"""Bounded pilot training primitives; corpus runs require a frozen ready manifest.""" import math from pathlib import Path import numpy as np import torch from s2_models.network import weighted_mse def part_weights(counts, parts_in_component, component_count, learning_parts): c=np.asarray(counts,dtype=np.float64) if c.ndim!=1 or not len(c) or np.any(c<=0) or not np.isfinite(c).all():raise ValueError('Invalid common query counts') if min(parts_in_component,component_count,learning_parts)<=0:raise ValueError('Empty training hierarchy') # Uniform shuffled part steps estimate the full component-balanced objective. return c/c.sum()/parts_in_component/component_count*learning_parts def train_part(model,optimizer,base,ids,gaps,edge,target,weights,*,scaler=None,amp=False): optimizer.zero_grad(set_to_none=True) with torch.autocast(device_type=base.device.type,dtype=torch.float16,enabled=amp): prediction=model(base,ids,gaps,edge if model.arm=='E' else None) loss=weighted_mse(prediction,target,weights) if not torch.isfinite(loss):raise FloatingPointError('Nonfinite training loss') if scaler is None: loss.backward() else: scaler.scale(loss).backward();scaler.unscale_(optimizer) if any(p.grad is not None and not torch.isfinite(p.grad).all() for p in model.parameters()): raise FloatingPointError('Nonfinite gradient; no silent retry or precision fallback') if scaler is None:optimizer.step() else:scaler.step(optimizer);scaler.update() if any(not torch.isfinite(p).all() for p in model.parameters()):raise FloatingPointError('Nonfinite updated parameter') return float(loss.detach()) def component_score(part_errors,component_ids): if not part_errors or set(part_errors)!=set(component_ids):raise ValueError('Incomplete validation roster') groups={} for part,value in part_errors.items(): if value is None or not math.isfinite(value):raise ValueError('Invalid validation part score') groups.setdefault(component_ids[part],[]).append(value) return float(np.mean([np.mean(values) for values in groups.values()])) def select_capacity(scores): # Exactly four run scores, no choice from an incomplete/failed run matrix. if set(scores)!={(arm,width) for arm in ['N','E'] for width in [128,256]}:raise ValueError('All four pilot runs required') if any(v is None or not math.isfinite(v) for v in scores.values()):raise ValueError('All four pilot validation scores must be finite') return min([128,256],key=lambda w:((scores['N',w]+scores['E',w])/2,w))