Skip to content
This repository has been archived by the owner on Oct 25, 2024. It is now read-only.

Commit

Permalink
fix comment on test.py, merge latest lieonn.
Browse files Browse the repository at this point in the history
  • Loading branch information
@bitsofcotton
bitsofcotton authored Jun 6, 2024
1 parent b0155d2 commit 7c59673
Show file tree
Hide file tree
Showing 3 changed files with 54 additions and 61 deletions.
2 changes: 2 additions & 0 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -95,4 +95,6 @@ Please refer bitsofcotton/p8 for this.
2024/04/16 move command chg, no arguments, with ddpmopt:[pq]redg.
2024/04/19 add cleansc additional command, change synonim cleans... to clean... , eliminate cleans command.
2024/04/23 add predC, predCbond command.
2024/05/29 cleanC, cleanc change.
2024/06/07 fix comment on test.py.

111 changes: 51 additions & 60 deletions lieonn.hh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3161,7 +3161,7 @@ template <typename T> inline T P012L<T>::next(const SimpleVector<T>& d) {
ii < 2 * int(- log(SimpleMatrix<T>().epsilon()) / log(T(int(2))) )
&& sqrt(work.dot(work) * SimpleMatrix<T>().epsilon()) <
abs(work[work.size() - 1] - last); ii ++) {
last = work[work.size() - 1];
last = work[work.size() - 1];
const auto vdp(makeProgramInvariant<T>(work));
avg = avg0 * sqrt(vdp.first.dot(vdp.first) / avg0.dot(avg0));
work[work.size() - 1] =
Expand Down Expand Up @@ -3488,7 +3488,7 @@ public:
SimpleMatrix<T> invariants(3, nonlinear ? varlen + 2 : varlen);
invariants.O();
for(int i0 = 0; i0 < invariants.rows(); i0 ++) {
SimpleMatrix<T> toeplitz((in.size() - varlen - step + 2) / skip
SimpleMatrix<T> toeplitz((in.size() - varlen - step + 2 - 1) / skip
- invariants.rows() + 1, invariants.cols());
for(int i = i0; i < toeplitz.rows() + i0; i ++) {
auto work(in.subVector(i * skip, varlen));
Expand All @@ -3501,7 +3501,7 @@ public:
SimpleVector<T> invariant(invariants.cols());
invariant.O();
for(int i = 0; i < invariants.cols(); i ++)
invariant[i] = P0maxRank0<T>(step).next(invariants.col(i));
invariant[i] = P0maxRank0<T>().next(invariants.col(i));
if(invariant[varlen - 1] == zero) return zero;
SimpleVector<T> work(varlen);
for(int i = 1; i < work.size(); i ++)
Expand Down Expand Up @@ -4123,17 +4123,9 @@ template <typename T> pair<pair<vector<SimpleVector<T> >, vector<T> >, pair<vect
SimpleVector<T> init(3);
for(int i = 0; i < init.size(); i ++)
init[i] = T(int(i));
cerr << "P0 initialize: " << P0maxRank0<T>(1).next(init) << endl;
// N.B. we need rich internal status.
int p0(floor(sqrt(T(int(in.size() - 4 - 1 + 2 - 4 - 2 - 3)) ) / T(skip) ));
// N.B. we need large accuracy to continue much more.
for(int i = 0; i < p0; i ++) {
const auto pp(pnextcacher<T>(3, i + 1, 4));
if(T(int(in.size())) < sqrt(pp.dot(pp))) {
p0 = i + 1;
break;
}
cerr << "P0 initialize: " << P0maxRank0<T>(i + 1).next(init) << endl;
}
auto p0(min(int(floor(sqrt(T(int(in.size() - 4 - 1 + 2 - 4 - 2 - 3)) ) / T(skip) )), int(in.size() - 2) / skip / 26));
vector<SimpleVector<T> > p;
vector<T> psec;
if(p0 < 1) return make_pair(make_pair(p, psec), make_pair(p, psec));
Expand Down Expand Up @@ -4174,8 +4166,8 @@ template <typename T> pair<pair<vector<SimpleVector<T> >, vector<T> >, pair<vect
pb.next(pf.res[pf.res.size() - 1 - k]);
assert(pb.full);
for(int i = 0; i < p0; i ++) {
q[i][j] += P01<T, false>(4, i + 1).next(pb.res, skip);
p[i][j] += P01<T, false>(4, i + 1).next(pf.res, skip);
q[i][j] = P01<T, false>(4, i + 1).next(pb.res, skip);
p[i][j] = P01<T, false>(4, i + 1).next(pf.res, skip);
}
}
#if defined(_OPENMP)
Expand Down Expand Up @@ -4223,16 +4215,18 @@ template <typename T> pair<vector<vector<SimpleVector<T> > >, vector<vector<Simp
rres[m] = p.first.first[i][j * in0[0][0].size() +
getImgPt<int>(k + m - rres.size() / 2, in0[0][0].size())];
sort(rres.begin(), rres.end());
res.first[i][j][k] = revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.first.second[i]), true);
res.first[i][j][k] =
revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.first.second[i]), true);
}
for(int k = 0; k < in0[0][0].size(); k ++) {
for(int m = 0; m < rres.size(); m ++)
rres[m] = p.second.first[i][j * in0[0][0].size() +
getImgPt<int>(k + m - rres.size() / 2, in0[0][0].size())];
sort(rres.begin(), rres.end());
res.second[i][j][k] = revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.second.second[i]), true);
res.second[i][j][k] =
revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.second.second[i]), true);
}
}
}
Expand Down Expand Up @@ -4281,8 +4275,9 @@ template <typename T> pair<vector<vector<SimpleMatrix<T> > >, vector<vector<Simp
in0[0][0].cols() +
getImgPt<int>(m + mm - ccj / 2, in0[0][0].cols())];
sort(rres.begin(), rres.end());
res.first[i][j](k, m) = revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.first.second[i]), true);
res.first[i][j](k, m) =
revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.first.second[i]), true);
}
for(int k = 0; k < in0[0][0].rows(); k ++)
for(int m = 0; m < in0[0][0].cols(); m ++) {
Expand All @@ -4294,71 +4289,67 @@ template <typename T> pair<vector<vector<SimpleMatrix<T> > >, vector<vector<Simp
in0[0][0].cols() +
getImgPt<int>(m + mm - ccj / 2, in0[0][0].cols())];
sort(rres.begin(), rres.end());
res.second[i][j](k, m) = revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.second.second[i]), true);
res.second[i][j](k, m) =
revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.second.second[i]), true);
}
}
}
return res;
}

template <typename T> pair<vector<SimpleSparseTensor<T> >, vector<SimpleSparseTensor<T> > > predSTen(const vector<SimpleSparseTensor<T> >& in0, const vector<int>& idx, const int& skip = 1, const int& cj = 11) {
template <typename T> pair<vector<SimpleSparseTensor<T> >, vector<SimpleSparseTensor<T> > > predSTen(const vector<SimpleSparseTensor<T> >& in0, const vector<int>& idx, const int& skip = 1) {
assert(idx.size() && 0 < skip && in0.size() / skip);
// N.B.: we don't do input scaling.
// const int ccj(ceil(pow(T(cj), T(int(1)) / T(int(3)) )) );
const int ccj(1);
assert(0 < ccj);
if(ccj * ccj * ccj * 19683 < idx.size() * idx.size() * idx.size())
cerr << "predSTen : elements larger than 19683, exceeds function entropy." << endl;
// N.B. we don't do input scaling.
// N.B. the data we target is especially string stream corpus.
// they are incontinuous one, so complementing with continuous stream
// shouldn't improve outputs.
vector<SimpleVector<T> > in;
vector<pair<int, pair<int, int> > > absent;
in.resize(in0.size());
for(int i = 0; i < in0.size(); i ++) {
in[i].resize(idx.size() * idx.size() * idx.size());
absent.reserve(idx.size() * idx.size() * idx.size());
for(int i = 0; i < idx.size(); i ++)
for(int j = 0; j < idx.size(); j ++)
for(int k = 0; k < idx.size(); k ++) {
for(int ii = 0; ii < in0.size(); ii ++)
if(in0[ii][idx[i]][idx[j]][idx[k]] != T(int(0)))
goto next;
absent.emplace_back(make_pair(i, make_pair(j, k)));
next:
}
sort(absent.begin(), absent.end());
if(19683 < idx.size() * idx.size() * idx.size() - absent.size())
cerr << "predSTen : elements larger than 19683, exceeds function entropy." << endl;
for(int i = 0; i < in0.size(); i ++) {
in[i].resize(idx.size() * idx.size() * idx.size() - absent.size());
for(int j = 0, cnt = 0; j < idx.size(); j ++)
for(int k = 0; k < idx.size(); k ++)
for(int m = 0; m < idx.size(); m ++)
in[i][j * idx.size() * idx.size() + k * idx.size() + m] =
(in0[i][idx[j]][idx[k]][idx[m]] + T(int(1))) / T(int(2));
if(! binary_search(absent.begin(), absent.end(),
make_pair(j, make_pair(k, m))))
in[i][cnt ++] =
(in0[i][idx[j]][idx[k]][idx[m]] + T(int(1))) / T(int(2));
}
auto p(predv<T>(in, skip));
in.resize(0);
pair<vector<SimpleSparseTensor<T> >, vector<SimpleSparseTensor<T> > > res;
res.first.resize( p.first.first.size() );
res.second.resize(p.second.first.size());
vector<T> rres(ccj * ccj * ccj);
for(int i = 0; i < p.first.first.size(); i ++)
for(int j = 0; j < idx.size(); j ++)
for(int j = 0, cnt = 0; j < idx.size(); j ++)
for(int k = 0; k < idx.size(); k ++)
for(int m = 0; m < idx.size(); m ++) {
for(int jj = 0; jj < ccj; jj ++)
for(int kk = 0; kk < ccj; kk ++)
for(int mm = 0; mm < ccj; mm ++)
rres[jj * ccj * ccj + kk * ccj + mm] =
p.first.first[i][getImgPt<int>(j + jj - ccj / 2, idx.size()) *
idx.size() * idx.size() +
getImgPt<int>(k + kk - ccj / 2, idx.size()) * idx.size() +
getImgPt<int>(m + mm - ccj / 2, idx.size())] * T(int(2)) -
T(int(1));
sort(rres.begin(), rres.end());
res.first[i][ idx[j]][idx[k]][idx[m]] =
if(binary_search(absent.begin(), absent.end(),
make_pair(j, make_pair(k, m))))
continue;
res.first[i][idx[j]][idx[k]][idx[m]] =
revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.first.second[i]), true);
for(int jj = 0; jj < ccj; jj ++)
for(int kk = 0; kk < ccj; kk ++)
for(int mm = 0; mm < ccj; mm ++)
rres[jj * ccj * ccj + kk * ccj + mm] =
p.second.first[i][getImgPt<int>(j + jj - ccj / 2, idx.size()) *
idx.size() * idx.size() +
getImgPt<int>(k + kk - ccj / 2, idx.size()) * idx.size() +
getImgPt<int>(m + mm - ccj / 2, idx.size())] * T(int(2)) -
T(int(1));
sort(rres.begin(), rres.end());
res.second[i][ idx[j]][idx[k]][idx[m]] =
p.first.first[i][cnt] * T(int(2)) - T(int(1)),
p.first.second[i]), true);
res.second[i][idx[j]][idx[k]][idx[m]] =
revertProgramInvariant<T>(make_pair(
rres[rres.size() / 2], p.second.second[i]), true);
p.second.first[i][cnt ++] * T(int(2)) - T(int(1)),
p.second.second[i]), true);
}
return res;
}
Expand Down
2 changes: 1 addition & 1 deletion test.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -300,8 +300,8 @@
h = int(a[1])
# N.B.
# With ddpmopt/README.md, we have 20kbit upper limit on function entropy.
# We expect 8 bit color for each pixel.
# In practical, 2 bit from MSB expecting input is enough.
# However, we expect 8 bit color for each pixel.
upix = pow(19683 / 8., .5)
upixr = upix / pow(w * h, .5)
if(argv[2][- 1] == "c"): upixr /= pow(3., .5)
Expand Down

0 comments on commit 7c59673

Please sign in to comment.