File size: 17,533 Bytes
a1ceb4f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 | #!/usr/bin/env python3
"""
Large-scale experiment runner for NeurIPS-quality results.
Runs synthetic (full grid), model-family ablation, and real-data experiments.
Includes sanity checks and bootstrap CIs.
"""
import os, sys, json, time, yaml, argparse
import numpy as np
from datetime import datetime
from collections import defaultdict
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from src.data import (generate_gamma_poisson_data, generate_gaussian_gaussian_data,
generate_gaussian_gamma_data, sample_deletions)
from src.model import PoissonGammaVI, GaussianGaussianVI, GaussianGammaMAP
from src.graph_utils import build_adjacency, compute_graph_stats
from src.metrics import (compute_all_metrics, compute_deletion_influence_by_distance,
fit_exponential_decay, compute_local_error)
from src.unlearning import one_step_downdate_poisson_gamma
from src.utils import FitResult, generate_run_id, generate_config_id, save_jsonl, ensure_dir
# ===========================================================
# Sanity checks
# ===========================================================
def run_sanity_checks(model, edges, full_params, edge_to_del, exact_params,
local_params_by_R, ws_params, model_family, N, M, K):
"""Run all sanity checks, return dict of results."""
checks = {}
# 1. ELBO / objective improvement check
if hasattr(model, 'compute_elbo'):
try:
elbo_full = model.compute_elbo(edges, full_params)
checks['full_objective'] = elbo_full
checks['objective_finite'] = bool(np.isfinite(elbo_full))
except:
checks['objective_finite'] = False
# 2. Parameters positive (for Gamma models)
if model_family == 'poisson_gamma':
checks['params_positive'] = bool(
np.all(full_params['a'] > 0) and np.all(full_params['b'] > 0) and
np.all(full_params['c'] > 0) and np.all(full_params['d'] > 0))
checks['params_no_nan'] = bool(
not np.any(np.isnan(full_params['a'])) and not np.any(np.isnan(full_params['b'])))
# 3. Responsibilities sum to 1 (for PG)
if model_family == 'poisson_gamma':
from scipy.special import digamma
a, b, c, d = full_params['a'], full_params['b'], full_params['c'], full_params['d']
# Check a few random edges
resp_ok = True
for edge in edges[:min(20, len(edges))]:
i, j, x = edge
if x > 0:
log_r = digamma(a[i]) - np.log(b[i]) + digamma(c[j]) - np.log(d[j])
log_r -= log_r.max()
r = np.exp(log_r)
r_sum = r.sum()
r /= r_sum
if abs(r.sum() - 1.0) > 1e-6:
resp_ok = False
break
checks['responsibilities_sum_to_one'] = resp_ok
# 4. Exact deletion differs from full
from src.metrics import compute_all_param_vector
v_full = compute_all_param_vector(full_params, model_family)
v_exact = compute_all_param_vector(exact_params, model_family)
diff = np.linalg.norm(v_full - v_exact)
checks['exact_differs_from_full'] = bool(diff > 1e-10)
checks['exact_full_diff_norm'] = float(diff)
# 5. Local error decreases with R
errors_by_R = {}
for R, lp in sorted(local_params_by_R.items()):
err = compute_local_error(lp, exact_params, model_family)
errors_by_R[R] = err['relative_error']
checks['errors_by_R'] = errors_by_R
if len(errors_by_R) >= 2:
R_list = sorted(errors_by_R.keys())
checks['error_decreases_with_R'] = bool(errors_by_R[R_list[-1]] <= errors_by_R[R_list[0]])
# 6. Warm-start matches large-R local (approximate)
if ws_params is not None and max(local_params_by_R.keys()) >= 4:
ws_err = compute_local_error(ws_params, exact_params, model_family)
r4_err = errors_by_R.get(4, None)
if r4_err is not None:
checks['ws_error'] = ws_err['relative_error']
checks['ws_close_to_R4'] = bool(ws_err['relative_error'] <= r4_err * 5 + 0.01)
return checks
# ===========================================================
# Config builders
# ===========================================================
def build_full_synthetic_configs():
"""Full synthetic grid: 3 graph × 3 degree × 3 count × 2 prior × 3 K = 162 configs."""
configs = []
N, M = 300, 300
radii = [1, 2, 3, 4]
num_del = 50
for K in [5, 10, 20]:
for gt in ['bounded_degree', 'erdos_renyi', 'power_law']:
for deg_name, deg in [('low', 5), ('medium', 10), ('high', 20)]:
for cs_name, cs in [('low', 0.5), ('medium', 1.0), ('high', 3.0)]:
for ps_name, ps in [('strong', {'a0':1.0,'b0':1.0,'c0':1.0,'d0':1.0}),
('weak', {'a0':0.1,'b0':0.1,'c0':0.1,'d0':0.1})]:
configs.append({
'N': N, 'M': M, 'K': K,
'graph_type': gt, 'avg_degree': deg,
'count_scale': cs, 'count_scale_label': cs_name,
'prior_strength': ps_name, 'prior_config': ps,
'num_deletions': num_del, 'radii': radii, 'seed': 42,
'model_family': 'poisson_gamma', 'max_iter': 300, 'tol': 1e-5,
})
return configs
def build_model_family_configs():
"""Model-family ablation: balanced across 3 families."""
configs = []
N, M = 200, 200
radii = [1, 2, 3, 4]
num_del = 30
for K in [5, 10]:
for gt in ['bounded_degree', 'erdos_renyi', 'power_law']:
for deg in [5, 15]:
# Poisson-Gamma
for cs_name, cs in [('low', 0.5), ('medium', 1.0), ('high', 3.0)]:
for ps_name, ps in [('strong', {'a0':1.0,'b0':1.0,'c0':1.0,'d0':1.0}),
('weak', {'a0':0.3,'b0':0.3,'c0':0.3,'d0':0.3})]:
configs.append({
'N': N, 'M': M, 'K': K,
'graph_type': gt, 'avg_degree': deg,
'count_scale': cs, 'count_scale_label': cs_name,
'prior_strength': ps_name, 'prior_config': ps,
'num_deletions': num_del, 'radii': radii, 'seed': 42,
'model_family': 'poisson_gamma', 'max_iter': 300, 'tol': 1e-5,
})
# Gaussian-Gaussian
for sx_name, sx in [('high_noise', 2.0), ('medium_noise', 1.0), ('low_noise', 0.3)]:
for sp_name, sp in [('strong_prior', 0.5), ('weak_prior', 3.0)]:
configs.append({
'N': N, 'M': M, 'K': K,
'graph_type': gt, 'avg_degree': deg,
'sigma_x': sx, 'sigma_x_label': sx_name,
'sigma_U': sp, 'sigma_V': sp,
'prior_strength': sp_name,
'num_deletions': num_del, 'radii': radii, 'seed': 42,
'model_family': 'gaussian_gaussian', 'max_iter': 300, 'tol': 1e-5,
})
# Gaussian-Gamma MAP (with fixed optimizer settings)
for sx_name, sx in [('high_noise', 2.0), ('medium_noise', 1.0), ('low_noise', 0.3)]:
for gp_name, gp in [('strong', {'a0':2.0,'b0':2.0,'c0':2.0,'d0':2.0}),
('weak', {'a0':0.3,'b0':0.3,'c0':0.3,'d0':0.3})]:
configs.append({
'N': N, 'M': M, 'K': K,
'graph_type': gt, 'avg_degree': deg,
'sigma_x': sx, 'sigma_x_label': sx_name,
'prior_strength': gp_name, 'prior_config': gp,
'num_deletions': num_del, 'radii': radii, 'seed': 42,
'model_family': 'gaussian_gamma_map',
'lr': 0.05, 'max_iter': 2000, 'tol': 1e-6, 'grad_clip': 10.0,
})
return configs
# ===========================================================
# Run single config
# ===========================================================
def run_config(config):
"""Run one configuration end-to-end with sanity checks."""
model_family = config['model_family']
gt = config['graph_type']
N, M, K = config['N'], config['M'], config['K']
avg_degree = config['avg_degree']
radii = config.get('radii', [1, 2, 3, 4])
num_del = config.get('num_deletions', 50)
seed = config.get('seed', 42)
max_iter = config.get('max_iter', 300)
tol = config.get('tol', 1e-5)
config_id = generate_config_id(config)
run_id = generate_run_id()
prior_cfg = config.get('prior_config', {})
a0 = prior_cfg.get('a0', 0.3)
b0 = prior_cfg.get('b0', 1.0)
c0 = prior_cfg.get('c0', 0.3)
d0 = prior_cfg.get('d0', 1.0)
count_scale = config.get('count_scale', 1.0)
prior_strength = config.get('prior_strength', 'strong')
# Generate data
if model_family == 'poisson_gamma':
edges, U_true, V_true, ge = generate_gamma_poisson_data(
N, M, K, gt, avg_degree, count_scale, a0, b0, c0, d0, seed=seed)
elif model_family == 'gaussian_gaussian':
sigma_U = config.get('sigma_U', 1.0)
sigma_V = config.get('sigma_V', 1.0)
sigma_x = config.get('sigma_x', 1.0)
edges, U_true, V_true, ge = generate_gaussian_gaussian_data(
N, M, K, gt, avg_degree, sigma_U, sigma_V, sigma_x, seed=seed)
elif model_family == 'gaussian_gamma_map':
sigma_x = config.get('sigma_x', 1.0)
edges, U_true, V_true, ge = generate_gaussian_gamma_data(
N, M, K, gt, avg_degree, a0, b0, c0, d0, sigma_x, seed=seed)
if len(edges) < 10:
print(f" SKIP: only {len(edges)} edges")
return []
# Create model
if model_family == 'poisson_gamma':
model = PoissonGammaVI(N, M, K, a0, b0, c0, d0, max_iter=max_iter, tol=tol, seed=seed)
elif model_family == 'gaussian_gaussian':
model = GaussianGaussianVI(N, M, K, sigma_U=config.get('sigma_U', 1.0),
sigma_V=config.get('sigma_V', 1.0),
sigma_x=config.get('sigma_x', 1.0),
max_iter=max_iter, tol=tol, seed=seed)
elif model_family == 'gaussian_gamma_map':
model = GaussianGammaMAP(N, M, K, a0, b0, c0, d0,
sigma_x=config.get('sigma_x', 1.0),
lr=config.get('lr', 0.05),
max_iter=max_iter, tol=tol, seed=seed,
grad_clip=config.get('grad_clip', 10.0))
# Fit
t0 = time.time()
full_result = model.fit_full(edges)
t_full = time.time() - t0
full_params = full_result.params
# Deletions
u2i, i2u, ed = build_adjacency(edges, N, M)
dels = sample_deletions(edges, u2i, i2u, num_del, seed=seed)
records = []
sanity_results = []
for del_idx, (edge_to_del, del_type) in enumerate(dels):
i_del, j_del, x_del = edge_to_del
# Exact
exact_result = model.fit_without_edge(edges, edge_to_del, init_params=full_params)
# Local
local_results = {}
local_params = {}
for R in radii:
lr = model.fit_local(edges, edge_to_del, R, init_params=full_params)
local_results[R] = lr
local_params[R] = lr.params
# Warm-start
ws_result = model.fit_warm_start_global(edges, edge_to_del, init_params=full_params)
# One-step (PG only)
one_step_params = None
one_step_runtime = None
if model_family == 'poisson_gamma':
os_res = one_step_downdate_poisson_gamma(
edges, edge_to_del, full_params, N, M, K, a0, b0, c0, d0)
one_step_params = os_res.params
one_step_runtime = os_res.runtime_sec
# Metrics
model_kwargs = {}
if model_family == 'poisson_gamma':
model_kwargs = {'a0': a0, 'b0': b0, 'c0': c0, 'd0': d0}
else:
model_kwargs = {'sigma_x': config.get('sigma_x', 1.0)}
metrics = compute_all_metrics(
full_params, exact_result.params, local_params,
ws_result.params, one_step_params,
edge_to_del, edges, N, M, K,
model_family, radii=radii, model_kwargs=model_kwargs)
# Sanity checks (first 3 deletions only)
if del_idx < 3:
sanity = run_sanity_checks(
model, edges, full_params, edge_to_del,
exact_result.params, local_params, ws_result.params,
model_family, N, M, K)
sanity_results.append(sanity)
# Build record
record = {
'run_id': run_id, 'config_id': config_id,
'dataset_type': 'synthetic', 'dataset_name': f'synthetic_{model_family}',
'model_family': model_family,
'inference_type': 'vi' if model_family != 'gaussian_gamma_map' else 'map',
'likelihood': 'poisson' if model_family == 'poisson_gamma' else 'gaussian',
'prior': 'gamma' if 'gamma' in model_family else 'gaussian',
'graph_type': gt, 'seed': seed, 'N': N, 'M': M, 'K': K,
'avg_degree': avg_degree,
'count_scale': count_scale if model_family == 'poisson_gamma' else None,
'prior_strength': prior_strength,
'deletion_edge': [int(i_del), int(j_del), float(x_del)],
'deletion_type': del_type, 'deletion_index': del_idx,
'runtime_full': t_full, 'runtime_exact': exact_result.runtime_sec,
'runtime_warm_start': ws_result.runtime_sec,
'runtime_one_step': one_step_runtime,
'exact_converged': exact_result.converged,
'exact_iterations': exact_result.n_iterations,
'full_converged': full_result.converged,
}
for R in radii:
record[f'runtime_local_R{R}'] = local_results[R].runtime_sec
record[f'local_R{R}_converged'] = local_results[R].converged
record[f'local_R{R}_iterations'] = local_results[R].n_iterations
record.update(metrics)
if 'influence_by_distance' in record:
for d_str, val in record['influence_by_distance'].items():
record[f'influence_d{d_str}'] = val
record['regime'] = f"{gt}_{prior_strength}_deg{avg_degree}"
if model_family == 'poisson_gamma':
record['regime'] += f"_cs{count_scale}"
record['a0'] = a0; record['b0'] = b0; record['c0'] = c0; record['d0'] = d0
if model_family in ('gaussian_gaussian', 'gaussian_gamma_map'):
record['sigma_x'] = config.get('sigma_x', 1.0)
if model_family == 'gaussian_gaussian':
record['sigma_U'] = config.get('sigma_U', 1.0)
record['sigma_V'] = config.get('sigma_V', 1.0)
records.append(record)
return records, sanity_results
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type=str, default='full_synthetic',
choices=['full_synthetic', 'model_family', 'both'])
parser.add_argument('--max_configs', type=int, default=None)
args = parser.parse_args()
if args.mode in ('full_synthetic', 'both'):
configs = build_full_synthetic_configs()
label = 'full_synthetic'
elif args.mode == 'model_family':
configs = build_model_family_configs()
label = 'model_family_v2'
if args.mode == 'both':
configs += build_model_family_configs()
label = 'all'
if args.max_configs:
configs = configs[:args.max_configs]
print(f"Running {len(configs)} configs ({args.mode})")
output_dir = ensure_dir('results/raw')
ts = datetime.now().strftime("%Y%m%d_%H%M%S")
output_file = os.path.join(output_dir, f'{label}_{ts}.jsonl')
sanity_file = os.path.join(output_dir, f'sanity_{label}_{ts}.jsonl')
total_records = 0
all_sanity = []
for idx, config in enumerate(configs):
mf = config['model_family']
gt = config['graph_type']
K = config['K']
print(f"\n[{idx+1}/{len(configs)}] {mf} {gt} K={K} deg={config['avg_degree']} ps={config.get('prior_strength','')}")
try:
records, sanity = run_config(config)
total_records += len(records)
all_sanity.extend(sanity)
save_jsonl(records, output_file)
print(f" -> {len(records)} records (total: {total_records})")
except Exception as e:
print(f" ERROR: {e}")
import traceback; traceback.print_exc()
# Save sanity checks
save_jsonl(all_sanity, sanity_file)
print(f"\n{'='*60}")
print(f"Done. {total_records} records in {output_file}")
print(f"Sanity checks: {len(all_sanity)} in {sanity_file}")
if __name__ == '__main__':
main()
|