randomfolk's picture
Add anonymous scratch checkpoints, synthetic data, and source
bf71208 verified
Raw History Blame Contribute Delete
4.84 kB
"""Generate a paired seven-weight pilot, with separate observable/audit files."""
import argparse
from dataclasses import asdict
import hashlib
import json
from pathlib import Path
import numpy as np
from mechcal.generators.spatial_mixture import SpatialConfig, simulate_subject, observable_record, coordinates
WEIGHTS = [0., .1, .3, .5, .7, .9, 1.]
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--output',type=Path,default=Path('outputs/spatial_pilot_v1'))
parser.add_argument('--train',type=int,default=200)
parser.add_argument('--val',type=int,default=50)
parser.add_argument('--test',type=int,default=50)
parser.add_argument('--seed',type=int,default=20260918)
args=parser.parse_args()
if args.output.exists(): raise FileExistsError('Use a new output directory; never overwrite a pilot')
args.output.mkdir(parents=True)
config=SpatialConfig()
protocol=dict(config=asdict(config),weights=WEIGHTS,counts={s:getattr(args,s) for s in ('train','val','test')},
master_seed=args.seed,task='5x5 stationary spatial bandit, 8 independent rounds of 20 choices',
mixture='(1-lapse)*(w*softmax(GP_mean/tau_R)+(1-w)*softmax(-Manhattan_distance/tau_L))+lapse/25',
parameters=dict(gp_length=[.8,1.6],reward_temperature=[.12,.35],local_temperature=[.5,1.1]),
parameter_distribution='Independent uniform per participant; paired across weights',
rewards='Unbounded Gaussian landscape: mean 50, amplitude SD 15; Gaussian observation SD 2; no clipping/rounding',
information='Only cue and chosen rewards observable; no UCB, latent map and parameters audit-only',
split='Distinct participants and environment seeds across splits; test generated but not summarized',
pairing='Same environments, initial cues, potential-outcome noise, and choice uniforms across weights',
reference='Wu et al. 2018, doi:10.1038/s41562-018-0467-4; adapted task, not a replication',
status='Pilot generation only; not a locked confirmatory dataset or neural training run')
(args.output/'protocol.json').write_text(json.dumps(protocol,indent=2))
summaries={}
coords=coordinates(config.grid_size)
for split_idx,split in enumerate(('train','val','test')):
n=getattr(args,split)
rng=np.random.default_rng(np.random.SeedSequence([args.seed,split_idx,999]))
parameters=[dict(gp_length=float(rng.uniform(.8,1.6)),reward_temperature=float(rng.uniform(.12,.35)),
local_temperature=float(rng.uniform(.5,1.1))) for _ in range(n)]
seeds=[int(np.random.SeedSequence([args.seed,split_idx,i,888]).generate_state(1)[0]) for i in range(n)]
assert len(set(seeds))==n
for w in WEIGHTS:
folder=args.output/f'reward_w{round(w*100):03d}'
folder.mkdir(exist_ok=True)
episodes=[]
with (folder/f'{split}_observable.jsonl').open('w',encoding='utf-8') as f:
for i in range(n):
e=simulate_subject(config,parameters[i],w,seeds[i]);episodes.append(e)
f.write(json.dumps(observable_record(config,f'{split}_{i:04d}',e))+'\n')
arrays={k:np.stack([e[k] for e in episodes]) for k in episodes[0]}
arrays.update(participant=np.array([f'{split}_{i:04d}' for i in range(n)]),
subject_seed=np.array(seeds,dtype=np.uint32),
**{k:np.array([p[k] for p in parameters]) for k in parameters[0]})
np.savez_compressed(folder/f'{split}_audit.npz',**arrays)
if split!='test':
a=arrays['actions'];p=arrays['probabilities']
last=np.concatenate([arrays['cue_arm'][...,None],a[...,:-1]],axis=-1)
nll=-np.log(np.take_along_axis(p,a[...,None],axis=-1)[...,0])
summaries.setdefault(split,{})[str(w)]=dict(oracle_nll=float(nll.mean()),
mean_reward=float(arrays['rewards'].mean()),stay_rate=float((a==last).mean()),
mean_manhattan_step=float(np.abs(coords[a]-coords[last]).sum(-1).mean()),
unique_arms_per_round=float(np.mean([len(np.unique(row)) for row in a.reshape(-1,config.trials)])),
oracle_entropy=float((-p*np.log(p)).sum(-1).mean()))
print('GENERATED',split,w,n,flush=True)
(args.output/'summary.json').write_text(json.dumps(summaries,indent=2))
files=sorted(p for p in args.output.rglob('*') if p.is_file())
manifest={str(p.relative_to(args.output)):hashlib.sha256(p.read_bytes()).hexdigest() for p in files}
(args.output/'manifest.json').write_text(json.dumps(manifest,indent=2))
if __name__=='__main__': main()