sci-run / tasks /surface-code-decoding /surface-code-decoding-astra-xhigh-1h100-48h-20261020b /finalization /submission /conditional_edge_aux_encoder.py
| """Dense masked-event objective using the unchanged axial encoder equations.""" | |
| import jax | |
| import jax.numpy as jnp | |
| from conditional_edge_axial_base import DEFAULT,layer_norm,attention,dropout | |
| def features(dets,grid,basis,rounds,mask): | |
| valid=grid!=600 | |
| target=dets[:,grid].astype(jnp.float32)*valid[None] | |
| mask=mask&valid[None] | |
| events=jnp.where(mask,0.,target) | |
| measured=(jnp.cumsum(events.astype(jnp.int32),axis=1)&1).astype(jnp.float32) | |
| broadcast=lambda value:jnp.broadcast_to(value,events.shape) | |
| x=jnp.stack([events,measured,broadcast(valid[None]),broadcast(jnp.arange(26)[None,:,None]/25),broadcast(jnp.asarray(basis,jnp.float32)),broadcast(jnp.asarray(rounds,jnp.float32)/25),mask.astype(jnp.float32),jnp.zeros_like(events)],axis=-1) | |
| return x,valid,target,mask | |
| def encode(parameters,inputs,valid,*,training=False,key=None,config=None): | |
| c=DEFAULT if config is None else config;mask=valid[None,:,:,None] | |
| x=(inputs@parameters['input_w']+parameters['input_b']+parameters['space_embedding'][None,None]+parameters['time_embedding'][None,:,None])*mask | |
| if key is None:key=jax.random.key(0) | |
| keys=jax.random.split(key,3*c['depth']) | |
| for i in range(c['depth']): | |
| for rank,axis in enumerate(['time','space']): | |
| prefix=f'b{i}_{axis}';normalized=layer_norm(x,parameters[prefix+'_scale'],parameters[prefix+'_bias']);update=attention(parameters,prefix,normalized,valid,axis,c['heads']);x=(x+dropout(update,keys[3*i+rank],c['dropout'],training))*mask | |
| prefix=f'b{i}_ff';normalized=layer_norm(x,parameters[prefix+'_scale'],parameters[prefix+'_bias']);update=jax.nn.gelu(normalized@parameters[prefix+'_w1']+parameters[prefix+'_b1'])@parameters[prefix+'_w2']+parameters[prefix+'_b2'];x=(x+dropout(update,keys[3*i+2],c['dropout'],training))*mask | |
| return layer_norm(x,parameters['readout_scale'],parameters['readout_bias'])*mask | |
| def token_logits(parameters,dets,grid,basis,rounds,mask,*,training=False,key=None): | |
| x,valid,target,mask=features(dets,grid,basis,rounds,mask);hidden=encode(parameters,x,valid,training=training,key=key) | |
| return (hidden@parameters['mask_w'])[:,:, :,0]+parameters['mask_b'][0],target,mask | |
| def objective(parameters,dets,grid,basis,rounds,mask,*,training=False,key=None): | |
| z,y,mask=token_logits(parameters,dets,grid,basis,rounds,mask,training=training,key=key);count=jnp.maximum(mask.sum(),1) | |
| return jnp.sum(jnp.where(mask,jax.nn.softplus(z)-y*z,0.))/count | |
| def attach(trained,initial): | |
| result={key:value for key,value in trained.items() if not key.startswith('mask_')} | |
| result['input_w']=result['input_w'].at[6:8].set(initial['input_w'][6:8]) | |
| for key in ['readout_w1','readout_b1','readout_w2','readout_b2']:result[key]=initial[key] | |
| return result | |