def dataframe_to_gcn_input(self, input_data):
x_atoms_cold, x_bonds_cold, x_edges_cold = tensorise_smiles(
input_data['rdkit'],
max_degree=self.encoder_params['max_degree'],
max_atoms=self.encoder_params['max_atoms']
)
return [x_atoms_cold, x_bonds_cold, x_edges_cold]
def multistage_autoenc(smiles_x, num_layers, params, train_params):
# Create empty lists for outputs
#val_losses = []
# X_atom, X_bond, X_edge = tensorise_smiles(smiles_x[:2], max_degree=5, max_atoms=60)
print('Processing SMILES...')
#X, val = train_test_split(smiles_x, test_size=train_params["validation_split"], shuffle=True,
#random_state = np.random.randint(1, 10000))
X_atoms, X_bonds, X_edges = tensorise_smiles(smiles_x,
max_degree=5,
max_atoms=params['max_atoms'])
#X_atoms_val, X_bonds_val, X_edges_val = tensorise_smiles(val, max_degree=5, max_atoms=params['max_atoms'])
vni, vxi = vni_vxi(X_atoms, X_bonds, X_edges)
#vni_val, vxi_val = vni_vxi(X_atoms_val, X_bonds_val, X_edges_val)
# Iterate for every layer
for layer in range(1, num_layers + 1):
opt = Adam(lr=params["learning_rates"][layer - 1],
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=params['adam_decay'],
amsgrad=False)
#########################################################################
######################### STAGE I #######################################
#########################################################################
#gen = GraphDataGen(X, train_params['batch_size'], params, shuffle=False)
#valid = GraphDataGen(val, train_params['batch_size'], params, shuffle=False
if layer == 1:
stage_I_enc, _, stage_I_dec = stage_creator(params,
layer,
conv=True)
stage_I_dec.compile(optimizer=opt,
loss=params['losses_conv'],
metrics=['mse'])
stage_I_dec.fit(x=[X_atoms, X_bonds, X_edges],
y=[vni, vxi],
epochs=train_params['epochs'],
validation_split=0.1,
callbacks=params['callbacks'],
batch_size=train_params['batch_size'])
stage_I_enc = stage_I_dec.layers[3]
stage_I_enc.save_weights(
'layer_{}_stage_I_enc_weights.h5'.format(layer))
#val_losses.append(stage_I_dec.evaluate(x=[X_atoms_val, X_bonds_val, X_edges_val], y=[vni_val, vxi_val])[0])
else:
stage_I_dec, stage_I_enc = add_new_layer(
stage_I_enc,
params,
train_params,
layer,
X=[X_atoms, X_bonds, X_edges])
#########################################################################
######################### STAGE II ######################################
#########################################################################
stage_I_encodings = stage_I_enc.predict([X_atoms, X_bonds, X_edges])
_, vxi_II = vni_vxi(stage_I_encodings, X_bonds, X_edges)
stage_II_dec, stage_II_enc = stage_creator(params, layer, conv=False)
opt = Adam(lr=params["learning_rates_fp"][layer - 1],
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8,
decay=params['adam_decay'],
amsgrad=False)
stage_II_dec.compile(optimizer=opt,
loss=params['loss_fp'],
metrics=['mse'])
stage_II_dec.fit([stage_I_encodings, X_bonds, X_edges],
y=[vxi_II],
epochs=train_params['epochs'],
validation_split=train_params['validation_split'],
callbacks=params['callbacks'],
batch_size=train_params['batch_size'],
verbose=1)
stage_II_enc.save_weights(f'layer_{layer}_stage_II_enc_weights.h5')