cube barrel_distort_rgb(const cube &F, double offset_x) {
cube G(F.n_rows, F.n_cols, F.n_slices);
G.slice(0) = barrel_distort(F.slice(0), offset_x);
G.slice(1) = barrel_distort(F.slice(1), offset_x);
G.slice(2) = barrel_distort(F.slice(2), offset_x);
return G;
}
SEXP move_B(const mat& y, cube& B, const mat& kappa_star, const field<mat>& C,
//const field<sp_mat>& C,
mat& gamma, const mat& D, const ucolvec& s, double tau_e)
{
BEGIN_RCPP
// N x T matrix of standardized counts, y
// B = (B_1,...,B_K), where B_k is N x T matrix for iGMRF term k
// N x T matrix, gamma = sum_{k=1^K}(B_k)
// K x T, D, where row k contains T x 1 diagonal elements of Q_k
// sample cluster assignments, s(1), ..., s(N)
// Q = (Q_1,...,Q_K), where Q_k is a T x T de-scaled iGMRF precision matrix
// C = (C_1,...,C_K), where C_k = D_k^-1 * Omega_k,
// where Omega_k is the T x T adjacency matrix for iGMRF term, k
// D is a K x T matrix where row k contains T diagonal elements of Q_k
// K x M matrix, kappa_star records locations for each iGMRF term
int K = B.n_slices;
int N = y.n_rows;
int T = D.n_cols;
colvec bbar_ki(T); bbar_ki.zeros();
rowvec gammatilde_ki(T); gammatilde_ki.zeros();
rowvec ytilde_ki(T); ytilde_ki.zeros();
rowvec d_k(T);
vec zro(T); zro.zeros();
double e_ij, phi_ij, bkij, h_ij;
int k = 0, i = 0, j = 0; /* loop variables */
for( k = 0; k < K; k++ ) /* over iGMRF terms */
{
d_k = D.row(k);
for( i = 0; i < N; i++ ) /* over units */
{
/* take out T x 1, b_ki, to be sampled from gamma */
//gammatilde_ki = gamma.row(i) - B.slice(k).row(i); /* 1 x T */
//ytilde_ki = y.row(i) - gammatilde_ki;
gammatilde_ki = gamma.row(i); /* 1 x T */
// mean of univariate iGMRF, b_kij = 1/d_kj * (omega_kj(-j) * b_ki(-j))
bbar_ki = C(k,0) * B.slice(k).row(i).t(); /* T x 1 */
for( j = 0; j < T; j++ ) /* over time points */
{
gammatilde_ki(j) -= B.slice(k)(i,j);
ytilde_ki(j) = y(i,j) - gammatilde_ki(j);
// mean of univariate iGMRF, b_kij = 1/d_kj * (omega_kj(-j) * b_ki(-j))
B.slice(k)(i,j) = 0;
//bbar_ki(j) = dot( C(k,0).row(j), B.slice(k).row(i) );
e_ij = tau_e*ytilde_ki(j)
+ d_k(j)*kappa_star(k,s(i)) * bbar_ki(j);
phi_ij = tau_e + d_k(j)*kappa_star(k,s(i));
h_ij = (e_ij / phi_ij);
bkij = rnorm( 1, (h_ij), sqrt(1/phi_ij) )[0];
B.slice(k)(i,j) = bkij;
/* put back new sampled values for b_kij */
gammatilde_ki(j) += B.slice(k)(i,j);
} /* end loop j over time points */
/* put back new sampled values for b_ki */
//gamma.row(i) = gammatilde_ki + B.slice(k).row(i);
gamma.row(i) = gammatilde_ki;
} /* end loop i over units */
} /* end loop k over iGMRF terms */
END_RCPP
} /* end function ustep to sample B_1,...,B_K */
/**********************************************************************
* MSE Class
***********************************************************************/
double MSE::cost(cube pred, cube y){
dbg_assert(y.n_cols == 1 && y.n_slices == 1);
pred.reshape(pred.n_elem,1,1);
y.reshape(y.n_elem,1,1);
double retn = 0.0f;
for(int i=0; i < y.n_elem; i++) {
retn += pow(pred(i,0,0) - y(i,0,0),2);
}
retn /= y.n_elem;
retn *= 0.5f;
//return(retn);
return(0.5 * mean(mean(square(pred.slice(0) - y.slice(0)))));
}
cube Utils::conv3(cube body, cube kernel)
{
if (body.max() <= 0)
{
return zeros(body.n_rows, body.n_cols, body.n_slices);
}
cube Q(body);
for (int i = 1; i < body.n_slices - 1; i++)
{
Q.slice(i) = conv2(body.slice(i), kernel.slice(1), "same") +
conv2(body.slice(i - 1), kernel.slice(0), "same") +
conv2(body.slice(i + 1), kernel.slice(2), "same");
}
return Q;
}
cube noncont_slices(const cube & X, const uvec & index){
cube Xsub(X.n_rows, X.n_cols, index.n_elem);
for(unsigned int i=0; i<index.n_elem; i++){
Xsub.slice(i) = X.slice(index[i]);
}
return Xsub;
}
void lmMean(const cube& wX, const rowvec& vec_beta, mat& mean){
int n = wX.n_slices, P=wX.n_cols;
mean = zeros(n,P);
for(int i=0;i<n;i++){
mean.row(i) = vec_beta* wX.slice(i);
}
}
cube conv2(const cube &F, const mat &H) {
cube G(F.n_rows, F.n_cols, F.n_slices);
for (uword i = 0; i < F.n_slices; i++) {
G.slice(i) = conv2(F.slice(i), H);
}
return G;
}
cube imresize2(const cube &C, uword m, uword n) {
cube F(m, n, C.n_slices);
for (uword k = 0; k < C.n_slices; k++) {
F.slice(k) = imresize2(C.slice(k), m, n);
}
return F;
}
//! compute (trial)point-to-(model)point Mahalanobis distance
//! @param[in] index index of the points (in trial and model) to be compared
//! @param[in] &trial reference to the trial
//! @param[in] &model reference to the model
//! @param[in] &variance reference to the model variance
//! @return Mahalanobis distance between trial-point and model-point
float Classifier::mahalanobisDist(int index,mat &trial,mat &model,cube &variance)
{
mat difference = trial.col(index) - model.col(index);
mat distance = (difference.t() * (variance.slice(index)).i()) * difference;
return distance(0,0);
}
cube noncont_slices(const cube & X, const uvec & index, int r_start, int c_start, int r_stop, int c_stop){
cube Xsub(r_stop-r_start+1, c_stop-c_start+1, index.n_elem);
for(unsigned int i=0; i<index.n_elem; i++){
Xsub.slice(i) = X.slice(index[i]).submat(r_start, c_start, r_stop, c_stop);
}
return Xsub;
}
void white(const mat& y, const mat& y_center, const uvec& sti, const cube& matphi1, const cube& matphi2, mat& wy, mat& wy_center, const uvec& WTIME){
int L=matphi1.n_slices, wT = WTIME.n_elem, t=0;
mat multiply;
if(L==0){wy=y; wy_center = y_center;}
else{
wy = y.rows(WTIME);
wy_center = y_center.rows(WTIME);
for(int wt=0; wt<wT ; wt++){
t = WTIME[wt];
for(int l=0; l<L; l++){
multiply = (sti[t-l-1])?(matphi1.slice(l)):(matphi2.slice(l));
wy.row(wt) -= y.row(t-l-1)*multiply;
wy_center.row(wt) -= y_center.row(t-l-1)*multiply;
}
}
}
}
cube ConvLayer::backward(cube delta) {
// NOTE: delta may come from a linear layer
delta.reshape(_a.n_rows, _a.n_cols, _a.n_slices);
if(!_prev) {
dbg_print("null pointer to _prev in ConvLayer::backward");
return zeros<cube>(0,0,0);
}
// Compute Weight Updates
cube input = addPadding(_prev->getActivationCube(), _ksize);
for(int i = 0; i < _units; i++){
// dz_dw
for(int j=0; j < input.n_slices; j++) {
_dw(i).slice(j) = conv2d(input.slice(j), delta.slice(i));
}
// dz_db
_db(0,0,i) = accu(delta.slice(i));
}
// Compute next delta
int nr = _prev->getActivationCube().n_rows;
int nc = _prev->getActivationCube().n_cols;
int ns = _prev->getActivationCube().n_slices;
delta = addPadding(delta, _ksize);
cube next_delta = zeros<cube>(nr,nc,ns);
for(int i=0; i < ns; i++){
for(int j=0; j < _units; j++) {
next_delta.slice(i) += conv2d(
delta.slice(j),
fliplr(flipud(_w(j).slice(i)))
);
}
}
if(_prev) {
return(_prev->backward(next_delta));
}
return(next_delta);
}
void white(const mat& x1, const mat& x2, const uvec& sti, const cube& matphi1, const cube& matphi2, cube& wX, const uvec& WTIME){
int T = x1.n_rows, L = matphi1.n_slices, P = x1.n_cols, wT = WTIME.n_elem, t=0;
mat multiply(P,P);
cube X(3*P,P,T);
for(int t=0;t<T;t++){
X.slice(t)=join_cols(join_cols(diagmat(x1.row(t)),diagmat(x2.row(t))),diagmat(ones(1,P)));
}
wX = noncont_slices(X, WTIME);
if(L>0){
for(int wt=0; wt<wT ; wt++){
t = WTIME[wt];
for(int l=0; l<L; l++){
multiply = sti[t-l-1]?(matphi1.slice(l)):(matphi2.slice(l));
wX.slice(wt) -= X.slice(t-l-1)*multiply;
}
}
}
}
void lm2(const mat& wy, const cube& wX, const mat& Omega, rowvec& numer, mat& denom){
int P=wy.n_cols, n = wy.n_rows, k = wX.n_rows;
denom.zeros(), numer.zeros();
mat twX(P,k);
for(int t=0; t<n; t++){
twX = trans(wX.slice(t));
denom=denom+wX.slice(t)*Omega*twX;
numer=numer+wy.row(t)*Omega*twX;
}
}
void white(const mat& x1, const mat& x2, const uvec& sti, const cube& matphi1, const cube& matphi2, cube& wX, const uvec& End, const uvec & wEnd){
int T = x1.n_rows, L = matphi1.n_slices, P = x1.n_cols, m = End.n_elem-1;
mat multiply(P,P);
cube X(3*P,P,T);
for(int t=0;t<T;t++){
X.slice(t)=join_cols(join_cols(diagmat(x1.row(t)),diagmat(x2.row(t))),diagmat(ones(1,P)));
}
for(int run=0;run<m;run++){
wX.slices(wEnd[run],wEnd[run+1]-1) = X.slices(End[run]+L,End[run+1]-1);
if(L>0){
for(unsigned int t=End[run]+L;t<End[run+1];t++){
for(int l=0;l<L;l++){
multiply = sti[t-l-1]?(matphi1.slice(l)):(matphi2.slice(l));
wX.slice(t-L*(run+1)) -= X.slice(t-l-1)*multiply;
}
}
}
}
}
void white(const mat& y, const mat& y_center, const uvec& sti, const cube& matphi1, const cube& matphi2, mat& wwy, mat& wwy_center, const uvec& End, const uvec& wEnd){
int L=matphi1.n_slices, m = End.n_elem-1;
mat multiply, wy, wy_center;
if(L==0){wwy=y; wwy_center = y_center;}
else{
for(int run=0;run< m ;run++){
wy = y.rows(End[run]+L,End[run+1]-1);
wy_center = y_center.rows(End[run]+L,End[run+1]-1);
for(unsigned int t=End[run]+L;t<End[run+1];t++){
for(int l=0;l<L;l++){
multiply = (sti[t-l-1])?(matphi1.slice(l)):(matphi2.slice(l));
wy.row(t-End[run]-L) -= y.row(t-l-1)*multiply;
wy_center.row(t-End[run]-L) -= y_center.row(t-l-1)*multiply;
}
}
wwy.rows(wEnd[run],wEnd[run+1]-1) = wy;
wwy_center.rows(wEnd[run],wEnd[run+1]-1) = wy_center;
}
}
}
cube maxPoolLayer::forward (cube x) {
for(int i=0; i < x.n_slices; i++)
_a.slice(i) = maxDownSample(x.slice(i),
_sample_size,
_mask.slice(i));
if(_next)
return(_next->forward(_a));
return _a;
}
void white(const mat& Hc, const rowvec vec_beta, const uvec& sti, const cube& matphi1, const cube& matphi2, cube& wHX, const uvec& WTIME){
int T = Hc.n_rows, J = Hc.n_cols/2, L = matphi1.n_slices, P = matphi1.n_rows, wT = WTIME.n_elem, t=0;
mat multiply(P,P);
cube X = zeros(J*P,P,T);
for(int t=0;t<T;t++){
for(int p=0; p<P;p++){
X.slice(t).col(p).rows(J*p, J*(p+1)-1) = trans(Hc.submat(t, 0, t, J-1))*vec_beta[p] + trans(Hc.submat(t, J, t, 2*J-1))*vec_beta[P+p];
}
}
wHX = noncont_slices(X, WTIME);
if(L>0){
for(int wt=0; wt < wT; wt++){
t = WTIME[wt];
for(int l=0; l<L; l++){
multiply = sti[t-l-1]?(matphi1.slice(l)):(matphi2.slice(l));
wHX.slice(wt) -= X.slice(t-l-1)*multiply;
}
}
}
}
SEXP move_kappastar_alt(mat& kappa_star, const cube& B, const cube& Q,
const ucolvec& s, uvec& o,
int T, int a, int b, const vec& ipr)
{
BEGIN_RCPP
// K x M matrix, kappa_star, records location values
// N x T matrices, B_1,...,B_K contains the de-noised functions
// on posterior for kappa_star.
int K = kappa_star.n_rows; /* number of iGMRF terms */
int M = kappa_star.n_cols; /* number of clusters */
int k = 0, m = 0, i = 0, count_m;
double num_m;
uvec pos_m;
double a1_mk; /* posterior shape parameter */
double b1_mk; /* posterior rate parameter */
colvec b_ki(T);
for(m = 0; m < M; m++)
{
pos_m = find( s == m ); /* s is vector length N */
count_m = pos_m.n_elem;
num_m = sum( 1/ipr(pos_m) );
/* sample posterior for kappa_star(k,) for each iGMRF term, k = 1,...,K */
for( k = 0; k < K; k++ )
{
b1_mk = 0;
for( i = 0; i < count_m; i++ )
{
b_ki = B.slice(k).row(pos_m(i)).t(); /* T x 1 */
b1_mk += 0.5*( as_scalar(b_ki.t()*symmatl(Q.slice(k))*b_ki)
/ ipr(pos_m(i)) );
} /* end loop i over num_m weighted units in cluster m */
b1_mk += b;
a1_mk = 0.5*num_m*(double(T)-double(o(k))) + a;
kappa_star(k,m) = rgamma(1, a1_mk, (1/b1_mk))[0];
} /* end loop over K iGMRF terms */
} /* end loop m over clusters */
/* add a bumper */
END_RCPP
} /* end function move_kappastar() to sample cluster locations */
cube maxPoolLayer::backward (cube delta) {
cube next_delta(_mask);
for(int i=0; i < delta.n_slices; i++)
next_delta.slice(i) = maxUpSample(delta.slice(i),
_sample_size,
_mask.slice(i));
if(_prev)
return(_prev->backward(next_delta));
return next_delta;
}
void white(const mat& Hc, const rowvec vec_beta, const uvec& sti, const cube& matphi1, const cube& matphi2, cube& wHX, const uvec& End, const uvec & wEnd){
int T = Hc.n_rows, J = Hc.n_cols/2, L = matphi1.n_slices, P = matphi1.n_rows, m = End.n_elem-1;
mat multiply(P,P);
cube X = zeros(J*P,P,T);
for(int t=0;t<T;t++){
for(int p=0; p<P;p++){
X.slice(t).col(p).rows(J*p, J*(p+1)-1) = trans(Hc.submat(t, 0, t, J-1))*vec_beta[p] + trans(Hc.submat(t, J, t, 2*J-1))*vec_beta[P+p];
}
}
for(int run=0;run<m;run++){
wHX.slices(wEnd[run],wEnd[run+1]-1) = X.slices(End[run]+L,End[run+1]-1);
if(L>0){
for(unsigned int t=End[run]+L;t<End[run+1];t++){
for(int l=0;l<L;l++){
multiply = sti[t-l-1]?(matphi1.slice(l)):(matphi2.slice(l));
wHX.slice(t-L*(run+1)) -= X.slice(t-l-1)*multiply;
}
}
}
}
}
void phi_design(mat& U, const uvec& sti, const int L, umat& vxicat, cube& wZ, const uvec& WTIME){
int P=U.n_cols, wT = WTIME.n_elem, t;
int Psq=P*P;
wZ.zeros();
for(int wt=0; wt<wT; wt++){
t = WTIME[wt];
for(int l=1;l<=L;l++){
if(sti[t-l]){
for(int p=0;p<P;p++){
uvec cond= (vxicat.col(p) >= l);
wZ.slice(wt).col(p).subvec((l-1)*Psq+ p*P,(l-1)*Psq+(p+1)*P-1) = (cond)%trans(U.row(t-l));
}
wZ.slice(wt).rows((L+l-1)*Psq,(L+l)*Psq-1).zeros();
}else{
wZ.slice(wt).rows((l-1)*Psq,l*Psq-1).zeros();
for(int p=0;p<P;p++){
uvec cond= (vxicat.col(p) >= l);
wZ.slice(wt).col(p).subvec((L+l-1)*Psq+p*P,(L+l-1)*Psq+(p+1)*P-1)=cond%trans(U.row(t-l));
}
}
}
}
}
// update vector of cluster membership indicators, s(i),....,s(N)
SEXP clusterstep(const cube& B, mat& kappa_star, mat& B1, const uvec& o,
const field<mat>& C, const mat& D, ucolvec& s,
//const field<sp_mat>& C,
ucolvec& num, unsigned int& M, double& conc, int a, int b,
const vec& ipr, colvec& Num)
{
BEGIN_RCPP
// sample cluster assignments, s(1), ..., s(N)
// B = (B_1,...,B_K), where B_k is N x T matrix for iGMRF term k
// Q = (Q_1,...,Q_K), where Q_k is a T x T de-scaled iGMRF precision matrix
// C = (C_1,...,C_K), where C_k = D_k^-1 * Omega_k,
// where Omega_k is the T x T adjacency matrix for iGMRF term, k
// D is a K x T matrix where row k contains T diagonal elements of Q_k
// K x M matrix, kappa_star records locations for each iGMRF term
// o = (o_1,...,o_k) is a vector where each entry denotes the order of term K.
// e.g. RW(1) -> o = 2, RW(2) -> o = 3, seas(3) -> o = 3
int N = B.slice(0).n_rows;
int T = B.slice(0).n_cols;
int K = C.n_rows;
double sweights = 0;
// zro is the zeros.T vector
colvec zro(T); zro.zeros();
uvec o_adjust = o;
//o_adjust.zeros();
// capture quadratic product for rate kernel of posterior gamma
// posterior for kappa_star(k,i).
// save B1 to latter (in another function) compute posterior for kappa_star
// mat B1(K,N);
double a1k; /* posterior shape for kappa_star(k,i) under 1 obs */
B1.zeros();
int i, j, k;
unsigned int l;
/*
mat D_k(T,T), Omega_k(T,T);
cube Q(T,T,K);
for(k = 0; k < k; k++)
{
D_k.zeros(); D_k.diag() = D.row(k);
Omega_k = D_k * C(k,0);
Q.slice(k) = D_k - Omega_k;
} // end loop K over iGMRF terms
*/
for(i = 0; i < N; i++)
{
// check if _i assigned to singleton cluster
// if so, remove the cluster associated to _i
// and decrement the cluster labels for m > s(i)
if(num(s(i)) == 1) /* remove singleton cluster */
{
kappa_star.shed_col(s(i));
num.shed_row(s(i));
Num.shed_row(s(i));
M -= 1; /* decrement cluster count */
//decrement cluster tracking values by 1 for tossed cluster
s( find(s > s(i)) ) -= 1;
} /* end cluster accounting adjustment for singleton cluster */
else /* cluster contains more than one unit */
{
num(s(i)) -= 1;
/* scale up num to population totals, Num, based on H-T inverse probability estimator */
Num(s(i)) -= 1/ipr(i);
} /* decrement non-singleton cluster count by one */
// construct normalization constant, q0i, to sample s(i)
// build loqq0 and exponentiate
colvec bki(T), bbar_ki(T); /* T x 1, D_k^-1*Omega_k*b_ki = C(k,0)*b_ki */
mat bbar_i(K,T); bbar_i.zeros();
double logd_dk = 0; /* set of T 0 mean gaussian densities for term k */
double logq0ki = 0, logq0i = 0, q0i = 0;
// accumulate weight, q0i, for s(i) over K iGMRF terms
for( k = 0; k < K; k++)
{
logq0ki = 0; /* reset k-indexed log-like on each k */
//a1k = 0.5*(double(T)) + a;
a1k = 0.5*(double(T)-double(o_adjust(k))) + a;
bki = B.slice(k).row(i).t();
bbar_ki = C(k,0) * bki; /* T x 1 */
bbar_i.row(k) = bbar_ki.t();
B1(k,i) = 0.5*dot( D.row(k), pow((bki-bbar_ki),2) ); /* no b */
logd_dk = 0; /* set of T gaussian densities for term k */
/* dmvn(zro|m,Q.slice(k),true) */
for( j = 0; j < T; j++ )
{
logd_dk += R::dnorm(0.0,0.0,double(1/sqrt(D(k,j))),true);
}
logq0ki = logd_dk + lgamma(a1k) + a*log(b) -
lgamma(a) - a1k*trunc_log(B1(k,i)+b);
logq0i += logq0ki;
} /* end loop k over iGMRF terms */
q0i = trunc_exp(logq0i);
// construct posterior sampling weights to sample s(i)
colvec weights(M+1); weights.zeros();
/* evaluate likelihood under kappa_star(k,i) */
double lweights_l;
for(l = 0; l < M; l++) /* cycle through all clusters for s(i) */
{
s(i) = l; /* will compute likelihoods for every cluster */
lweights_l = 0; /* hold log densities for K computations */
for(k = 0; k < K; k++)
{
bki = B.slice(k).row(i).t();
for( j = 0; j < T; j++ )
{
/* effectively making assignment, s(i) = l */
lweights_l += trunc_log(R::dnorm(bki(j),bbar_i(k,j),
double(1/sqrt(kappa_star(k,l)*D(k,j))),false));
} /* end loop j over time index */
} /* end loop k over iGMRF terms */
//if(lweights_l < -300){lweights_l = -300;}
weights(l) = trunc_exp(lweights_l);
weights(l) *= double(Num(s(i)))/(double(N) - 1/ipr(i) + conc);
} /* end loop l over existing or populated clusters */
/* M+1 or new component sampled from F_{0} */
weights(M) = conc/(double(N) - 1/ipr(i) + conc)*q0i;
// normalize weights
sweights = sum(weights);
if(sweights == 0)
{
weights.ones(); weights *= 1/(double(M)+1);
}
else
{
weights /= sweights;
}
// conduct discrete posterior draw for s(j)
unsigned long MplusOne = M + 1;
s(i) = rdrawone(weights, MplusOne);
// if new cluster chosen, generate new location
if(s(i) == M)
{
/* sample posterior of ksi_star[k,m] for 1 (vs. n_m) observation */
double a_star_k; /* shape for 1 obs */
double bstar_ki;
kappa_star.insert_cols(M,1); /* add K vector new location to kappa_star */
num.insert_rows(M,1);
Num.insert_rows(M,1);
for(k = 0; k < K; k++)
{
a_star_k = 0.5*(double(T) - double(o_adjust(k))) + a; /* shape for 1 obs */
bstar_ki = B1(k,i) + b; /* B1(k,i) is a scalar quadratic product */
/*
bki = B.slice(k).row(i).t();
bstar_ki = 0.5*( as_scalar(bki.t()*symmatl(Q.slice(k))*bki) ) + b;
*/
kappa_star(k,M) = rgamma(1, a_star_k, (1/bstar_ki))[0];
}
num(M) = 1;
Num(M) = 1/ipr(i);
M = MplusOne;
}
else
{
num(s(i)) += 1;
Num(s(i)) += 1/ipr(i);
}
} /* end loop i for cluster assignment to unit i = 1,...,N */
END_RCPP
} /* end function bstep for cluster assignments, s, and computing zb */
void draw_circle(cube &I, const vec &v, vec pt, double radius) {
for (uword k = 0; k < v.n_elem; k++) {
draw_circle(I.slice(k), v(k), pt, radius);
}
}
cube LinearLayer::backward(cube delta){
// we need to calculate two types
// of gradients here:
// 1. d(z) / d(theta)
// 2. d(z) / d(a_prev_layer)
// Phase 1: Consume
// compute d(z) / d(theta)
if(!prev){
cout << __LINE__ << ": _prev is null, this should not happen." << endl;
return zeros<cube>(0,0,0);
}
dbg_assert(delta.n_cols == 1);
dbg_assert(delta.n_slices == 1);
dbg_print("backward at linear layer(" << _units << ")");
mat delta_mat = delta.slice(0);
cube prev_a = prev->getActivationCube();
//mat dz_dw = reshape(prev_a, prev_a.n_elem, 1, 1);
mat dz_dw = vectorise(prev_a);
//TODO: this still looks jenky
// Add bias term to dz_dw
//mat bias = ones<mat>(dz_dw.n_cols, 1);
//dz_dw.insert_rows(0, bias); // Add bias term
// delta_weight = delta * d(z)/d(w)
// dz_dw is [nsamples, prev_nunits]
// delta is [nsamples, nunits] , delta is [nunits,1]
_dw.slice(0) = (momentum * _dw.slice(0)) + delta_mat * trans(dz_dw);
_db.slice(0) = (momentum * _db.slice(0)) + delta_mat;
// Phase 2: Produce
// compute grad = d(z) / d(a_prev_layer)
// NOTE: gradi here equals to just theta, however
// we should remove the bias column
// mat grad = _w.tail_cols(_w.n_cols - 1);
// compute delta_new = sum(delta * grad) over output units
// _w [_units x prev->_units]
// delta [_units x 1 ]
// new delta [prev->_units x 1]
mat new_delta = trans(_w.slice(0)) * delta_mat;
// TODO: reduce the amount of mat copies.
// as it stands, 5+ copies are performed...not good.
cube next_delta = zeros<cube>(new_delta.n_rows, 1, 1);
next_delta.slice(0) = new_delta;
if(_prev)
return(_prev->backward(next_delta));
return(next_delta);
}
void HMM::_fwdback(mat init_state_distrib, mat _transmat, mat obslik,
mat &alpha, mat &beta, mat& gamma, double &loglik, mat &xi_summed,
cube &gamma2, cube &obslik2,
bool fwd_only, bool compute_gamma2) {
/*
* Compute the posterior probs. in an HMM using the forwards backwards algo.
*
* Notation:
* Y(t) = observation, Q(t) = hidden state, M(t) = mixture variable (for MOG outputs)
* A(t) = discrete input (action) (for POMDP models)
*
* INPUT:
* init_state_distrib(i) = Pr(Q(1) = i)
* transmat(i,j) = Pr(Q(t) = j | Q(t-1)=i)
* or transmat{a}(i,j) = Pr(Q(t) = j | Q(t-1)=i, A(t-1)=a) if there are discrete inputs
* obslik(i,t) = Pr(Y(t)| Q(t)=i)
*
*/
bool scaled = true;
bool maximize = false;
bool compute_xi = true;
int Q = obslik.n_rows;
int T = obslik.n_cols;
mat mixmat;
mat act; // qui act è tutti zero, altrimenti potrebbe essere un input, TODO aggiungere &act negli input
mat scale;
if (obslik2.is_empty())
compute_gamma2 = false;
act = zeros(1,T); // TODO this could be a colvec
scale = ones(1,T);
field<mat> transmat(1,1);
transmat(0,0) = _transmat;
// scale(t) = Pr(O(t) | O(1:t-1)) = gamma21/c(t) as defined by Rabiner (1989).
// Hence prod_t scale(t) = Pr(O(1)) Pr(O(2)|O(1)) Pr(O(3) | O(1:2)) ... = Pr(O(1), ... ,O(T))
// or log P = sum_t log scale(t).
// Rabiner suggests multiplying beta(t) by scale(t), but we can instead
// normalise beta(t) - the constants will cancel when we compute gamma.
if (compute_xi)
xi_summed = zeros(Q,Q);
//else
// xi_summed = [];
//%%%%%%%%% Forwards %%%%%%%%%%
//cout << "fwdback > Forwards" << endl;
int t = 0;
alpha.col(0) = vectorize(init_state_distrib) % obslik.col(t);
if (scaled){
std::pair<mat,double> _tmp = normaliseC(alpha.col(t));
alpha.col(t) = _tmp.first;
scale(t) = _tmp.second;
}
for(int t=1; t<T; t++) {
mat trans;
mat m;
trans = transmat(act(t-1));
if (maximize){
//m = max_mult(trans.t(), alpha.col(t-1)); // TODO max_mult
} else {
m = trans.t() * alpha.col(t-1);
}
alpha.col(t) = vectorize(m) % obslik.col(t);
if (scaled) {
std::pair<mat,double> _tmp = normaliseC(alpha.col(t));
alpha.col(t) = _tmp.first;
scale(t) = _tmp.second;
}
if (compute_xi && fwd_only) {// useful for online EM
xi_summed = xi_summed + normalise((alpha.col(t-1) * obslik.col(t).t()) % trans);
}
}
if (scaled) {
uvec _s = find(scale); // se c'è almeno uno zero
// portando a logaritmo c'è almeno un infinito
// quindi somma tutto a infinito
if ( _s.is_empty() ) {
loglik = -std::numeric_limits<double>::max();
} else {
loglik = sum(sum(log(scale))); // nested arma::sum because sum(mat X) return a rowvec
}
} else {
loglik = log(sum(alpha.col(T)));
}
if (fwd_only) {
gamma = alpha;
return;
}
//%%%%%%%%% Backwards %%%%%%%%%%
//cout << "fwdback > Backwards" << endl;
int M;
mat trans;
mat denom;
beta = zeros(Q,T);
if (compute_gamma2) {
M = mixmat.n_cols;
gamma2 = zeros(Q,M,T);
} else {
//gamma2 = []
}
beta.col(T-1) = ones(Q,1);
gamma.col(T-1) = normalise(alpha.col(T-1) % beta.col(T-1));
t=T-1;
if (compute_gamma2) {
denom = obslik.col(t) + (obslik.col(t)==0); // replace 0s with 1s before dividing
gamma2.slice(t) = obslik2.slice(t) % mixmat % repmat(gamma.col(t), 1, M) % repmat(denom, 1, M);
}
for (int t=T-2; t>=0; t--) { // T-2 because there are some calls to t+1
// and col(T) will generate the error Mat::col(): out of bounds
// so we must assure the limit of col(T-1)
mat b = beta.col(t+1) % obslik.col(t+1);
trans = transmat(act(t));
if (maximize){
mat B = repmat(vectorize(b).t(), Q, 1);
beta.col(t) = max(trans % B, 1);
} else
beta.col(t) = trans * b;
if (scaled)
beta.col(t) = normalise( beta.col(t) );
gamma.col(t) = normalise(alpha.col(t) % beta.col(t));
if (compute_xi){
xi_summed = xi_summed + normalise((trans % (alpha.col(t) * b.t())));
}
if (compute_gamma2){
denom = obslik.col(t) + (obslik(t)==0); // replace 0s with 1s before dividing
gamma2.slice(t) = obslik2.slice(t) % mixmat % repmat(gamma.col(t), 1, M) % repmat(denom, 1, M);
}
}
}
void draw_rect(cube &I, const vec &v, vec topleft, vec btmright) {
for (uword k = 0; k < v.n_elem; k++) {
draw_rect(I.slice(k), v(k), topleft, btmright);
}
}
void draw_line(cube &I, const vec &v, vec pt1, vec pt2) {
for (uword k = 0; k < v.n_elem; k++) {
draw_line(I.slice(k), v(k), pt1, pt2);
}
}
vector<float> predict_list(const record_array & rcd_array) {
// predicting stage
unsigned int j = 0;
unsigned int train_start = 0;
unsigned int train_end = 0;
unsigned int test_start = 0;
unsigned int test_end = 0;
unsigned int train_user = ptr_train_data->data[0].user;
unsigned int test_user = ptr_test_data->data[0].user;
vec Hu = zeros<vec>(F);
vec Vum(K);
ivec scores = linspace<ivec>(1, 5, 5);
vector<float>results;
results.resize(rcd_array.size);
for (int i = 0; i < ptr_test_data->size; i++) {
record r_test = ptr_test_data->data[i];
if ((test_user != r_test.user) || i == ptr_test_data->size -1) {
// make prediction of test_user for movies in the test set
test_end = (i == ptr_test_data->size-1) ? (i + 1) : i;
int u_size = test_end - test_start;
// find train_start and train_end
// record r_train = ptr_train_data->data[j];
while (j < ptr_train_data->size) {
record r_train = ptr_train_data->data[j];
if (r_train.user < test_user) {
train_start = j + 1;
} else if (r_train.user > test_user) {
break;
}
j++;
}
train_end = j;
if (ptr_train_data->data[j-1].user == test_user) {
// positive phase to compute Hu
Hu = BH;
for (int f = 0; f < F; f++) {
for (int u = train_start; u < train_end; u++) {
record r_train = ptr_train_data->data[u];
unsigned int k = int(r_train.score) - 1;
double w = W(k, f, r_train.movie);
Hu(f) += w;
}
}
Hu = 1.0 / (1 + exp(-Hu));
// negative phase to predict score
for (int u = test_start; u < test_end; u++) {
record r_test = ptr_test_data->data[u];
Vum = normalise( exp(BV.col(r_test.movie) + W.slice(r_test.movie) * Hu), 1);
results[u] = dot(Vum, scores);
}
} else {
// TODO: predict all movies to be the averaged movie rating
double predict_score;
for (int u = test_start; u < test_end; u++) {
predict_score = 3.6;
results[u] = predict_score;
}
}
train_start = j;
test_start = i;
test_user = r_test.user;
}
}
return results;
}
void train(const record *data, unsigned int user_id, unsigned int size, int CD_K) {
// initialization
mat V0 = zeros<mat>(K, size);
mat Vt = zeros<mat>(K, size);
vec H0 = zeros<vec>(F);
vec Ht = zeros<vec>(F);
// set up V0 and Vt based on the input data.
for (int i = 0; i < size; i++) {
record r = data[i];
V0(int(r.score)-1, i) = 1; // score - 1 is the index
Vt(int(r.score)-1, i) = 1;
}
/*
/////////////////// set up H0 by V -> H //////////////////
H0(j) = sigma( BH(j) + sum_ik ( W(k, j, r.movie) * V0(k, i) ))
*/
H0 = BH;
for (int i = 0; i < size; i++) {
H0 += W.slice(data[i].movie).t() * V0.col(i);
}
H0 = 1.0 / (1 + exp(-H0));
/////////////////// Do the contrastive divergence ///////////////////
for (int n = 0; n < CD_K; n++) {
////////////// positive phase: V -> H /////////
Ht = BH;
for (int i = 0; i < size; i ++) {
Ht += W.slice(data[i].movie).t() * Vt.col(i);
}
Ht = 1.0 / (1 + exp(-Ht));
// negative phase: H -> V
for (int i = 0; i < size; i++) {
record r = data[i];
Vt.col(i) = exp(BV.col(r.movie) + W.slice(r.movie) * Ht);
}
// Normalize Vt -> sum_k (Vt(k, i)) = 1
Vt = normalise(Vt, 1);
}
// update W
for (int i = 0; i < size; i++) {
W.slice(data[i].movie) += lrate * (V0.col(i) * H0.t() - Vt.col(i) * Ht.t());
}
// update BH
BH += lrate * (H0 - Ht);
// update BV
for (int i = 0; i < size; i++) {
BV.col(data[i].movie) += lrate * (V0.col(i) - Vt.col(i));
}
}
