def prepare_predictor(input_file,
weights_file=None,
model_file=None,
mean_and_std_file=None):
predictor = Predictor()
predictor.load_input(input_file)
if model_file is not None:
predictor.load_architecture(model_file)
predictor.load_parameters(param_path=weights_file,
mean_and_std_path=mean_and_std_file)
return predictor
def test_read_input_file(self):
predictor_test = Predictor()
path = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'minimal_predictor', 'predictor_input.py')
read_input_file(path, predictor_test)
predictor_model = predictor_test.model
self.assertEqual(len(predictor_model.layers), 4)
self.assertTrue(isinstance(predictor_model.layers[1], MoleculeConv))
self.assertTrue(isinstance(predictor_model.layers[2], Dense))
self.assertEqual(predictor_model.layers[1].inner_dim, 38)
self.assertEqual(predictor_model.layers[1].units, 512)
def load_estimator(model_path):
estimator = Predictor()
input_file = os.path.join(model_path, 'predictor_input.py')
weights_file = os.path.join(model_path, 'full_train.h5')
model_file = os.path.join(model_path, 'full_train.json')
mean_and_std_file = os.path.join(model_path, 'full_train_mean_std.npz')
estimator.load_input(input_file)
if os.path.exists(model_file):
estimator.load_architecture(model_file)
uncertainty = True
else:
uncertainty = False
mean_and_std_file = mean_and_std_file if os.path.exists(
mean_and_std_file) else None
estimator.load_parameters(param_path=weights_file,
mean_and_std_path=mean_and_std_file)
return estimator, uncertainty
def load_estimator(model_path):
estimator = Predictor()
input_file = os.path.join(model_path, 'predictor_input.py')
weights_file = os.path.join(model_path, 'full_train.h5')
model_file = os.path.join(model_path, 'full_train.json')
mean_and_std_file = os.path.join(model_path, 'full_train_mean_std.npz')
estimator.load_input(input_file)
if os.path.exists(model_file):
estimator.load_architecture(model_file)
uncertainty = True
else:
uncertainty = False
mean_and_std_file = mean_and_std_file if os.path.exists(mean_and_std_file) else None
estimator.load_parameters(param_path=weights_file, mean_and_std_path=mean_and_std_file)
return estimator, uncertainty
level = logging.INFO
initialize_log(level, os.path.join(out_dir, 'train.log'))
# Log start timestamp
logging.info('CNN training initiated at ' + time.asctime() + '\n')
from rmgpy.rmg.main import RMG
rmg = RMG()
rmg.logHeader()
# Importing Keras should happen after setting random seed of Numpy
from dde.predictor import Predictor
predictor = Predictor(data_file=data_file,
save_tensors_dir=save_tensors_dir,
keep_tensors=keep_tensors,
out_dir=out_dir,
normalize=normalize)
predictor.load_input(input_file)
predictor.load_parameters(param_path=weights_file)
lr_func = "float({0} * np.exp(- epoch / {1}))".format(lr0, lr1)
save_model_path = os.path.join(out_dir, 'saved_model')
if not os.path.exists(save_model_path):
os.mkdir(save_model_path)
if train_mode == 'in_house':
predictor.kfcv_train(folds=folds,
batch_size=batch_size,
lr_func=lr_func,
save_model_path=save_model_path,
def setUp(self):
self.predictor = Predictor()
class TestPredictor(unittest.TestCase):
def setUp(self):
self.predictor = Predictor()
def test_model(self):
self.predictor.build_model()
predictor_model = self.predictor.model
self.assertEqual(len(predictor_model.layers), 4)
self.assertTrue(isinstance(predictor_model.layers[1], MoleculeConv))
self.assertTrue(isinstance(predictor_model.layers[2], Dense))
self.assertEqual(predictor_model.layers[1].inner_dim, 32)
self.assertEqual(predictor_model.layers[1].units, 512)
def test_load_input(self):
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'predictor_input.py')
self.predictor.load_input(test_predictor_input)
predictor_model = self.predictor.model
self.assertEqual(len(predictor_model.layers), 4)
self.assertTrue(isinstance(predictor_model.layers[1], MoleculeConv))
self.assertTrue(isinstance(predictor_model.layers[2], Dense))
self.assertTrue(isinstance(predictor_model.layers[3], Dense))
gfp = self.predictor.model.layers[1]
dense1 = self.predictor.model.layers[2]
dense2 = self.predictor.model.layers[3]
self.assertEqual(gfp.W_inner.shape.eval()[0], 4)
self.assertEqual(gfp.W_inner.shape.eval()[1], 38)
self.assertEqual(gfp.W_inner.shape.eval()[2], 38)
self.assertEqual(gfp.b_inner.shape.eval()[0], 4)
self.assertEqual(gfp.b_inner.shape.eval()[1], 1)
self.assertEqual(gfp.b_inner.shape.eval()[2], 38)
self.assertEqual(gfp.W_output.shape.eval()[0], 4)
self.assertEqual(gfp.W_output.shape.eval()[1], 38)
self.assertEqual(gfp.W_output.shape.eval()[2], 512)
self.assertEqual(gfp.b_output.shape.eval()[0], 4)
self.assertEqual(gfp.b_output.shape.eval()[1], 1)
self.assertEqual(gfp.b_output.shape.eval()[2], 512)
self.assertEqual(dense1.W.shape.eval()[0], 512)
self.assertEqual(dense1.W.shape.eval()[1], 50)
self.assertEqual(dense1.b.shape.eval()[0], 50)
self.assertEqual(dense2.W.shape.eval()[0], 50)
self.assertEqual(dense2.W.shape.eval()[1], 1)
self.assertEqual(dense2.b.shape.eval()[0], 1)
def test_specify_datasets(self):
"""
Test the datasets specification is done properly
"""
datasets_file = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'datasets.txt')
self.predictor.specify_datasets(datasets_file)
expected_datasets = [('rmg', 'sdata134k', 'polycyclic_2954_table',
0.1),
('rmg', 'sdata134k', 'cyclic_O_only_table', 0.1)]
self.assertEqual(self.predictor.datasets, expected_datasets)
def test_load_parameters(self):
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'predictor_input.py')
self.predictor.load_input(test_predictor_input)
param_path = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'minimal_predictor', 'weights.h5')
self.predictor.load_parameters(param_path)
gfp = self.predictor.model.layers[1]
dense1 = self.predictor.model.layers[2]
dense2 = self.predictor.model.layers[3]
self.assertAlmostEqual(gfp.W_inner.eval()[0][0][0], 1.000, 3)
self.assertAlmostEqual(gfp.b_inner.eval()[0][0][0], 0.000, 3)
self.assertAlmostEqual(gfp.W_output.eval()[0][0][0], 0.040, 3)
self.assertAlmostEqual(gfp.b_output.eval()[0][0][0], -0.561, 3)
self.assertAlmostEqual(dense1.W.eval()[0][0], -0.023, 3)
self.assertAlmostEqual(dense1.b.eval()[0], 1.517, 3)
self.assertAlmostEqual(dense2.W.eval()[0][0], -4.157, 3)
self.assertAlmostEqual(dense2.b.eval()[0], 1.515, 3)
def test_predict(self):
"""
Test predictor is predicting within a reasonable range
we should change weights.h5 every time we change feature space
"""
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'predictor_input.py')
self.predictor.load_input(test_predictor_input)
self.assertTrue(self.predictor.add_extra_atom_attribute)
self.assertTrue(self.predictor.add_extra_bond_attribute)
param_path = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'minimal_predictor', 'weights.h5')
self.predictor.load_parameters(param_path)
mol_test = Molecule().fromAdjacencyList(
"""1 C u0 p0 c0 {2,B} {6,B} {7,S}
2 C u0 p0 c0 {1,B} {3,B} {8,S}
3 C u0 p0 c0 {2,B} {4,B} {9,S}
4 C u0 p0 c0 {3,B} {5,B} {10,S}
5 C u0 p0 c0 {4,B} {6,B} {11,S}
6 C u0 p0 c0 {1,B} {5,B} {12,S}
7 H u0 p0 c0 {1,S}
8 H u0 p0 c0 {2,S}
9 H u0 p0 c0 {3,S}
10 H u0 p0 c0 {4,S}
11 H u0 p0 c0 {5,S}
12 H u0 p0 c0 {6,S}
""")
h298_predicted = self.predictor.predict(mol_test)
self.assertAlmostEqual(h298_predicted, 19.5, 0)
def test_normalize(self):
y1 = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0]
y2 = [2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]
y1_norm_expected = [-1.5, -1.0, -0.5, 0.0, 0.5, 1.0, 1.5]
y2_norm_expected = [-1.0, -0.5, 0.0, 0.5, 1.0, 1.5, 2.0]
y1_norm, y2_norm = self.predictor.normalize_output(y1, y2)
self.assertAlmostEqual(self.predictor.y_mean, 4.0)
self.assertAlmostEqual(self.predictor.y_std, 2.0)
self.assertTrue(np.allclose(y1_norm, y1_norm_expected))
self.assertTrue(np.allclose(y2_norm, y2_norm_expected))
y1 = [[1.0, 2.0], [3.0, 4.0]]
y2 = [[2.0, 3.0], [4.0, 5.0]]
mean_expected = [2.0, 3.0]
std_expected = [1.0, 1.0]
y1_norm_expected = [[-1.0, -1.0], [1.0, 1.0]]
y2_norm_expected = [[0.0, 0.0], [2.0, 2.0]]
y1_norm, y2_norm = self.predictor.normalize_output(y1, y2)
self.assertTrue(np.allclose(self.predictor.y_mean, mean_expected))
self.assertTrue(np.allclose(self.predictor.y_std, std_expected))
self.assertTrue(np.allclose(y1_norm, y1_norm_expected))
self.assertTrue(np.allclose(y2_norm, y2_norm_expected))
self.predictor.y_mean = None
self.predictor.y_std = None
def test_kfcv_train(self):
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'predictor_input.py')
self.predictor.load_input(test_predictor_input)
param_path = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'minimal_predictor', 'weights.h5')
self.predictor.load_parameters(param_path)
datafile = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'datafile.csv')
self.predictor.data_file = datafile
self.predictor.get_data_from_file = True
out_dir = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'test_out')
self.predictor.out_dir = out_dir
save_model_path = os.path.join(out_dir, 'saved_model')
if not os.path.exists(out_dir):
os.makedirs(out_dir)
if not os.path.exists(save_model_path):
os.mkdir(save_model_path)
lr_func = "float({0} * np.exp(- epoch / {1}))".format(0.0007, 30.0)
self.predictor.kfcv_train(2,
lr_func,
save_model_path,
nb_epoch=2,
patience=-1,
testing_ratio=0.1)
self.assertTrue(
not os.path.exists(os.path.join(save_model_path, 'best_model.h5')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'current_model.h5')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_0.h5')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_0.hist')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_0.json')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_0.png')))
self.assertTrue(
os.path.exists(
os.path.join(save_model_path, 'fold_0_loss_report.txt')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_1.h5')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_1.hist')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_1.json')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'fold_1.png')))
self.assertTrue(
os.path.exists(
os.path.join(save_model_path, 'fold_1_loss_report.txt')))
self.assertTrue(
os.path.exists(
os.path.join(save_model_path, 'full_folds_loss_report.txt')))
self.predictor.input_file = None
self.predictor.data_file = None
self.predictor.save_tensors_dir = None
self.predictor.keep_tensors = False
self.predictor.out_dir = None
shutil.rmtree(out_dir)
def test_full_train(self):
self.predictor.normalize = True
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'predictor_input.py')
self.predictor.load_input(test_predictor_input)
datafile = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'datafile.csv')
self.predictor.data_file = datafile
self.predictor.get_data_from_file = True
out_dir = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'test_out')
self.predictor.out_dir = out_dir
save_model_path = os.path.join(out_dir, 'saved_model')
if not os.path.exists(out_dir):
os.makedirs(out_dir)
if not os.path.exists(save_model_path):
os.mkdir(save_model_path)
lr_func = "float({0} * np.exp(- epoch / {1}))".format(0.0007, 30.0)
self.predictor.full_train(lr_func,
save_model_path,
nb_epoch=2,
training_ratio=1.0,
testing_ratio=0.0)
self.assertTrue(
os.path.exists(os.path.join(out_dir, 'identifiers_test.txt')))
self.assertTrue(
os.path.exists(os.path.join(out_dir, 'identifiers_train.txt')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'current_model.h5')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'full_train.h5')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'full_train.hist')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'full_train.json')))
self.assertTrue(
os.path.exists(os.path.join(save_model_path, 'full_train.png')))
self.assertTrue(
os.path.exists(
os.path.join(save_model_path, 'full_train_loss_report.txt')))
self.assertTrue(
os.path.exists(
os.path.join(save_model_path, 'full_train_mean_std.npz')))
self.predictor.input_file = None
self.predictor.data_file = None
self.predictor.save_tensors_dir = None
self.predictor.keep_tensors = False
self.predictor.out_dir = None
self.predictor.normalize = False
shutil.rmtree(out_dir)
def test_kfcv_batch_train(self):
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'minimal_predictor',
'predictor_input.py')
self.predictor.load_input(test_predictor_input)
datafile = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'datafile.csv')
self.predictor.data_file = datafile
self.predictor.get_data_from_file = True
out_dir = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'test_out')
self.predictor.out_dir = out_dir
save_model_path = os.path.join(out_dir, 'saved_model')
if not os.path.exists(out_dir):
os.makedirs(out_dir)
if not os.path.exists(save_model_path):
os.mkdir(save_model_path)
weights_file = os.path.join(os.path.dirname(dde.__file__), 'test_data',
'minimal_predictor', 'weights.h5')
self.predictor.kfcv_batch_train(3,
pretrained_weights=weights_file,
batch_size=2,
nb_epoch=2,
training_ratio=0.8,
testing_ratio=0.1)
self.assertTrue(
os.path.exists(os.path.join(out_dir, 'history.json_fold_0')))
self.assertTrue(
os.path.exists(os.path.join(out_dir, 'history.json_fold_1')))
self.assertTrue(
os.path.exists(os.path.join(out_dir, 'history.json_fold_2')))
self.predictor.input_file = None
self.predictor.data_file = None
self.predictor.save_tensors_dir = None
self.predictor.keep_tensors = False
self.predictor.out_dir = None
shutil.rmtree(out_dir)
def test_ensemble_predictor(self):
test_predictor_input = os.path.join(os.path.dirname(dde.__file__),
'test_data', 'ensemble_predictor',
'predictor_input.py')
test_predictor_architecture = os.path.join(
os.path.dirname(dde.__file__), 'test_data', 'ensemble_predictor',
'fold_0.json')
test_predictor_parameters = os.path.join(os.path.dirname(dde.__file__),
'test_data',
'ensemble_predictor',
'fold_0.h5')
self.predictor.load_input(test_predictor_input)
self.predictor.load_architecture(test_predictor_architecture)
self.predictor.load_parameters(test_predictor_parameters)
y_avg, y_std = self.predictor.predict(
Molecule().fromSMILES('CCC1CC(C)(C)C1'), sigma=True)
expected_y_avg = -21.4971179962
expected_y_std = 0.89726549387
self.assertAlmostEqual(expected_y_avg, y_avg, 2)
self.assertAlmostEqual(expected_y_std, y_std, 2)