void init(
const Shared & shared,
const Group & group,
rng_t &) {
Shared post = shared.plus_group(group);
score_coeff = -fast_log(1.f + post.inv_beta);
score = -fast_lgamma(post.alpha)
+ post.alpha * (fast_log(post.inv_beta) + score_coeff);
post_alpha = post.alpha;
}
float score_data(
const Shared & shared,
rng_t &) const {
Shared post = shared.plus_group(*this);
float score = fast_lgamma(post.alpha) - fast_lgamma(shared.alpha);
score += shared.alpha * fast_log(shared.inv_beta)
- post.alpha * fast_log(post.inv_beta);
score += -log_prod;
return score;
}
/* Update embeddings & return likelihood */
real Update(real *vec_u, real *vec_v, real *vec_error, int label)
{
real x = 0, f, g;
for (int c = 0; c != dim; c++) x += vec_u[c] * vec_v[c];
f = FastSigmoid(x);
g = (label - f) * rho;
for (int c = 0; c != dim; c++) vec_error[c] += g * vec_v[c];
for (int c = 0; c != dim; c++) vec_v[c] += g * vec_u[c];
return label > 0? fast_log(f+LOG_MIN): fast_log(1-f+LOG_MIN);
}
inline float fast_score_student_t (
float x,
float nu,
float mu,
float lambda)
{
// \cite{murphy2007conjugate}, Eq. 304
float p = 0.f;
p += fast_lgamma_nu(nu);
p += 0.5f * fast_log(lambda / (M_PIf * nu));
p += (-0.5f * nu - 0.5f) * fast_log(1.f + (lambda * sqr(x - mu)) / nu);
return p;
}
float score_data(
const Shared & shared,
const std::vector<Group> & groups,
rng_t &) const {
const float alpha_part = fast_lgamma(shared.alpha);
const float beta_part = shared.alpha * fast_log(shared.inv_beta);
float score = 0;
for (auto const & group : groups) {
if (group.count) {
Shared post = shared.plus_group(group);
score += fast_lgamma(post.alpha) - alpha_part;
score += beta_part - post.alpha * fast_log(post.inv_beta);
score += -group.log_prod;
}
}
return score;
}
/* This routine will scan over the sieve for values that look promising, and then attempt to factor
* them over the factor base. If it finds one that factors (or is a partial), it will add it to the
* list that we're accumulating for this poly_group.
*/
void extract_relations (block_data_t *data, poly_group_t *pg, poly_t *p, nsieve_t *ns, int block_start){
/* The sieve now contains estimates of the size (in bits) of the unfactored portion of the
* polynomial values. We scan the sieve for values less than this cutoff, and trial divide
* the ones that pass the test. */
mpz_t temp;
mpz_init (temp);
poly(temp, p, block_start + BLOCKSIZE/2);
mpz_abs(temp, temp);
uint8_t logQ = (uint8_t) mpz_sizeinbase (temp, 2);
int cutoff = (int) (fast_log(ns->lp_bound) * ns->T);
/* To accelerate the sieve scanning for promising values, instead of comparing each 8-bit entry
* one at a time, we instead look at 64 bits at a time. We are looking for sieve values such that
* sieve[x] > logQ - cutoff. We can use as a first approximation to this the test that
* sieve[x] > S, where S is the smallest power of 2 larger than (logQ - cutoff). Create an 8-bit
* mask - say S is 32 - the mask is 11100000. Then if you AND the mask with any sieve value, and
* the result is nonzero, it cannot possibly be below the cutoff. Make 8 copies of the mask, put
* it in a 64-bit int, cast the pointer to the sieve block, and do this AND on 64-bit chunks (8
* sieve locations at a time). Since it will be relatively rare for a value to pass the cutoff,
* most of the time this test will immediately reject all 8 values. If it doesn't, test each one
* against the cutoff individually.
*/
uint64_t *chunk = (uint64_t *) data->sieve;
uint32_t nchunks = BLOCKSIZE/8;
uint8_t maskchar = 1; // this is the 8-bit version of the mask
while (maskchar < logQ - cutoff){
maskchar *= 2;
}
maskchar /= 2;
maskchar = ~ (maskchar - 1);
uint64_t mask = maskchar;
for (int i=0; i<8; i++){ // make 8 copies of it
mask = (mask << 8) | maskchar;
}
// now loop over the sieve, a chunk at a time
for (int i=0; i < nchunks; i++){
if (mask & chunk[i]){ // then some value *might* have passed the test
for (int j=0; j<8; j++){
if (logQ - data->sieve[i*8+j] < cutoff){ // check them all
poly (temp, p, block_start + i*8 + j);
construct_relation (temp, block_start + i*8 + j, p, ns);
}
}
}
}
mpz_clear (temp);
}
void score_value(const Model & model, AlignedFloats scores) const {
if (DIST_DEBUG_LEVEL >= 1) {
DIST_ASSERT_EQ(scores.size(), counts().size());
}
const size_t size = counts().size();
const float shift = -fast_log(sample_size() + model.alpha);
const float * __restrict__ in = VectorFloat_data(shifted_scores_);
float * __restrict__ out = VectorFloat_data(scores);
for (size_t i = 0; i < size; ++i) {
out[i] = in[i] + shift;
}
}
float score_add_value(
count_t group_size,
count_t nonempty_group_count,
count_t sample_size,
count_t empty_group_count = 1) const {
// What is the probability (score) of adding a customer
// to a table which currently has:
//
// group_size people sitting at it (can be zero)
// nonempty_group_count tables that have people sitting at them
// sample_size people seated total
//
// In particular, if group_size == 0, this is the prob of sitting
// at a new table. In that case, nonempty_group_count does not
// include this "new" table, as it is obviously unoccupied.
if (group_size == 0) {
float numer = alpha + d * nonempty_group_count;
float denom = (sample_size + alpha) * empty_group_count;
return fast_log(numer / denom);
} else {
return fast_log((group_size - d) / (sample_size + alpha));
}
}
/* update the switching posterior weights */
void SkipCTS::updatePosteriors(SkipNode &n, int n_submodels, double alpha,
double log_alpha, double log_stop_mul, double log_split_mul) {
// update switching log-posteriors
double dn = static_cast<double>(n_submodels);
double K = (1.0 - alpha) * dn - 1.0;
double log_K = fast_log(K);
double log_scale = -s_log_tbl[n_submodels-1];
posteriorUpdate(n, log_scale, log_alpha, log_K, log_stop_mul, n.m_log_prob_est);
posteriorUpdate(n, log_scale, log_alpha, log_K, log_split_mul, n.m_log_prob_split);
for (int k=0; k < m_log_skip_preds.size(); k++) {
posteriorUpdate(n, log_scale, log_alpha, log_K, m_log_skip_preds[k], n.m_log_skip_lik[k]);
}
}
static int set_auto_gain(sensor_t *sensor, int enable, float gain_db, float gain_db_ceiling)
{
uint8_t reg;
int ret = cambus_readb(sensor->slv_addr, BANK_SEL, ®);
ret |= cambus_writeb(sensor->slv_addr, BANK_SEL, reg | BANK_SEL_SENSOR);
ret |= cambus_readb(sensor->slv_addr, COM8, ®);
ret |= cambus_writeb(sensor->slv_addr, COM8, (reg & (~COM8_AGC_EN)) | ((enable != 0) ? COM8_AGC_EN : 0));
if ((enable == 0) && (!isnanf(gain_db)) && (!isinff(gain_db))) {
float gain = IM_MAX(IM_MIN(fast_expf((gain_db / 20.0) * fast_log(10.0)), 32.0), 1.0);
int gain_temp = fast_roundf(fast_log2(IM_MAX(gain / 2.0, 1.0)));
int gain_hi = 0xF >> (4 - gain_temp);
int gain_lo = IM_MIN(fast_roundf(((gain / (1 << gain_temp)) - 1.0) * 16.0), 15);
ret |= cambus_writeb(sensor->slv_addr, GAIN, (gain_hi << 4) | (gain_lo << 0));
} else if ((enable != 0) && (!isnanf(gain_db_ceiling)) && (!isinff(gain_db_ceiling))) {
static int set_auto_gain(sensor_t *sensor, int enable, float gain_db, float gain_db_ceiling)
{
uint8_t reg;
int ret = cambus_readb(sensor->slv_addr, REG_COM8, ®);
ret |= cambus_writeb(sensor->slv_addr, REG_COM8, (reg & (~REG_COM8_AGC)) | ((enable != 0) ? REG_COM8_AGC : 0));
if ((enable == 0) && (!isnanf(gain_db)) && (!isinf(gain_db))) {
float gain = IM_MAX(IM_MIN(fast_expf((gain_db / 20.0) * fast_log(10.0)), 128.0), 1.0);
int gain_temp = fast_roundf(fast_log2(IM_MAX(gain / 2.0, 1.0)));
int gain_hi = 0x3F >> (6 - gain_temp);
int gain_lo = IM_MIN(fast_roundf(((gain / (1 << gain_temp)) - 1.0) * 16.0), 15);
ret |= cambus_writeb(sensor->slv_addr, REG_GAIN, ((gain_hi & 0x0F) << 4) | (gain_lo << 0));
ret |= cambus_readb(sensor->slv_addr, REG_VREF, ®);
ret |= cambus_writeb(sensor->slv_addr, REG_VREF, ((gain_hi & 0x30) << 2) | (reg & 0x3F));
} else if ((enable != 0) && (!isnanf(gain_db_ceiling)) && (!isinf(gain_db_ceiling))) {
JNIEXPORT jfloat JNICALL Java_jist_swans_misc_Util_fast_1log
(JNIEnv *env, jclass cl, jfloat n)
{
return fast_log(n);
}
void imlib_illuminvar(image_t *img) // http://ai.stanford.edu/~alireza/publication/cic15.pdf
{
switch(img->bpp) {
case IMAGE_BPP_BINARY: {
break;
}
case IMAGE_BPP_GRAYSCALE: {
break;
}
case IMAGE_BPP_RGB565: {
for (int y = 0, yy = img->h; y < yy; y++) {
uint16_t *row_ptr = IMAGE_COMPUTE_RGB565_PIXEL_ROW_PTR(img, y);
for (int x = 0, xx = img->w; x < xx; x++) {
int pixel = IMAGE_GET_RGB565_PIXEL_FAST(row_ptr, x);
#ifdef IMLIB_ENABLE_INVARIANT_TABLE
int rgb565 = invariant_table[pixel];
#else
float r_lin = xyz_table[COLOR_RGB565_TO_R8(pixel)] + 1.0;
float g_lin = xyz_table[COLOR_RGB565_TO_G8(pixel)] + 1.0;
float b_lin = xyz_table[COLOR_RGB565_TO_B8(pixel)] + 1.0;
float r_lin_sharp = (r_lin * 0.9968f) + (g_lin * 0.0228f) + (b_lin * 0.0015f);
float g_lin_sharp = (r_lin * -0.0071f) + (g_lin * 0.9933f) + (b_lin * 0.0146f);
float b_lin_sharp = (r_lin * 0.0103f) + (g_lin * -0.0161f) + (b_lin * 0.9839f);
float lin_sharp_avg = r_lin_sharp * g_lin_sharp * b_lin_sharp;
lin_sharp_avg = (lin_sharp_avg > 0.0f) ? fast_cbrtf(lin_sharp_avg) : 0.0f;
float r_lin_sharp_div = 0.0f;
float g_lin_sharp_div = 0.0f;
float b_lin_sharp_div = 0.0f;
if (lin_sharp_avg > 0.0f) {
lin_sharp_avg = 1.0f / lin_sharp_avg;
r_lin_sharp_div = r_lin_sharp * lin_sharp_avg;
g_lin_sharp_div = g_lin_sharp * lin_sharp_avg;
b_lin_sharp_div = b_lin_sharp * lin_sharp_avg;
}
float r_lin_sharp_div_log = (r_lin_sharp_div > 0.0f) ? fast_log(r_lin_sharp_div) : 0.0f;
float g_lin_sharp_div_log = (g_lin_sharp_div > 0.0f) ? fast_log(g_lin_sharp_div) : 0.0f;
float b_lin_sharp_div_log = (b_lin_sharp_div > 0.0f) ? fast_log(b_lin_sharp_div) : 0.0f;
float chi_x = (r_lin_sharp_div_log * 0.7071f) + (g_lin_sharp_div_log * -0.7071f) + (b_lin_sharp_div_log * 0.0000f);
float chi_y = (r_lin_sharp_div_log * 0.4082f) + (g_lin_sharp_div_log * 0.4082f) + (b_lin_sharp_div_log * -0.8164f);
float e_t_x = 0.9326f;
float e_t_y = -0.3609f;
float p_th_00 = e_t_x * e_t_x;
float p_th_01 = e_t_x * e_t_y;
float p_th_10 = e_t_y * e_t_x;
float p_th_11 = e_t_y * e_t_y;
float x_th_x = (p_th_00 * chi_x) + (p_th_01 * chi_y);
float x_th_y = (p_th_10 * chi_x) + (p_th_11 * chi_y);
float r_chi = (x_th_x * 0.7071f) + (x_th_y * 0.4082f);
float g_chi = (x_th_x * -0.7071f) + (x_th_y * 0.4082f);
float b_chi = (x_th_x * 0.0000f) + (x_th_y * -0.8164f);
float r_chi_invariant = fast_expf(r_chi);
float g_chi_invariant = fast_expf(g_chi);
float b_chi_invariant = fast_expf(b_chi);
float chi_invariant_sum = r_chi_invariant + g_chi_invariant + b_chi_invariant;
float r_chi_invariant_m = 0.0f;
float g_chi_invariant_m = 0.0f;
float b_chi_invariant_m = 0.0f;
if (chi_invariant_sum > 0.0f) {
chi_invariant_sum = 1.0f / chi_invariant_sum;
r_chi_invariant_m = r_chi_invariant * chi_invariant_sum;
g_chi_invariant_m = g_chi_invariant * chi_invariant_sum;
b_chi_invariant_m = b_chi_invariant * chi_invariant_sum;
}
int r_chi_invariant_m_int = IM_MAX(IM_MIN(r_chi_invariant_m * 255.0f, COLOR_R8_MAX), COLOR_R8_MIN);
int g_chi_invariant_m_int = IM_MAX(IM_MIN(g_chi_invariant_m * 255.0f, COLOR_G8_MAX), COLOR_G8_MIN);
int b_chi_invariant_m_int = IM_MAX(IM_MIN(b_chi_invariant_m * 255.0f, COLOR_B8_MAX), COLOR_B8_MIN);
int rgb565 = COLOR_R8_G8_B8_TO_RGB565(r_chi_invariant_m_int, g_chi_invariant_m_int, b_chi_invariant_m_int);
#endif
IMAGE_PUT_RGB565_PIXEL_FAST(row_ptr, x, rgb565);
}
}
break;
}
default: {
break;
}
}
}
/* compute the logarithm of the KT-estimator update multiplier */
inline double SkipNode::logKTMul(bit_t b) const {
return fast_log(ktMul(b));
}
/* process a new piece of sensory experience */
void SkipCTS::update(bit_t b) {
getContext();
double alpha = switchRate(m_history.size());
double log_alpha = fast_log(alpha);
zobhash_t hash = 0;
int skips_left, last_idx;
// update nodes from deepest to shallowest
for (int i=m_depth; i >= 0; i--) {
// update the KT statistics, then the weighted
// probability for every node on this level
const indices_list_t &il = m_indices[i];
for (int j=0; j < il.size(); j++) {
m_log_skip_preds.clear();
getContextInfo(hash, il[j]);
skips_left = m_auxinfo[i][j].skips_left;
last_idx = m_auxinfo[i][j].last_idx;
// update the node
int n_submodels = numSubmodels(last_idx, skips_left);
SkipNode &n = getNode(hash, i, n_submodels);
n.m_buf = n.m_log_prob_weighted;
// lazy allocation of skipping prior weights
if (n_submodels > 2 && n.m_log_skip_lik == NULL)
lazyAllocate(n, n_submodels, skips_left);
// handle the stop case
double log_est_mul = n.logKTMul(b);
if (n_submodels == 1) {
n.updateKT(b, log_est_mul);
n.m_log_prob_weighted += log_est_mul;
n.m_buf = log_est_mul;
continue;
}
double log_acc = n.m_log_prob_est + log_est_mul;
n.updateKT(b, log_est_mul);
// handle the split case
zobhash_t delta = s_zobtbl[last_idx+1][m_context[last_idx+1]];
const SkipNode &nn = getNode(hash ^ delta, i+1, numSubmodels(last_idx+1, skips_left));
double log_split_pred = nn.m_buf;
log_acc = fast_logadd(log_acc, n.m_log_prob_split + log_split_pred);
// handle the skipping case
if (n_submodels > 2) {
// update the skipping models
for (int k=last_idx+2; k < m_depth; k++) {
zobhash_t h = hash ^ s_zobtbl[k][m_context[k]];
SkipNode &sn = getNode(h, i+1, numSubmodels(k, skips_left - 1));
double log_skip_pred = sn.m_buf;
m_log_skip_preds.push_back(log_skip_pred);
int z = k - last_idx - 2;
log_acc = fast_logadd(log_acc, n.m_log_skip_lik[z] + log_skip_pred);
}
}
// store the weighted probability
n.m_log_prob_weighted = log_acc;
assert(n.m_log_prob_weighted < n.m_buf);
// Store the *difference* in log probability in m_buf
n.m_buf = n.m_log_prob_weighted - n.m_buf;
updatePosteriors(n, n_submodels, alpha, log_alpha, log_est_mul, log_split_pred);
}
}
m_history.push_back(b != 0);
}
