// Args:
// rng: The random number generator.
// binary_inputs: The value of the inputs to the terminal layer (i.e. the
// outputs from the final hidden layer). These will be updated by the
// imputation.
// logprob: On input this is a vector giving the marginal (un-logged)
// probability that each input node is active. These values will be
// over-written by their logarithms.
// logprob_complement: On input this is any vector with size matching
// logprob. On output its elements contain log(1 - exp(logprob)).
//
// Effects:
// The latent data for the terminal layer is imputed, and the sufficient
// statistics for the latent regression model in the terminal layer are
// updated to included the imputed data.
void GFFPS::impute_terminal_layer_inputs(
RNG &rng,
double response,
std::vector<bool> &binary_inputs,
Vector &logprob,
Vector &logprob_complement) {
for (int i = 0; i < logprob.size(); ++i) {
logprob_complement[i] = log(1 - logprob[i]);
logprob[i] = log(logprob[i]);
}
Vector terminal_layer_inputs(binary_inputs.size());
Nnet::to_numeric(binary_inputs, terminal_layer_inputs);
double logp_original = terminal_inputs_log_full_conditional(
response, terminal_layer_inputs, logprob, logprob_complement);
for (int i = 0; i < terminal_layer_inputs.size(); ++i) {
terminal_layer_inputs[i] = 1 - terminal_layer_inputs[i];
double logp = terminal_inputs_log_full_conditional(
response, terminal_layer_inputs, logprob, logprob_complement);
double log_input_prob = logp - lse2(logp, logp_original);
double logu = log(runif_mt(rng));
if (logu < log_input_prob) {
logp_original = logp;
} else {
terminal_layer_inputs[i] = 1 - terminal_layer_inputs[i];
}
}
model_->terminal_layer()->suf()->add_mixture_data(
response, terminal_layer_inputs, 1.0);
Nnet::to_binary(terminal_layer_inputs, binary_inputs);
}
double logit_loglike_1(const Vec & beta, bool y, const Vec &x,
Vec *g, Mat *h, double mix_wgt){
double eta = x.dot(beta);
double lognc = lse2(0, eta);
double ans = y? eta : 0;
ans -= lognc;
if(g){
double p = exp(eta-lognc);
g->axpy(x, mix_wgt* (y-p));
if(h){
double q = 1-p;
h->add_outer( x,x, -mix_wgt * p*q);}}
return mix_wgt * ans;
}
void MlvsDataImputer::impute_latent_data_point(const ChoiceData &dp,
SufficientStatistics *suf,
RNG &rng) {
model_->fill_eta(dp, eta); // eta+= downsampling_logprob
if (downsampling_) eta += log_sampling_probs_; //
uint M = model_->Nchoices();
uint y = dp.value();
assert(y < M);
double loglam = lse(eta);
double logzmin = rlexp_mt(rng, loglam);
u[y] = -logzmin;
for (uint m = 0; m < M; ++m) {
if (m != y) {
double tmp = rlexp_mt(rng, eta[m]);
double logz = lse2(logzmin, tmp);
u[m] = -logz;
}
uint k = unmix(rng, u[m] - eta[m]);
u[m] -= mu_[k];
wgts[m] = sigsq_inv_[k];
}
suf->update(dp, wgts, u);
}