"""A small contextual network reweights hyperedges before exact decoding."""
import numpy as np
import jax
import jax.numpy as jnp


def initialize(seed,input_width,hidden=32):
    rng=np.random.default_rng(seed)
    return {'w1':jnp.asarray(rng.normal(0,1/np.sqrt(input_width),(input_width,hidden)),jnp.float32),
            'b1':jnp.zeros(hidden),'w2':jnp.asarray(rng.normal(0,1/np.sqrt(hidden),(hidden,hidden)),jnp.float32),
            'b2':jnp.zeros(hidden),'w3':jnp.asarray(rng.normal(0,.001,(hidden,1)),jnp.float32),'b3':jnp.zeros(1)}


def probabilities(parameters,dets,context_indices,static,base_logits,valid):
    observed=dets[:,context_indices].astype(jnp.float32)
    features=jnp.concatenate([observed,jnp.broadcast_to(static[None,:,:],(len(dets),*static.shape))],axis=-1)
    hidden=jax.nn.silu(features@parameters['w1']+parameters['b1'])
    hidden=jax.nn.silu(hidden@parameters['w2']+parameters['b2'])
    delta=2*jnp.tanh((hidden@parameters['w3']+parameters['b3'])[...,0])
    p=jax.nn.sigmoid(base_logits[None,:]+delta)*valid[None,:]
    return p.T