def test_generate_all_predictions():
"""Spot-check that the predicted free energies are reasonable
(right shape, finite, RMSE (in kcal/mol) in a reasonable range)"""
theta = np.ones(2 * n_types)
from time import time
print('predicting hydration free energies with theta=ones...')
t0 = time()
predictions = generate_all_predictions(theta)
t1 = time()
print('that took {:.3f} s'.format(t1 - t0))
# sanity check that the predictions are finite and that there's the right number of them
assert (len(predictions) == len(molecules))
assert (np.isfinite(predictions).all())
# sanity check that the numerical values of the predictions aren't grossly wrong...
pred_kcal_mol = unreduce(predictions)
expt_kcal_mol = unreduce(expt_means)
rmse = np.sqrt(np.mean((pred_kcal_mol - expt_kcal_mol)**2))
print('RMSE for theta all ones: {:.3f} kcal/mol'.format(rmse))
# first time I ran this test, it was ~7.154 kcal/mol
assert (rmse < 10.)
assert (rmse > 2.)
# check that running with mbondi radii and scales gives an RMSE closer to like 2.5 kcal/mol
theta = pack(radii=mbondi_model.get_radii(),
scales=mbondi_model.get_scale_factors())
print('predicting hydration free energies with theta from mbondi model...')
t0 = time()
predictions = generate_all_predictions(theta)
t1 = time()
print('that took {:.3f} s'.format(t1 - t0))
# sanity check that the predictions are finite and that there's the right number of them
assert (len(predictions) == len(molecules))
assert (np.isfinite(predictions).all())
# sanity check that the numerical values of the predictions aren't grossly wrong...
pred_kcal_mol = unreduce(predictions)
expt_kcal_mol = unreduce(expt_means)
rmse = np.sqrt(np.mean((pred_kcal_mol - expt_kcal_mol)**2))
print('RMSE for mbondi model: {:.3f} kcal/mol'.format(rmse))
# I think it's around 2.4 kcal/mol, but this test is saying something like 2.628 kcal/mol
assert (rmse > 2.)
assert (rmse < 3.)
assert (np.isfinite(predictions).all())
# sanity check that the numerical values of the predictions aren't grossly wrong...
pred_kcal_mol = unreduce(predictions)
expt_kcal_mol = unreduce(expt_means)
rmse = np.sqrt(np.mean((pred_kcal_mol - expt_kcal_mol)**2))
print('RMSE for mbondi model: {:.3f} kcal/mol'.format(rmse))
# I think it's around 2.4 kcal/mol, but this test is saying something like 2.628 kcal/mol
assert (rmse > 2.)
assert (rmse < 3.)
np.random.seed(0)
theta0 = pack(radii=mbondi_model.get_radii(),
scales=mbondi_model.get_scale_factors())
n_start = 10
x0s = [(np.random.rand(len(theta0)) * 0.05) - 0.025 + theta0
for _ in range(n_start)]
from collections import namedtuple
Experiment = namedtuple('Experiment', ['x0', 'cv_fold', 'll', 'stepsize'])
gd_stepsize = dict(gaussian=1e-8, student_t=1e-6)
experiments = []
for x0 in x0s:
for cv_fold in range(N_folds):
for ll in ll_dict:
experiments.append(
Experiment(x0=x0,
cv_fold=cv_fold,
L = log_prior(theta)
if not (L > -np.inf):
return L
else:
parameterized_list = construct_arrays(theta)
for i, mol in enumerate(mols):
radii, scale_factors = parameterized_list[i]
L += mol.log_prob(radii, scale_factors)
return L
if __name__ == '__main__':
n_types = mbondi_model.number_of_nodes
print('n_types: {}'.format(n_types))
initial_radii = np.array(mbondi_model.get_radii())
initial_scales = np.array(mbondi_model.get_scale_factors())
theta0 = np.hstack((initial_radii, initial_scales))
print('initial theta', theta0)
initial_log_prob = log_prob(theta0)
minimized_theta_fname = os.path.join(
data_path,
'elemental_types_l-bfgs-finite-difference_{}.npy'.format(name))
print('minimizing...')
from scipy.optimize import minimize
bounds = [(min_r, max_r)] * n_types + [(min_scale, max_scale)] * n_types