def call(self, x_batch):
### GAT with embedding:
target_features, feature_dist_graph, rij_dist_pairs, b_scope, start_end_env, l_scope, scope_update, scope_update_lig = x_batch
target_features = self.embedding(target_features)
rij_dist_pairs = tf.reshape(rij_dist_pairs,
(tf.shape(rij_dist_pairs)[0], 1))
message = tf.concat([feature_dist_graph, rij_dist_pairs], axis=1)
message = self.dist_embedding(message)
target_features = tf.concat(
[[np.zeros(tf.shape(target_features)[1])], target_features],
axis=0)
for i in range(0, self.num_iters):
multible_entry_f_lig = getConnectedFeatures(
target_features, start_end_env)
target_features = self.GATLayer(multible_entry_f_lig, message,
b_scope)
message = updateConnectedDict(target_features, scope_update_lig,
scope_update)
cmp_enc = tf.gather(target_features, indices=(l_scope), axis=0)
mol_vecs = tf.reduce_sum(cmp_enc, 1)
if self.output_nn:
return tf.reshape(self.output_nn(mol_vecs), [-1])
return mol_vecs
def call(self, x_batch):
### GAT with embedding:
target_features_orig, feature_dist_graph, rij_dist_pairs, b_scope, start_end_env, l_scope, scope_update, scope_update_lig = x_batch
target_features_orig = self.embedding(target_features_orig)
rij_dist_pairs = tf.reshape(rij_dist_pairs,
(tf.shape(rij_dist_pairs)[0], 1))
message = tf.concat([feature_dist_graph, rij_dist_pairs], axis=1)
message = self.dist_embedding(message)
target_features = tf.concat(
[[np.zeros(tf.shape(target_features_orig)[1])],
target_features_orig],
axis=0)
target_features_orig = tf.concat(
[[np.zeros(tf.shape(target_features_orig)[1])],
target_features_orig],
axis=0)
for i in range(0, self.num_heads - 1):
multible_entry_f_lig = getConnectedFeatures(
target_features, start_end_env)
target_features = self.heads[i](multible_entry_f_lig, message,
b_scope)
target_features = self._update_target_features(
target_features, target_features_orig)
message = updateConnectedDict(target_features, scope_update_lig,
scope_update)
cmp_enc = tf.gather(target_features, indices=(l_scope), axis=0)
mol_vecs = tf.reduce_sum(cmp_enc, 1)
try:
if self.output_nn and self.regression:
return tf.reshape(self.output_nn(mol_vecs), [-1])
elif self.output_nn:
return self.output_nn(mol_vecs)
return mol_vecs
except Exception as e:
log.error(
'There seems to be an issue with the input dimensions.' +
'Please check the input dimension of the output_nn you defined outside.'
+ 'Currently the NN output dimension is:' +
str(self.num_heads * self.emb_dim), e)
def call(self, x_batch):
target_features, feature_dist_graph, rij_dist_pairs, b_scope, start_end_env, l_scope, scope_update, scope_update_lig = x_batch
target_features = createVar(target_features)
feature_dist_graph = createVar(feature_dist_graph)
rij_dist_pairs = createVar(rij_dist_pairs)
target_features = self.W_i_a(target_features)
feature_dist_graph = self.W_i_a(feature_dist_graph)
rij_dist_pairs = tf.reshape(rij_dist_pairs,
(tf.shape(rij_dist_pairs)[0], 1))
rij_dist_pairs = self.W_i_b(rij_dist_pairs)
rij_dist_pairs = tf.keras.activations.relu(rij_dist_pairs)
message = tf.concat([feature_dist_graph, rij_dist_pairs], axis=1)
message = self.W_h(message)
message = tf.concat([[np.zeros(32)], message], axis=0)
message = tf.gather(message, indices=(b_scope))
message = tf.reduce_sum((message), 1)
for i in range(self.depth):
target_features = self.gru(message, target_features)
message = tf.convert_to_tensor(
updateConnectedDict(target_features, scope_update_lig,
scope_update))
message = tf.concat([message, rij_dist_pairs], axis=1)
message = self.W_h(message)
message = tf.concat([[np.zeros(32)], message], axis=0)
message = tf.gather(message, indices=(b_scope))
message = tf.reduce_sum(message, 1)
feature_hiddens = tf.concat([[np.zeros(32)], target_features], axis=0)
mol_vecs = tf.gather(feature_hiddens, indices=(l_scope), axis=0)
mol_vecs = tf.reduce_sum(mol_vecs, 1)
if self.output_nn:
return tf.reshape(self.output_nn(mol_vecs), [-1])
return mol_vecs