plumed_pytorch_gnn/PytorchModelGNN.cpp

/* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Copyright (c) 2022-2024 of Luigi Bonati and Enrico Trizio.



The pytorch module is free software: you can redistribute it and/or modify

it under the terms of the GNU Lesser General Public License as published by

the Free Software Foundation, either version 3 of the License, or

(at your option) any later version.



The pytorch module 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 Lesser General Public License for more details.



You should have received a copy of the GNU Lesser General Public License

along with plumed.  If not, see <http://www.gnu.org/licenses/>.

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */

#ifdef __PLUMED_HAS_LIBTORCH

#include <cmath>
#include <memory>
#include <fstream>
#include <type_traits>
#include <torch/torch.h>
#include <torch/script.h>
#include <torch/csrc/jit/jit_log.h>

#include "core/PlumedMain.h"
#include "config/Config.h"
#include "colvar/Colvar.h"
#include "colvar/ActionRegister.h"
#include "tools/NeighborList.h"
#include "tools/Communicator.h"
#include "tools/OpenMP.h"
#include "tools/File.h"
#include "tools/PDB.h"

// NOTE: Freezing a ScriptModule (torch::jit::freeze) works only in >=1.11
// For 1.8 <= versions <=1.10 we need a hack
// (see https://discuss.pytorch.org/t/how-to-check-libtorch-version/77709/4 and also
// https://github.com/pytorch/pytorch/blob/dfbd030854359207cb3040b864614affeace11ce/torch/csrc/jit/api/module.cpp#L479)
// adapted from NequIP https://github.com/mir-group/nequip
#if (TORCH_VERSION_MAJOR == 1 && TORCH_VERSION_MINOR <= 10)
#define DO_TORCH_FREEZE_HACK
// For the hack, need more headers:
#include <torch/csrc/jit/passes/freeze_module.h>
#include <torch/csrc/jit/passes/frozen_graph_optimizations.h>
#endif

using namespace std;

namespace PLMD {

class NeighborList;

namespace colvar {

namespace pytorch_gnn {

template <typename Main, typename Atoms>
auto getUsingNaturalUnits(Main& main, Atoms&, int)

  -> decltype(main.usingNaturalUnits()) {
  return main.usingNaturalUnits();
}

template <typename Main, typename Atoms>
auto getUsingNaturalUnits(Main&, Atoms& atoms, long)

  -> decltype(atoms.usingNaturalUnits()) {
  return atoms.usingNaturalUnits();
}

template <typename Main, typename Atoms>
auto getLengthUnit(Main& main, Atoms&, int)

  -> decltype(main.getUnits().getLength()) {
  return main.getUnits().getLength();
}

template <typename Main, typename Atoms>
auto getLengthUnit(Main&, Atoms& atoms, long)

  -> decltype(atoms.getUnits().getLength()) {
  return atoms.getUnits().getLength();
}

//+PLUMEDOC PYTORCH_GNN PYTORCH_GNN
/*

Load a Graph Neural Network (GNN) model compiled with TorchScript.



This module uses a fixed length unit of _Angstrom_. Thus, the GNN model read by

this module should be trained under the same unit convention. Besides, the

module constructs node attributes w.r.t to the atomic types. As a result, this

module require a PDB file which records names of _ALL_ atoms in the system

(the STRUCTURE keyword). Note that the atom names in this PDB file could

_ONLY_ be element symbols, e.g.:

\auxfile{plumed_topo.pdb}

ATOM      1  H   ACE A   1      15.100  12.940  29.390  1.00  0.00           H

ATOM      2  C   ACE A   1      14.970  13.860  29.960  1.00  0.00           C

ATOM      3  H   ACE A   1      15.720  13.820  30.760  1.00  0.00           H

ATOM      4  H   ACE A   1      13.980  13.920  30.410  1.00  0.00           H

ATOM      5  C   ACE A   1      15.300  15.070  29.100  1.00  0.00           C

\endauxfile



The module constructs graph edges between neighbors inside the selected atom

group, using the cutoff value recorded in the model file. By default, such an 

atom group is defined by the single `SYSTEM_SELECTION` keyword. In this case, the number 

of nodes in the input graph is fixed, while the number of edges 

changes according to the relative positions of the atoms. 

If the `ENVIRONMENT_SELECTION` parameter is given, the graph instead contains all atoms

in `SYSTEM_SELECTION` and all atoms in `ENVIRONMENT_SELECTION` that are within a cutoff

radius from _any_ atom in `SYSTEM_SELECTION`. Such a radius equals to the cutoff recorded in the

model file _plus_ the buffer size, also recorded in the model file. 

Thus, when `ENVIRONMENT_SELECTION` is given, the node number of the input 

graph can change dynamically during the simulation. Besides, when SUBSYSTEM_SELECTION is defined

the module will add long edges bewteen such a group. Cutoff radius of these

long edges will equal to the long_range_cutoff attribute recorded in the model file.



The outputs are exposed as `node-0`, `node-1`, etc.



Note that this function requires \ref installation-libtorch LibTorch C++ library.

Check the instructions in the \ref PYTORCH page to enable the module.

Specifically, we encourage the user to install the GPU-enabled version of

LibTorch, when dealing with large input graphs.



\par Examples

The following example instructs plumed to evaluate the GNN model using the atoms 1-10. The neighbor list for determining the edges will be updated every 100 steps.

\plumedfile

PYTORCH_GNN ...

  SYSTEM_SELECTION=1-10

  MODEL=model.ptc

  STRUCTURE=plumed_topo.pdb

  NL_STRIDE=100

  LABEL=gnn

... PYTORCH_GNN

\endplumedfile



The following example instructs plumed to do the same calculation as the above example, but will evaluate the model on CUDA using double precision, and add an OPES bias potential on the CV.

\plumedfile

PYTORCH_GNN ...

  SYSTEM_SELECTION=1-10

  MODEL=model.ptc

  STRUCTURE=plumed_topo.pdb

  NL_STRIDE=100

  CUDA

  FLOAT64

  LABEL=gnn

... PYTORCH_GNN



OPES_METAD ...

  LABEL=opes

  ARG=gnn.node-0

  FILE=KERNELS

  PACE=500

  TEMP=300

  BARRIER=35

... OPES_METAD

\endplumedfile





The following example instructs plumed to evaluate the GNN model using 

the atoms 1-10 as system atoms, and atoms 11-100 as the environment atoms.

In addition, long-range edges will be added between the subsystem atoms (1-10). 

The neighbor list for determining the edges will be updated every 2 steps.

\plumedfile

PYTORCH_GNN ...

  SYSTEM_SELECTION=1-10

  SUBSYSTEM_SELECTION=1-10

  ENVIRONMENT_SELECTION=11-100

  MODEL=model.ptc

  STRUCTURE=plumed_topo.pdb

  NL_STRIDE=2

  LABEL=gnn

... PYTORCH_GNN

\endplumedfile



*/
//+ENDPLUMEDOC


class PytorchGNN: public Colvar
{
  int n_out = 0;
  bool pbc = true;
  bool serial = false;
  bool firsttime = true;
  bool invalidate_list = true;
  bool bailout_fusion = false;
  double r_max = 0.0; // In PLUMED length unit
  double buffer = 0.0; // In PLUMED length unit
  double r_max_l = -1.0; // In PLUMED length unit
  std::string model_file_name;
  std::string structure_file_name;
  std::vector<int> system_node_types;
  std::vector<int> model_atomic_numbers;
  std::vector<AtomNumber> atom_list_a;
  std::vector<AtomNumber> atom_list_b;
  std::vector<AtomNumber> atom_list_sub_a;
  std::vector<int> atom_list_active; // local_ids
  std::vector<int> atom_list_active_subgroup; // local_ids
  std::unique_ptr<NeighborList> neighbor_list;
  torch::jit::script::Module model;
  torch::ScalarType torch_float_dtype = torch::kFloat32;
  torch::Device device = c10::Device(torch::kCPU);
  const std::array<std::string, 118> periodic_table = {
     "h", "he",
     "li", "be",                                                              "b",  "c",  "n",  "o",  "f", "ne",
     "na", "mg",                                                             "al", "si",  "p",  "s", "cl", "ar",
     "k",  "ca", "sc", "ti",  "v", "cr", "mn", "fe", "co", "ni", "cu", "zn", "ga", "ge", "as", "se", "br", "kr",
     "rb", "sr",  "y", "zr", "nb", "mo", "tc", "ru", "rh", "pd", "ag", "cd", "in", "sn", "sb", "te",  "i", "xe",
     "cs", "ba", "la", "ce", "pr", "nd", "pm", "sm", "eu", "gd", "tb", "dy", "ho", "er", "tm", "yb", "lu",
                       "hf", "ta",  "w", "re", "os", "ir", "pt", "au", "hg", "tl", "pb", "bi", "po", "at", "rn",
     "fr", "ra", "ac", "th", "pa",  "u", "np", "pu", "am", "cm", "bk", "cf", "es", "fm", "md", "no", "lr",
                       "rf", "db", "sg", "bh", "hs", "mt", "ds", "rg", "cn", "nh", "fl", "mc", "lv", "ts", "og"
  }; // TODO: add ghost atoms
  std::string model_summary(

    std::string model_name, torch::jit::Module module, int level_max, int level

  );
  int atomic_number_from_name(std::string name);
  bool groups_have_intersection(void);
  bool subgroup_is_in_group_a(void);
  void find_active_atoms(int n_threads);
  void find_active_subgroup_atoms(void);

public:
  explicit PytorchGNN(const ActionOptions&);
  ~PytorchGNN();
  static void registerKeywords(Keywords& keys);
  void calculate() override;
  void prepare() override;
}; // class PytorchGNN

PLUMED_REGISTER_ACTION(PytorchGNN, "PYTORCH_GNN")

void PytorchGNN::registerKeywords(Keywords& keys)

{
  Colvar::registerKeywords(keys);

  keys.add(
    "atoms",
    "SYSTEM_SELECTION",
    "First list of atoms (corresponding to the `system_selection` in mlcolvar)"
  );

  keys.add(
    "atoms",
    "ENVIRONMENT_SELECTION",
    "Second list of atoms (corresponding to the `environment_selection` in mlcolvar`)"
  );

  keys.add(
    "atoms",
    "SUBSYSTEM_SELECTION",
    "List of subsystem atoms (corresponding to the `subsystem_selection in mlcolvar`)"
  );

  keys.add(
    "compulsory",
    "MODEL",
    "Filename of the PyTorch compiled model"
  );

  keys.add(
    "compulsory",
    "STRUCTURE",
    "PDB file name that contains the whole simulated system, with correct atom names and orders"
  );

  keys.add(
    "optional",
    "NL_STRIDE",
    "The frequency with which we are updating the atoms in the neighbor list"
  );

  keys.addFlag(
    "CUDA",
    false,
    "Perform the calculation on CUDA"
  );

  keys.addFlag(
    "SERIAL",
    false,
    "Perform the calculation in serial - for debug purpose"
  );

  keys.addFlag(
    "BAILOUTFUSION",
    false,
    "Use a faster LibTorch fusion strategy (experimental), by default false. Set it to false for debugging and older version compatibility."
  );

  keys.addFlag(
    "FLOAT64",
    false,
    "Evaluate the model in double precision"
  );

  keys.addOutputComponent(
    "node",
    "default",
    "Model outputs"
  );

}

PytorchGNN::PytorchGNN(const ActionOptions& ao):
  PLUMED_COLVAR_INIT(ao)
{// print libtorch version
  std::stringstream ss;
  ss << TORCH_VERSION_MAJOR << "." \
     << TORCH_VERSION_MINOR << "." \
     << TORCH_VERSION_PATCH;
  std::string version;
  ss >> version; // extract into the string.
  std::string version_info = "  LibTorch version: " + version + "\n";
  log.printf(version_info.data());

  // parse input
  parseAtomList("SYSTEM_SELECTION", atom_list_a);
  parseAtomList("ENVIRONMENT_SELECTION", atom_list_b);
  parseAtomList("SUBSYSTEM_SELECTION", atom_list_sub_a);

  parse("MODEL", model_file_name);

  parse("STRUCTURE", structure_file_name);

  int neighbor_list_stride = 1;
  parse("NL_STRIDE", neighbor_list_stride);
  if (neighbor_list_stride <= 0)
    plumed_merror("NL_STRIDE should be positive!");

  bool use_cuda = false;
  bool required_cuda = false;
  parseFlag("CUDA", required_cuda);

  parseFlag("SERIAL", serial);
  if (required_cuda and serial)
    plumed_merror("Can not enable CUDA with SERIAL at the same time!");

  bool use_float64 = false;
  parseFlag("FLOAT64", use_float64);

  bool nopbc = !pbc;
  parseFlag("NOPBC", nopbc);
  pbc = !nopbc;

  parseFlag("BAILOUTFUSION", bailout_fusion);

  checkRead();

  // check groups
  if (atom_list_b.size() > 0) {
    if (groups_have_intersection())
      plumed_merror("SYSTEM_SELECTION can't intersect with ENVIRONMENT_SELECTION!");
    atom_list_active.resize(atom_list_a.size() + atom_list_b.size());
    atom_list_active.clear();
  } else {
    atom_list_active.resize(atom_list_a.size());
    atom_list_active.clear();
    find_active_atoms(1);
  }

  if (atom_list_sub_a.size() > 0)
    find_active_subgroup_atoms();

  if (atom_list_sub_a.size() > 0)
    if (!subgroup_is_in_group_a())
      plumed_merror("Not all atoms in SUBSYSTEM_SELECTION present in SYSTEM_SELECTION!");

  // check precision to be used
  if (use_float64)
    torch_float_dtype = torch::kFloat64;

  // check whether to use CUDA
  if (required_cuda && torch::cuda::is_available()) {
    device = c10::Device(torch::kCUDA);
    use_cuda = true;
  } else if (required_cuda) {
    use_cuda = false;
  }

  // check structure file
  PDB pdb;
  FILE *fp = fopen(structure_file_name.c_str(), "r");
  if (fp != NULL) {
    pdb.readFromFilepointer(
      fp,
      getUsingNaturalUnits(plumed, plumed.getAtoms(), 0),
      0.1 / getLengthUnit(plumed, plumed.getAtoms(), 0)
    );
    fclose(fp);
  } else {
    plumed_merror("Can not open PDB file: '" + structure_file_name + "'");
  }

  // deserialize the model from file
  try {
    model = torch::jit::load(model_file_name, device);
  } catch (const c10::Error& e) {
    plumed_merror(
      "Can't load model file: '" + model_file_name + "'. Reason: " + e.what()
    );
  }

  // disable parameter grads
  for (auto p: model.parameters())
    p.requires_grad_(false);

  // set up model precision
  model.to(torch_float_dtype);

  // summary
  std::string model_architecture = model_summary("CV", model, 3, 0);

  // get CV length
  if (!model.hasattr("n_out"))
    plumed_merror(
      "Can not find model attribute 'n_out'! This has to be set during the compilation of the model!"
    );
  if (model.hasattr("n_out"))
    n_out = model.attr("n_out").toTensor().item<int>();
  
  // get cutoff radius
  if (!model.hasattr("r_max") && !model.hasattr("cutoff") )
    plumed_merror(
      "Can not find model attribute: 'r_max' or 'cutoff'! One of these attributes has to be set during the compilation of the model!"
    );
  else if (model.hasattr("r_max") && model.hasattr("cutoff") )
    plumed_merror(
      "Both model attribute: 'r_max' and 'cutoff' are defined!"
    );

  // TODO: now, the `r_max` parameter in the model file is defined in unit of Angstrom.
  // We should warn the users about this default
  if (model.hasattr("cutoff"))
    r_max = model.attr("cutoff").toTensor().item<double>();
  else
    r_max = model.attr("r_max").toTensor().item<double>();
  r_max = r_max / getLengthUnit(plumed, plumed.getAtoms(), 0) * 0.1;

  // get buffer size
  if (model.hasattr("buffer")) {
    if (atom_list_b.size() == 0)
      plumed_merror(
        "Model attribute 'buffer' is defined but no ENVIRONMENT_SELECTION given!"
      );
    buffer = model.attr("buffer").toTensor().item<double>();
    } else
    buffer = 0.0;
  buffer = buffer / getLengthUnit(plumed, plumed.getAtoms(), 0) * 0.1; 

  // get long cutoff radius
  if (atom_list_sub_a.size() > 0) {
    if (!model.hasattr("long_range_cutoff")) {
      plumed_merror(
        "Can not find model attribute: 'long_range_cutoff'! Such an attributes is required for defining the subsystem group (SUBSYSTEM_SELECTION)!"
      );
    } else if (model.attr("long_range_cutoff").toTensor().item<double>() < 0) {
      plumed_merror(
        "Model attribute: 'long_range_cutoff' is negative! A positive long cutoff radius is required for defining the subsystem group (SUBSYSTEM_SELECTION)!"
      );
    } else {
      r_max_l = model.attr("long_range_cutoff").toTensor().item<double>();
      r_max_l = r_max_l / getLengthUnit(plumed, plumed.getAtoms(), 0) * 0.1;
    }
  } else if (
    model.hasattr("long_range_cutoff")
    && model.attr("long_range_cutoff").toTensor().item<double>() > 0
  ) {
    plumed_merror(
      "Found model attribute: 'long_range_cutoff'! Such an attributes requires defining the subsystem group (SUBSYSTEM_SELECTION)!"
    );
  }

  // get atomic numbers
  if (!model.hasattr("atomic_numbers"))
    plumed_merror(
      "Can't find model attribute: 'atomic_numbers'! This attribute has to be set during the compilation of the model!"
    );
  auto atomic_numbers = model.attr("atomic_numbers").toTensor();
  for (int64_t i = 0; i < atomic_numbers.size(0); i++)
    model_atomic_numbers.push_back(atomic_numbers[i].item<int64_t>());

// https://stackoverflow.com/questions/77102532/libtorch-performance-issue-when-using-multiple-gpus-in-multiple-threads
  if (bailout_fusion) {
    torch::jit::FusionStrategy bailout = {
      {torch::jit::FusionBehavior::STATIC,  0},
      {torch::jit::FusionBehavior::DYNAMIC, 0},
    };
    torch::jit::setFusionStrategy(bailout);
  }

  // optimize model
  model.eval();
#ifdef DO_TORCH_FREEZE_HACK
  // NOTE: do the hack
  // copied from the implementation of torch::jit::freeze,
  // except without the broken check
  // see https://github.com/pytorch/pytorch/blob/dfbd030854359207cb3040b864614affeace11ce/torch/csrc/jit/api/module.cpp
  bool optimize_numerics = true;  // the default
  // the {} is preserved_attrs
  auto out_mod = torch::jit::freeze_module(model, {});
  // see 1.11 bugfix in https://github.com/pytorch/pytorch/pull/71436
  auto graph = out_mod.get_method("forward").graph();
  OptimizeFrozenGraph(graph, optimize_numerics);
  model = out_mod;
#else
  // do it normally
  model = torch::jit::freeze(model);
#endif

  // optimize model for inference
  if (TORCH_VERSION_MAJOR == 2 || (TORCH_VERSION_MAJOR == 1 && TORCH_VERSION_MINOR >= 10)) {
    model = torch::jit::optimize_for_inference(model);
  }

  // send the model to device
  model.to(device);

  // create system atomic numbers
  std::vector<int> atom_is_required(pdb.getAtomNumbers().size());
  for (size_t i = 0; i < atom_list_a.size(); i++) {
    int index = atom_list_a[i].index();
    atom_is_required[index] = 1;
  }
  for (size_t i = 0; i < atom_list_b.size(); i++) {
    int index = atom_list_b[i].index();
    atom_is_required[index] = 1;
  }

  for (size_t i = 0; i < pdb.getAtomNumbers().size(); i++) {
    AtomNumber index = pdb.getAtomNumbers()[i];
    std::string name = pdb.getAtomName(index);
    int number = atomic_number_from_name(name);
    auto iter = std::find(
      model_atomic_numbers.begin(),
      model_atomic_numbers.end(),
      number
    );
    if (iter == model_atomic_numbers.end()) {
      if (atom_is_required[i])
        plumed_merror(
          "Element '" + name + "' does not present in model " + model_file_name
        );
      else
        system_node_types.push_back(-1);
    } else {
      int node_type = std::distance(model_atomic_numbers.begin(), iter);
      system_node_types.push_back(node_type);
    }
  }

  // create components
  for (int i = 0; i < n_out; i++) {
    string name_comp = "node-" + std::to_string(i);
    addComponentWithDerivatives(name_comp);
    componentIsNotPeriodic(name_comp);
  }

  // initialize the neighbor list
  if (atom_list_b.size() > 0)
    neighbor_list = Tools::make_unique<NeighborList>(
      atom_list_a,
      atom_list_b,
      serial,
      false,
      pbc,
      getPbc(),
      comm,
      r_max + buffer,
      neighbor_list_stride
    );
  else
    neighbor_list = Tools::make_unique<NeighborList>(
      atom_list_a,
      serial,
      pbc,
      getPbc(),
      comm,
      r_max + buffer,
      neighbor_list_stride
    );
  requestAtoms(neighbor_list->getFullAtomList());

  // print log
  std::string thename = getLabel();
  if(atom_list_b.size() > 0) {
    log.printf(
      "  Will build graphs using %u system and %u environment atoms\n",
      static_cast<unsigned>(atom_list_a.size()),
      static_cast<unsigned>(atom_list_b.size())
    );
    log.printf("  System atom list (SYSTEM_SELECTION):\n");
    for (unsigned int i = 0; i < atom_list_a.size(); i++) {
      if (((i + 1) % 10) == 0)
        log.printf("\n");
      log.printf("  %d", atom_list_a[i].serial());
    }
    log.printf("\n");
    log.printf("  Environment atom list (ENVIRONMENT_SELECTION):\n");
    for (unsigned int i = 0; i < atom_list_b.size(); i++) {
      if (((i + 1) % 10) == 0)
        log.printf("\n");
      log.printf("  %d", atom_list_b[i].serial());
    }
    log.printf("\n");
  } else {
    log.printf(
      "  Will build graphs using %u atoms\n",
      static_cast<unsigned>(atom_list_a.size())
    );
    log.printf("  Atom list:\n");
    for (unsigned int i = 0; i < atom_list_a.size(); i++) {
      if (((i + 1) % 10) == 0)
        log.printf("\n");
      log.printf("  %d", atom_list_a[i].serial());
    }
    log.printf("\n");
  }
  if (atom_list_sub_a.size() > 0) {
    log.printf(
      "  Will add long-range edges between %u atoms\n",
      static_cast<unsigned>(atom_list_sub_a.size())
    );
    log.printf("  Subsystem atom list:\n");
    for (unsigned int i = 0; i < atom_list_sub_a.size(); i++) {
      if (((i + 1) % 10) == 0)
        log.printf("\n");
      log.printf("  %d", atom_list_sub_a[i].serial());
    }
    log.printf("\n");
  }
  log << "  Model atomic numbers: " << model_atomic_numbers;
  log.printf("\n");
  log.printf("  Boundary conditions: ");
  if (pbc)
    log.printf("periodic\n");
  else
    log.printf("non-periodic\n");
  log.printf("  Neighbor List update stride: %d\n", neighbor_list_stride);
  log.printf("  Graph cutoff radius: %f (PLUMED length unit)\n", r_max);
  if(atom_list_b.size() > 0)
    log.printf("  Environment buffer size: %f (PLUMED length unit)\n", buffer);
  if (atom_list_sub_a.size() > 0)
    log.printf("  Subsystem long-range cutoff radius: %f (PLUMED length unit)\n", r_max_l);
  log.printf("  Number of outputs: %d \n", n_out);
  log.printf("  Will run on device: ");
  if (use_cuda)
    log.printf("CUDA\n");
  else if (required_cuda)
    log.printf("CPU (CUDA device not found/LibTorch does not support CUDA)\n");
  else
    log.printf("CPU (as required)\n");
  log << "  Model architecture: \n";
  log << model_architecture;
  log.printf("  Bibliography: ");
  log << plumed.cite("Bonati, Trizio, Rizzi and Parrinello, J. Chem. Phys. 159, 014801 (2023)");
  log << plumed.cite("Zhang et al., J. Chem. Theory Comput. 20, 24, 10787–10797 (2024)");
  log.printf("\n");
}

PytorchGNN::~PytorchGNN()
{
  return;
}

void PytorchGNN::prepare()

{
  if (neighbor_list->getStride() > 0) {
    if (firsttime || ((getStep() % neighbor_list->getStride()) == 0)) {
      requestAtoms(neighbor_list->getFullAtomList());
      invalidate_list = true;
      firsttime = false;
    } else {
      requestAtoms(neighbor_list->getReducedAtomList());
      invalidate_list = false;
      if (getExchangeStep())
        plumed_merror(
          "Neighbor lists should be updated on exchange steps - choose a NL_STRIDE which divides the exchange stride!"
        );
    }
    if (getExchangeStep())
      firsttime = true;
  }
}


void PytorchGNN::calculate()

{
  // get some common data
  auto pbc_tools = getPbc();
  int n_atoms = getNumberOfAtoms();
  std::vector<PLMD::Vector> x_local = getPositions();

  // threads
  int n_threads = OpenMP::getNumThreads();
  if (!serial)
    n_threads = std::min(n_threads, n_atoms);
  else
    n_threads = 1;

  // perform the size check
  if (system_node_types.size() != (size_t)plumed.getAtoms().getNatoms()) // TODO: this is now deprecated from plumed 2.10
    plumed_merror(
      "Structure file '" +
      structure_file_name +
      "' has different number of atoms with the simulated system!"
    );

  // update the neighbor list and number of atoms
  if (neighbor_list->getStride() > 0 && invalidate_list)
    neighbor_list->update(x_local);
  if (atom_list_b.size() > 0)
    find_active_atoms(n_threads);
  n_atoms = (int)atom_list_active.size();
  n_threads = std::min(n_threads, n_atoms);

  // get the unit
  double to_ang = 10 * getLengthUnit(plumed, plumed.getAtoms(), 0);

  // get the positions
  // TODO: now, the positions used by the model file is in unit of Angstrom.
  // We should warn the users about this default
  std::vector<float> positions_vector(n_atoms * 3);
  #pragma omp parallel for num_threads(n_threads)
  for (int i = 0; i < n_atoms; i++) {
    int index = atom_list_active[i]; 
    positions_vector[i * 3 + 0] = x_local[index][0] * to_ang;
    positions_vector[i * 3 + 1] = x_local[index][1] * to_ang;
    positions_vector[i * 3 + 2] = x_local[index][2] * to_ang;
  }  

  torch::Tensor positions = torch::from_blob(  
    positions_vector.data(),
    n_atoms * 3,
    torch::TensorOptions().dtype(torch::kFloat32)
  );
  positions = positions.to(device).to(torch_float_dtype);
  positions = positions.reshape({n_atoms, 3});

  // cell
  // TODO: now, the box data used by the model file is in unit of Angstrom.
  // We should warn the users about this default
  PLMD::Tensor box = getBox();
  std::vector<float> cell_vector(9);
  for (int i = 0; i < 3; i++) {
    for (int j = 0; j < 3; j++)
      cell_vector[i * 3 + j] = box[i][j] * to_ang;
  }
  torch::Tensor cell = torch::from_blob(
    cell_vector.data(),
    9,
    torch::TensorOptions().dtype(torch::kFloat32)
  );
  cell = cell.to(device).to(torch_float_dtype);
  cell = cell.reshape({3, 3});

  // build node attributes
  // TODO: now, the node attributes are in MACE's format.
  // We should try to give more options, or warn the users about this default
  int n_node_feats = (int)model_atomic_numbers.size();
  std::vector<float> node_attrs_vector(n_node_feats * n_atoms);
  #pragma omp parallel for num_threads(n_threads)
  for (int i = 0; i < n_atoms; i++) {
    int index = atom_list_active[i];
    int node_type = system_node_types[getAbsoluteIndex(index).index()];
    node_attrs_vector[i * n_node_feats + node_type] = 1.0;
  }
  torch::Tensor node_attrs = torch::from_blob(
    node_attrs_vector.data(),
    n_node_feats * n_atoms,
    torch::TensorOptions().dtype(torch::kFloat32)
  );
  node_attrs = node_attrs.to(device).to(torch_float_dtype);
  node_attrs = node_attrs.reshape({n_atoms, n_node_feats});

  // build edges
  int n_edges = 0;
  int n_edges_l = 0;
  torch::Tensor edge_index;

  if (atom_list_b.size() > 0) {
    n_edges = n_atoms * (n_atoms - 1);
    std::vector<float> distance_vector(n_edges);
    std::vector<std::vector<int64_t>> edge_index_vector;
    edge_index_vector.resize(2, std::vector<int64_t>(n_edges));

    #pragma omp parallel for num_threads(n_threads)
    for (int i = 0; i < n_atoms; i++) {
      int count = 0;
      for (int j = 0; j < n_atoms; j++) {
        if (i != j) {
          edge_index_vector[0][i * (n_atoms - 1) + count] = i;
          edge_index_vector[1][i * (n_atoms - 1) + count] = j;
          count++;
        }
      }
    }

    #pragma omp parallel for num_threads(n_threads)
    for (int i = 0; i < n_edges; i++) {
      distance_vector[i] = pbc_tools.distance(
        true,
        x_local[atom_list_active[edge_index_vector[0][i]]],
        x_local[atom_list_active[edge_index_vector[1][i]]]
      );
    }

    torch::Tensor distances = torch::from_blob(
      distance_vector.data(),
      n_edges,
      torch::TensorOptions().dtype(torch::kFloat32)
    );
    torch::Tensor senders = torch::from_blob(
      edge_index_vector[0].data(),
      n_edges,
      torch::TensorOptions().dtype(torch::kInt64)
    );
    torch::Tensor receivers = torch::from_blob(
      edge_index_vector[1].data(),
      n_edges,
      torch::TensorOptions().dtype(torch::kInt64)
    );

    const torch::Tensor mask = distances <= r_max;
    senders = senders.index({mask});
    receivers = receivers.index({mask});
    edge_index = torch::vstack({senders, receivers});
    n_edges = (int)edge_index.size(1);
  } else {
    n_edges = (int)neighbor_list->size() * 2;
    int n_pairs = (int)neighbor_list->size();
    std::vector<std::vector<int64_t>> edge_index_vector;
    edge_index_vector.resize(2, std::vector<int64_t>(n_edges));

    #pragma omp parallel for num_threads(n_threads)
    for (int i = 0; i < n_pairs; i++) {
        auto pair = neighbor_list->getClosePair(i);
        edge_index_vector[0][i] = pair.first;
        edge_index_vector[1][i] = pair.second;
        edge_index_vector[0][n_pairs + i] = pair.second;
        edge_index_vector[1][n_pairs + i] = pair.first;
    }

    torch::Tensor senders = torch::from_blob(
      edge_index_vector[0].data(),
      n_edges,
      torch::TensorOptions().dtype(torch::kInt64)
    );
    torch::Tensor receivers = torch::from_blob(
      edge_index_vector[1].data(),
      n_edges,
      torch::TensorOptions().dtype(torch::kInt64)
    );
    edge_index = torch::vstack({senders, receivers});
  }

  if (atom_list_sub_a.size() > 0) {
    torch::Tensor edge_index_l;
    int n_atoms_l = atom_list_sub_a.size();
    n_edges_l = n_atoms_l * (n_atoms_l - 1);
    std::vector<float> distance_vector_l(n_edges_l);
    std::vector<std::vector<int64_t>> edge_index_vector_l;
    edge_index_vector_l.resize(2, std::vector<int64_t>(n_edges_l));

    #pragma omp parallel for num_threads(n_threads)
    for (int i = 0; i < n_atoms_l; i++) {
      int count = 0;
      for (int j = 0; j < n_atoms_l; j++) {
        if (i != j) {
          // NOTE: build edge index using the partial local index list
          edge_index_vector_l[0][
            i * (n_atoms_l - 1) + count
          ] = atom_list_active_subgroup[i];
          edge_index_vector_l[1][
            i * (n_atoms_l - 1) + count
          ] = atom_list_active_subgroup[j];
          count++;
        }
      }
    }

    #pragma omp parallel for num_threads(n_threads)
    for (int i = 0; i < n_edges_l; i++) {
      distance_vector_l[i] = pbc_tools.distance(
        pbc,
        x_local[edge_index_vector_l[0][i]],
        x_local[edge_index_vector_l[1][i]]
      );
    }

    torch::Tensor distances_l = torch::from_blob(
      distance_vector_l.data(),
      n_edges_l,
      torch::TensorOptions().dtype(torch::kFloat32)
    );
    const torch::Tensor mask_l = distances_l <= r_max_l;

    if (mask_l.size(0) > 0) {
      torch::Tensor senders_l = torch::from_blob(
        edge_index_vector_l[0].data(),
        n_edges_l,
        torch::TensorOptions().dtype(torch::kInt64)
      );
      torch::Tensor receivers_l = torch::from_blob(
        edge_index_vector_l[1].data(),
        n_edges_l,
        torch::TensorOptions().dtype(torch::kInt64)
      );
      senders_l = senders_l.index({mask_l});
      receivers_l = receivers_l.index({mask_l});
      edge_index_l = torch::vstack({senders_l, receivers_l});
      n_edges_l = (int)edge_index_l.size(1);

      edge_index = torch::hstack({edge_index, edge_index_l});
      n_edges = n_edges + n_edges_l;
    }
  }

  edge_index = edge_index.to(device);

  // edge shifts
  torch::Tensor shifts;
  torch::Tensor unit_shifts;

  if (pbc) {
    if (pbc_tools.isOrthorombic()) {
      auto deltas = (
        torch::index_select(positions, 0, edge_index[0])
        - torch::index_select(positions, 0, edge_index[1])
      );
      unit_shifts = torch::round(deltas / torch::diagonal(cell, 0));
      shifts = unit_shifts * torch::diagonal(cell, 0);
    } else {
      auto cell_inv = torch::linalg_pinv(cell.transpose(1, 0));
      auto positions_s = torch::matmul(
        cell_inv, positions.transpose(1, 0)
      );
      auto deltas = (
        torch::index_select(positions_s, 1, edge_index[0])
        - torch::index_select(positions_s, 1, edge_index[1])
      );
      unit_shifts = torch::round(deltas);
      shifts = torch::matmul(
        cell.transpose(1, 0), unit_shifts
      ).transpose(1, 0);
    }
  } else {
    shifts = torch::zeros({n_edges, 3}, torch_float_dtype);
    unit_shifts = torch::zeros({n_edges, 3}, torch_float_dtype);
  }

  shifts = shifts.to(device);
  unit_shifts = unit_shifts.to(device);

  // other things
  // TODO: some of these things are required by MACE. We should disable
  // some of them when not using MACE, maybe by distinguishing the MACE model.
  auto batch = torch::zeros({n_atoms}, torch::dtype(torch::kInt64));
  auto ptr = torch::empty({2}, torch::dtype(torch::kInt64));
  auto weight = torch::empty({1}, torch_float_dtype);
  ptr[0] = 0;
  ptr[1] = n_atoms;
  weight[0] = 1.0;

  // load data to device
  // TODO: some of these things are required by MACE. We should disable
  // some of them when not using MACE, maybe by distinguishing the MACE model.
  batch = batch.to(device);
  ptr = ptr.to(device);
  weight = weight.to(device);

  // pack the input, call the model
  // TODO: some of these things are required by MACE. We should disable
  // some of them when not using MACE, maybe by distinguishing the MACE model.
  c10::Dict<std::string, torch::Tensor> input;
  input.insert("batch", batch);
  input.insert("cell", cell);
  input.insert("edge_index", edge_index);
  input.insert("node_attrs", node_attrs);
  positions.set_requires_grad(true);
  input.insert("positions", positions);
  input.insert("ptr", ptr);
  input.insert("weight", weight);
  input.insert("shifts", shifts);
  input.insert("unit_shifts", unit_shifts);

  // create and insert system mask and n_system
  torch::Tensor system_masks = torch::zeros({n_atoms, 1}, torch::dtype(torch::kBool));
  for (size_t i = 0; i < atom_list_a.size(); i++)
    system_masks[i] = true;
  system_masks = system_masks.to(device);
  input.insert("system_masks", system_masks);

  torch::Tensor n_system = torch::ones({1, 1}, torch::dtype(torch::kInt64));
  n_system = n_system.to(device);
  n_system[0][0] = (int64_t)atom_list_a.size();
  input.insert("n_system", n_system);

  if (atom_list_sub_a.size() > 0){
    auto edge_masks_lr = torch::vstack({
      torch::zeros({n_edges - n_edges_l, 1}, torch::dtype(torch::kBool)),
      torch::ones({n_edges_l, 1}, torch::dtype(torch::kBool)),
    });
    edge_masks_lr = edge_masks_lr.to(device);
    input.insert("edge_masks_lr", edge_masks_lr);
  }
  
  // TODO: figure out how to enable virials. Maybe we could port MACE's python
  // code to our python module.
  auto output = model.forward({input}).toTensor();

  // helper variables
  std::vector<PLMD::Vector> derivatives(n_atoms);
  auto grad_output = torch::ones({1}).expand({1, 1}).to(device);

  // Here we simply compute the output and its derivatives
  for (int i = 0; i < n_out; i++) {
    // set CV values
    string name_comp = "node-" + std::to_string(i);
    getPntrToComponent(name_comp)->set(output[0][i].cpu().item<double>());
    // set derivatives
    auto gradients = torch::autograd::grad(
      {output.slice(1, i, (i + 1))},
      {positions},
      {grad_output}, // grad_outputs
      true,          // retain_graph
      false          // create_graph
    )[0].cpu();
    #pragma omp parallel for num_threads(n_threads)
    for (int j = 0; j < n_atoms; j++) {
      derivatives[j][0] = gradients[j][0].item<double>() * to_ang;
      derivatives[j][1] = gradients[j][1].item<double>() * to_ang;
      derivatives[j][2] = gradients[j][2].item<double>() * to_ang;
    }
    #pragma omp parallel for num_threads(n_threads)
    for (int j = 0; j < n_atoms; j++) {
      int index = atom_list_active[j];
      setAtomsDerivatives(
        getPntrToComponent(name_comp), index, derivatives[j]
      );
  
    }
  }
}


int PytorchGNN::atomic_number_from_name(std::string name)

{
  std::transform(
    name.begin(),
    name.end(),
    name.begin(),
    [](unsigned char c){return std::tolower(c);}
  );
  auto iter = std::find(periodic_table.begin(), periodic_table.end(), name);
  if (iter == periodic_table.end())
    plumed_merror(
      "Can not find element name '" + name + "' from the periodic table!"
    );
  return std::distance(periodic_table.begin(), iter) + 1;
}

std::string PytorchGNN::model_summary(

    std::string model_name, torch::jit::Module module, int level_max, int level

) {
  std::stringstream ss;

  std::string model_type = module.type()->name()->name();
  ss << "  (" << model_name << "): " << model_type;

  if (module.named_children().size() != 0) {
    if (level <= level_max) {
      ss << " {\n";
      for (const torch::jit::NameModule& s : module.named_children())
          ss << torch::jit::jit_log_prefix(
              "  ",
              model_summary(s.name, s.value, level_max, level + 1)
          );
      ss << "  }\n";
    } else {
      ss << " { ... }";
    }
  } else {
    ss << "\n";
  }

  return ss.str();
}

bool PytorchGNN::groups_have_intersection(void) {
  std::vector<AtomNumber> intersections;
  std::vector<AtomNumber> atom_list_a_copy(atom_list_a);
  std::vector<AtomNumber> atom_list_b_copy(atom_list_b);

  std::sort(atom_list_a_copy.begin(), atom_list_a_copy.end());
  std::sort(atom_list_b_copy.begin(), atom_list_b_copy.end());

  std::set_intersection(
    atom_list_a_copy.begin(),
    atom_list_a_copy.end(),
    atom_list_b_copy.begin(),
    atom_list_b_copy.end(),
    back_inserter(intersections)
  );

  return intersections.size() > 0;
}

bool PytorchGNN::subgroup_is_in_group_a(void) {
  std::vector<AtomNumber> atom_list_a_copy(atom_list_a);
  std::vector<AtomNumber> atom_list_sub_a_copy(atom_list_sub_a);
  for (auto atom_elt: atom_list_sub_a_copy)
    if (
      std::find(atom_list_a_copy.begin(),
      atom_list_a_copy.end(), atom_elt) == atom_list_a_copy.end()
    )
      return false;
  return true;
}

void PytorchGNN::find_active_atoms(int n_threads) {
  if (atom_list_b.size() > 0) {
    atom_list_active.clear();
    std::vector<int> neighbors(neighbor_list->size());

    #pragma omp parallel for num_threads(n_threads)
    for (size_t i = 0; i < neighbor_list->size(); i++)
      neighbors[i] = neighbor_list->getClosePair(i).second;

    // TODO: make this faster
    std::unordered_set<int> neighbors_set;
    for (int i : neighbors)
        neighbors_set.insert(i);
    neighbors.assign(neighbors_set.begin(), neighbors_set.end());

    for (size_t i = 0; i < atom_list_a.size(); i++)
      atom_list_active.push_back(i);
    for (size_t i = 0; i < neighbors.size(); i++)
      atom_list_active.push_back(neighbors[i]);
  } else if (atom_list_active.size() == 0) {
    atom_list_active.clear();

    for (size_t i = 0; i < atom_list_a.size(); i++)
      atom_list_active.push_back(i);
  }
}

void PytorchGNN::find_active_subgroup_atoms(void) {
  atom_list_active_subgroup.clear();
  // NOTE: since system atoms (atom_list_a) always appear at the head of the
  // local indices, we simply find subsystem indices in atom_list_a.
  for (auto atom_sub: atom_list_sub_a) {
    int index = (int)std::distance(
      atom_list_a.begin(),
      find(atom_list_a.begin(), atom_list_a.end(), atom_sub)
    );
    atom_list_active_subgroup.push_back(index);
  }
}

} // pytorch_gnn

} // colvar

} // PLMD

#endif // PLUMED_HAS_LIBTORCH