CharlesZhang-USTC's picture
Add Control v2 incremental experiment results (evidence, visuals, 42 fixed samples)
a64600b verified
Raw History Blame Contribute Delete
5.74 kB
"""Exploratory fixed-state coupling probe for masked generators.
The score is a nonlinear KL proxy, NOT Gaussian total_I or a conditional
covariance theorem. No production sample/alpha curve is used to fit the score.
Reference generation (when requested) is counted separately from probe NFE.
"""
import argparse
import json
from pathlib import Path
import numpy as np
import torch
from .design import ROOT,read_tasks
from .results import atomic_json
@torch.no_grad()
def run(model,weights,tasks_file,output,groups=16,limit=8):
from .adapters import LOADERS,target_mask
from huggingface_hub import snapshot_download
cfg=json.loads((ROOT/'suite/models.json').read_text())[model]
if model!='momask':
lock=json.loads((Path(weights)/'weights.lock.json').read_text())
for k,v in list(cfg.items()):
if k.endswith('repo'):cfg[k]=snapshot_download(v,revision=lock[v],local_files_only=True)
else:
original=torch.load
def load(*a,**kw):kw.setdefault('weights_only',False);return original(*a,**kw)
torch.load=load
adapter=LOADERS[model](cfg,weights);out=Path(output);out.mkdir(parents=True,exist_ok=True);records=[]
for task in [t for t in read_tasks(tasks_file) if model in t.get('models',[]) and t.get('split','test')=='test'][:limit]:
_,ref=adapter.reference(task);ref=ref.reshape(1,-1)
if model=='showo':
mask_id=adapter.model.config.mask_token_id;side=int(np.sqrt(ref.numel()));shape=(side,side)
predict=lambda ids:adapter.predict(task['prompt'],ids)
elif model=='momask':
mask_id=adapter.model.mask_id;shape=(ref.shape[-1],);vector=adapter.model.encode_text([task['prompt']]);padding=torch.zeros_like(ref,dtype=torch.bool)
predict=lambda ids:adapter.model.forward_with_cond_scale(ids,cond_vector=vector,padding_mask=padding,cond_scale=cfg['guidance']).permute(0,2,1)
else:
shape=(int(np.sqrt(ref.numel())),)*2;mask_id=adapter.pipe.scheduler.config.mask_token_id;holder={}
class Captured(Exception):pass
def capture(module,args,kw):holder.update(kw);raise Captured()
handle=adapter.pipe.transformer.register_forward_pre_hook(capture,with_kwargs=True)
try:adapter.pipe(prompt=task['prompt'],height=cfg['resolution'],width=cfg['resolution'],guidance_scale=cfg['guidance'],num_inference_steps=cfg['steps'])
except Captured:pass
finally:handle.remove()
def predict(ids):
kw=dict(holder);x=ids.reshape(1,*shape);kw['hidden_states']=torch.cat([x,x]);kw['timestep']=torch.tensor([cfg['steps']//2],device=x.device)
u,c=adapter.pipe.transformer(**kw).chunk(2);logits=u+cfg['guidance']*(c-u)
return logits.reshape(1,logits.shape[1],-1).permute(0,2,1)
support=target_mask(shape,task).reshape(-1);n=ref.numel();count=min(groups,n)
chunks=[torch.tensor(a,device=ref.device) for a in np.array_split(np.arange(n),count)]
masked=torch.full_like(ref,mask_id);base=predict(masked).float().log_softmax(-1);confidence=base.exp().max(-1).values[0]
coupling=np.zeros((count,count),dtype='float64')
for i,index in enumerate(chunks):
visible=masked.clone();visible[:,index]=ref[:,index];changed=predict(visible).float().log_softmax(-1)
kl=(base.exp()*(base-changed)).sum(-1).clamp_min(0)[0]
for j,targets in enumerate(chunks):coupling[i,j]=float(kl[targets].mean())
target_fraction=np.array([float(support[c].mean()) for c in chunks]);conf=np.array([float(confidence[c].mean()) for c in chunks])
random=np.random.default_rng(5678).random(count)
priorities=dict(confidence=-conf,random=random,target_first=-target_fraction+random*.001,target_last=target_fraction+random*.001)
scores={}
for order,priority in priorities.items():
permutation=np.argsort(priority,kind='stable');rank=np.argsort(permutation)
# Source group committed before target group; weight by target support.
scores[order]=float(sum(coupling[i,j]*target_fraction[j]*(rank[i]<rank[j]) for i in range(count) for j in range(count)))
folder=out/task['id'];folder.mkdir(exist_ok=True);np.save(folder/'coupling.npy',coupling)
import matplotlib;matplotlib.use('Agg')
import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(5,4));im=ax.imshow(coupling);fig.colorbar(im,ax=ax,label='Fixed-state KL influence');ax.set(xlabel='Affected group',ylabel='Revealed reference group');fig.tight_layout();fig.savefig(folder/'coupling.png',dpi=180);plt.close(fig)
records.append(dict(task=task['id'],model=model,score_name='fixed_state_KL_order_proxy',scores=scores,probe_model_evaluations=1+count,
reference_kind=task.get('reference_kind'),conditioning='all MASK, one revealed reference group; Meissonic fixed middle timestep',
caveat='confidence order uses initial confidence only; this is not exact total_I or a measured prior covariance'))
atomic_json(out/'probe.json',records)
if not records:raise ValueError('No matching test tasks')
if __name__=='__main__':
p=argparse.ArgumentParser();p.add_argument('--model',choices=('meissonic','momask','showo'),required=True);p.add_argument('--weights',required=True);p.add_argument('--tasks',required=True);p.add_argument('--output',required=True);p.add_argument('--groups',type=int,default=16);p.add_argument('--limit',type=int,default=8);a=p.parse_args()
if a.groups<2 or a.limit<1:raise ValueError('groups >=2 and limit >=1 required')
run(a.model,a.weights,a.tasks,a.output,a.groups,a.limit)