Add a constructor to TorchForce that takes a torch::jit::Module (#97)

* Add version number as a member to TorchForceProxy

* Encode the model file contents when serializing TorchForce

* Add tests for new TorchForce serialization

* Fix test not finding Python executable

* Format include directives correctly

* Hardcode TorchForceProxy version number

* Fix formatting issues

* Move Python serialization test to the correct place

* Make function encodeFromFileName static

* Update serialization python test to correctly remove temporary files after executing

* Use the base64 encoding capabilities of openssl to serialize model file

* Update TorchForce serializer

* Add a constructor to TorchForce that takes a torch::jit::Module.
 TorchForce(string fileName) is implemented by delegating to the new
 constructor.
 Update serialization test accordingly to compare the module file name
 and the module itself.

* Remove unnecessary include

* Change i_file to file in TorchForce constructor

* Add swig typemaps to new TorchForce constructor

* Add setup.py as a dependency for the PythonInstall CMake rule

* Fix swig out typemap for torch::jit::Module
 Now it is possible to call getModule() on a TorchForce object from
 Python, which will return a module of the same type as, for instance, torch.jit.load()

* Remove commented line in CMakeLists.txt

* Remove unnecessary dependency in setup.py

* Add more tests for new constructor

* Add some comments for the new constructor

* Updates to TorchForce serialization

* Use hex encoding instead of base64 for serialization.
SSL no longer a direct dependency.

* Remove unnecessary header

* Update Python serialization test

* Minor changes

* Improve temporary path handling in python serialization tests

* More informative exception when failing to serialize TorchForce

* Remove unnecessary check in TorchForce serialization

* Changes to C++ serialization tests

* Changes to C++ serialization tests

CMakeLists.txt

@@ -13,7 +13,6 @@ LINK_DIRECTORIES("${OPENMM_DIR}/lib" "${OPENMM_DIR}/lib/plugins")
 SET(PYTORCH_DIR "" CACHE PATH "Where the PyTorch C++ API is installed")
 SET(CMAKE_PREFIX_PATH "${PYTORCH_DIR}")
 FIND_PACKAGE(Torch REQUIRED)
-#LINK_DIRECTORIES("${TENSORFLOW_DIR}/lib")
 
 # Specify the C++ version we are building for.
 SET (CMAKE_CXX_STANDARD 14)

openmmapi/include/TorchForce.h

@@ -35,13 +35,18 @@
 #include "openmm/Context.h"
 #include "openmm/Force.h"
 #include <string>
+#include <torch/torch.h>
 #include "internal/windowsExportTorch.h"
 
 namespace TorchPlugin {
 
 /**
  * This class implements forces that are defined by user-supplied neural networks.
- * It uses the PyTorch library to perform the computations. */
+ * It uses the PyTorch library to perform the computations.
+ * The PyTorch module can either be  passed directly as an argument to
+ * the  constructor  or loaded  from  a  file.   In either  case,  the
+ * constructor makes a copy of the  module in memory. Later changes to
+ * the original module or to the file do not affect it.*/
 
 class OPENMM_EXPORT_NN TorchForce : public OpenMM::Force {
 public:
@@ -52,10 +57,24 @@ class OPENMM_EXPORT_NN TorchForce : public OpenMM::Force {
      * @param file   the path to the file containing the network
      */
     TorchForce(const std::string& file);
+    /**
+     * Create a TorchForce.  The network is defined by a PyTorch ScriptModule
+     * Note that this constructor makes a copy of the provided module.
+     * Any changes to the module  after calling this constructor will be ignored by TorchForce.
+     *
+     * @param module   an instance of the torch module
+     */
+    TorchForce(const torch::jit::Module &module);
     /**
      * Get the path to the file containing the network.
+     * If the TorchForce instance was constructed with a module, instead of a filename,
+     * this function returns an empty string.
      */
     const std::string& getFile() const;
+    /**
+     * Get the torch module currently in use.
+     */
+    const torch::jit::Module & getModule() const;
     /**
      * Set whether this force makes use of periodic boundary conditions.  If this is set
      * to true, the network must take a 3x3 tensor as its second input, which
@@ -128,6 +147,7 @@ class OPENMM_EXPORT_NN TorchForce : public OpenMM::Force {
     std::string file;
     bool usePeriodic, outputsForces;
     std::vector<GlobalParameterInfo> globalParameters;
+    torch::jit::Module module;
 };
 
 /**

openmmapi/include/internal/TorchForceImpl.h

@@ -65,7 +65,6 @@ class OPENMM_EXPORT_NN TorchForceImpl : public OpenMM::ForceImpl {
 private:
     const TorchForce& owner;
     OpenMM::Kernel kernel;
-    torch::jit::script::Module module;
 };
 
 } // namespace TorchPlugin

openmmapi/src/TorchForce.cpp

@@ -34,18 +34,28 @@
 #include "openmm/OpenMMException.h"
 #include "openmm/internal/AssertionUtilities.h"
 #include <fstream>
+#include <torch/torch.h>
+#include <torch/csrc/jit/serialization/import.h>
 
 using namespace TorchPlugin;
 using namespace OpenMM;
 using namespace std;
 
-TorchForce::TorchForce(const std::string& file) : file(file), usePeriodic(false), outputsForces(false) {
+TorchForce::TorchForce(const torch::jit::Module& module) : file(), usePeriodic(false), outputsForces(false), module(module) {
+}
+
+TorchForce::TorchForce(const std::string& file) : TorchForce(torch::jit::load(file)) {
+    this->file = file;
 }
 
 const string& TorchForce::getFile() const {
     return file;
 }
 
+const torch::jit::Module& TorchForce::getModule() const {
+    return this->module;
+}
+
 ForceImpl* TorchForce::createImpl() const {
     return new TorchForceImpl(*this);
 }

openmmapi/src/TorchForceImpl.cpp

@@ -46,12 +46,8 @@ TorchForceImpl::~TorchForceImpl() {
 }
 
 void TorchForceImpl::initialize(ContextImpl& context) {
-    // Load the module from the file.
-
-    module = torch::jit::load(owner.getFile());
-
+    auto module = owner.getModule();
     // Create the kernel.
-
     kernel = context.getPlatform().createKernel(CalcTorchForceKernel::Name(), context);
     kernel.getAs<CalcTorchForceKernel>().initialize(context.getSystem(), owner, module);
 }

platforms/cuda/tests/TestCudaTorchForce.cpp

@@ -42,6 +42,8 @@
 #include "sfmt/SFMT.h"
 #include <cmath>
 #include <iostream>
+#include <torch/torch.h>
+#include <torch/csrc/jit/serialization/import.h>
 #include <vector>
 
 using namespace TorchPlugin;
@@ -52,7 +54,7 @@ extern "C" OPENMM_EXPORT void registerTorchCudaKernelFactories();
 
 void testForce(bool outputsForces) {
     // Create a random cloud of particles.
-    
+
     const int numParticles = 10;
     System system;
     vector<Vec3> positions(numParticles);
@@ -62,20 +64,21 @@ void testForce(bool outputsForces) {
         system.addParticle(1.0);
         positions[i] = Vec3(genrand_real2(sfmt), genrand_real2(sfmt), genrand_real2(sfmt))*10;
     }
-    TorchForce* force = new TorchForce(outputsForces ? "tests/forces.pt" : "tests/central.pt");
+    auto model = torch::jit::load(outputsForces ? "tests/forces.pt" : "tests/central.pt");
+    TorchForce* force = new TorchForce(model);
     force->setOutputsForces(outputsForces);
     system.addForce(force);
-    
+
     // Compute the forces and energy.
 
     VerletIntegrator integ(1.0);
     Platform& platform = Platform::getPlatformByName("CUDA");
     Context context(system, integ, platform);
     context.setPositions(positions);
     State state = context.getState(State::Energy | State::Forces);
-    
+
     // See if the energy is correct.  The network defines a potential of the form E(r) = |r|^2
-    
+
     double expectedEnergy = 0;
     for (int i = 0; i < numParticles; i++) {
         Vec3 pos = positions[i];

python/CMakeLists.txt

@@ -2,21 +2,25 @@ set(WRAP_FILE TorchPluginWrapper.cpp)
 set(MODULE_NAME openmmtorch)
 
 # Execute SWIG to generate source code for the Python module.
+foreach(dir ${TORCH_INCLUDE_DIRS})
+    set(torchincs "${torchincs}" "-I${dir}")
+endforeach()
 
 add_custom_command(
     OUTPUT "${WRAP_FILE}"
     COMMAND "${SWIG_EXECUTABLE}"
         -python -c++
         -o "${WRAP_FILE}"
         "-I${OPENMM_DIR}/include"
+	${torchincs}
         "${CMAKE_CURRENT_SOURCE_DIR}/openmmtorch.i"
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/openmmtorch.i"
     WORKING_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}"
 )
 
 # Compile the Python module.
 
-add_custom_target(PythonInstall DEPENDS "${WRAP_FILE}")
+add_custom_target(PythonInstall DEPENDS "${WRAP_FILE}"  "${CMAKE_CURRENT_SOURCE_DIR}/setup.py")
 set(NN_PLUGIN_HEADER_DIR "${CMAKE_SOURCE_DIR}/openmmapi/include")
 set(NN_PLUGIN_LIBRARY_DIR "${CMAKE_BINARY_DIR}")
 configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py ${CMAKE_CURRENT_BINARY_DIR}/setup.py)

python/openmmtorch.i

@@ -12,6 +12,7 @@
 #include "OpenMMDrude.h"
 #include "openmm/RPMDIntegrator.h"
 #include "openmm/RPMDMonteCarloBarostat.h"
+#include <torch/csrc/jit/python/module_python.h>
 %}
 
 /*
@@ -26,12 +27,33 @@
     }
 }
 
+%typemap(in) const torch::jit::Module&(torch::jit::Module module) {
+    py::object o = py::reinterpret_borrow<py::object>($input);
+    module = torch::jit::as_module(o).value();
+    $1 = &module;
+}
+
+%typemap(out) const torch::jit::Module& {
+    auto fileName = std::tmpnam(nullptr);
+    $1->save(fileName);
+    $result = py::module::import("torch.jit").attr("load")(fileName).release().ptr();
+    //This typemap assumes that torch does not require the file to exist after construction
+    std::remove(fileName);
+}
+
+%typecheck(SWIG_TYPECHECK_POINTER) const torch::jit::Module& {
+    py::object o = py::reinterpret_borrow<py::object>($input);
+    $1 = torch::jit::as_module(o).has_value() ? 1 : 0;
+}
+
 namespace TorchPlugin {
 
 class TorchForce : public OpenMM::Force {
 public:
     TorchForce(const std::string& file);
+    TorchForce(const torch::jit::Module& module);
     const std::string& getFile() const;
+    const torch::jit::Module& getModule() const;
     void setUsesPeriodicBoundaryConditions(bool periodic);
     bool usesPeriodicBoundaryConditions() const;
     void setOutputsForces(bool);

python/setup.py

@@ -5,12 +5,13 @@
 import platform
 
 openmm_dir = '@OPENMM_DIR@'
+torch_include_dirs = '@TORCH_INCLUDE_DIRS@'.split(';')
 nn_plugin_header_dir = '@NN_PLUGIN_HEADER_DIR@'
 nn_plugin_library_dir = '@NN_PLUGIN_LIBRARY_DIR@'
 torch_dir, _ = os.path.split('@TORCH_LIBRARY@')
 
 # setup extra compile and link arguments on Mac
-extra_compile_args = ['-std=c++11']
+extra_compile_args = ['-std=c++14']
 extra_link_args = []
 
 if platform.system() == 'Darwin':
@@ -20,9 +21,9 @@
 extension = Extension(name='_openmmtorch',
                       sources=['TorchPluginWrapper.cpp'],
                       libraries=['OpenMM', 'OpenMMTorch'],
-                      include_dirs=[os.path.join(openmm_dir, 'include'), nn_plugin_header_dir],
+                      include_dirs=[os.path.join(openmm_dir, 'include'), nn_plugin_header_dir] + torch_include_dirs,
                       library_dirs=[os.path.join(openmm_dir, 'lib'), nn_plugin_library_dir],
-                      runtime_library_dirs=[os.path.join(openmm_dir, 'lib'), torch_dir],
+                      runtime_library_dirs=[os.path.join(openmm_dir, 'lib')],
                       extra_compile_args=extra_compile_args,
                       extra_link_args=extra_link_args
                      )
@@ -32,4 +33,3 @@
       py_modules=['openmmtorch'],
       ext_modules=[extension],
      )
-

python/tests/TestSerializeTorchForce.py

@@ -0,0 +1,48 @@
+import torch
+import shutil
+import pytest
+from openmm import XmlSerializer, OpenMMException
+from openmmtorch import TorchForce
+import os
+import tempfile
+
+class ForceModule(torch.nn.Module):
+    """A simple module that can be serialized"""
+    def forward(self, positions):
+        return torch.sum(positions**2)
+
+
+class ForceModule2(torch.nn.Module):
+    """A dummy module distict from ForceModule"""
+    def forward(self, positions):
+        return torch.sum(positions**3)
+
+
+def createAndSerialize(model_filename, serialized_filename):
+    module = torch.jit.script(ForceModule())
+    module.save(model_filename)
+    torch_force = TorchForce(model_filename)
+    stored = XmlSerializer.serialize(torch_force)
+    with open(serialized_filename, 'w') as f:
+        f.write(stored)
+
+def readXML(filename):
+    with open(filename, 'r') as f:
+        fileContents = f.read()
+    return fileContents
+
+def deserialize(filename):
+    other_force = XmlSerializer.deserialize(readXML(filename))
+
+def test_serialize():
+    with tempfile.TemporaryDirectory() as tempdir:
+        model_filename = os.path.join(tempdir, 'model.pt')
+        serialized_filename = os.path.join(tempdir, 'stored.xml')
+        createAndSerialize(model_filename, serialized_filename)
+
+def test_deserialize():
+    with tempfile.TemporaryDirectory() as tempdir:
+        model_filename = os.path.join(tempdir, 'model.pt')
+        serialized_filename = os.path.join(tempdir, 'stored.xml')
+        createAndSerialize(model_filename, serialized_filename)
+        deserialize(serialized_filename)

python/tests/TestTorchForce.py

@@ -6,12 +6,23 @@
 import torch as pt
 from tempfile import NamedTemporaryFile
 
-@pytest.mark.parametrize('model_file, output_forces,',
-                        [('../../tests/central.pt', False),
-                         ('../../tests/forces.pt', True)])
+@pytest.mark.parametrize('model_file,',
+                        ['../../tests/central.pt',
+                         '../../tests/forces.pt'])
+def testConstructors(model_file):
+    force = ot.TorchForce(model_file)
+    model = pt.jit.load(model_file)
+    force = ot.TorchForce(pt.jit.load(model_file))
+    model = force.getModule()
+    force = ot.TorchForce(model)
+
+@pytest.mark.parametrize('model_file, output_forces, use_module_constructor',
+                        [('../../tests/central.pt', False, False,),
+                         ('../../tests/forces.pt', True, False),
+                         ('../../tests/forces.pt', True, True)])
 @pytest.mark.parametrize('use_cv_force', [True, False])
 @pytest.mark.parametrize('platform', ['Reference', 'CPU', 'CUDA', 'OpenCL'])
-def testForce(model_file, output_forces, use_cv_force, platform):
+def testForce(model_file, output_forces, use_module_constructor, use_cv_force, platform):
 
     if pt.cuda.device_count() < 1 and platform == 'CUDA':
         pytest.skip('A CUDA device is not available')
@@ -24,7 +35,11 @@ def testForce(model_file, output_forces, use_cv_force, platform):
         system.addParticle(1.0)
 
     # Create a force
-    force = ot.TorchForce(model_file)
+    if use_module_constructor:
+        model = pt.jit.load(model_file)
+        force = ot.TorchForce(model)
+    else:
+        force = ot.TorchForce(model_file)
     assert not force.getOutputsForces() # Check the default
     force.setOutputsForces(output_forces)
     assert force.getOutputsForces() == output_forces