row, A_row_T, tm, vm, tem, epochs=200, lr=0.01, hidden=64): torch.manual_seed(seed); np.random.seed(seed); torch.cuda.manual_seed_all(seed) data = { @@ -97,7 +97,7 @@ def graft_one(L, seed, d, A_hat, A_row, A_row_T, tm, vm, tem, 'num_features': d.x.shape[1], 'num_classes': int(d.y.max())+1, 'num_nodes': d.num_nodes, 'traces': {}, } - trainer = GraphGrAPETrainer( + trainer = KAFTTrainer( data=data, hidden_dim=hidden, lr=lr, weight_decay=5e-4, lr_feedback=0.5, num_probes=64, topo_mode='fixed_A', max_topo_power=3, diffusion_alpha=0.5, diffusion_iters=10, @@ -133,13 +133,13 @@ def main(): t0 = time.time() bp = bp_one(L, s, d, tm, vm, tem) t1 = time.time() - gf = graft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem) + gf = kaft_one(L, s, d, A_hat, A_row, A_row_T, tm, vm, tem) t2 = time.time() bp_a.append(bp); gf_a.append(gf) - print(f' L={L} s={s}: BP={bp:.4f}({t1-t0:.0f}s) GRAFT={gf:.4f}({t2-t1:.0f}s)', flush=True) + print(f' L={L} s={s}: BP={bp:.4f}({t1-t0:.0f}s) KAFT={gf:.4f}({t2-t1:.0f}s)', flush=True) bp_m, bp_sd = np.mean(bp_a), np.std(bp_a) gf_m, gf_sd = np.mean(gf_a), np.std(gf_a) - print(f'>>> L={L}: BP {bp_m:.4f}±{bp_sd:.4f} GRAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True) + print(f'>>> L={L}: BP {bp_m:.4f}±{bp_sd:.4f} KAFT {gf_m:.4f}±{gf_sd:.4f} Δ={gf_m-bp_m:+.3f}', flush=True) if __name__ == '__main__': |