Merge pull request #88 from issp-center-dev/save_tensors_te

enable to save tensors in time-evolution / finite-temperature modes

src/iTPS/CMakeLists.txt

@@ -11,6 +11,7 @@ add_library(
   iTPS_impl STATIC
   iTPS.cpp
   tensors.cpp
+  saveload_tensors.cpp
   measure.cpp
   onesite_obs.cpp
   twosite_obs.cpp

src/iTPS/main.cpp

@@ -173,6 +173,7 @@ int run_timeevolution(MPI_Comm comm, PEPS_Parameters peps_parameters,
                    onesite_operators, twosite_operators, multisite_operators,
                    corparam, clength_param);
   tns.time_evolution();
+  tns.save_tensors();
   tns.summary();
   return 0;
 }
@@ -191,6 +192,7 @@ int run_finitetemperature(MPI_Comm comm, PEPS_Parameters peps_parameters,
                    onesite_operators, twosite_operators, multisite_operators,
                    corparam, clength_param);
   tns.finite_temperature();
+  tns.save_tensors();
   tns.summary();
   return 0;
 }

src/iTPS/saveload_tensors.cpp

@@ -0,0 +1,348 @@
+/* TeNeS - Massively parallel tensor network solver /
+/ Copyright (C) 2019- The University of Tokyo */
+
+/* This program is free software: you can redistribute it and/or modify /
+/ it under the terms of the GNU General Public License as published by /
+/ the Free Software Foundation, either version 3 of the License, or /
+/ (at your option) any later version. */
+
+/* This program is distributed in the hope that it will be useful, /
+/ but WITHOUT ANY WARRANTY; without even the implied warranty of /
+/ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the /
+/ GNU General Public License for more details. */
+
+/* You should have received a copy of the GNU General Public License /
+/ along with this program. If not, see http://www.gnu.org/licenses/. */
+
+#define _USE_MATH_DEFINES
+#include <algorithm>
+#include <complex>
+#include <random>
+#include <string>
+#include <array>
+#include <cstdlib>
+
+#include "iTPS.hpp"
+
+#include "../tensor.hpp"
+
+#include "../printlevel.hpp"
+#include "../util/file.hpp"
+#include "../util/string.hpp"
+
+using std::size_t;
+
+namespace tenes {
+namespace itps {
+
+template <class ptensor>
+void iTPS<ptensor>::save_tensors() const {
+  std::string const &save_dir = peps_parameters.tensor_save_dir;
+  if (save_dir.empty()) {
+    return;
+  }
+  {
+    // metadata
+    std::string filename = save_dir + "/params.dat";
+    std::ofstream ofs(filename.c_str());
+
+    constexpr int tensor_format_version = 1;
+    ofs << tensor_format_version << " # Format_Version\n";
+    ofs << N_UNIT << " # N_UNIT\n";
+    ofs << CHI << " # CHI\n";
+    for (int i = 0; i < N_UNIT; ++i) {
+      for (int j = 0; j < nleg; ++j) {
+        ofs << lattice.virtual_dims[i][j] << " ";
+      }
+      ofs << lattice.physical_dims[i] << " # Shape of Tn[" << i << "]\n";
+    }
+  }
+  for (int i = 0; i < N_UNIT; ++i) {
+    std::string filename = save_dir + "/";
+    std::string suffix = "_" + std::to_string(i) + ".dat";
+    Tn[i].save((filename + "T" + suffix).c_str());
+    eTt[i].save((filename + "Et" + suffix).c_str());
+    eTr[i].save((filename + "Er" + suffix).c_str());
+    eTb[i].save((filename + "Eb" + suffix).c_str());
+    eTl[i].save((filename + "El" + suffix).c_str());
+    C1[i].save((filename + "C1" + suffix).c_str());
+    C2[i].save((filename + "C2" + suffix).c_str());
+    C3[i].save((filename + "C3" + suffix).c_str());
+    C4[i].save((filename + "C4" + suffix).c_str());
+  }
+  if (mpirank == 0) {
+    for (int i = 0; i < N_UNIT; ++i) {
+      std::ofstream ofs(save_dir + "/lambda_" + std::to_string(i) + ".dat");
+      for (int j = 0; j < nleg; ++j) {
+        for (int k = 0; k < lattice.virtual_dims[i][j]; ++k) {
+          ofs << lambda_tensor[i][j][k] << std::endl;
+        }
+      }
+    }
+  }
+  if (peps_parameters.print_level >= PrintLevel::info) {
+    std::cout << "Tensors saved in " << save_dir << std::endl;
+  }
+}
+
+template <class ptensor>
+void iTPS<ptensor>::load_tensors() {
+  std::string const &load_dir = peps_parameters.tensor_load_dir;
+
+  if (!util::isdir(load_dir)) {
+    std::string msg = load_dir + " does not exists.";
+    throw tenes::load_error(msg);
+  }
+
+  int tensor_format_version = 0;
+  if (mpirank == 0) {
+    std::string filename = load_dir + "/params.dat";
+    std::string line;
+    if (util::path_exists(filename)) {
+      std::ifstream ifs(filename.c_str());
+      std::getline(ifs, line);
+      tensor_format_version = std::stoi(util::drop_comment(line));
+    }
+  }
+  bcast(tensor_format_version, 0, comm);
+  if (tensor_format_version == 0) {
+    load_tensors_v0();
+  } else if (tensor_format_version == 1) {
+    load_tensors_v1();
+  } else {
+    std::stringstream ss;
+    ss << "ERROR: Unknown saved tensor format version: "
+       << tensor_format_version;
+    throw tenes::load_error(ss.str());
+  }
+}
+
+template <class ptensor>
+void load_tensor(ptensor &A, std::string const& name, std::string const& directory, int iunit){
+  std::string filename = directory + "/" + name + "_" + std::to_string(iunit) + ".dat";
+  ptensor temp;
+  temp.load(filename.c_str());
+  if (A.rank() != temp.rank()){
+    std::stringstream ss;
+    ss << "ERROR: rank mismatch in load_tensor: ";
+    ss << name << "[" << iunit << "] has " << A.rank() << " legs, but";
+    ss << "loaded one has " << temp.rank() << " legs." << std::endl;
+    ss << "HINT: check the calculation mode. The number of legs differs between ground state calculation and finite temperature calculation.";
+    throw tenes::load_error(ss.str());
+  }
+  A = resize_tensor(temp, A.shape());
+}
+
+template <class ptensor>
+void iTPS<ptensor>::load_tensors_v1() {
+  std::string const &load_dir = peps_parameters.tensor_load_dir;
+
+  int loaded_CHI = 1;
+  std::vector<std::vector<int>> loaded_shape(N_UNIT,
+                                             std::vector<int>(nleg + 1));
+  if (mpirank == 0) {
+    std::string filename = load_dir + "/params.dat";
+    std::string line;
+    std::ifstream ifs(filename.c_str());
+
+    std::getline(ifs, line);
+    const int format_version = std::stoi(util::drop_comment(line));
+    assert(format_version == 1);
+
+    std::getline(ifs, line);
+    const int loaded_N_UNIT = std::stoi(util::drop_comment(line));
+    if (N_UNIT != loaded_N_UNIT) {
+      std::stringstream ss;
+      ss << "ERROR: N_UNIT is " << N_UNIT << " but loaded N_UNIT has "
+         << loaded_N_UNIT << std::endl;
+      throw tenes::load_error(ss.str());
+    }
+
+    std::getline(ifs, line);
+    loaded_CHI = std::stoi(util::drop_comment(line));
+    if (CHI != loaded_CHI) {
+      if (peps_parameters.print_level >= PrintLevel::info) {
+        std::cout << "WARNING: parameters.ctm.dimension is " << CHI
+                  << " but loaded tensors have CHI = " << loaded_CHI
+                  << std::endl;
+      }
+    }
+
+    for (int i = 0; i < N_UNIT; ++i) {
+      std::getline(ifs, line);
+      const auto shape = util::split(util::drop_comment(line));
+      for (int j = 0; j < nleg; ++j) {
+        loaded_shape[i][j] = std::stoi(shape[j]);
+        const int vd_param = lattice.virtual_dims[i][j];
+        if (vd_param != loaded_shape[i][j]) {
+          if (peps_parameters.print_level >= PrintLevel::info) {
+            std::cout << "WARNING: virtual dimension of the leg " << j
+                      << " of the tensor " << i << " is " << vd_param
+                      << " but loaded tensor has " << loaded_shape[i][j]
+                      << std::endl;
+          }
+        }
+      }
+      loaded_shape[i][nleg] = std::stoi(shape[nleg]);
+      const int pdim = lattice.physical_dims[i];
+      if (pdim != loaded_shape[i][nleg]) {
+        std::stringstream ss;
+        ss << "ERROR: dimension of the physical bond of the tensor " << i
+           << " is " << pdim << " but loaded tensor has "
+           << loaded_shape[i][nleg] << std::endl;
+        throw tenes::load_error(ss.str());
+      }
+    }
+  }
+  for (int i = 0; i < N_UNIT; ++i) {
+    bcast(loaded_shape[i], 0, comm);
+  }
+
+// #define LOAD_TENSOR_(A, name)                      \
+//   do {                                             \
+//     ptensor temp;                                  \
+//     temp.load((filename + name + suffix).c_str()); \
+//     A = resize_tensor(temp, A.shape());            \
+//   } while (false)
+
+  for (int i = 0; i < N_UNIT; ++i) {
+    std::string filename = load_dir + "/";
+    std::string suffix = "_" + std::to_string(i) + ".dat";
+
+    load_tensor(Tn[i], "T", load_dir, i);
+    load_tensor(eTl[i], "El", load_dir, i);
+    load_tensor(eTt[i], "Et", load_dir, i);
+    load_tensor(eTr[i], "Er", load_dir, i);
+    load_tensor(eTb[i], "Eb", load_dir, i);
+    load_tensor(C1[i], "C1", load_dir, i);
+    load_tensor(C2[i], "C2", load_dir, i);
+    load_tensor(C3[i], "C3", load_dir, i);
+    load_tensor(C4[i], "C4", load_dir, i);
+    // LOAD_TENSOR_(Tn[i], "T");
+    // LOAD_TENSOR_(eTl[i], "El");
+    // LOAD_TENSOR_(eTt[i], "Et");
+    // LOAD_TENSOR_(eTr[i], "Er");
+    // LOAD_TENSOR_(eTb[i], "Eb");
+    // LOAD_TENSOR_(C1[i], "C1");
+    // LOAD_TENSOR_(C2[i], "C2");
+    // LOAD_TENSOR_(C3[i], "C3");
+    // LOAD_TENSOR_(C4[i], "C4");
+  }
+// #undef LOAD_TENSOR_
+
+  std::vector<double> ls;
+  if (mpirank == 0) {
+    for (int i = 0; i < N_UNIT; ++i) {
+      std::ifstream ifs(load_dir + "/lambda_" + std::to_string(i) + ".dat");
+      for (int j = 0; j < nleg; ++j) {
+        for (int k = 0; k < loaded_shape[i][j]; ++k) {
+          double temp = 0.0;
+          ifs >> temp;
+          ls.push_back(temp);
+        }
+      }
+    }
+  }
+  bcast(ls, 0, comm);
+  int index = 0;
+  for (int i = 0; i < N_UNIT; ++i) {
+    const auto vdim = lattice.virtual_dims[i];
+    for (int j = 0; j < nleg; ++j) {
+      lambda_tensor[i][j].clear();
+      for (int k = 0; k < loaded_shape[i][j]; ++k) {
+        lambda_tensor[i][j].push_back(ls[index]);
+        ++index;
+      }
+      lambda_tensor[i][j].resize(vdim[j]);
+    }
+  }
+}
+
+template <class ptensor>
+void iTPS<ptensor>::load_tensors_v0() {
+  using mptensor::Shape;
+  std::string const &load_dir = peps_parameters.tensor_load_dir;
+
+  // load from the checkpoint
+  if (!util::isdir(load_dir)) {
+    std::string msg = load_dir + " does not exists.";
+    throw tenes::load_error(msg);
+  }
+  for (int i = 0; i < N_UNIT; ++i) {
+    std::string filename = load_dir + "/";
+    std::string suffix = "_" + std::to_string(i) + ".dat";
+    Tn[i].load((filename + "T" + suffix).c_str());
+    eTt[i].load((filename + "Et" + suffix).c_str());
+    eTr[i].load((filename + "Er" + suffix).c_str());
+    eTb[i].load((filename + "Eb" + suffix).c_str());
+    eTl[i].load((filename + "El" + suffix).c_str());
+    C1[i].load((filename + "C1" + suffix).c_str());
+    C2[i].load((filename + "C2" + suffix).c_str());
+    C3[i].load((filename + "C3" + suffix).c_str());
+    C4[i].load((filename + "C4" + suffix).c_str());
+  }
+  std::vector<double> ls;
+  if (mpirank == 0) {
+    for (int i = 0; i < N_UNIT; ++i) {
+      const auto vdim = lattice.virtual_dims[i];
+      std::ifstream ifs(load_dir + "/lambda_" + std::to_string(i) + ".dat");
+      for (int j = 0; j < nleg; ++j) {
+        for (int k = 0; k < vdim[j]; ++k) {
+          double temp = 0.0;
+          ifs >> temp;
+          ls.push_back(temp);
+        }
+      }
+    }
+  }
+  bcast(ls, 0, comm);
+  int index = 0;
+  for (int i = 0; i < N_UNIT; ++i) {
+    const auto vdim = lattice.virtual_dims[i];
+    for (int j = 0; j < nleg; ++j) {
+      for (int k = 0; k < vdim[j]; ++k) {
+        lambda_tensor[i][j][k] = ls[index];
+        ++index;
+      }
+    }
+  }
+
+  // overwrite dimensions
+  const Shape Cshape = C1[0].shape();
+  if (CHI != Cshape[0]) {
+    if (peps_parameters.print_level >= PrintLevel::info) {
+      std::cout << "WARNING: parameters.ctm.dimension is " << CHI
+                << " but loaded tensors have CHI = " << Cshape[0] << std::endl;
+    }
+  }
+  for (int i = 0; i < N_UNIT; ++i) {
+    const Shape Tshape = Tn[i].shape();
+    const int pdim = lattice.physical_dims[i];
+    if (pdim != Tshape[4]) {
+      std::stringstream ss;
+      ss << "ERROR: dimension of the physical bond of the tensor " << i
+         << " is " << pdim << " but loaded tensor has " << Tshape[4]
+         << std::endl;
+      throw tenes::input_error(ss.str());
+    }
+
+    for (int l = 0; l < nleg; ++l) {
+      const int vd_param = lattice.virtual_dims[i][l];
+      const int vd_loaded = Tshape[l];
+      if (vd_param != vd_loaded) {
+        if (peps_parameters.print_level >= PrintLevel::info) {
+          std::cout << "WARNING: virtual dimension of the leg " << l
+                    << " of the tensor " << i << " is " << vd_param
+                    << " but loaded tensor has " << vd_loaded << std::endl;
+        }
+      }
+    }
+  }
+}
+
+// template specialization
+template class iTPS<real_tensor>;
+template class iTPS<complex_tensor>;
+
+}  // namespace itps
+}  // namespace tenes