Source code for jax_md_mod.model.neural_networks

# Copyright 2023 Multiscale Modeling of Fluid Materials, TU Munich
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Neural network models for potential energy and molecular property
prediction.
 """
from functools import partial
from typing import Callable, Dict, Any, Tuple, Union

import haiku as hk
from jax import numpy as jnp, nn as jax_nn
from jax_md import smap, space, partition, nn, util

from jax_md_mod.model import dropout, layers, sparse_graph


[docs] class DimeNetPP(hk.Module): """DimeNet++ for molecular property prediction. This model takes as input a sparse representation of a molecular graph - consisting of pairwise distances and angular triplets - and predicts per-atom properties. Global properties can be obtained by summing over per-atom predictions. The default values correspond to the orinal values of DimeNet++. This custom implementation follows the original DimeNet [#Gasteiger2022Dimenet]_ / DimeNet++ [#Gasteiger2022Dimenetpp]_ (Directional Message Passing Neural Network), while correcting for known issues (see https://github.com/klicperajo/dimenet). References: .. [#Gasteiger2022Dimenet] Gasteiger, J.; Groß, J.; Günnemann, S. Directional Message Passing for Molecular Graphs. arXiv April 5, 2022. https://doi.org/10.48550/arXiv.2003.03123. .. [#Gasteiger2022Dimenetpp] Gasteiger, J.; Giri, S.; Margraf, J. T.; Günnemann, S. Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules. http://arxiv.org/abs/2011.14115. """
[docs] def __init__(self, r_cutoff: float, n_species: int, num_targets: int, kbt_dependent: bool = False, embed_size: int = 128, n_interaction_blocks: int = 4, num_residual_before_skip: int = 1, num_residual_after_skip: int = 2, out_embed_size: int = None, type_embed_size: int = None, angle_int_embed_size: int = None, basis_int_embed_size: int = 8, num_dense_out: int = 3, num_rbf: int = 6, num_sbf: int = 7, activation: Callable = jax_nn.swish, envelope_p: int = 6, init_kwargs: Dict[str, Any] = None, dropout_mode: Dict[str, Any] = None, name: str = 'DimeNetPP', outscale: bool = False): """Initializes the DimeNet++ model The default values correspond to the orinal values of DimeNet++. Args: r_cutoff: Radial cut-off distance of edges n_species: Number of different atom species the network is supposed to process. num_targets: Number of different atomic properties to predict kbt_dependent: True, if DimeNet explicitly depends on temperature. In this case 'kT' needs to be provided as a kwarg during the model call to the energy_fn. Default False results in a model independent of temperature. embed_size: Size of message embeddings. Scale interaction and output embedding sizes accordingly, if not specified explicitly. n_interaction_blocks: Number of interaction blocks num_residual_before_skip: Number of residual blocks before the skip connection in the Interaction block. num_residual_after_skip: Number of residual blocks after the skip connection in the Interaction block. out_embed_size: Embedding size of output block. If None is set to 2 * embed_size. type_embed_size: Embedding size of atom type embeddings. If None is set to 0.5 * embed_size. angle_int_embed_size: Embedding size of Linear layers for down-projected triplet interation. If None is 0.5 * embed_size. basis_int_embed_size: Embedding size of Linear layers for interation of RBS/ SBF basis in interaction block num_dense_out: Number of final Linear layers in output block num_rbf: Number of radial Bessel embedding functions num_sbf: Number of spherical Bessel embedding functions activation: Activation function envelope_p: Power of envelope polynomial init_kwargs: Kwargs for initializaion of Linear layers dropout_mode: A dict defining which fully connected layers to apply dropout and at which rate (see dropout.dimenetpp_setup). If None, no Dropout is applied. name: Name of DimeNet++ model outscale: Scale the output initially return zero energy """ super().__init__(name=name) self.dropout_setup = dropout.dimenetpp_setup(dropout_mode, num_dense_out, n_interaction_blocks, num_residual_before_skip, num_residual_after_skip) if init_kwargs is None: init_kwargs = { 'w_init': layers.OrthogonalVarianceScalingInit(scale=1.), 'b_init': hk.initializers.Constant(0.), } # input representation: self.r_cutoff = r_cutoff self._rbf_layer = layers.RadialBesselLayer(r_cutoff, num_rbf, envelope_p) self._sbf_layer = layers.SphericalBesselLayer(r_cutoff, num_sbf, num_rbf, envelope_p) # build GNN structure self._n_interactions = n_interaction_blocks self._output_blocks = [] self._int_blocks = [] self._embedding_layer = layers.EmbeddingBlock( embed_size, n_species, type_embed_size, activation, init_kwargs, kbt_dependent) self._output_blocks.append(layers.OutputBlock( embed_size, out_embed_size, num_dense_out, num_targets, activation, init_kwargs, outscale=outscale) ) for _ in range(n_interaction_blocks): self._int_blocks.append(layers.InteractionBlock( embed_size, num_residual_before_skip, num_residual_after_skip, activation, init_kwargs, angle_int_embed_size, basis_int_embed_size) ) self._output_blocks.append(layers.OutputBlock( embed_size, out_embed_size, num_dense_out, num_targets, activation, init_kwargs, outscale=outscale) )
[docs] def __call__(self, graph: sparse_graph.SparseDirectionalGraph, **dyn_kwargs) -> jnp.ndarray: """Predicts per-atom quantities for a given molecular graph. Args: graph: An instance of sparse_graph.SparseDirectionalGraph defining the molecular graph connectivity. **dyn_kwargs: Kwargs supplied on-the-fly, such as 'kT' for temperature-dependent models or 'dropout_key' for Dropout. Returns: An (n_partciles, num_targets) array of predicted per-atom quantities """ dropout_key = dyn_kwargs.get('dropout_key', None) dropout_params = dropout.construct_dropout_params(dropout_key, self.dropout_setup) n_particles = graph.species.size # cutoff all non-existing edges: are encoded as 0 by rbf envelope # non-existing triplets will be masked explicitly in DimeNet++ pair_distances = jnp.where(graph.edge_mask, graph.distance_ij, 2. * self.r_cutoff) rbf = self._rbf_layer(pair_distances) # explicitly masked via mask array in angular_connections sbf = self._sbf_layer(pair_distances, graph.angles, graph.triplet_mask, graph.expand_to_kj) messages = self._embedding_layer(rbf, graph.species, graph.idx_i, graph.idx_j, dropout_params, **dyn_kwargs) per_atom_quantities = self._output_blocks[0]( messages, rbf, graph.idx_i, n_particles, dropout_params) for i in range(self._n_interactions): messages = self._int_blocks[i]( messages, rbf, sbf, graph.reduce_to_ji, graph.expand_to_kj, dropout_params) per_atom_quantities += self._output_blocks[i + 1]( messages, rbf, graph.idx_i, n_particles, dropout_params) return per_atom_quantities
[docs] def dimenetpp_neighborlist(displacement: space.DisplacementFn, r_cutoff: float, n_species: int = 100, positions_test: jnp.ndarray = None, neighbor_test: partition.NeighborList = None, max_triplet_multiplier: float = 1.25, max_edge_multiplier: float = 1.25, max_edges=None, max_triplets=None, mode: Union[str, Tuple[int]] = "energy", per_particle: bool = False, **dimenetpp_kwargs ) -> Tuple[nn.InitFn, Callable[[Any, util.Array], Tuple[util.Array]]]: """DimeNet++ energy function for Jax, M.D. This function provides an interface for the :class:`DimeNetPP` haiku model to be used as a jax_md energy_fn. Analogous to jax_md energy_fns, the initialized DimeNet++ energy_fn requires particle positions and a dense neighbor list as input - plus an array for species or other dynamic kwargs, if applicable. From particle positions and neighbor list, the sparse graph representation with edges and angle triplets is computed. Due to the constant shape requirement of jit of the neighborlist in jax_md, the neighbor list contains many masked edges, i.e. pairwise interactions that only "fill" the neighbor list, but are set to 0 during computation. This translates to masked edges and triplets in the sparse graph representation. For improved computational efficiency during jax_md simulations, the maximum number of edges and triplets can be estimated during model initialization. Edges and triplets beyond this maximum estimate can be capped to reduce computational and memory requirements. Capping is enabled by providing sample inputs (positions_test and neighbor_test) at initialization time. However, beware that currently, an overflow of max_edges and max_angles is not caught, as this requires passing an error code throgh jax_md simulators - analogous to the overflow detection in jax_md neighbor lists. If in doubt, increase the max edges/angles multipliers or disable capping. Args: displacement: Jax_md displacement function r_cutoff: Radial cut-off distance of DimeNetPP and the neighbor list n_species: Number of different atom species the network is supposed to process. positions_test: Sample positions to estimate max_edges / max_angles. Needs to be provided to enable capping. neighbor_test: Sample neighborlist to estimate max_edges / max_angles. Needs to be provided to enable capping. max_edge_multiplier: Multiplier for initial estimate of maximum edges. max_triplet_multiplier: Multiplier for initial estimate of maximum triplets. max_edges: Expected maximum of valid edges. max_triplets: Expected maximum of valid triplets. mode: Select what the model should return. If set to 'energy' (default), returns a single scalar quantity, i.e., the potential energy. If set to tuple ``(n_global, n_local)``, returns a tuple of global and local atomic properties. per_particle: If True, returns the energy per particle. dimenetpp_kwargs: Kwargs to change the default structure of DimeNet++. For definition of the kwargs, see DimeNetPP. Returns: A tuple of 2 functions: A init_fn that initializes the model parameters and an energy function that computes the energy for a particular state given model parameters. The energy function requires the same input as other energy functions with neighbor lists in jax_md.energy. """ r_cutoff = jnp.array(r_cutoff, dtype=util.f32) if positions_test is not None and neighbor_test is not None: print('Capping edges and triplets. Beware of overflow, which is' ' currently not being detected.') testgraph, _ = sparse_graph.sparse_graph_from_neighborlist( displacement, positions_test, neighbor_test, r_cutoff) max_triplets = jnp.int32(jnp.ceil(testgraph.n_triplets * max_triplet_multiplier)) max_edges = jnp.int32(jnp.ceil(testgraph.n_edges * max_edge_multiplier)) # cap maximum edges and angles to avoid overflow from multiplier n_particles, n_neighbors = neighbor_test.idx.shape max_edges = min(max_edges, n_particles * n_neighbors) max_triplets = min(max_triplets, n_particles * n_neighbors**2) print(f"Estimated max. {max_edges} edges and max. {max_triplets} triplets.") @hk.without_apply_rng @hk.transform def model(positions: jnp.ndarray, neighbor: partition.NeighborList, species: jnp.ndarray = None, **dynamic_kwargs) -> Union[jnp.ndarray, Tuple[jnp.ndarray, jnp.ndarray]]: """Evalues the DimeNet++ model and predicts the potential energy. Args: positions: Jax_md state-position. (N_particles x dim) array of particle positions neighbor: Jax_md dense neighbor list corresponding to positions species: (N_particles,) Array encoding atom types. If None, assumes all particles to belong to the same species **dynamic_kwargs: Dynamic kwargs, such as 'box' or 'kT'. Returns: If ``n_global = 1`` and ``n_local = 0`` (default), returns the potential energy. If ``n_global > 0`` xor ``n_local > 0``, returns a vector of global, respectively local, features. Otherwise, returns the tuple ``(global_feats, local_feats)``. """ # dynamic box necessary for pressure computation dynamic_displacement = partial(displacement, **dynamic_kwargs) graph_rep, overflow = sparse_graph.sparse_graph_from_neighborlist( dynamic_displacement, positions, neighbor, r_cutoff, species, max_edges, max_triplets ) # TODO: return overflow to detect possible overflow del overflow if mode == "energy": n_global = 1 n_local = 0 else: n_global, n_local = mode net = DimeNetPP(r_cutoff, n_species, num_targets=n_global + n_local, **dimenetpp_kwargs) features = net(graph_rep, **dynamic_kwargs) if "mask" in dynamic_kwargs: features *= dynamic_kwargs["mask"][:, jnp.newaxis] # Default behaviour if mode == "energy": if per_particle: return features.squeeze(axis=-1) else: return util.high_precision_sum(features) global_features = util.high_precision_sum( features[:, :n_global], axis=0) local_features = features[:, n_global:] return global_features, local_features return dropout.model_init_apply(model, dimenetpp_kwargs)
[docs] def dimenetpp_property_prediction( r_cutoff: float, n_targets: int = 1, n_species: int = 100, n_per_atom: int = None, **model_kwargs) -> Tuple[nn.InitFn, Callable[[Any, jnp.ndarray], jnp.ndarray]]: """Initializes DimeNet++ to predict molecular properties. This function provides an interface to the :class:`DimeNetPP` haiku model. Args: r_cutoff: Radial cut-off distance of DimeNetPP and the neighbor list. n_targets: Number of different molecular properties to predict. n_species: Number of different atom species the network is supposed to process. n_per_atom: Number of per-atom predictions, i.e., predicted quantity equals the number of atoms. The remaining predictions are considered global, i.e., one prediction per molecule. **model_kwargs: Kwargs to change the default structure of DimeNet++. Returns: A tuple of 2 functions: A init_fn that initializes the model parameters and an apply_function that predicts global molecular properties. """ @hk.without_apply_rng @hk.transform def property_predictor( mol_graph: sparse_graph.SparseDirectionalGraph, **dynamic_kwargs): """Predicts global quantities for a given molecular graph. Args: mol_graph: An instance of sparse_graph.SparseDirectionalGraph defining the molecular graph connectivity. dynamic_kwargs: Dynamic kwargs for DimeNet++. Returns: An (n_targets,) array of predicted global quantities. """ model = DimeNetPP(r_cutoff, n_species, n_targets, **model_kwargs) per_atom_predictions = model(mol_graph, **dynamic_kwargs) # Split the global predictions from the per-atom predictions. # Assume that the global predictions are a sum of the per-atom # predictions. n_predicted = per_atom_predictions.shape[1] n_global = n_predicted - n_per_atom per_atom_properties = per_atom_predictions[:, n_global:] global_properties = jnp.sum(per_atom_predictions[:, :n_global], axis=0) return global_properties, per_atom_properties return dropout.model_init_apply(property_predictor, model_kwargs)
[docs] class PairwiseNN(hk.Module): """A neural network predicting pairwise edge quantities Can be used for energy prediction for pairwise interactions. """
[docs] def __init__(self, r_cutoff: float, hidden_layers, init_kwargs: Dict = None, activation: Callable = jax_nn.swish, num_rbf: int = 6, envelope_p: int = 6, name: str = 'PairNN'): super().__init__(name=name) if init_kwargs is None: init_kwargs = { 'w_init': layers.OrthogonalVarianceScalingInit(scale=1.), 'b_init': hk.initializers.Constant(0.), } assert num_rbf > 0, 'The number of RBF embeddings but be at least 1.' self.cutoff = r_cutoff self.embedding = layers.RadialBesselLayer(r_cutoff, num_radial=num_rbf, envelope_p=envelope_p) self.pair_nn = hk.nets.MLP(hidden_layers, activation=activation, **init_kwargs) self.rbf_transform = hk.Linear(1, with_bias=False, name='RBF_Transform', **init_kwargs)
[docs] def __call__(self, distances, species=None, **kwargs): # ensure differentiability construction: rbf=0 for r > r_cut rbf = self.embedding(distances) predicted_energies = self.pair_nn(rbf) # rbf_transform has no bias: masked pairs remain 0 and counteract # possible non-zero contribution from biases in MPL (in a # continuously differentiable manner) per_pair_energy = predicted_energies * self.rbf_transform(rbf) return per_pair_energy
[docs] def pair_interaction_nn(displacement: space.DisplacementFn, r_cutoff: float, hidden_layers, **pair_net_kwargs): """An MLP acting on pairwise distances independently and summing the contributions. Embeds pairwise distances via radial Bessel functions (RBF). The RBF is also used to enforce a differentiable cut-off. Args: displacement: Displacement function r_cutoff: Radial cut-off of pairwise interactions and neighbor list hidden_layers: A list (or scalar in the case of a single hidden layer) of number of neurons for each hidden layer in the MLP pair_net_kwargs: Kwargs to change the default structure of PairwiseNN. For definition of the kwargs, see PairwiseNN. Returns: A tuple of 2 functions: A init_fn that initializes the model parameters and an energy function that computes the energy for a particular state given model parameters. """ if jnp.isscalar(hidden_layers): hidden_layers = [hidden_layers] hidden_layers.append(1) # output layer is scalar energy @hk.without_apply_rng @hk.transform def model(position, neighbor, species=None, **dynamic_kwargs): n_particles, _ = neighbor.idx.shape if species is not None: smap._check_species_dtype(species) # assert species are int raise NotImplementedError('Add species embedding to distance ' 'embedding.') dynamic_displacement = partial(displacement, **dynamic_kwargs) dyn_neighbor_displacement_fn = space.map_neighbor(dynamic_displacement) # compute pairwise distances neighbor_mask = neighbor.idx != n_particles r_neigh = position[neighbor.idx] pair_displacement = dyn_neighbor_displacement_fn(position, r_neigh) pair_distances = space.distance(pair_displacement) pair_distances = jnp.where(neighbor_mask, pair_distances, 2. * r_cutoff) net = PairwiseNN(r_cutoff, hidden_layers, **pair_net_kwargs) per_pair_energy = net(pair_distances, species, **dynamic_kwargs) # pairs are counted twice pot_energy = util.high_precision_sum(per_pair_energy) / 2. return pot_energy return model.init, model.apply