/* tag_viterbi:
* This function implement the Viterbi algorithm in order to decode the most
* probable sequence of labels according to the model. Some part of this code
* is very similar to the computation of the gradient as expected.
*
* And like for the gradient, the caller is responsible to ensure there is
* enough stack space.
*/
void tag_viterbi(mdl_t *mdl, const seq_t *seq,
uint32_t out[], double *sc, double psc[]) {
const uint32_t Y = mdl->nlbl;
const uint32_t T = seq->len;
double *vpsi = xvm_new(T * Y * Y);
uint32_t *vback = xmalloc(sizeof(uint32_t) * T * Y);
double (*psi) [T][Y][Y] = (void *)vpsi;
uint32_t (*back)[T][Y] = (void *)vback;
double *cur = xmalloc(sizeof(double) * Y);
double *old = xmalloc(sizeof(double) * Y);
// We first compute the scores for each transitions in the lattice of
// labels.
int op;
if (mdl->type == 1)
op = tag_memmsc(mdl, seq, vpsi);
else if (mdl->opt->lblpost)
op = tag_postsc(mdl, seq, vpsi);
else
op = tag_expsc(mdl, seq, vpsi);
if (mdl->opt->force)
tag_forced(mdl, seq, vpsi, op);
// Now we can do the Viterbi algorithm. This is very similar to the
// forward pass
// | α_1(y) = Ψ_1(y,x_1)
// | α_t(y) = max_{y'} α_{t-1}(y') + Ψ_t(y',y,x_t)
// We just replace the sum by a max and as we do the computation in the
// logarithmic space the product become a sum. (this also mean that we
// don't have to worry about numerical problems)
//
// Next we have to walk backward over the α in order to find the best
// path. In order to do this efficiently, we keep in the 'back' array
// the indice of the y value selected by the max. This also mean that
// we only need the current and previous value of the α vectors, not
// the full matrix.
for (uint32_t y = 0; y < Y; y++)
cur[y] = (*psi)[0][0][y];
for (uint32_t t = 1; t < T; t++) {
for (uint32_t y = 0; y < Y; y++)
old[y] = cur[y];
for (uint32_t y = 0; y < Y; y++) {
double bst = -HUGE_VAL;
uint32_t idx = 0;
for (uint32_t yp = 0; yp < Y; yp++) {
double val = old[yp];
if (op)
val *= (*psi)[t][yp][y];
else
val += (*psi)[t][yp][y];
if (val > bst) {
bst = val;
idx = yp;
}
}
(*back)[t][y] = idx;
cur[y] = bst;
}
}
// We can now build the sequence of labels predicted by the model. For
// this we search in the last α vector the best value. Using this index
// as a starting point in the back-pointer array we finally can decode
// the best sequence.
uint32_t bst = 0;
for (uint32_t y = 1; y < Y; y++)
if (cur[y] > cur[bst])
bst = y;
if (sc != NULL)
*sc = cur[bst];
for (uint32_t t = T; t > 0; t--) {
const uint32_t yp = (t != 1) ? (*back)[t - 1][bst] : 0;
const uint32_t y = bst;
out[t - 1] = y;
if (psc != NULL)
psc[t - 1] = (*psi)[t - 1][yp][y];
bst = yp;
}
free(old);
free(cur);
free(vback);
xvm_free(vpsi);
}
/* tag_nbviterbi:
* This function implement the Viterbi algorithm in order to decode the N-most
* probable sequences of labels according to the model. It can be used to
* compute only the best one and will return the same sequence than the
* previous function but will be slower to do it.
*/
void tag_nbviterbi(mdl_t *mdl, const seq_t *seq, uint32_t N,
uint32_t **out, double sc[], double **psc) {
const uint32_t Y = mdl->nlbl;
const uint32_t T = seq->len;
double *vpsi = xvm_new(T * Y * Y);
uint32_t *vback = xmalloc(sizeof(uint32_t) * T * Y * N);
double (*psi) [T][Y ][Y] = (void *)vpsi;
uint32_t (*back)[T][Y * N] = (void *)vback;
double *cur = xmalloc(sizeof(double) * Y * N);
double *old = xmalloc(sizeof(double) * Y * N);
// We first compute the scores for each transitions in the lattice of
// labels.
int op;
if (mdl->type == 1)
op = tag_memmsc(mdl, seq, vpsi);
else if (mdl->opt->lblpost)
op = tag_postsc(mdl, seq, (double *)psi);
else
op = tag_expsc(mdl, seq, (double *)psi);
if (mdl->opt->force)
tag_forced(mdl, seq, vpsi, op);
// Here also, it's classical but we have to keep the N best paths
// leading to each nodes of the lattice instead of only the best one.
// This mean that code is less trivial and the current implementation is
// not the most efficient way to do this but it works well and is good
// enough for the moment.
// We first build the list of all incoming arcs from all paths from all
// N-best nodes and next select the N-best one. There is a lot of room
// here for later optimisations if needed.
for (uint32_t y = 0, d = 0; y < Y; y++) {
cur[d++] = (*psi)[0][0][y];
for (uint32_t n = 1; n < N; n++)
cur[d++] = -DBL_MAX;
}
for (uint32_t t = 1; t < T; t++) {
for (uint32_t d = 0; d < Y * N; d++)
old[d] = cur[d];
for (uint32_t y = 0; y < Y; y++) {
// 1st, build the list of all incoming
double lst[Y * N];
for (uint32_t yp = 0, d = 0; yp < Y; yp++) {
for (uint32_t n = 0; n < N; n++, d++) {
lst[d] = old[d];
if (op)
lst[d] *= (*psi)[t][yp][y];
else
lst[d] += (*psi)[t][yp][y];
}
}
// 2nd, init the back with the N first
uint32_t *bk = &(*back)[t][y * N];
for (uint32_t n = 0; n < N; n++)
bk[n] = n;
// 3rd, search the N highest values
for (uint32_t i = N; i < N * Y; i++) {
// Search the smallest current value
uint32_t idx = 0;
for (uint32_t n = 1; n < N; n++)
if (lst[bk[n]] < lst[bk[idx]])
idx = n;
// And replace it if needed
if (lst[i] > lst[bk[idx]])
bk[idx] = i;
}
// 4th, get the new scores
for (uint32_t n = 0; n < N; n++)
cur[y * N + n] = lst[bk[n]];
}
}
// Retrieving the best paths is similar to classical Viterbi except that
// we have to search for the N bet ones and there is N time more
// possibles starts.
for (uint32_t n = 0; n < N; n++) {
uint32_t bst = 0;
for (uint32_t d = 1; d < Y * N; d++)
if (cur[d] > cur[bst])
bst = d;
if (sc != NULL)
sc[n] = cur[bst];
cur[bst] = -DBL_MAX;
for (uint32_t t = T; t > 0; t--) {
const uint32_t yp = (t != 1) ? (*back)[t - 1][bst] / N: 0;
const uint32_t y = bst / N;
out[t - 1][n] = y;
if (psc != NULL)
psc[t - 1][n] = (*psi)[t - 1][yp][y];
bst = (*back)[t - 1][bst];
}
}
free(old);
free(cur);
free(vback);
xvm_free(vpsi);
}
void trn_lbfgs(mdl_t *mdl) {
const size_t F = mdl->nftr;
const int K = mdl->opt->maxiter;
const int C = mdl->opt->objwin;
const int M = mdl->opt->lbfgs.histsz;
const size_t W = mdl->opt->nthread;
const bool l1 = mdl->opt->rho1 != 0.0;
double *x, *xp; // Current and previous value of the variables
double *g, *gp; // Current and previous value of the gradient
double *pg; // The pseudo-gradient (only for owl-qn)
double *d; // The search direction
double *s[M]; // History value s_k = Δ(x,px)
double *y[M]; // History value y_k = Δ(g,pg)
double p[M]; // ρ_k
double fh[C]; // f(x) history
grd_t *grds[W];
// Initialization: Here, we have to allocate memory on the heap as we
// cannot request so much memory on the stack as this will have a too
// big impact on performance and will be refused by the system on non-
// trivial models.
x = mdl->theta;
xp = xvm_new(F); g = xvm_new(F);
gp = xvm_new(F); d = xvm_new(F);
for (int m = 0; m < M; m++) {
s[m] = xvm_new(F);
y[m] = xvm_new(F);
}
pg = l1 ? xvm_new(F) : NULL;
grds[0] = grd_new(mdl, g);
for (size_t w = 1; w < W; w++)
grds[w] = grd_new(mdl, xvm_new(F));
// Minimization: This is the heart of the function. (a big heart...) We
// will perform iterations until one these conditions is reached
// - the maximum iteration count is reached
// - we have converged (upto numerical precision)
// - the report function return false
// - an error happen somewhere
double fx = grd_gradient(mdl, g, grds);
for (int k = 0; !uit_stop && k < K; k++) {
// We first compute the pseudo-gradient of f for owl-qn. It is
// defined in [3, pp 335(4)]
// | ∂_i^- f(x) if ∂_i^- f(x) > 0
// ◇_i f(x) = | ∂_i^+ f(x) if ∂_i^+ f(x) < 0
// | 0 otherwise
// with
// ∂_i^± f(x) = ∂/∂x_i l(x) + | Cσ(x_i) if x_i ≠ 0
// | ±C if x_i = 0
if (l1) {
const double rho1 = mdl->opt->rho1;
for (unsigned f = 0; f < F; f++) {
if (x[f] < 0.0)
pg[f] = g[f] - rho1;
else if (x[f] > 0.0)
pg[f] = g[f] + rho1;
else if (g[f] < -rho1)
pg[f] = g[f] + rho1;
else if (g[f] > rho1)
pg[f] = g[f] - rho1;
else
pg[f] = 0.0;
}
}
// 1st step: We compute the search direction. We search in the
// direction who minimize the second order approximation given
// by the Taylor series which give
// d_k = - H_k^{-1} g_k
// But computing the inverse of the hessian is intractable so
// the l-bfgs only approximate it's diagonal. The exact
// computation is well described in [1, pp 779].
// The only special thing for owl-qn here is to use the pseudo
// gradient instead of the true one.
xvm_neg(d, l1 ? pg : g, F);
if (k != 0) {
const int km = k % M;
const int bnd = (k <= M) ? k : M;
double alpha[M], beta;
// α_i = ρ_j s_j^T q_{i+1}
// q_i = q_{i+1} - α_i y_i
for (int i = bnd; i > 0; i--) {
const int j = (k - i + M + 1) % M;
alpha[i - 1] = p[j] * xvm_dot(s[j], d, F);
xvm_axpy(d, -alpha[i - 1], y[j], d, F);
}
// r_0 = H_0 q_0
// Scaling is described in [2, pp 515]
// for k = 0: H_0 = I
// for k > 0: H_0 = I * y_k^T s_k / ||y_k||²
// = I * 1 / ρ_k ||y_k||²
const double y2 = xvm_dot(y[km], y[km], F);
const double v = 1.0 / (p[km] * y2);
for (size_t f = 0; f < F; f++)
d[f] *= v;
// β_j = ρ_j y_j^T r_i
// r_{i+1} = r_i + s_j (α_i - β_i)
for (int i = 0; i < bnd; i++) {
const int j = (k - i + M) % M;
beta = p[j] * xvm_dot(y[j], d, F);
xvm_axpy(d, alpha[i] - beta, s[j], d, F);
}
}
// For owl-qn, we must remain in the same orthant than the
// pseudo-gradient, so we have to constrain the search
// direction as described in [3, pp 35(3)]
// d^k = π(d^k ; v^k)
// = π(d^k ; -◇f(x^k))
if (l1)
for (size_t f = 0; f < F; f++)
if (d[f] * pg[f] >= 0.0)
d[f] = 0.0;
// 2nd step: we perform a linesearch in the computed direction,
// we search a step value that satisfy the constrains using a
// backtracking algorithm. Much elaborated algorithm can perform
// better in the general case, but for CRF training, bactracking
// is very efficient and simple to implement.
// For quasi-Newton, the natural step is 1.0 so we start with
// this one and reduce it only if it fail with an exception for
// the first step where a better guess can be done.
// We have to keep track of the current point and gradient as we
// will need to compute the delta between those and the found
// point, and perhaps need to restore them if linesearch fail.
memcpy(xp, x, sizeof(double) * F);
memcpy(gp, g, sizeof(double) * F);
double sc = (k == 0) ? 0.1 : 0.5;
double stp = (k == 0) ? 1.0 / xvm_norm(d, F) : 1.0;
double gd = l1 ? 0.0 : xvm_dot(g, d, F); // gd = g_k^T d_k
double fi = fx;
bool err = false;
for (int ls = 1; !uit_stop; ls++, stp *= sc) {
// We compute the new point using the current step and
// search direction
xvm_axpy(x, stp, d, xp, F);
// For owl-qn, we have to project back the point in the
// current orthant [3, pp 35]
// x^{k+1} = π(x^k + αp^k ; ξ)
if (l1) {
for (size_t f = 0; f < F; f++) {
double or = xp[f];
if (or == 0.0)
or = -pg[f];
if (x[f] * or <= 0.0)
x[f] = 0.0;
}
}
// And we ask for the value of the objective function
// and its gradient.
fx = grd_gradient(mdl, g, grds);
// Now we check if the step satisfy the conditions. For
// l-bfgs, we check the classical decrease and curvature
// known as the Wolfe conditions [2, pp 506]
// f(x_k + α_k d_k) ≤ f(x_k) + β' α_k g_k^T d_k
// g(x_k + α_k d_k)^T d_k ≥ β g_k^T d_k
//
// And for owl-qn we check a variant of the Armijo rule
// described in [3, pp 36]
// f(π(x^k+αp^k;ξ)) ≤ f(x^k) - γv^T[π(x^k+αp^k;ξ)-x^k]
if (!l1) {
if (fx > fi + stp * gd * 1e-4)
sc = 0.5;
else if (xvm_dot(g, d, F) < gd * 0.9)
sc = 2.1;
else
break;
} else {
double vp = 0.0;
for (size_t f = 0; f < F; f++)
vp += (x[f] - xp[f]) * d[f];
if (fx < fi + vp * 1e-4)
break;
}
// If we reach the maximum number of linesearsh steps
// without finding a good one, we just fail.
if (ls == mdl->opt->lbfgs.maxls) {
warning("maximum linesearch reached");
err = true;
break;
}
}
// If linesearch failed or user interupted training, we return
// to the last valid point and stop the training. The model is
// probably not fully optimized but we let the user decide what
// to do with it.
if (err || uit_stop) {
memcpy(x, xp, sizeof(double) * F);
break;
}
if (uit_progress(mdl, k + 1, fx) == false)
break;
// 3rd step: we update the history used for approximating the
// inverse of the diagonal of the hessian
// s_k = x_{k+1} - x_k
// y_k = g_{k+1} - g_k
// ρ_k = 1 / y_k^T s_k
const int kn = (k + 1) % M;
xvm_sub(s[kn], x, xp, F);
xvm_sub(y[kn], g, gp, F);
p[kn] = 1.0 / xvm_dot(y[kn], s[kn], F);
// And last, we check for convergence. The convergence check is
// quite simple [2, pp 508]
// ||g|| / max(1, ||x||) ≤ ε
// with ε small enough so we stop when numerical precision is
// reached. For owl-qn we just have to check against the pseudo-
// gradient instead of the true one.
const double xn = xvm_norm(x, F);
const double gn = xvm_norm(l1 ? pg : g, F);
if (gn / max(xn, 1.0) <= 1e-5)
break;
if (k + 1 == K)
break;
// Second stoping criterion tested is a check for improvement of
// the function value over the past W iteration. When this come
// under an epsilon, we also stop the minimization.
fh[k % C] = fx;
double dlt = 1.0;
if (k >= C) {
const double of = fh[(k + 1) % C];
dlt = fabs(of - fx) / of;
if (dlt < mdl->opt->stopeps)
break;
}
}
// Cleanup: We free all the vectors we have allocated.
xvm_free(xp); xvm_free(g);
xvm_free(gp); xvm_free(d);
for (int m = 0; m < M; m++) {
xvm_free(s[m]);
xvm_free(y[m]);
}
if (l1)
xvm_free(pg);
for (size_t w = 1; w < W; w++)
xvm_free(grds[w]->g);
for (size_t w = 0; w < W; w++)
grd_free(grds[w]);
}
