def load_NN():
"""
Loads the Keras NN model for a user-trained metric.
"""
nn = KerasNN(name)
nn.load('../data/nns/{}.h5'.format(name))
return ('nn', nn)
def load_mp():
"""
Loads the Keras NN model for melting point.
"""
cnn_mp = KerasNN('mp')
cnn_mp.load('../data/nns/mp.h5')
return ('cnn', cnn_mp)
def load_bandgap():
"""
Loads the Keras NN model for H**O-LUMO energy
difference.
"""
cnn_bandgap = KerasNN('bandgap')
cnn_bandgap.load('../data/nns/bandgap.h5')
return ('cnn', cnn_bandgap)
def load_PCE():
"""
Loads the Keras NN model for Power Conversion
Efficiency.
"""
cnn_pce = KerasNN('pce')
cnn_pce.load('../data/nns/pce.h5')
return ('cnn', cnn_pce)
def train_nn_as_metric(self, name, train_x, train_y, nepochs=1000):
"""Sets up a metric with a neural network trained on
a dataset.
Arguments.
-----------
- name. String used to identify the metric.
- train_x. List of SMILES identificators.
- train_y. List of property values.
- nepochs. Number of epochs for training.
Note.
-----------
A name.h5 file is generated in the data/nns directory,
and this metric can be loaded in the future using the
load_prev_user_metric() method through the name.pkl
file generated in the data/ dir.
load_prev_user_metric('name.pkl')
"""
cnn = KerasNN(name)
cnn.train(train_x,
train_y,
500,
nepochs,
earlystopping=True,
min_delta=0.001)
K.clear_session()
def batch_NN(smiles, train_smiles=None, nn=None):
"""
User-trained neural network.
"""
if nn == None:
raise ValueError(
'The user-trained NN metric was not properly loaded.')
fsmiles = []
zeroindex = []
for k, sm in enumerate(smiles):
if mm.verify_sequence(sm):
fsmiles.append(sm)
else:
fsmiles.append('c1ccccc1')
zeroindex.append(k)
vals = np.asarray(nn.predict(fsmiles))
for k in zeroindex:
vals[k] = 0.0
vals = np.squeeze(np.stack(vals, axis=1))
return vals
def load_NN():
"""
Loads the Keras NN model for a user-trained metric.
"""
nn = KerasNN(name)
nn.load('../data/nns/{}.h5'.format(name))
return ('nn', nn)
self.AV_METRICS[name] = batch_NN
self.LOADINGS[name] = load_NN
if self.verbose:
print('Defined metric {}'.format(name))
metric = [batch_NN, load_NN]
with open('../data/{}.pkl'.format(name), 'wb') as f:
pickle.dump(metric, f)