Release 4.0.0: repository is the replication package (protocol, records, audit, conventions, fixed toolkit)
b890fab verified | """Run real experiments to fill paper gaps. | |
| 1. Rank sensitivity: LowRankGNN r sweep at d=500 | |
| 2. DAGMA d=150 5-seed benchmark | |
| """ | |
| import time, json, numpy as np, sys | |
| sys.stdout.reconfigure(encoding='utf-8') | |
| # causalscale installed via pip install -e . | |
| import torch, causalscale as cs | |
| from sklearn.metrics import f1_score | |
| def gen_sem_data(d, n, seed=42): | |
| np.random.seed(seed) | |
| W = np.zeros((d,d)); p=2/(d-1) | |
| for i in range(d): | |
| for j in range(d): | |
| if i!=j and np.random.random()<p: | |
| W[j,i]=np.random.uniform(0.5,1.0)*np.random.choice([-1,1]) | |
| for j in range(d): W[j,:j]=0 | |
| X=np.linalg.inv(np.eye(d)-W.T)@np.random.randn(d,n)*0.5 | |
| return X.T.astype(np.float32), W | |
| print("=== Experiment 1: Rank Sensitivity (LowRankGNN, d=500) ===\n") | |
| d, n = 500, 1000 | |
| X, W_true = gen_sem_data(d, n) | |
| X = (X - X.mean(0))/(X.std(0)+1e-8) | |
| results_rank = [] | |
| for r in [8, 16, 32, 64, 128]: | |
| print(f" r={r:3d}... ", end='', flush=True) | |
| torch.cuda.empty_cache() | |
| t0 = time.time() | |
| try: | |
| res = cs.train_lowrank_gnn(X, rank=r, epochs=300, lr=0.01, device='cuda', verbose=False) | |
| t1 = time.time() | |
| mem = torch.cuda.max_memory_allocated()/1e9 | |
| entry = {'r':r, 'time_s':round(t1-t0,1), 'vram_gb':round(mem,2), | |
| 'gnn_edges':res['gnn_edges'], 'f1_corr':res['f1']} | |
| results_rank.append(entry) | |
| print(f"{res['gnn_edges']} edges, F1_corr={res['f1']:.3f}, {t1-t0:.1f}s, {mem:.2f}GB") | |
| except Exception as e: | |
| print(f"Error: {e}") | |
| print("\n=== Experiment 2: DAGMA d=150, 5 seeds ===\n") | |
| from dagma.linear import DagmaLinear | |
| d, n = 150, 300 | |
| f1s, times = [], [] | |
| for seed in range(5): | |
| print(f" seed={seed}... ", end='', flush=True) | |
| X, W_true = gen_sem_data(d, n, seed=100+seed) | |
| X = X.astype(np.float32) | |
| t0 = time.time() | |
| W_est = DagmaLinear(loss_type='l2').fit(X, lambda1=0.01) | |
| t1 = time.time() | |
| W_bin = (np.abs(W_est)>0.3).astype(int) | |
| W_tbin = (np.abs(W_true)>0).astype(int) | |
| f1 = f1_score(W_tbin.flatten(), W_bin.flatten()) | |
| f1s.append(f1); times.append(t1-t0) | |
| print(f"F1={f1:.4f}, {t1-t0:.1f}s") | |
| f1s=np.array(f1s); times=np.array(times) | |
| print(f" Mean: F1={f1s.mean():.4f}+-{f1s.std():.4f}, time={times.mean():.1f}+-{times.std():.1f}s") | |
| # Save | |
| json.dump({'rank_sensitivity': results_rank, | |
| 'dagma_d150': {'f1_mean':round(float(f1s.mean()),4), 'f1_std':round(float(f1s.std()),4), | |
| 'time_mean':round(float(times.mean()),1), 'time_std':round(float(times.std()),1)}}, | |
| open(r'results/ablation_results.json','w'), indent=2) | |
| print("\nSaved ablation_results.json") | |