File size: 10,105 Bytes
bf71208 | 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 | """Development-only A/B pairing, cross-fitted calibration and oracle diagnostics."""
import argparse
import json
from pathlib import Path
import numpy as np
import torch
from scipy.optimize import minimize_scalar
from scipy.special import logsumexp
from mechcal.models import CausalTransformer, TransformerConfig
from mechcal.models.gru import CausalGRU, GRUConfig
from analyze_trainfit_baselines import fit as fit_baselines, predict as baseline_predict
from run_reference_followup import ROOT, WEIGHTS, load, params
from run_weight_curve_a import replay, estimate
OUT=Path('outputs/task1_diagnostics_20260916')
FULL=Path('outputs/weight_neural_full_20260916')
CO=Path('outputs/weight_neural_choice_only_20260916')
MODES=('donor','local','suffix_donor')
SEEDS=(11,22,33)
def loss(p,a):
return -np.log(np.take_along_axis(p,a[...,None],-1)[...,0])
def temper(p,t):
lp=np.log(np.clip(p,1e-30,1))/t
return np.exp(lp-logsumexp(lp,axis=-1,keepdims=True))
def calibrate(p,a,folds):
"""One temperature per training fold, selected only on intact probabilities."""
result=np.empty_like(p,dtype=float);ts=[]
for k in range(5):
train=folds!=k;val=~train
fit=minimize_scalar(lambda x:loss(temper(p[train],np.exp(x)),a[train]).mean(),
bounds=(np.log(.05),np.log(20)),method='bounded')
assert fit.success
t=float(np.exp(fit.x));ts.append(t);result[val]=temper(p[val],t)
return result,ts
def response_metrics(p,q,op,oq):
d=q-p[None];o=oq-op[None]
active=(np.abs(o).sum(-1)>1e-10)
dot=(d*o).sum(-1)
oe=(o*o).sum(-1);me=(d*d).sum(-1)*active
oe=oe.sum((0,2));me=me.sum((0,2));dot=dot.sum((0,2))
gain=np.divide(dot,oe,out=np.full_like(dot,np.nan),where=oe>0)
cosine=np.divide(dot,np.sqrt(oe*me),out=np.full_like(dot,np.nan),where=(oe*me)>0)
return dict(response_error=(.5*np.abs(d-o).sum(-1)).mean((0,2)),
response_tv=(.5*np.abs(d).sum(-1)).mean((0,2)),
projection_gain=gain,response_cosine=cosine,
oracle_active_fraction=active.mean((0,2)))
def summarize(v,sign):
result={}
for g,mask in [('all',np.ones(len(sign),bool)),('positive',sign>0),('negative',sign<0)]:
result[g]={}
for k,x in v.items():
valid=mask&np.isfinite(x)
result[g][k]=dict(estimate(x,valid),n=int(valid.sum())) if valid.any() else dict(mean=None,ci95=None,n=0)
return result
@torch.no_grad()
def export_choice():
torch.set_num_threads(4)
for w in WEIGHTS:
name=f'reward_w{round(w*100):03d}';d=load(w)
tok=torch.tensor(np.concatenate([d['tokens'][:,:1],d['tokens'][:,1::2]],1).astype(np.int64))
assert np.array_equal(tok[:,1:].numpy()-1,d['action'])
for family in ('gru','transformer'):
for seed in SEEDS:
target=OUT/f'{family}_{name}_seed{seed}_choice.npz'
if target.exists():continue
run=CO/f'{family}_{name}_choice_only_seed{seed}'
assert (run/'metrics.json').exists()
ck=torch.load(run/'best.pt',map_location='cpu',weights_only=True)
model=(CausalGRU(GRUConfig(**ck['model_config'])) if family=='gru' else CausalTransformer(TransformerConfig(**ck['model_config']))).cuda().eval()
model.load_state_dict(ck['model_state'])
p=np.concatenate([model(b[:,:-1].cuda()).float()[...,1:5].softmax(-1).cpu().numpy() for b in tok.split(32)])
assert p.shape==(250,200,4)
# Predictions through choice 100 cannot depend on later tokens.
assert np.allclose(p[:2,:100],model(tok[:2,:100].cuda()).float()[...,1:5].softmax(-1).cpu().numpy(),atol=2e-5)
np.savez_compressed(target,participants=d['base_participant_id'],probabilities=p)
del model;torch.cuda.empty_cache()
print('CHOICE_EXPORTED',w,flush=True)
def analyze():
scale=json.loads((ROOT/'protocol.json').read_text())['q_scale']
report={};baseline_specs={};temperatures={}
for w in WEIGHTS:
name=f'reward_w{round(w*100):03d}';d=load(w);a=d['action'][:,150:]
train=dict(np.load(ROOT/name/'train_000.npz'))
assert not set(train['base_participant_id'])&set(d['base_participant_id'])
fitted=fit_baselines(train['action']);baseline_specs[str(w)]=fitted
bp=baseline_predict(d['action'],fitted)
changed=d['action'].copy();changed[:,175:]=(changed[:,175:]+1)%4
for k,x in bp.items():assert np.array_equal(x[:,:175],baseline_predict(changed,fitted)[k][:,:175])
vectors={}
for k,x in bp.items():
p=x[:,149:];vectors['baseline_'+k+'_nll']=loss(p,a).mean(1)
vectors['baseline_'+k+'_accuracy']=(p.argmax(-1)==a).mean(1)
maps=np.load(ROOT/name/'validation_assay.npz')['donor_mappings']
op=replay(d,d['reward'],scale)[:,150:];ob=loss(op,a)
vectors['oracle_nll']=ob.mean(1);oracle_changed={}
for mode in MODES:
qs=[]
for m in maps:
r=d['reward'][m].copy()
if mode=='suffix_donor':r[:,:150]=d['reward'][:,:150]
qs.append(replay(d,r,scale,mode=='local')[:,150:])
oq=np.stack(qs);oracle_changed[mode]=oq
active=(np.abs(oq-op[None]).sum(-1)>1e-10)
kl=(op[None]*np.log(op[None]/oq)).sum(-1)
empirical=loss(oq,a[None])-ob[None]
vectors['oracle_'+mode+'_winner_set_change']=active.mean((0,2))
vectors['oracle_'+mode+'_expected_kl']=kl.mean((0,2))
vectors['oracle_'+mode+'_delta_nll']=empirical.mean((0,2))
vectors['oracle_'+mode+'_tv']=(.5*np.abs(oq-op[None]).sum(-1)).mean((0,2))
# For unique-winner pairs, KL = .9 log(37) iff winner changes.
unique=((op==op.max(-1,keepdims=True)).sum(-1)==1)[None]&((oq==oq.max(-1,keepdims=True)).sum(-1)==1)
assert np.allclose(kl[unique],(.9*np.log(37)*active)[unique],atol=1e-10)
folds=np.empty(250,int);folds[np.random.default_rng(20260916).permutation(250)]=np.arange(250)%5
for family in ('gru','transformer'):
runs=[]
for seed in SEEDS:
z=dict(np.load(FULL/f'{family}_{name}_full_seed{seed}'/'assay_full.npz'))
co=dict(np.load(OUT/f'{family}_{name}_seed{seed}_choice.npz'))
assert np.array_equal(z['participants'],co['participants']) and np.array_equal(z['donor_mappings'],maps)
p=z['intact_probabilities'][:,150:].astype(float);p/=p.sum(-1,keepdims=True)
cp=co['probabilities'][:,150:].astype(float);cp/=cp.sum(-1,keepdims=True)
calp,temps=calibrate(p,a,folds);temperatures[f'{w}_{family}_{seed}']=temps
bl=loss(p,a);cb=loss(calp,a)
v=dict(full_nll=bl.mean(1),choice_nll=loss(cp,a).mean(1),
full_accuracy=(p.argmax(-1)==a).mean(1),choice_accuracy=(cp.argmax(-1)==a).mean(1),
reward_predictive_gain=(loss(cp,a)-bl).mean(1),calibrated_full_nll=cb.mean(1))
for mode in MODES:
q=z[mode+'_probabilities'].astype(float);q/=q.sum(-1,keepdims=True)
cq=np.empty_like(q)
for k,t in enumerate(temps):cq[:,folds==k]=temper(q[:,folds==k],t)
for label,pp,qq,ll in [('raw',p,q,bl),('calibrated',calp,cq,cb)]:
v[label+'_'+mode+'_delta_nll']=(loss(qq,a[None])-ll[None]).mean((0,2))
for k,x in response_metrics(pp,qq,op,oracle_changed[mode]).items():v[label+'_'+mode+'_'+k]=x
v[mode+'_ablation_minus_retraining_gap']=v['raw_'+mode+'_delta_nll']-v['reward_predictive_gain']
v['choice_minus_train_lag_nll']=v['choice_nll']-vectors['baseline_train_selected_lag_nll']
runs.append(v)
for k in runs[0]:
arr=np.stack([r[k] for r in runs]);vectors[family+'_'+k]=arr.mean(0)
np.savez_compressed(OUT/f'{name}_diagnostics.npz',participants=d['base_participant_id'],sign=d['kernel_sign'],folds=folds,**vectors)
report[str(w)]=summarize(vectors,d['kernel_sign'])
print('DIAGNOSTICS',w, 'lag',fitted['selected']['lag'],flush=True)
(OUT/'summary.json').write_text(json.dumps(report,indent=2))
(OUT/'baseline_fits.json').write_text(json.dumps(baseline_specs,indent=2))
(OUT/'temperatures.json').write_text(json.dumps(temperatures,indent=2))
def common():
scale=json.loads((ROOT/'protocol.json').read_text())['q_scale'];report={}
for w in WEIGHTS:
collected={}
for source in WEIGHTS:
d=load(source);maps=np.load(ROOT/f'reward_w{round(source*100):03d}'/'validation_assay.npz')['donor_mappings']
changed=params(d,d['alpha'],np.full(250,w),d['kernel_sign'])
p=replay(changed,d['reward'],scale)[:,150:]
for mode in MODES:
kl=[];events=[]
for m in maps:
r=d['reward'][m].copy()
if mode=='suffix_donor':r[:,:150]=d['reward'][:,:150]
q=replay(changed,r,scale,mode=='local')[:,150:]
kl.append((p*np.log(p/q)).sum(-1).mean(1))
events.append((np.abs(q-p).sum(-1)>1e-10).mean(1))
collected.setdefault(mode+'_expected_kl',[]).append(np.mean(kl,0))
collected.setdefault(mode+'_winner_set_change',[]).append(np.mean(events,0))
vectors={k:np.mean(v,0) for k,v in collected.items()}
report[str(w)]=summarize(vectors,d['kernel_sign'])
np.savez_compressed(OUT/f'common_w{round(w*100):03d}.npz',**vectors)
(OUT/'common_history.json').write_text(json.dumps(report,indent=2))
print('COMMON_SUFFIX',w,flush=True)
if __name__=='__main__':
ap=argparse.ArgumentParser();ap.add_argument('phase',choices=['export','analyze','common']);args=ap.parse_args()
OUT.mkdir(exist_ok=True)
{'export':export_choice,'analyze':analyze,'common':common}[args.phase]()
|