Archive CIDM-v2 reproducible b2_r11_backward_attribution.py
Browse files
cidm_v2/source/B2_R11_P0/b2_r11_backward_attribution.py
ADDED
|
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""CIDM-v2 B2-R11/P0: exact frozen 12-step backward attribution, no training.
|
| 2 |
+
|
| 3 |
+
Uses the archived B2-R10 FIX2 objective, model, control heads and normalization.
|
| 4 |
+
Loss uses TRAIN labels only. Attribution does not implement an inference policy.
|
| 5 |
+
Logs (1) each controller invocation's contribution to the full parameter
|
| 6 |
+
gradient, (2) each core operator's VJP contribution to latent adjoint,
|
| 7 |
+
including its normalized local-control basis path, (3) exact replay.
|
| 8 |
+
"""
|
| 9 |
+
from __future__ import annotations
|
| 10 |
+
import math, time, json, zipfile, csv, hashlib
|
| 11 |
+
from pathlib import Path
|
| 12 |
+
import torch
|
| 13 |
+
from benchmark import Config,Windows,UnifiedModel,make_synthetic_cache
|
| 14 |
+
from b2_r6_causal_spatial_controller import TASK_LIST,SEED,PARENT_SHA,make_cache_checked,orthogonal_matrix,sha
|
| 15 |
+
from b2_r7_frozen_controller_gain_audit import load_heads,write_csv,write_json
|
| 16 |
+
from b2_r1_flux_control import check_zip
|
| 17 |
+
import b2_r10_gradient_credit_audit as credit
|
| 18 |
+
|
| 19 |
+
PREV_FIX2_SHA='bfc54b95d54714eee6ddd1ef3f1d158418f26f8e00a5fc42faf0b8cc5eeba6bc'
|
| 20 |
+
LEADS=(1,3,9,12)
|
| 21 |
+
BUDGET=.003
|
| 22 |
+
|
| 23 |
+
def l2(v):
|
| 24 |
+
return float(torch.linalg.vector_norm(v.detach().double()))
|
| 25 |
+
|
| 26 |
+
def finite_tensor(t,name):
|
| 27 |
+
if not torch.isfinite(t).all():raise RuntimeError('NONFINITE_TENSOR '+name)
|
| 28 |
+
|
| 29 |
+
def concat_grad(params):
|
| 30 |
+
return torch.cat([torch.zeros_like(p).reshape(-1) if p.grad is None else p.grad.detach().reshape(-1) for p in params])
|
| 31 |
+
|
| 32 |
+
def full_forward(core,head,h,truth,scale,Q=None):
|
| 33 |
+
"""R9/R10 exact forward with retained local adjoints and controller outputs."""
|
| 34 |
+
mm=core.model
|
| 35 |
+
one_d=(h.ndim==4)
|
| 36 |
+
hh=h.unsqueeze(-2) if one_d else h
|
| 37 |
+
B,T,C,H,W=hh.shape
|
| 38 |
+
co=mm.coefficients()
|
| 39 |
+
z=mm.encode(hh.reshape(B,T*C,H,W)).detach().requires_grad_(True)
|
| 40 |
+
prev=hh[:,-1].detach().requires_grad_(True)
|
| 41 |
+
initial=prev
|
| 42 |
+
skew=mm.skew-mm.skew.T
|
| 43 |
+
records=[];pred=[]
|
| 44 |
+
for lead in range(1,13):
|
| 45 |
+
z_old=z
|
| 46 |
+
z_old.retain_grad()
|
| 47 |
+
f=co['flux']*mm.flux(z_old)
|
| 48 |
+
local=co['local']*torch.tanh(mm.local(z_old))
|
| 49 |
+
sk=.035*torch.einsum('ij,bjhw->bihw',skew,z_old)
|
| 50 |
+
base=z_old+f+local+sk
|
| 51 |
+
basis,amps=credit.stable_exact_effect_basis(f,local,sk,Q)
|
| 52 |
+
gates=head(base,prev,initial,amps,lead)
|
| 53 |
+
raw=(gates.unsqueeze(2)*basis).sum(dim=1)/math.sqrt(3)
|
| 54 |
+
rr=credit.stable_exact_rms(raw)
|
| 55 |
+
change=BUDGET*credit.stable_exact_rms(base).clamp_min(1e-4)*raw/rr.clamp_min(1e-9)
|
| 56 |
+
change=torch.where(rr>=1e-9,change,torch.zeros_like(change))
|
| 57 |
+
z=base+change
|
| 58 |
+
prev_new=prev+co['out']*torch.tanh(mm.decode(z))
|
| 59 |
+
for elem in (f,local,sk,base,gates,z,prev_new):elem.retain_grad()
|
| 60 |
+
records.append({'lead':lead,'z_old':z_old,'prev_old':prev,'initial':initial,
|
| 61 |
+
'f':f,'l':local,'s':sk,'base':base,'amps':amps,
|
| 62 |
+
'gates':gates,'z_new':z,'prev_new':prev_new})
|
| 63 |
+
prev=prev_new
|
| 64 |
+
pred.append(prev.squeeze(-2) if one_d else prev)
|
| 65 |
+
pred=torch.stack(pred,dim=1)
|
| 66 |
+
shape=(1,1,scale.numel())+(1,)*(pred.ndim-3)
|
| 67 |
+
w=scale.reshape(shape)
|
| 68 |
+
losses={j:((pred[:,j-1]-truth[:,j-1])*w.squeeze(1)).square().mean() for j in LEADS}
|
| 69 |
+
loss=.5*(losses[9]+losses[12])+.125*(losses[1]+losses[3])
|
| 70 |
+
return pred,loss,records
|
| 71 |
+
|
| 72 |
+
def analyze_case(task,core,cache,r6zip,device,head_types=('operator','rotated'),max_train_examples=2):
|
| 73 |
+
core.requires_grad_(False).eval()
|
| 74 |
+
ch=int(cache['splits']['train']['h'].shape[2]);d=int(core.model.skew.shape[0])
|
| 75 |
+
head_task= 'swe_2d' if task=='smoke_2d' else ('burgers_1d' if task=='smoke_1d' else task)
|
| 76 |
+
heads=load_heads(r6zip,head_task,ch,device)
|
| 77 |
+
Q=orthogonal_matrix(d,32207,device,torch.float32)
|
| 78 |
+
train=Windows(cache['splits']['train']['h'],cache['splits']['train']['y'],cache['mean'],cache['std'])
|
| 79 |
+
ids=list(range(min(max_train_examples,len(train))))
|
| 80 |
+
arr=[train[i] for i in ids]
|
| 81 |
+
h=torch.stack([t[0] for t in arr]).to(device)
|
| 82 |
+
y=torch.stack([t[1] for t in arr]).to(device)
|
| 83 |
+
scale=train.sd.to(device)
|
| 84 |
+
result=[];steps=[];layers=[]
|
| 85 |
+
for name in head_types:
|
| 86 |
+
head=heads[name].to(device).train()
|
| 87 |
+
for p in head.parameters():p.requires_grad_(True)
|
| 88 |
+
params=tuple(head.parameters())
|
| 89 |
+
head.zero_grad(set_to_none=True)
|
| 90 |
+
rot=Q if name=='rotated' else None
|
| 91 |
+
tic=time.perf_counter()
|
| 92 |
+
out,loss,records=full_forward(core,head,h,y,scale,rot)
|
| 93 |
+
finite_tensor(out,'forward '+task+' '+name)
|
| 94 |
+
with torch.no_grad():
|
| 95 |
+
parity=float((out.detach()-credit.original_frozen_rollout(core,head,h,rotation=rot)).abs().max())
|
| 96 |
+
if parity>2e-6:raise RuntimeError(f'REPLAY_MISMATCH {task} {name} {parity}')
|
| 97 |
+
finite_tensor(loss,'loss '+task+' '+name)
|
| 98 |
+
loss.backward()
|
| 99 |
+
full=concat_grad(params).double()
|
| 100 |
+
finite_tensor(full,'full_gradient '+task+' '+name)
|
| 101 |
+
fullnorm=l2(full)
|
| 102 |
+
pieces=[]
|
| 103 |
+
state_u=[]
|
| 104 |
+
for r in records:
|
| 105 |
+
lead=r['lead']
|
| 106 |
+
gg=r['gates'].grad
|
| 107 |
+
if gg is None:raise RuntimeError(f'MISSING_GATE_ADJOINT {task} {name} {lead}')
|
| 108 |
+
finite_tensor(gg,f'gate_adjoint_{lead}')
|
| 109 |
+
# Replay frozen input to isolate direct per-call contribution to head theta.
|
| 110 |
+
g_replay=head(r['base'].detach(),r['prev_old'].detach(),
|
| 111 |
+
r['initial'].detach(),r['amps'].detach(),lead)
|
| 112 |
+
if not torch.allclose(g_replay.detach(),r['gates'].detach(),atol=1e-6,rtol=1e-6):
|
| 113 |
+
raise RuntimeError(f'GATE_REPLAY_MISMATCH_{lead}')
|
| 114 |
+
grads=torch.autograd.grad(g_replay,params,grad_outputs=gg.detach(),allow_unused=True)
|
| 115 |
+
part=torch.cat([torch.zeros_like(p).flatten() if x is None else x.detach().flatten() for p,x in zip(params,grads)]).double()
|
| 116 |
+
finite_tensor(part,f'head_gradient_part_{lead}')
|
| 117 |
+
pieces.append(part)
|
| 118 |
+
state_u.append(l2(r['z_old'].grad))
|
| 119 |
+
steps.append({'task':task,'controller':name,'lead':lead,
|
| 120 |
+
'gate_adjoint_l2':l2(gg),'parameter_part_l2':l2(part),
|
| 121 |
+
'parameter_part_signed_projection':float(torch.dot(part,full)/max(fullnorm,1e-30)),
|
| 122 |
+
'parameter_part_cosine_full':float(torch.dot(part,full)/max(l2(part)*fullnorm,1e-30)),
|
| 123 |
+
'latent_state_adjoint_l2':l2(r['z_old'].grad),
|
| 124 |
+
'base_adjoint_l2':l2(r['base'].grad),
|
| 125 |
+
'flux_effect_adjoint_l2':l2(r['f'].grad),
|
| 126 |
+
'local_effect_adjoint_l2':l2(r['l'].grad),
|
| 127 |
+
'skew_effect_adjoint_l2':l2(r['s'].grad),
|
| 128 |
+
'forward_parity_max':parity})
|
| 129 |
+
recon=torch.stack(pieces).sum(dim=0)
|
| 130 |
+
reconstruction_error=l2(recon-full)/max(fullnorm,1e-30)
|
| 131 |
+
if reconstruction_error>1e-4:raise RuntimeError(f'HEAD_PARTITION_GRADIENT_FAILED {task} {name} {reconstruction_error}')
|
| 132 |
+
for r in records:
|
| 133 |
+
z=r['z_old'].detach().requires_grad_(True)
|
| 134 |
+
co=core.model.coefficients();mm=core.model;skew=mm.skew-mm.skew.T
|
| 135 |
+
f=co['flux']*mm.flux(z)
|
| 136 |
+
local=co['local']*torch.tanh(mm.local(z))
|
| 137 |
+
sk=.035*torch.einsum('ij,bjhw->bihw',skew,z)
|
| 138 |
+
gf,gl,gs=(r[k].grad.detach() for k in ('f','l','s'))
|
| 139 |
+
gbase=r['base'].grad.detach()
|
| 140 |
+
vjp_f,=torch.autograd.grad(f,z,gf,retain_graph=True)
|
| 141 |
+
vjp_l,=torch.autograd.grad(local,z,gl,retain_graph=True)
|
| 142 |
+
vjp_s,=torch.autograd.grad(sk,z,gs,retain_graph=True)
|
| 143 |
+
components=[gbase.double(),vjp_f.double(),vjp_l.double(),vjp_s.double()]
|
| 144 |
+
recovered=sum(components)
|
| 145 |
+
# Critical: z_old participates in the *previous* step's decoder branch,
|
| 146 |
+
# so z_old.grad of the original full trajectory includes additional
|
| 147 |
+
# graph shortcuts. It is NOT the local transition VJP reference.
|
| 148 |
+
# Use an independently evaluated local VJP with the correct
|
| 149 |
+
# f/l/sk direct-basis adjoints minus their already counted base path.
|
| 150 |
+
other=torch.autograd.grad(
|
| 151 |
+
(z+f+local+sk,f,local,sk),z,
|
| 152 |
+
grad_outputs=(gbase,gf-gbase,gl-gbase,gs-gbase),allow_unused=False)[0]
|
| 153 |
+
actual=other.detach().double()
|
| 154 |
+
relative=l2(recovered-actual)/max(sum(l2(t) for t in components),1e-30)
|
| 155 |
+
if relative>1e-3:raise RuntimeError(f'OPERATOR_VJP_PARTITION_FAILED {task} {name} lead={r["lead"]} relative={relative}')
|
| 156 |
+
layers.append({'task':task,'controller':name,'lead':r['lead'],
|
| 157 |
+
'direct_identity_vjp_l2':l2(gbase),
|
| 158 |
+
'flux_vjp_l2':l2(vjp_f),'local_vjp_l2':l2(vjp_l),
|
| 159 |
+
'skew_vjp_l2':l2(vjp_s),
|
| 160 |
+
'latent_vjp_sum_l2':l2(recovered),'latent_local_reference_vjp_l2':l2(actual),'original_graph_latent_adjoint_l2':l2(r['z_old'].grad),
|
| 161 |
+
'local_vjp_partition_error':relative})
|
| 162 |
+
values=[l2(v) for v in pieces]
|
| 163 |
+
key_idx=sorted(range(12),key=lambda i:values[i],reverse=True)
|
| 164 |
+
result.append({'task':task,'controller':name,'examples':len(ids),'TRAIN_indices':ids,
|
| 165 |
+
'loss':float(loss.detach()),'forward_parity_max':parity,
|
| 166 |
+
'full_gradient_l2_fp64':fullnorm,
|
| 167 |
+
'head_step_reconstruction_rel_error':reconstruction_error,
|
| 168 |
+
'top_step_by_norm':int(key_idx[0]+1),'top_step_gradient_l2':values[key_idx[0]],
|
| 169 |
+
'early_steps_1_3_sum_norm':float(sum(values[:3])),
|
| 170 |
+
'late_steps_9_12_sum_norm':float(sum(values[8:])),
|
| 171 |
+
'elapsed_s':time.perf_counter()-tic})
|
| 172 |
+
for p in head.parameters():p.requires_grad_(False)
|
| 173 |
+
return result,steps,layers
|
| 174 |
+
|
| 175 |
+
def run(parent,r6zip,r9zip,fix2zip,work,output,smoke=False,device='cpu'):
|
| 176 |
+
work=Path(work);output=Path(output);work.mkdir(parents=True,exist_ok=True);output.mkdir(parents=True,exist_ok=True)
|
| 177 |
+
if sha(fix2zip)!=PREV_FIX2_SHA:raise RuntimeError('B2R10_FIX2_EXACT_EVIDENCE_SHA_MISMATCH')
|
| 178 |
+
with zipfile.ZipFile(fix2zip) as z:
|
| 179 |
+
if z.testzip():raise RuntimeError('B2R10_FIX2_EVIDENCE_CRC_ERROR')
|
| 180 |
+
rep=json.loads(z.read('report.json'))
|
| 181 |
+
if rep['status']!='FROZEN_GRADIENT_CREDIT_AUDIT_COMPLETED_NO_TRAINING' or rep['rows']!=18:
|
| 182 |
+
raise RuntimeError('B2R10_FIX2_EVIDENCE_NOT_FULL')
|
| 183 |
+
gate=json.loads(z.read('gradient_qualification_gate.json'))
|
| 184 |
+
if gate['numerical_status']!='SEVERE_GRADIENT_AMPLIFICATION_NO_TRAINING':raise RuntimeError('FIX2_EVIDENCE_NOT_MATCHED')
|
| 185 |
+
credit.validate_r6(r6zip);credit.validate_r9(r9zip)
|
| 186 |
+
rows=[];perlead=[];ops=[];cases=[]
|
| 187 |
+
if smoke:
|
| 188 |
+
cfg=Config(profile='smoke')
|
| 189 |
+
a=make_synthetic_cache(cfg);b=make_synthetic_cache(cfg)
|
| 190 |
+
for split in ('train','val','test'):
|
| 191 |
+
a['splits'][split]['h']=a['splits'][split]['h'].mean(dim=-2)
|
| 192 |
+
a['splits'][split]['y']=a['splits'][split]['y'].mean(dim=-2)
|
| 193 |
+
for task,cache in [('smoke_1d',a),('smoke_2d',b)]:
|
| 194 |
+
core=UnifiedModel('V2_Ported_R5',1,8,12,4).to(device)
|
| 195 |
+
res,st,op=analyze_case(task,core,cache,r6zip,device,max_train_examples=1)
|
| 196 |
+
rows.extend(res);perlead.extend(st);ops.extend(op)
|
| 197 |
+
status='PROGRAM_ONLY_SMOKE_NO_SCIENCE'
|
| 198 |
+
else:
|
| 199 |
+
if parent is None or sha(parent)!=PARENT_SHA:raise RuntimeError('PARENT_SHA_MISMATCH')
|
| 200 |
+
with zipfile.ZipFile(parent) as z:
|
| 201 |
+
manifest,contract=check_zip(z);cfg=Config(**manifest['config'])
|
| 202 |
+
for task in TASK_LIST:
|
| 203 |
+
core,cache,replay=make_cache_checked(task,z,cfg,work,device)
|
| 204 |
+
try:
|
| 205 |
+
res,st,op=analyze_case(task,core,cache,r6zip,device)
|
| 206 |
+
except Exception as e:
|
| 207 |
+
write_json(output/'failure_case.json',{'task':task,'exception':repr(e),'already_completed':cases})
|
| 208 |
+
raise
|
| 209 |
+
rows.extend(res);perlead.extend(st);ops.extend(op)
|
| 210 |
+
cases.append({'task':task,'parent_replay':replay,'rows':len(res),'exact_model_frozen':True})
|
| 211 |
+
write_csv(output/'temporal_gradient_contributions.csv',perlead)
|
| 212 |
+
write_csv(output/'operator_local_vjp.csv',ops)
|
| 213 |
+
write_json(output/'progress.json',{'completed':[x['task'] for x in cases]})
|
| 214 |
+
status='FROZEN_BACKWARD_PATH_ATTRIBUTION_COMPLETED_NO_TRAINING'
|
| 215 |
+
if not smoke and (len(rows)!=6 or len(perlead)!=72 or len(ops)!=72):
|
| 216 |
+
raise RuntimeError(f'INCOMPLETE_BACKWARD_PATH_AUDIT_{len(rows)}_{len(perlead)}_{len(ops)}')
|
| 217 |
+
write_csv(output/'temporal_gradient_contributions.csv',perlead)
|
| 218 |
+
write_csv(output/'operator_local_vjp.csv',ops)
|
| 219 |
+
write_csv(output/'case_summary.csv',rows)
|
| 220 |
+
risk=[r for r in rows if r['full_gradient_l2_fp64']>1e8]
|
| 221 |
+
gate={'status':'DIAGNOSTIC_ONLY_NO_TRAINING','completed_groups':len(rows),
|
| 222 |
+
'gradient_large_cases':risk,'train_allowed':False,
|
| 223 |
+
'note':'Relative VJP norms and parameter contribution amplitudes are diagnostic, not causal branch ablations or Jacobian spectral norms.'}
|
| 224 |
+
write_json(output/'gradient_path_gate.json',gate)
|
| 225 |
+
report={'experiment':'CIDM_v2_B2_R11_P0_BackwardPathAttribution',
|
| 226 |
+
'status':status,'cases':cases,'number_of_controllers':len(rows),
|
| 227 |
+
'temporal_rows':len(perlead),'vjp_rows':len(ops),
|
| 228 |
+
'original_fix2_sha256':PREV_FIX2_SHA,
|
| 229 |
+
'training_updates':0,'no_outer_holdout':True,'train_allowed':False,
|
| 230 |
+
'attribution':'exact decomposition of the frozen head gradient and local VJPs; no weights changed'}
|
| 231 |
+
write_json(output/'report.json',report)
|
| 232 |
+
return report
|