jiri/LA_library
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Tucked tested on compressed tensors, flattening implemented

Browse Source
This commit is contained in:
Jiri Pittner 2025-10-23 15:29:48 +02:00
parent a3ace7c757
commit cd09d93c27
3 changed files with 151 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

59
t.cc
View File

@ -3619,7 +3619,7 @@ cout << "Error "<<(u*sdiag*vt-abak).norm()<<endl;
}
if(1)
if(0)
{
//tucker of a flat tensor
int r,n;
@ -3647,5 +3647,62 @@ cout <<"invTucker\n"<<y;
cout <<"Error = "<<(x0-y).norm()<<endl;
}
if(0)
{
//tucker of a non-flat non-symmetric tensor
int r,n;
cin>>r>>n;
INDEXGROUP shape;
{
shape.number=r;
shape.symmetry=0;
shape.range=n;
shape.offset=0;
}
Tensor<double> x(shape);
x.randomize(1.);
cout<<x;
Tensor<double> x0(x);
x0.copyonwrite();
bool inv=true;
NRVec<NRMat<double> > dec=x.Tucker(1e-12,inv);
cout<<"Tucker\n"<<x<<endl;
cout<<dec;
Tensor<double> y = x.inverseTucker(dec,inv);
cout <<"invTucker\n"<<y;
x0.split_index_group(0);
cout <<"Error = "<<(x0-y).norm()<<endl;
}
if(1)
{
//tucker of a non-flat symmetric tensor
int r,n;
cin>>r>>n;
INDEXGROUP shape;
{
shape.number=r;
shape.symmetry= -1;
shape.range=n;
shape.offset=0;
}
Tensor<double> x(shape);
x.randomize(1.);
cout<<x;
Tensor<double> x0(x);
x0.copyonwrite();
bool inv=true;
NRVec<NRMat<double> > dec=x.Tucker(1e-12,inv);
cout<<"Tucker\n"<<x<<endl;
cout<<dec;
Tensor<double> y = x.inverseTucker(dec,inv);
cout <<"invTucker\n"<<y;
Tensor<double> x1=x0.flatten();
cout <<"Error = "<<(x1-y).norm()<<endl;
}
}//main

91
tensor.cc
View File

@ -516,7 +516,7 @@ template<typename T>
void Tensor<T>::grouploopover(void (*callback)(const GROUPINDEX &, T *))
{
GROUPINDEX I(shape.size());
T *pp=&data[0];
T *pp= &data[0];
loopovergroups(*this,shape.size()-1,&pp,I,callback);
}
@ -649,7 +649,7 @@ for(int i=0; i<shape.size(); ++i)
else flatindex += shape[i].number;
}
std::cout <<"unwind new shape = "<<newshape<<std::endl;
//std::cout <<"unwind new shape = "<<newshape<<std::endl;
Tensor<T> r(newshape);
if(r.rank()!=rank()) laerror("internal error 2 in unwind_index");
@ -670,7 +670,7 @@ if(!indexperm.is_valid())
laerror("internal error 3 in unwind_index");
}
std::cout <<"unwind permutation = "<<indexperm<<std::endl;
//std::cout <<"unwind permutation = "<<indexperm<<std::endl;
//loop recursively and do the unwinding
help_tt<T> = this;
@ -680,6 +680,79 @@ return r;
}
template<typename T>
static void flatten_callback(const SUPERINDEX &I, T *v)
{
FLATINDEX J = superindex2flat(I);
//std::cout <<"TEST flatten_callback: from "<<JP<<" TO "<<J<<std::endl;
*v = (*help_tt<T>)(J); //rhs operator() generates the redundant elements for the unwinded lhs tensor
}
//
template<typename T>
Tensor<T> Tensor<T>::flatten(int group) const
{
if(group>=shape.size()) laerror("too high group number in flatten");
if(is_flat()) return *this;
if(group>=0) //single group
{
if(shape[group].number==1) return *this;
if(shape[group].symmetry==0)
{
Tensor<T> r(*this);
r.split_index_group(group);
return r;
}
}
if(group<0 && !is_compressed())
{
Tensor<T> r(*this);
for(int g=0; g<shape.size(); ++g) if(shape[g].number>1) r.split_index_group(g);
return r;
}
//general case
int newsize;
if(group<0) newsize=rank();
else newsize=shape.size()+shape[group].number-1;
//build new shape
NRVec<indexgroup> newshape(newsize);
int gg=0;
for(int g=0; g<shape.size(); ++g)
{
if((group<0 ||g==group) && shape[g].number>1) //flatten this group
{
for(int i=0; i<shape[g].number; ++i)
{
newshape[gg].symmetry=0;
newshape[gg].number=1;
newshape[gg].range=shape[g].range;
#ifndef LA_TENSOR_ZERO_OFFSET
newshape[gg].offset = shape[g].offset;
#endif
gg++;
}
}
else //preserve this group
{
newshape[gg++] = shape[g];
}
}
std::cout <<"Flatten new shape = "<<newshape<<std::endl;
//decompress the tensor data
Tensor<T> r(newshape);
help_tt<T> = this;
r.loopover(flatten_callback);
return r;
}
template<typename T>
Tensor<T> Tensor<T>::unwind_indices(const INDEXLIST &il) const
{
@ -1012,7 +1085,7 @@ void Tensor<T>::split_index_group(int group)
{
if(group<0||group >= shape.size()) laerror("illegal index group number");
if(shape[group].number==1) return; //nothing to split
if(shape[group].symmetry!=0) laerror("only non-symmetric index group can be splitted");
if(shape[group].symmetry!=0) laerror("only non-symmetric index group can be splitted, use flatten instead");
NRVec<indexgroup> newshape(shape.size()+shape[group].number-1);
int gg=0;
@ -1133,16 +1206,19 @@ for(int i=0; i<r; ++i)
//std::cout << "resulting U "<<u<<std::endl;
//std::cout << "resulting W "<<w<<std::endl;
//std::cout << "resulting VT "<<vt<<std::endl;
int umnr=um.nrows();
int umnc=um.ncols();
um.resize(0,0); //deallocate
int preserve=mini;
for(int k=0; k<mini; ++k) if(w[k]<thr) {preserve=k; break;}
if(preserve==0) laerror("singular tensor in Tucker decomposition");
NRMat<T> umnew;
//std::cout <<"TEST "<<i<<" mini preserve "<<mini<<" "<<preserve<<std::endl;
if(preserve<mini)
{
vt=vt.submatrix(0,preserve-1,0,um.ncols()-1);
vt=vt.submatrix(0,preserve-1,0,umnc-1);
w=w.subvector(0,preserve-1);
umnew=u.submatrix(0,um.nrows()-1,0,preserve-1);
umnew=u.submatrix(0,umnr-1,0,preserve-1);
}
else umnew=u;
ret[(inverseorder? r-i-1 : i)]=vt.transpose(true);
@ -1197,6 +1273,9 @@ else
template class Tensor<double>;
template class Tensor<std::complex<double> >;
template std::ostream & operator<<(std::ostream &s, const Tensor<double> &x);

11
tensor.h
View File

@ -40,14 +40,16 @@
//TODO:
//@@@!!!!!! - implement index names and contractions, unwinding etc. by named index list
//@@@index names flat or in groups ?
//
//@@@contraction inside one tensor - compute resulting shape, loopover the shape, create index into the original tensor + loop over the contr. index, do the summation, store result
//@@@ will need to store vector of INDEX to the original tensor for the result's flatindex
//@@@ will not be particularly efficient
//
//@@@conversions to/from fourindex
//@@@conversions to/from fourindex, optional negarive rande for beta spin handling
//@@@ optional distinguish covariant and contravariant check in contraction
//
//@@@!!!!!!!!!!!const loopover and grouploopover
//maybe const loopover and grouploopover to avoid problems with shallowly copied tensors
//
//@@@?general permutation of individual indices - check the indices in sym groups remain adjacent, calculate result's shape, loopover the result and permute using unwind_callback
//
@ -162,6 +164,7 @@ public:
NRMat<T> matrix() const {return NRMat<T>(data,data.size()/groupsizes[0],groupsizes[0],0);}; //reinterpret as matrix with column index being the tensor's leftmost index group (typically the unwound single index)
bool is_flat() const {for(int i=0; i<shape.size(); ++i) if(shape[i].number>1) return false; return true;};
bool is_compressed() const {for(int i=0; i<shape.size(); ++i) if(shape[i].number>1&&shape[i].symmetry!=0) return true; return false;};
void clear() {data.clear();};
int rank() const {return myrank;};
int calcrank(); //is computed from shape
@ -240,9 +243,11 @@ public:
// Note that *this tensor can be e.g. antisymmetric while rhs is not and is being antisymmetrized by the PermutationAlgebra
// The efficiency is not optimal, even when avoiding the outer product, the calculation is done indexing element by element
// More efficient would be applying permutation algebra symbolically and efficiently computing term by term
void split_index_group(int group); //formal split of a non-symmetric index group WITHOUT the need for data reorganization
void split_index_group(int group); //formal in-place split of a non-symmetric index group WITHOUT the need for data reorganization
void merge_adjacent_index_groups(int groupfrom, int groupto); //formal merge of non-symmetric index groups WITHOUT the need for data reorganization
Tensor merge_index_groups(const NRVec<int> &groups) const;
Tensor flatten(int group= -1) const; //split and uncompress a given group or all of them
NRVec<NRMat<T> > Tucker(typename LA_traits<T>::normtype thr=1e-12, bool inverseorder=true); //HOSVD-Tucker decomposition, return core tensor in *this, flattened
Tensor inverseTucker(const NRVec<NRMat<T> > &x, bool inverseorder=true) const; //rebuild the original tensor from Tucker
};