def test_fit_1():
"""
This function tests the first way of fitting the descriptor: the data is passed by first creating compounds and then
the descriptors are created from the compounds.
"""
test_dir = os.path.dirname(os.path.realpath(__file__))
filenames = glob.glob(test_dir + "/CN_isobutane/*.xyz")
energies = np.loadtxt(test_dir + '/CN_isobutane/prop_kjmol_training.txt',
usecols=[1])
filenames.sort()
estimator = ARMP(representation="acsf")
estimator.generate_compounds(filenames[:50])
estimator.set_properties(energies[:50])
estimator.generate_representation()
idx = np.arange(0, 50)
estimator.fit(idx)
}
# Generate estimator
estimator = ARMP(iterations=10,
l1_reg=0.0001,
l2_reg=0.005,
learning_rate=0.0005,
representation_name='acsf',
representation_params=acsf_params,
tensorboard=True,
store_frequency=2,
hidden_layer_sizes=(50, 30, 10),
batch_size=200)
estimator.set_properties(ene_isopent)
estimator.generate_representation(pad_xyz, pad_zs, method='fortran')
print("Generated the representations")
print(estimator.representation.shape)
idx = list(range(n_samples))
idx_train, idx_test = modsel.train_test_split(idx,
random_state=42,
shuffle=True)
estimator.fit(idx_train)
data_squal = h5py.File(
"/Volumes/Transcend/data_sets/CN_squalane/dft/squalane_cn_dft.hdf5", "r")
xyz_squal = np.array(data_squal.get("xyz")[:10])
"zeta": 220.127,
"eta": 30.8065
}
estimator = ARMP(iterations=6000,
representation_name='acsf',
representation_params=acsf_params,
l1_reg=0.0,
l2_reg=0.0,
scoring_function="rmse",
tensorboard=False,
store_frequency=10,
learning_rate=0.075)
estimator.set_properties(energies[:100])
estimator.generate_compounds(filenames[:100])
estimator.generate_representation(method="tf")
print(estimator.representation.shape)
idx = list(range(100))
idx_train, idx_test = modsel.train_test_split(idx,
test_size=0,
random_state=42,
shuffle=True)
estimator.fit(idx_train)
score = estimator.score(idx_train)
print("The RMSE is %s kcal/mol." % (str(score)))
ene_pred = estimator.predict(idx_train)
from qml.aglaia.aglaia import ARMP
import glob
import numpy as np
from sklearn import model_selection as modsel
test_dir = "/Volumes/Transcend/repositories/my_qml_fork/qml/test/"
filenames = glob.glob(test_dir + "/qm7/*.xyz")
energies = np.loadtxt(test_dir + '/data/hof_qm7.txt', usecols=[1])
filenames.sort()
n_samples = 500
estimator = ARMP(representation_name="acsf", iterations=100)
estimator.generate_compounds(filenames[:n_samples])
estimator.set_properties(energies[:n_samples])
estimator.generate_representation(method="fortran")
idx = np.arange(0, n_samples)
idx_train, idx_test = modsel.train_test_split(idx,
random_state=42,
shuffle=True,
test_size=0.1)
estimator.fit(idx_train)
estimator.score(idx_train)
"angular_rs": np.arange(0, 10, 0.5),
"theta_s": np.arange(0, 3.14, 0.25)
}
estimator = ARMP(iterations=2000,
batch_size=256,
l1_reg=0.0001,
l2_reg=0.005,
learning_rate=0.00015,
representation='acsf',
representation_params=acsf_params,
tensorboard=True,
store_frequency=50)
estimator.set_properties(ene)
estimator.generate_representation(xyz, zs)
idx = list(range(n_samples))
idx_train, idx_test = modsel.train_test_split(idx,
test_size=0.15,
random_state=42,
shuffle=True)
all_scores = []
for lr in learning_rate:
for l1 in l1_reg:
for l2 in l2_reg:
estimator.fit(idx_train)
score = estimator.score(idx_test)
"zeta": 100.06564927139748,
"eta": 39.81824764370754
}
estimator = ARMP(iterations=2633,
batch_size=22,
l1_reg=1.46e-05,
l2_reg=0.0001,
learning_rate=0.0013,
representation_name='acsf',
representation_params=acsf_params,
tensorboard=True,
store_frequency=25,
hidden_layer_sizes=(185, ))
estimator.set_properties(ene_isopent)
estimator.generate_representation(xyz_isopent, zs_isopent, method="fortran")
# Training the model on 3 folds of n data points
for n in n_samples:
cv_idx = idx_train[:n]
splitter = modsel.KFold(n_splits=3, random_state=42, shuffle=True)
indices = splitter.split(cv_idx)
scores_per_fold = []
traj_scores_per_fold = []
for item in indices:
idx_train_fold = cv_idx[item[0]]
idx_test_fold = cv_idx[item[1]]
l1_reg=1.46e-05,
l2_reg=0.0001,
learning_rate=0.0013,
representation_name='acsf',
representation_params=acsf_params,
tensorboard=True,
store_frequency=25,
hidden_layer_sizes=(185, ))
# Loading the model previously trained
estimator.load_nn("../trained_nn/vr-nn")
estimator.set_properties(ene_surface)
# Generating the representation
start = time.time()
estimator.generate_representation(xyz_surface, zs_surface, method="fortran")
end = time.time()
print("The time taken to generate the representations is %s s" %
(str(end - start)))
print("The shape of the representations is %s" %
(str(estimator.representation.shape)))
# Predicting the energies
idx = list(range(n_samples))
predictions = estimator.predict(idx)
# Printing the mean absolute error
mae = mean_absolute_error(ene_surface, predictions)
print("The MAE is %.2f kJ/mol" % mae)
# Saving the results to a HDF5 file
## ------------- ** Loading the data ** ---------------
current_dir = os.path.dirname(os.path.realpath(__file__))
filenames = glob.glob(current_dir + '/../test/CN_isobutane/*.xyz')
energies = np.loadtxt(current_dir + '/../test/CN_isobutane/prop_kjmol_training.txt', usecols=[1])
filenames.sort()
## ------------- ** Setting up the estimator ** ---------------
estimator = ARMP(iterations=10, representation='acsf', representation_params={"radial_rs": np.arange(0, 10, 1), "angular_rs": np.arange(0.5, 10.5, 1),
"theta_s": np.arange(0, 5, 1)}, tensorboard=False)
estimator.generate_compounds(filenames)
estimator.set_properties(energies)
estimator.generate_representation()
## ------------- ** Fitting to the data ** ---------------
idx = np.arange(0,100)
estimator.fit(idx)
## ------------- ** Predicting and scoring ** ---------------
score = estimator.score(idx)
print("The mean absolute error is %s kJ/mol." % (str(-score)))
energies_predict = estimator.predict(idx)
# Getting the dataset
data = h5py.File("/Volumes/Transcend/data_sets/CN_isopentane/pruned_dft_with_forces/pruned_isopentane_cn_dft.hdf5", "r")
n_samples = 500
xyz = np.array(data.get("xyz")[-n_samples:])
ene = np.array(data.get("ene")[-n_samples:])*2625.50
ene = ene - data.get("ene")[0]*2625.50
zs = np.array(data["zs"][-n_samples:], dtype=np.int32)
# Creating the estimator
acsf_param = {"nRs2": 5, "nRs3": 5, "nTs": 5, "rcut": 5, "acut": 5, "zeta": 220.127, "eta": 30.8065}
estimator = ARMP(iterations=1000, batch_size=512, l1_reg=0.0, l2_reg=0.0, learning_rate=0.001, representation_name='acsf',
representation_params=acsf_param, tensorboard=False, store_frequency=50)
estimator.set_properties(ene)
estimator.generate_representation(xyz, zs, method='fortran')
print(estimator.g.shape)
# Doing cross validation
idx = list(range(n_samples))
idx_train, idx_test = modsel.train_test_split(idx, test_size=0.15, random_state=42, shuffle=False)
print("Starting the fitting...")
estimator.fit(idx_train)
# estimator.save_nn("saved_model")
pred1 = estimator.predict(idx_train)
pred2 = estimator.predict_from_xyz(xyz[idx_train], zs[idx_train])