import tensorflow.keras.backend as kb
from megnet.layers.graph.base import GraphNetworkLayer
from megnet.activations import softplus2
import tensorflow as tf
class InteractionLayer(GraphNetworkLayer):
"""
The Continuous filter InteractionLayer in Schnet
Schütt et al. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions
Args:
activation (str): Default: None. The activation function used for each sub-neural network. Examples include
'relu', 'softmax', 'tanh', 'sigmoid' and etc.
use_bias (bool): Default: True. Whether to use the bias term in the neural network.
kernel_initializer (str): Default: 'glorot_uniform'. Initialization function for the layer kernel weights,
bias_initializer (str): Default: 'zeros'
activity_regularizer (str): Default: None. The regularization function for the output
kernel_constraint (str): Default: None. Keras constraint for kernel values
bias_constraint (str): Default: None .Keras constraint for bias values
Methods:
call(inputs, mask=None): the logic of the layer, returns the final graph
compute_output_shape(input_shape): compute static output shapes, returns list of tuple shapes
build(input_shape): initialize the weights and biases for each function
phi_e(inputs): update function for bonds and returns updated bond attribute e_p
rho_e_v(e_p, inputs): aggregate updated bonds e_p to per atom attributes, b_e_p
phi_v(b_e_p, inputs): update the atom attributes by the results from previous step b_e_p and all the inputs
returns v_p.
rho_e_u(e_p, inputs): aggregate bonds to global attribute
rho_v_u(v_p, inputs): aggregate atom to global attributes
get_config(): part of keras interface for serialization
"""
def __init__(self,
activation=softplus2,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
**kwargs):
super().__init__(activation=activation,
use_bias=use_bias,
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer,
activity_regularizer=activity_regularizer,
kernel_constraint=kernel_constraint,
bias_constraint=bias_constraint,
**kwargs)
def build(self, input_shapes):
vdim = input_shapes[0][2]
edim = input_shapes[1][2]
with kb.name_scope(self.name):
with kb.name_scope('phi_e'):
e_shapes = [[edim, vdim]] + [[vdim, vdim]] * 2
self.phi_e_weights = [self.add_weight(shape=i,
initializer=self.kernel_initializer,
name='weight_v_%d' % j,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint)
for j, i in enumerate(e_shapes)]
if self.use_bias:
self.phi_e_biases = [self.add_weight(shape=(i[-1],),
initializer=self.bias_initializer,
name='bias_v_%d' % j,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint)
for j, i in enumerate(e_shapes)]
else:
self.phi_e_biases = None
with kb.name_scope(self.name):
with kb.name_scope('phi_v'):
v_shapes = [[vdim, vdim]] + [[vdim, vdim]] * 2
self.phi_v_weights = [self.add_weight(shape=i,
initializer=self.kernel_initializer,
name='weight_v_%d' % j,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint)
for j, i in enumerate(v_shapes)]
if self.use_bias:
self.phi_v_biases = [self.add_weight(shape=(i[-1],),
initializer=self.bias_initializer,
name='bias_v_%d' % j,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint)
for j, i in enumerate(v_shapes)]
else:
self.phi_v_biases = None
self.built = True
def compute_output_shape(self, input_shape):
return input_shape
def phi_e(self, inputs):
nodes, edges, u, index1, index2, gnode, gbond = inputs
return edges
def rho_e_v(self, e_p, inputs):
"""
Reduce edge attributes to node attribute, eqn 5 in the paper
Args:
e_p: updated bond
inputs: the whole input list
Returns: summed tensor
"""
nodes, edges, u, index1, index2, gnode, gbond = inputs
atomwise1 = self._mlp(nodes, self.phi_v_weights[0], self.phi_v_biases[0])
cfconv1 = self.activation(self._mlp(edges, self.phi_e_weights[0], self.phi_e_biases[0]))
cfconv2 = self.activation(self._mlp(cfconv1, self.phi_e_weights[1], self.phi_e_biases[1]))
cfconv_out = self._mlp(cfconv2, self.phi_e_weights[2], self.phi_e_biases[2])
index1 = tf.reshape(index1, (-1,))
index2 = tf.reshape(index2, (-1,))
fr = tf.gather(atomwise1, index2, axis=1)
after_cfconv = atomwise1 + \
tf.transpose(a=tf.math.segment_sum(tf.transpose(a=fr * cfconv_out, perm=[1, 0, 2]), index1), perm=[1, 0, 2])
atomwise2 = self.activation(self._mlp(after_cfconv, self.phi_v_weights[1], self.phi_v_biases[1]))
atomwise3 = self._mlp(atomwise2, self.phi_v_weights[2], self.phi_v_biases[2])
return atomwise3
def phi_v(self, b_ei_p, inputs):
nodes, edges, u, index1, index2, gnode, gbond = inputs
return nodes + b_ei_p
def rho_e_u(self, e_p, inputs):
return 0
def rho_v_u(self, v_p, inputs):
return 0
def phi_u(self, b_e_p, b_v_p, inputs):
return inputs[2]
def _mlp(self, input_, weights, bias):
output = kb.dot(input_, weights) + bias
return output
def get_config(self):
base_config = super().get_config()
return base_config
