data/model/util.py

# Copyright 2023 Google LLC
#
# 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 Primitives."""

import functools
from typing import Callable, Dict, Optional, Tuple, Union
import flax.linen as nn
from jax import vmap
from jax.nn import initializers
import jax.numpy as jnp
from jax_md import energy
from jax_md import partition
from jax_md import util
import jraph
from jraph import GraphsTuple
from ml_collections import ConfigDict


Array = util.Array
FeaturizerFn = Callable[
    [GraphsTuple, Array, Array, Optional[Array]], GraphsTuple
]

f32 = jnp.float32

partial = functools.partial
normal = lambda var: initializers.variance_scaling(var, 'fan_in', 'normal')


# Nonlinearities:


UnaryFn = Callable[[Array], Array]


class BetaSwish(nn.Module):

  @nn.compact
  def __call__(self, x):
    features = x.shape[-1]
    beta = self.param('Beta', nn.initializers.ones, (features,))
    return x * nn.sigmoid(beta * x)


NONLINEARITY = {
    'none': lambda x: x,
    'relu': nn.relu,
    'swish': BetaSwish(),
    'raw_swish': nn.swish,
    'tanh': nn.tanh,
    'sigmoid': nn.sigmoid,
    'silu': nn.silu,
}


def get_nonlinearity_by_name(name: str) -> UnaryFn:
  if name in NONLINEARITY:
    return NONLINEARITY[name]
  raise ValueError(f'Nonlinearity "{name}" not found.')


# Fully-Connected Networks


class MLP(nn.Module):
  """Multilayer Perceptron."""
  
  features: Tuple[int, ...]
  nonlinearity: str

  use_bias: bool = True
  scalar_mlp_std: Optional[float] = None

  @nn.compact
  def __call__(self, x):
    features = self.features

    dense = partial(nn.Dense, use_bias=self.use_bias)
    phi = get_nonlinearity_by_name(self.nonlinearity)

    kernel_init = normal(self.scalar_mlp_std)

    for h in features[:-1]:
      x = phi(dense(h, kernel_init=kernel_init)(x))

    return dense(features[-1], kernel_init=normal(1.0))(x)


def mlp(
    hidden_features: Union[int, Tuple[int, ...]], nonlinearity: str, **kwargs
) -> Callable[..., Array]:
  if isinstance(hidden_features, int):
    hidden_features = (hidden_features,)

  def mlp_fn(*args):
    fn = MLP(hidden_features, nonlinearity, **kwargs)
    return jraph.concatenated_args(fn)(*args)

  return mlp_fn


# Featurizers


def neighbor_list_featurizer(displacement_fn):
  def featurize(atoms, position, neighbor, **kwargs):
    N = position.shape[0]
    graph = partition.to_jraph(neighbor, nodes=atoms)
    mask = partition.neighbor_list_mask(neighbor, True)

    Rb = position[graph.senders]
    Ra = position[graph.receivers]

    d = vmap(partial(displacement_fn, **kwargs))
    dR = d(Ra, Rb)
    dR = jnp.where(mask[:, None], dR, 1)

    return graph._replace(edges=dR)

  return featurize


# Bessel Functions


# Similar to the original Behler-Parinello features. Used by Nequip [1] and
# Schnet [2] to encode distance information.


def bessel(r_c, frequencies, r):
  rp = jnp.where(r > f32(1e-5), r, f32(1000.0))
  b = 2 / r_c * jnp.sin(frequencies * rp / r_c) / rp
  return jnp.where(r > f32(1e-5), b, 0)


class BesselEmbedding(nn.Module):
  count: int
  inner_cutoff: float
  outer_cutoff: float

  @nn.compact
  def __call__(self, rs: Array) -> Array:
    def init_fn(key, shape):
      del key
      assert len(shape) == 1
      n = shape[0]
      return jnp.arange(1, n + 1) * jnp.pi

    frequencies = self.param('frequencies', init_fn, (self.count,))
    bessel_fn = partial(bessel, self.outer_cutoff, frequencies)
    bessel_fn = vmap(
        energy.multiplicative_isotropic_cutoff(
            bessel_fn, self.inner_cutoff, self.outer_cutoff
        )
    )
    return bessel_fn(rs)


# Scale and Shifts
DATASET_SHIFT_SCALE = {'harder_silicon': (2.2548, 0.8825)}


def get_shift_and_scale(cfg: ConfigDict) -> Tuple[float, float]:
  if hasattr(cfg, 'scale') and hasattr(cfg, 'shift'):
    return cfg.shift, cfg.scale
  elif hasattr(cfg, 'train_dataset'):
    return DATASET_SHIFT_SCALE[cfg.train_dataset[0]]
  else:
    raise ValueError()