void show_below_above(const AzIntArr &ia_below, const AzIntArr &ia_above, AzBytArr &s) const {
s << "(";
for (int ix = 0; ix < ia_below.size(); ++ix) {
if (ix > 0) s << ",";
s << ia_below.get(ix);
}
s << ") -> (";
for (int ix = 0; ix < ia_above.size(); ++ix) {
if (ix > 0) s << ",";
s << ia_above.get(ix);
}
s << ")";
}
/*-------------------------------------------------------------------------*/
void AzPrepText2::gen_Y(const AzIntArr &ia_tokno, int dic_sz,
const AzIntArr &ia_pos,
int xpch_sz, /* patch size used to generate X */
int min_dist, int max_dist,
AzSmat *m_y) const
{
AzX::throw_if_null(m_y, "AzPrepText2::gen_Y", "m_y");
int t_num;
const int *tokno = ia_tokno.point(&t_num);
m_y->reform(dic_sz*2, ia_pos.size()); /* *2 for left and right */
for (int ix = 0; ix < ia_pos.size(); ++ix) {
int xtx0 = ia_pos.get(ix);
int xtx1 = xtx0 + xpch_sz;
AzIntArr ia_ctx0, ia_ctx1;
for (int tx = MAX(0,xtx0+min_dist); tx < MIN(t_num,xtx0); ++tx) if (tokno[tx] >= 0) ia_ctx0.put(tokno[tx]);
ia_ctx0.unique();
for (int tx = MAX(0,xtx1); tx < MIN(t_num,xtx1+max_dist); ++tx) if (tokno[tx] >= 0) ia_ctx1.put(tokno[tx]);
ia_ctx1.unique();
ia_ctx1.add(dic_sz);
AzIntArr ia_ctx;
ia_ctx.concat(&ia_ctx0);
ia_ctx.concat(&ia_ctx1);
if (ia_ctx.size() > 0) {
ia_ctx.unique();
m_y->col_u(ix)->load(&ia_ctx, 1);
}
}
}
/*-------------------------------------------------------------------------*/
void AzPrepText2::gen_Y_ifeat(int top_num_each, int top_num_total, const AzSmat *m_feat,
const AzIntArr &ia_tokno, const AzIntArr &ia_pos,
int xpch_sz, int min_dist, int max_dist,
bool do_nolr,
int f_pch_sz, int f_pch_step, int f_padding,
AzSmat *m_y,
feat_info fi[2]) const
{
const char *eyec = "AzPrepText2::gen_neigh_topfeat";
AzX::throw_if_null(m_feat, eyec, "m_feat");
AzX::throw_if_null(m_y, eyec, "m_y");
int t_num;
const int *tokno = ia_tokno.point(&t_num);
int feat_sz = m_feat->rowNum();
int f_pch_num = DIVUP(t_num+f_padding*2-f_pch_sz, f_pch_step) + 1;
if (m_feat->colNum() != f_pch_num) {
AzBytArr s("#patch mismatch: Expcected: "); s << f_pch_num << " Actual: " << m_feat->colNum();
AzX::throw_if(true, AzInputError, eyec, s.c_str());
}
if (do_nolr) m_y->reform(feat_sz, ia_pos.size());
else m_y->reform(feat_sz*2, ia_pos.size());
for (int ix = 0; ix < ia_pos.size(); ++ix) {
int xtx0 = ia_pos[ix];
int xtx1 = xtx0 + xpch_sz;
AzIFarr ifa_ctx;
int offs = 0;
for (int tx = xtx0+min_dist; tx < xtx0; ++tx) {
if (tx + f_pch_sz > xtx0) break;
set_ifeat(m_feat, top_num_each, (tx+f_padding)/f_pch_step, offs, &ifa_ctx, fi);
}
if (!do_nolr) offs = feat_sz;
for (int tx = xtx1; tx < xtx1+max_dist; ++tx) {
if (tx + f_pch_sz > xtx1+max_dist) break;
set_ifeat(m_feat, top_num_each, (tx+f_padding)/f_pch_step, offs, &ifa_ctx, fi);
}
ifa_ctx.squeeze_Max();
if (top_num_total > 0 && ifa_ctx.size() > top_num_total) {
ifa_ctx.sort_Float(false);
ifa_ctx.cut(top_num_total);
}
m_y->col_u(ix)->load(&ifa_ctx);
}
}
/*-------------------------------------------------------------------------*/
void AzTools_text::tokenize(AzByte *buff, int &len,
const AzDic *dic,
AzIntArr &ia_nn,
bool do_lower, bool do_utf8dashes,
/*--- output ---*/
AzDataArr<AzIntArr> &aia_tokno)
{
const char *eyec = "AzTools_text::tokenize(multi n)";
AzStrPool sp_tok;
tokenize(buff, len, do_utf8dashes, do_lower, &sp_tok);
int t_num = sp_tok.size();
aia_tokno.reset(ia_nn.size());
for (int ix = 0; ix < ia_nn.size(); ++ix) {
identify_tokens(&sp_tok, ia_nn[ix], dic, aia_tokno(ix));
if (aia_tokno[ix]->size() != t_num) throw new AzException(eyec, "Conflict in #tokens");
}
}
inline void rewind(AzCursor &cur) const {
if (_shouldDoBackward) {
cur.set(ia_index.size());
}
else {
cur.set(0);
}
}
virtual void format(AzBytArr &s, bool do_reset=false) const {
if (do_reset) s.reset();
// s.c("(");
int ix;
for (ix = 0; ix < ia_sz.size(); ++ix) {
if (ix > 0) s.c(" x ");
s.cn(ia_sz.get(ix));
}
// s.c(")");
}
/*-------------------------------------------------------------------------*/
void AzPrepText2::gen_Y_ngram_bow(const AzIntArr &ia_nn,
const AzDataArr<AzIntArr> &aia_tokno, int dic_sz,
const AzIntArr &ia_pos,
int xpch_sz, /* patch size used to generate X */
int min_dist, int max_dist,
bool do_nolr,
AzSmat *m_y) const
{
const char *eyec = "AzPrepText2::gen_Y_ngram_bow";
int t_num = aia_tokno[0]->size();
if (do_nolr) m_y->reform(dic_sz, ia_pos.size());
else m_y->reform(dic_sz*2, ia_pos.size()); /* *2 for left and right */
for (int ix = 0; ix < ia_pos.size(); ++ix) {
int xtx0 = ia_pos.get(ix);
int xtx1 = xtx0 + xpch_sz;
AzIntArr ia_ctx;
int base = xtx0+min_dist;
for (int nx = 0; nx < aia_tokno.size(); ++nx) {
const AzIntArr *ia_tokno = aia_tokno[nx];
int nn = ia_nn[nx];
for (int tx = MAX(0,base); tx <= MIN(t_num,xtx0)-nn; ++tx) {
int tokno = ia_tokno->get(tx);
if (tokno >= 0) ia_ctx.put(tokno);
}
}
base = xtx1;
for (int nx = 0; nx < aia_tokno.size(); ++nx) {
const AzIntArr *ia_tokno = aia_tokno[nx];
int nn = ia_nn[nx];
for (int tx = MAX(0,base); tx <= MIN(t_num,xtx1+max_dist)-nn; ++tx) {
int tokno = ia_tokno->get(tx);
if (tokno >= 0) {
if (do_nolr) ia_ctx.put(tokno);
else ia_ctx.put(dic_sz+tokno);
}
}
}
ia_ctx.unique();
m_y->col_u(ix)->load(&ia_ctx, 1);
}
}
virtual int sz(int dx) const {
AzX::throw_if((dx < 0 || dx >= ia_sz.size()), "AzxD::sz", "dim is out of range");
return ia_sz.get(dx);
}
virtual int get_dim() const { return ia_sz.size(); }
virtual bool is_valid() const {
if (ia_sz.size() <= 0) return false;
int ix;
for (ix = 0; ix < ia_sz.size(); ++ix) if (ia_sz.get(ix) <= 0) return false;
return true;
}
virtual int top_lay_ind() const {
return ia_layer_order.get(ia_layer_order.size()-1);
}
/*------------------------------------------------------------*/
void AzpCNet3_multi::insert_connectors(AzIntArr &ia_order, /* inout */
AzDataArr<AzIntArr> &aia_below, /* inout */
AzDataArr<AzIntArr> &aia_above, /* inout */
AzDataArr<AzpLayerConn> &conn)
const
{
const char *eyec = "AzpCNet3_multi::insert_connectors";
int layer_num = ia_order.size();
/*--- count connectors to be inserted ---*/
int conn_num = 0;
for (int lx = 0; lx < layer_num; ++lx) {
if (aia_below[lx]->size() > 1) ++conn_num;
if (aia_above[lx]->size() > 1) ++conn_num;
}
/*--- copy the current edges ---*/
AzDataArr<AzIntArr> aia_b(layer_num + conn_num);
AzDataArr<AzIntArr> aia_a(layer_num + conn_num);
conn.reset(layer_num + conn_num);
for (int lx = 0; lx < layer_num; ++lx) {
aia_b(lx)->reset(aia_below[lx]);
aia_a(lx)->reset(aia_above[lx]);
}
/*--- insert connection where multiple input/output ---*/
AzIntArr ia_o;
int cx = layer_num;
for (int ix = 0; ix < ia_order.size(); ++ix) {
int lx = ia_order.get(ix);
if (aia_b[lx]->size() > 1) { /* multiple inputs */
aia_b(cx)->reset(aia_b[lx]);
aia_a(cx)->put(lx);
for (int ix = 0; ix < aia_b[cx]->size(); ++ix) {
int below = aia_b[cx]->get(ix);
int count = aia_a(below)->replace(lx, cx);
AzX::throw_if((count != 1), eyec, "something is wrong");
}
aia_b(lx)->reset();
aia_b(lx)->put(cx);
ia_o.put(cx);
++cx;
}
ia_o.put(lx);
if (aia_above[lx]->size() > 1) {
aia_b(cx)->put(lx);
aia_a(cx)->reset(aia_above[lx]);
for (int ix = 0; ix < aia_a[cx]->size(); ++ix) {
int above = aia_a[cx]->get(ix);
int count = aia_b(above)->replace(lx, cx);
AzX::throw_if((count != 1), eyec, "something is wrong-2");
}
aia_a(lx)->reset();
aia_a(lx)->put(cx);
ia_o.put(cx);
++cx;
}
}
/*--- output ---*/
aia_below.reset(&aia_b);
aia_above.reset(&aia_a);
ia_order.reset(&ia_o);
}