Thibaut Forest
Claude Fable 5
Phase 62: synthetic scaling study (1x/2x/4x, official-mechanism EOLs); count rule scales cleanly; fixed-ref vmin7 dial (default off); alpha-1.0 edge holds at every scale
34c30a6 | """Synthetic fleet generator for the SCALING evaluation. | |
| Fleet at scale k = k time-shifted clones of each real building. Each clone | |
| building gets ONE shift (preserves within-building death correlation, | |
| decorrelates clone sets). Ground-truth EOL is DERIVED from the shifted | |
| telemetry with the official mechanism (7d-median < 2.40, validated 81/81 | |
| by the team repo) - labels stay mechanically consistent with the data. | |
| Travel: distance between clones = distance between base buildings; | |
| same-base clone pairs get the median off-diagonal hop. | |
| Writes /tmp/c5/fleet_k{K}.pkl with: timeseries, locations, eol, travel fn. | |
| """ | |
| import sys, pickle, time | |
| from pathlib import Path | |
| import numpy as np | |
| import pandas as pd | |
| TEAM = Path('/tmp/teamrepo') | |
| OURS = Path('/Users/THIF/repos/BatterySwapAI2026-Example') | |
| sys.path.insert(0, str(TEAM)) | |
| sys.path.insert(0, str(OURS)) | |
| from batteryswap_public.utils import load_dataset | |
| from features import daily_panel, eol_from_panel | |
| K = int(sys.argv[1]) if len(sys.argv) > 1 else 2 | |
| SEED = 7 | |
| locations, ts, eol_times, scenarios = load_dataset(OURS / 'dataset/train') | |
| df = ts.reset_index() | |
| df['device_id'] = df['device_id'].astype(str) | |
| dev = pd.read_csv(OURS / 'dataset/train/devices.csv').drop_duplicates('device_id') | |
| span_lo = df['end_time'].min() | |
| span_hi = df['end_time'].max() | |
| rng = np.random.default_rng(SEED) | |
| frames, loc_rows = [], [] | |
| for c in range(K): | |
| for b, sub in dev.groupby('building_id'): | |
| # clone 0 of each building = the untouched original (scale anchor) | |
| shift = 0 if c == 0 else int(rng.integers(-120, 121)) | |
| sb = b if c == 0 else f'{b}_c{c}' | |
| for _, r in sub.iterrows(): | |
| sd = r['device_id'] if c == 0 else f"{r['device_id']}_c{c}" | |
| d = df[df['device_id'] == r['device_id']].copy() | |
| d['device_id'] = sd | |
| d['end_time'] = d['end_time'] + pd.Timedelta(days=shift) | |
| d = d[(d['end_time'] >= span_lo) & (d['end_time'] <= span_hi)] | |
| if len(d) < 50: | |
| continue | |
| frames.append(d) | |
| loc_rows.append(dict(battery=sd, building=sb, | |
| room=(r['room_id'] if c == 0 | |
| else f"{r['room_id']}_c{c}"), | |
| base_building=b, | |
| start_time=pd.Timestamp(r['start_time']) | |
| + pd.Timedelta(days=shift), | |
| end_time=pd.Timestamp(r['end_time']) | |
| + pd.Timedelta(days=shift))) | |
| sts = pd.concat(frames, ignore_index=True) | |
| locs = pd.DataFrame(loc_rows) | |
| for col in ('start_time', 'end_time'): | |
| locs[col] = pd.to_datetime(locs[col], utc=True) | |
| t0 = time.time() | |
| panel = daily_panel(sts.set_index(['device_id', 'end_time'])) | |
| eol = eol_from_panel(panel) | |
| print(f'k={K}: {locs.battery.nunique()} devices, ' | |
| f'{locs.building.nunique()} buildings, {len(eol)} EOLs ' | |
| f'(panel {time.time()-t0:.0f}s)') | |
| # sanity: clone-0 EOLs must match the official file exactly | |
| official = pd.read_csv(OURS / 'dataset/train/eol_times.csv') | |
| col = [c for c in official.columns if c != 'device_id'][0] | |
| truth = pd.to_datetime(official.set_index('device_id')[col]).dt.tz_localize(None) | |
| c0 = eol[~eol.index.str.contains('_c')] | |
| both = truth.dropna().index.intersection(c0.index) | |
| mism = int((pd.to_datetime(c0.reindex(both)).dt.normalize() | |
| != truth.reindex(both).dt.normalize()).sum()) | |
| print(f'clone-0 EOL check: {len(both)} deaths, {mism} mismatches') | |
| # travel-cost template: base 24x24 matrix from scenario 0 | |
| tc0 = scenarios[0]['travel_costs'] | |
| base_h = tc0.set_index(['from', 'to'])['hours'] | |
| med_hop = tc0.loc[tc0['from'] != tc0['to'], 'hours'].median() | |
| blds = sorted(locs['building'].unique()) | |
| base_of = locs.drop_duplicates('building').set_index('building')['base_building'] | |
| rows = [] | |
| for i in blds: | |
| for j in blds: | |
| if i == j: | |
| h = 0.0 | |
| else: | |
| bi, bj = base_of[i], base_of[j] | |
| h = med_hop if bi == bj else float(base_h.get((bi, bj), med_hop)) | |
| rows.append(dict(**{'from': i, 'to': j}, hours=h)) | |
| travel = pd.DataFrame(rows) | |
| with open(f'/tmp/c5/fleet_k{K}.pkl', 'wb') as f: | |
| pickle.dump(dict(ts=sts, locations=locs, eol=pd.to_datetime(eol), | |
| travel=travel), f) | |
| print(f'wrote /tmp/c5/fleet_k{K}.pkl') | |