for _ in range(3):
atom_state, bond_state = message_block(atom_state, bond_state, connectivity)
atom_state = Dense(atom_features//2, activation='softplus')(atom_state)
atom_state = Dense(1)(atom_state)
atom_state = Add()([atom_state, atomwise_energy])
output = ReduceAtomToMol(reducer='mean')([atom_state, snode_graph_indices])
model = GraphModel([
node_graph_indices, atom_types, distance_rbf, connectivity], [output])
lr = 1E-4
epochs = 500
model.compile(optimizer=keras.optimizers.Adam(lr=lr, decay=1E-5), loss='mae')
model.summary()
if not os.path.exists(model_name):
os.makedirs(model_name)
with open('{}/schnet_preprocessor.p'.format(model_name), 'wb') as f:
pickle.dump(preprocessor, f)
filepath = model_name + "/best_model.hdf5"
checkpoint = ModelCheckpoint(filepath, save_best_only=True, period=10, verbose=1)
csv_logger = CSVLogger(model_name + '/log.csv')
hist = model.fit_generator(train_sequence, validation_data=valid_sequence,
epochs=epochs, verbose=1,
callbacks=[checkpoint, csv_logger])
X = BatchNormalization(momentum=0.9)(mol_out)
X = Dense(512, activation='relu')(X)
X = BatchNormalization(momentum=0.9)(X)
X = Dense(256, activation='relu')(X)
X = Dense(num_output)(X)
model = GraphModel([node_graph_indices, atom_types, bond_types, connectivity],
[X])
epochs = 500
lr = 1E-3
decay = lr / epochs
model.compile(optimizer=keras.optimizers.Adam(lr=lr, decay=decay),
loss=masked_mean_squared_error)
model.summary()
filepath = model_name + "/best_model.hdf5"
checkpoint = ModelCheckpoint(filepath,
save_best_only=True,
period=10,
verbose=1)
csv_logger = CSVLogger(model_name + '/log.csv')
hist = model.fit_generator(train_generator,
validation_data=valid_generator,
verbose=1,
epochs=epochs,
callbacks=[checkpoint, csv_logger])
atom_rnn_layer = GRUStep(atom_features)
message_layer = MessageLayer(reducer='sum')
message_steps = 3
# Perform the message passing
for _ in range(message_steps):
# Get the message updates to each atom
message = message_layer([atom_state, bond_matrix, connectivity])
# Update memory and atom states
atom_state = atom_rnn_layer([message, atom_state])
atom_fingerprint = Dense(64, activation='relu')(atom_state)
mol_fingerprint = GraphOutput(reducer='sum')(
[snode_graph_indices, atom_fingerprint])
mol_fingerprint = BatchNormalization()(mol_fingerprint)
out = Dense(1)(mol_fingerprint)
model = GraphModel(
[node_graph_indices, atom_types, bond_types, connectivity], [out])
model.compile(optimizer=keras.optimizers.Adam(lr=0.001), loss='mse')
model.summary()
with warnings.catch_warnings():
warnings.simplefilter("ignore")
hist = model.fit_generator(train_generator,
validation_data=test_generator,
epochs=50,
verbose=2)
epochs = 1200
if args.restart:
model = load_model(filepath,
custom_objects={
'GraphModel': GraphModel,
'Squeeze': Squeeze,
'GatherAtomToBond': GatherAtomToBond,
'ReduceBondToAtom': ReduceBondToAtom,
'ReduceAtomToPro': ReduceAtomToPro
})
else:
model = GraphModel(
[atom_index, atom_types, distance_rbf, connectivity, n_pro], [output])
model.compile(optimizer=keras.optimizers.Adam(lr=lr), loss='mae')
model.summary()
checkpoint = ModelCheckpoint(filepath,
save_best_only=True,
period=10,
verbose=1)
csv_logger = CSVLogger('log.csv')
def decay_fn(epoch, learning_rate):
""" Jorgensen decays to 0.96*lr every 100,000 batches, which is approx
every 28 epochs """
if (epoch % 70) == 0:
return 0.96 * learning_rate
def test_save_and_load_model(get_2d_sequence, tmpdir):
preprocessor, sequence = get_2d_sequence
node_graph_indices = Input(shape=(1, ),
name='node_graph_indices',
dtype='int32')
atom_types = Input(shape=(1, ), name='atom', dtype='int32')
bond_types = Input(shape=(1, ), name='bond', dtype='int32')
connectivity = Input(shape=(2, ), name='connectivity', dtype='int32')
squeeze = Squeeze()
snode_graph_indices = squeeze(node_graph_indices)
satom_types = squeeze(atom_types)
sbond_types = squeeze(bond_types)
atom_features = 5
atom_state = Embedding(preprocessor.atom_classes,
atom_features,
name='atom_embedding')(satom_types)
bond_matrix = Embedding2D(preprocessor.bond_classes,
atom_features,
name='bond_embedding')(sbond_types)
atom_rnn_layer = GRUStep(atom_features)
message_layer = MessageLayer(reducer='sum', dropout=0.1)
# Perform the message passing
for _ in range(2):
# Get the message updates to each atom
message = message_layer([atom_state, bond_matrix, connectivity])
# Update memory and atom states
atom_state = atom_rnn_layer([message, atom_state])
atom_fingerprint = Dense(64, activation='sigmoid')(atom_state)
mol_fingerprint = ReduceAtomToMol(reducer='sum')(
[atom_fingerprint, snode_graph_indices])
out = Dense(1)(mol_fingerprint)
model = GraphModel(
[node_graph_indices, atom_types, bond_types, connectivity], [out])
with warnings.catch_warnings():
warnings.simplefilter("ignore")
model.compile(optimizer=keras.optimizers.Adam(lr=1E-4), loss='mse')
hist = model.fit_generator(sequence, epochs=1)
loss = model.evaluate_generator(sequence)
_, fname = tempfile.mkstemp('.h5')
model.save(fname)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
model = load_model(fname, custom_objects=custom_layers)
loss2 = model.evaluate_generator(sequence)
assert_allclose(loss, loss2)
