Add script for DMFF model saving. (#109)

* Add issue templates for feature request and bug-report

* Add script for dmff model saving.

* Remove issue template from devel branch.

* debug workflow

* remove debug

* Update ut.yml. Install mdtraj by conda.

.github/workflows/ut.yml

@@ -19,10 +19,10 @@ jobs:
       - name: Install Dependencies
         run: |
           source $CONDA/bin/activate
-          conda create -n dmff python=${{ matrix.python-version }} numpy openmm==7.7.0 pytest rdkit biopandas openbabel -c conda-forge
+          conda create -n dmff -y python=${{ matrix.python-version }} numpy openmm==7.7.0 mdtraj=1.9.7 pytest rdkit biopandas openbabel -c conda-forge
           conda activate dmff
           pip install --upgrade pip
-          pip install jax==0.3.15 jaxlib==0.3.15 jax_md==0.2.0 mdtraj==1.9.7 pymbar==4.0.1 tqdm
+          pip install jax==0.3.15 jaxlib==0.3.15 jax_md==0.2.0 pymbar==4.0.1 chex==0.1.4 dm-haiku==0.0.7 tqdm
       - name: Install DMFF
         run: |
           source $CONDA/bin/activate dmff && pip install .

backend/save_dmff2tf.py

@@ -0,0 +1,76 @@
+import dmff
+from dmff import NeighborList
+import jax
+import jax.numpy as jnp
+from jax.experimental import jax2tf
+# The model is saved in double precision by default. 
+# Since forces accuracy in double precision is needed in molecular dynamics simulations, 
+# we need to enable double precision in JAX.
+from jax import config
+config.update("jax_enable_x64", True)
+import openmm.app as app
+import openmm.unit as unit
+import tensorflow as tf
+
+import os
+import argparse
+
+gpus = tf.config.experimental.list_physical_devices('GPU')
+for gpu in gpus:
+  tf.config.experimental.set_memory_growth(gpu, True)
+
+def create_dmff_potential(input_pdb_file, ff_xml_files):
+    pdb = app.PDBFile(input_pdb_file)
+    h = dmff.Hamiltonian(*ff_xml_files)
+    pot = h.createPotential(pdb.topology,
+                            nonbondedMethod=app.PME,
+                            nonbondedCutoff=1.2 *
+                            unit.nanometer)
+    pot_func = pot.getPotentialFunc()
+    a, b, c = pdb.topology.getPeriodicBoxVectors()
+    a = a.value_in_unit(unit.nanometer)
+    b = b.value_in_unit(unit.nanometer)
+    c = c.value_in_unit(unit.nanometer)
+
+    engrad = jax.value_and_grad(pot_func, 0)
+
+    covalent_map = h.getGenerators()[-1].covalent_map
+
+    def potential_engrad(positions, box, pairs):
+        if jnp.shape(pairs)[-1] == 2:
+            nbond = covalent_map[pairs[:, 0], pairs[:, 1]]
+            pairs = jnp.concatenate([pairs, nbond[:, None]], axis=1)
+
+        return engrad(positions, box, pairs, h.paramtree)
+
+    return pdb, potential_engrad, covalent_map, pot, h
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input_pdb", dest="input_pdb", help="input pdb file. Box information is required in the pdb file.")
+    parser.add_argument("--xml_files", dest="xml_files", nargs="+", help=".xml files with parameters are derived from DMFF.")
+    parser.add_argument("--output", dest="output", help="output directory")
+    args = parser.parse_args()
+
+    input_pdb = args.input_pdb
+    ff_xml_files = args.xml_files
+    output_dir = args.output
+    if output_dir[-1] == "/":
+        output_dir = output_dir[:-1]
+    if not os.path.exists(output_dir):
+        os.mkdir(output_dir)
+
+    pdb, pot_grad, covalent_map, pot, h = create_dmff_potential(input_pdb, ff_xml_files)
+
+    natoms = pdb.getTopology().getNumAtoms()
+
+    f_tf = jax2tf.convert(
+        jax.jit(pot_grad),
+        polymorphic_shapes=["("+str(natoms)+", 3)", "(3, 3)", "(b, 2)"]
+    )
+    dmff_model = tf.Module()
+    dmff_model.f = tf.function(f_tf, autograph=False,
+                            input_signature=[tf.TensorSpec(shape=[natoms,3], dtype=tf.float64), tf.TensorSpec(shape=[3,3], dtype=tf.float64), tf.TensorSpec(shape=tf.TensorShape([None, 2]), dtype=tf.int32)])
+
+    tf.saved_model.save(dmff_model, output_dir, options=tf.saved_model.SaveOptions(experimental_custom_gradients=True))