File size: 1,482 Bytes
32d1b4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Frozen reconstruction features at fault endpoints, with no logical inputs."""
import numpy as np
import jax
import jax.numpy as jnp
from conditional_edge_aux_encoder import features as masked_features,encode

def endpoint_plan(context,static,valid,measurement_indices):
    context=np.asarray(context);static=np.asarray(static);valid=np.asarray(valid);mapping=np.asarray(measurement_indices);support=static[:,:120]>0;counts=support.sum(axis=1);assert counts.max()<=4
    indices=np.full((len(context),4),624,np.int32);weights=np.zeros((len(context),4),np.float32)
    for op in range(len(context)):
        selected=context[op,support[op]]
        if not valid[op] or not len(selected):continue
        assert np.all(selected<600);tokens=mapping[selected];assert np.all(tokens<624) and len(np.unique(tokens))==len(tokens)
        indices[op,:len(tokens)]=tokens;weights[op,:len(tokens)]=1/len(tokens)
    return indices,weights

def token_embeddings(encoder,dets,grid,basis,rounds):
    from conditional_edge_tokens import conditional_tokens,COLORS
    # Explicit float32 precision avoids compile-context-dependent encoder approximations.
    with jax.default_matmul_precision('highest'):
        return conditional_tokens(encoder,dets,grid,basis,rounds,jnp.asarray(COLORS,jnp.int32))

def pool_endpoints(tokens,indices,weights):
    flat=tokens.reshape(len(tokens),624,64);flat=jnp.pad(flat,((0,0),(0,1),(0,0)))
    return (flat[:,indices,:]*weights[None,:,:,None]).sum(axis=2)