def test_converged_zetas(self):
"""
Testing convergence of zetas.
"""
friction = 10.0
dt = 0.08
temp = 0.0
num_atoms = self.x0.shape[0]
x_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
num_steps = 100 # with a temp of 10.0 should converge really quickly
with tf.variable_scope("reference"):
ref_intg = integrator.LangevinIntegrator(self.masses, x_ph,
[self.ha, self.hb], dt,
friction, temp)
# ref_intg.vscale = 0.45 -> so we should converge fully to 16 decimals after 47 steps
with tf.variable_scope("test"):
test_intg = integrator.LangevinIntegrator(self.masses,
x_ph, [self.ha, self.hb],
dt,
friction,
temp,
buffer_size=50)
ref_dx, ref_dxdps = ref_intg.step_op()
test_dx, test_dxdps = test_intg.step_op()
sess = tf.Session()
sess.run(tf.initializers.global_variables())
x_ref = np.copy(self.x0)
x_test = np.copy(self.x0)
for step in range(num_steps):
ref_dx_val, ref_dxdp_val = sess.run([ref_dx, ref_dxdps],
feed_dict={x_ph: x_ref})
test_dx_val, test_dxdp_val = sess.run([test_dx, test_dxdps],
feed_dict={x_ph: x_test})
np.testing.assert_array_almost_equal(ref_dx_val,
test_dx_val,
decimal=14)
np.testing.assert_array_almost_equal(
ref_dxdp_val, test_dxdp_val, decimal=14) # BAD WTF CONVERGENCE
x_ref += ref_dx_val
x_test += test_dx_val
def test_optimize_single_structure(self):
"""
Testing optimization of a single structure.
"""
masses = np.array([8.0, 1.0, 1.0])
x0 = np.array([
[-0.0070, -0.0100, 0.0000],
[-1.1426, 0.5814, 0.0000],
[0.4728, -0.2997, 0.0000],
],
dtype=np.float64) # starting geometry
x0.setflags(write=False)
x_opt = np.array([
[-0.0070, -0.0100, 0.0000],
[-0.1604, 0.4921, 0.0000],
[0.5175, 0.0128, 0.0000],
],
dtype=np.float64) # idealized geometry
x_opt.setflags(write=False)
bonds = x_opt - x_opt[0, :]
bond_lengths = np.linalg.norm(bonds[1:, :], axis=1)
num_atoms = len(masses)
starting_bond = 0.8 # Guessestimate starting (true x_opt: 0.52)
starting_angle = 2.1 # Guessestimate ending (true x_opt: 1.81)
bond_params = [
tf.get_variable("OH_kb",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(100.0)),
tf.get_variable(
"OH_b0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(starting_bond)),
]
hb = bonded_force.HarmonicBondForce(
params=bond_params,
bond_idxs=np.array([[0, 1], [0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1], [0, 1]], dtype=np.int32))
angle_params = [
tf.get_variable("HOH_ka",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(75.0)),
tf.get_variable(
"HOH_a0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(starting_angle)),
]
ha = bonded_force.HarmonicAngleForce(
params=angle_params,
angle_idxs=np.array([[1, 0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1]], dtype=np.int32))
friction = 10.0
dt = 0.005
# temp = 50.0
temp = 0.0
x_ph = tf.placeholder(name="input_geom",
dtype=tf.float64,
shape=(num_atoms, 3))
intg = integrator.LangevinIntegrator(masses, x_ph, [hb, ha], dt,
friction, temp)
dx_op, dxdp_op = intg.step_op()
num_steps = 500
# param_optimizer = tf.train.AdamOptimizer(0.02)
param_optimizer = tf.train.RMSPropOptimizer(0.01)
def loss(pred_x):
# Compute pairwise distances
def dij(x):
v01 = x[0] - x[1]
v02 = x[0] - x[2]
v12 = x[1] - x[2]
return tf.stack([tf.norm(v01), tf.norm(v02), tf.norm(v12)])
return tf.norm(dij(x_opt) - dij(pred_x))
# geometry we arrive at at time t=inf
x_final_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
dLdx = tf.gradients(loss(x_final_ph), x_final_ph)
grads_and_vars = intg.grads_and_vars(dLdx)
train_op = param_optimizer.apply_gradients(grads_and_vars)
sess = tf.Session()
sess.run(tf.initializers.global_variables())
num_epochs = 100
for e in range(num_epochs):
params = sess.run(bond_params + angle_params)
print("starting epoch", e, "current params", params)
# converged
if np.abs(params[1] - 0.52) < 0.05 and np.abs(params[3] -
1.81) < 0.05:
return
x = np.copy(x0)
intg.reset(sess) # clear integration buffers
for step in range(num_steps):
dx_val, dxdp_val = sess.run([dx_op, dxdp_op],
feed_dict={x_ph: x})
x += dx_val
sess.run(train_op, feed_dict={x_final_ph: x})
# failed to converge
assert 0
def test_optimize_water_frequencies(self):
masses = np.array([8.0, 1.0, 1.0])
x_opt = np.array([
[-0.0070, -0.0100, 0.0000],
[-0.1604, 0.4921, 0.0000],
[0.5175, 0.0128, 0.0000],
],
dtype=np.float64)
x_opt.setflags(write=False) # idealized geometry
bonds = x_opt - x_opt[0, :]
bond_lengths = np.linalg.norm(bonds[1:, :], axis=1)
num_atoms = len(masses)
starting_bond = 1.47 # Guessestimate starting (true x_opt: 0.52)
starting_angle = 0.07 # Guessestimate ending (true x_opt: 1.81)
bond_params = [
tf.get_variable("OH_kb",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(100.0)),
tf.get_variable(
"OH_b0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(starting_bond)),
]
hb = bonded_force.HarmonicBondForce(
params=bond_params,
bond_idxs=np.array([[0, 1], [0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1], [0, 1]], dtype=np.int32))
angle_params = [
tf.get_variable("HOH_ka",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(75.0)),
tf.get_variable(
"HOH_a0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(starting_angle)),
]
ha = bonded_force.HarmonicAngleForce(
params=angle_params,
angle_idxs=np.array([[1, 0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1]], dtype=np.int32))
friction = 10.0
dt = 0.005
temp = 0.0
x_ph = tf.placeholder(name="input_geom",
dtype=tf.float64,
shape=(num_atoms, 3))
intg = integrator.LangevinIntegrator(masses, x_ph, [hb, ha], dt,
friction, temp)
dx_op, dxdp_op = intg.step_op()
num_steps = 500
# param_optimizer = tf.train.AdamOptimizer(0.02)
param_optimizer = tf.train.RMSPropOptimizer(0.02)
def loss(pred_x):
test_eigs = observable.vibrational_eigenvalues(
pred_x, masses, [hb, ha])
true_freqs = [
0, 0, 0, 40.63, 59.383, 66.44, 1799.2, 3809.46, 3943
] # from http://gaussian.com/vib/
true_eigs = [(x / VIBRATIONAL_CONSTANT)**2 for x in true_freqs]
return tf.sqrt(tf.reduce_sum(tf.pow(true_eigs - test_eigs,
2))), test_eigs
# geometry we arrive at at time t=inf
x_final_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
loss_op, test_eigs_op = loss(x_final_ph)
dLdx = tf.gradients(loss_op, x_final_ph)
grads_and_vars = intg.grads_and_vars(dLdx[0])
train_op = param_optimizer.apply_gradients(grads_and_vars)
sess = tf.Session()
sess.run(tf.initializers.global_variables())
num_epochs = 750
for e in range(num_epochs):
x = np.copy(x_opt)
intg.reset(sess) # clear integration buffers
for step in range(num_steps):
dx_val, dxdp_val = sess.run([dx_op, dxdp_op],
feed_dict={x_ph: x})
x += dx_val
_, loss, evs = sess.run([train_op, loss_op, test_eigs_op],
feed_dict={x_final_ph: x})
print("starting epoch", e, "loss", loss, "current params",
sess.run(bond_params + angle_params), evs)
params = sess.run(bond_params + angle_params)
np.testing.assert_almost_equal(params[1], 0.52, decimal=2)
np.testing.assert_almost_equal(params[3], 1.81, decimal=1)
def test_optimize_ensemble(self):
"""
Testing that we can optimize ensembles with different integrator and ff settings relative to a canonical one.
"""
# 1. Generate a water ensemble by doing 5000 steps of MD starting from an idealized geometry.
# 2. Reservoir sample along the trajectory.
# 3. Generate a loss function used sum of square distances.
# 4. Compute parameter derivatives.
x_opt = np.array([
[-0.0070, -0.0100, 0.0000],
[-0.1604, 0.4921, 0.0000],
[0.5175, 0.0128, 0.0000],
],
dtype=np.float64) # idealized geometry
x_opt.setflags(write=False)
masses = np.array([8.0, 1.0, 1.0], dtype=np.float64)
num_atoms = len(masses)
ideal_bond = 0.52
ideal_angle = 1.81
bond_params = [
tf.get_variable("OH_kb",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(100.0)),
tf.get_variable("OH_b0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(ideal_bond)),
]
hb = bonded_force.HarmonicBondForce(
params=bond_params,
bond_idxs=np.array([[0, 1], [0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1], [0, 1]], dtype=np.int32))
angle_params = [
tf.get_variable("HOH_ka",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(75.0)),
tf.get_variable("HOH_a0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(ideal_angle)),
]
ha = bonded_force.HarmonicAngleForce(
params=angle_params,
angle_idxs=np.array([[1, 0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1]], dtype=np.int32))
# standard MD used to generate a canonical ensemble
friction = 1.0
dt = 0.0025
temp = 300.0
num_steps = 20000
x_ph = tf.placeholder(name="input_geom",
dtype=tf.float64,
shape=(num_atoms, 3))
intg = integrator.LangevinIntegrator(masses, x_ph, [hb, ha], dt,
friction, temp)
reservoir_size = 200
ref_dx_op, _ = intg.step_op(inference=True)
sess = tf.Session()
sess.run(tf.initializers.global_variables())
def reference_generator():
x = np.copy(
x_opt
) # (ytz): Do not remove this copy else you'll spend hours tracking down bugs.
for step in range(num_steps):
if step % 100 == 0:
print(step)
dx_val = sess.run(ref_dx_op, feed_dict={x_ph: x})
x += dx_val
# this copy is important here, otherwise we're just adding the same reference
# since += modifies the object in_place
yield np.copy(x), step # wip: decouple
def generate_reference_ensemble():
intg.reset(sess)
rs = ReservoirSampler(reference_generator(), reservoir_size)
rs.sample_all()
ensemble = []
for x, t in rs.R:
ensemble.append(x)
stacked_ensemble = np.stack(ensemble, axis=0)
ensemble_ph = tf.placeholder(shape=(reservoir_size, num_atoms, 3),
dtype=np.float64)
ref_d2ij_op = observable.sorted_squared_distances(ensemble_ph)
return ref_d2ij_op, stacked_ensemble, ensemble_ph
a1, inp1, ensemble_ph1 = generate_reference_ensemble()
a2, inp2, ensemble_ph2 = generate_reference_ensemble()
# compute MSE of two _identical_ simulations except for the RNG
loss = tf.reduce_sum(tf.pow(a1 - a2, 2)) / reservoir_size # 0.003
mutual_MSE = sess.run(loss,
feed_dict={
ensemble_ph1: inp1,
ensemble_ph2: inp2
})
print("Optimal MSE", mutual_MSE)
# on average, two identical ensembles should yield the same result
assert mutual_MSE < 0.05
# Use completely different integration parameters
friction = 10.0
dt = 0.05
temp = 100
num_steps = 500 # we need to make the num_steps variable and only stop once we've converged
x_ph = tf.placeholder(name="input_geom",
dtype=tf.float64,
shape=(num_atoms, 3))
with tf.variable_scope("bad"):
bad_intg = integrator.LangevinIntegrator(masses,
x_ph, [hb, ha],
dt,
friction,
temp,
buffer_size=None)
bad_dx_op, bad_dxdp_op = bad_intg.step_op()
d0_ph = tf.placeholder(shape=tuple(), dtype=tf.float64)
d1_ph = tf.placeholder(shape=tuple(), dtype=tf.float64)
d2_ph = tf.placeholder(shape=tuple(), dtype=tf.float64)
d3_ph = tf.placeholder(shape=tuple(), dtype=tf.float64)
grads_and_vars = []
for dp, var in zip([d0_ph, d1_ph, d2_ph, d3_ph],
bond_params + angle_params):
grads_and_vars.append((dp, var))
# param_optimizer = tf.train.AdamOptimizer(0.02)
# param_optimizer = tf.train.AdamOptimizer(0.1) # unstable
param_optimizer = tf.train.RMSPropOptimizer(0.1)
train_op = param_optimizer.apply_gradients(grads_and_vars)
sess.run(tf.initializers.global_variables())
# it turns out the force constants don't really matter a whole lot in this case, but the
# ideal lengths/angles do matter.
# Use completely different forcefield parameters, changing bond constants and lengths.
# See if we can recover the original parameters again.
sess.run([
tf.assign(bond_params[0], 60),
tf.assign(bond_params[1], 1.3),
tf.assign(angle_params[0], 43),
tf.assign(angle_params[1], 2.1)
])
print("Starting params", sess.run(bond_params + angle_params))
for epoch in range(10000):
def sub_optimal_generator():
x = np.copy(x_opt)
for step in range(num_steps):
dx_val, dxdp_val = sess.run([bad_dx_op, bad_dxdp_op],
feed_dict={x_ph: x})
x += dx_val
yield np.copy(x), dxdp_val, step # wip: decouple
bad_intg.reset(sess)
rs = ReservoirSampler(sub_optimal_generator(), reservoir_size)
rs.sample_all()
bad_ensemble = []
bad_ensemble_grads = []
for x, dxdp, t in rs.R:
bad_ensemble.append(x)
bad_ensemble_grads.append(dxdp)
stacked_bad_ensemble = np.stack(bad_ensemble, axis=0)
stacked_bad_ensemble_grads = np.stack(bad_ensemble_grads, axis=0)
# compute ensemble's average angle and bond length
bad_ensemble_ph = tf.placeholder(shape=(reservoir_size, num_atoms,
3),
dtype=np.float64)
ref_d2ij_op = observable.sorted_squared_distances(bad_ensemble_ph)
# b1, bnp1, bbad_ensemble_ph1 = generate_sub_optimal_bad_ensemble()
loss_op = tf.reduce_sum(tf.pow(a1 - ref_d2ij_op,
2)) / reservoir_size # 0.003
dLdx_op = tf.gradients(loss_op, bad_ensemble_ph)
loss, dLdx_val = sess.run([loss_op, dLdx_op],
feed_dict={
ensemble_ph1: inp1,
bad_ensemble_ph: stacked_bad_ensemble
}) # MSE 12.953122852970827
dLdx_dxdp = np.multiply(np.expand_dims(dLdx_val[0], 1),
stacked_bad_ensemble_grads)
reduced_dLdp = np.sum(dLdx_dxdp, axis=tuple([0, 2, 3]))
sess.run(train_op,
feed_dict={
d0_ph: reduced_dLdp[0],
d1_ph: reduced_dLdp[1],
d2_ph: reduced_dLdp[2],
d3_ph: reduced_dLdp[3],
})
if loss < mutual_MSE * 5:
# succesfully converged (should take about 600 epochs)
return
print("loss", loss, "epoch", epoch, "current params",
sess.run(bond_params + angle_params))
assert 0
def test_linear_chain(self):
"""
Testing minimization of HCN into a linear geometry.
"""
masses = np.array([6.0, 7.0, 1.0])
x0 = np.array(
[
[-0.0055, 0.0000, -0.0349], # C
[-1.0973, 0.0000, 0.3198], # N
[1.1213, 1.1250, -0.3198] # H
],
dtype=np.float64)
num_atoms = len(masses)
bond_params = [
tf.get_variable("HC_kb",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(100.0)),
tf.get_variable("HC_b0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(1.0)),
tf.get_variable("CN_kb",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(120.0)),
tf.get_variable("CN_b0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(1.2)),
]
hb = bonded_force.HarmonicBondForce(
params=bond_params,
bond_idxs=np.array([[0, 1], [0, 2]], dtype=np.int32),
param_idxs=np.array([[2, 3], [0, 1]], dtype=np.int32))
ideal_angle = np.pi
angle_params = [
tf.get_variable("HCN_ka",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(75.0)),
tf.get_variable("HCN_a0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(ideal_angle)),
]
ha = bonded_force.HarmonicAngleForce(
params=angle_params,
angle_idxs=np.array([[1, 0, 2]], dtype=np.int32),
param_idxs=np.array([[0, 1]], dtype=np.int32))
friction = 10.0
dt = 0.05
temp = 0.0 # disables noise
x_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
intg = integrator.LangevinIntegrator(masses, x_ph, [hb, ha], dt,
friction, temp)
dx_op, dxdp_op = intg.step_op()
num_steps = 400
sess = tf.Session()
sess.run(tf.initializers.global_variables())
x = x0
for step in range(num_steps):
dx_val, dxdp_val = sess.run([dx_op, dxdp_op], feed_dict={x_ph: x})
x += dx_val
# test idealized bond distances
bonds = x - x[0, :]
bond_lengths = np.linalg.norm(bonds[1:, :], axis=1)
np.testing.assert_almost_equal(bond_lengths,
np.array([1.2, 1.0]),
decimal=4)
cj = np.take(x, ha.angle_idxs[:, 0], axis=0)
ci = np.take(x, ha.angle_idxs[:, 1], axis=0)
ck = np.take(x, ha.angle_idxs[:, 2], axis=0)
vij = cj - ci
vik = ck - ci
top = np.sum(vij * vik, -1)
bot = np.linalg.norm(vij, axis=-1) * np.linalg.norm(vik, axis=-1)
angles = np.arccos(top / bot)
np.testing.assert_almost_equal(angles,
np.array([ideal_angle] * 1),
decimal=2)
def test_methane(self):
"""
Testing minimization of CH4 into a tetrahedral geometry.
"""
masses = np.array([6.0, 1.0, 1.0, 1.0, 1.0])
x0 = np.array([[0.0637, 0.0126, 0.2203], [1.0573, -0.2011, 1.2864],
[2.3928, 1.2209, -0.2230], [-0.6891, 1.6983, 0.0780],
[-0.6312, -1.6261, -0.2601]],
dtype=np.float64)
num_atoms = len(masses)
bond_params = [
tf.get_variable("HH_kb",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(100.0)),
tf.get_variable("HH_b0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(1.0)),
]
hb = bonded_force.HarmonicBondForce(
params=bond_params,
bond_idxs=np.array([[0, 1], [0, 2], [0, 3], [0, 4]],
dtype=np.int32),
param_idxs=np.array([[0, 1], [0, 1], [0, 1], [0, 1]],
dtype=np.int32))
ideal_angle = 1.9111355
angle_params = [
tf.get_variable("HCH_ka",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(75.0)),
tf.get_variable("HCH_a0",
shape=tuple(),
dtype=tf.float64,
initializer=tf.constant_initializer(ideal_angle)),
]
ha = bonded_force.HarmonicAngleForce(
params=angle_params,
angle_idxs=np.array([[1, 0, 2], [1, 0, 3], [1, 0, 4], [2, 0, 3],
[2, 0, 4], [3, 0, 4]],
dtype=np.int32),
param_idxs=np.array(
[[0, 1], [0, 1], [0, 1], [0, 1], [0, 1], [0, 1]],
dtype=np.int32))
friction = 10.0
dt = 0.005
temp = 0.0
x_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
intg = integrator.LangevinIntegrator(masses, x_ph, [hb, ha], dt,
friction, temp)
dx_op, dxdp_op = intg.step_op()
num_steps = 1000
sess = tf.Session()
sess.run(tf.initializers.global_variables())
x = x0
for step in range(num_steps):
dx_val, dxdp_val = sess.run([dx_op, dxdp_op], feed_dict={x_ph: x})
x += dx_val
# test idealized bond distances
bonds = x - x[0, :]
bond_lengths = np.linalg.norm(bonds[1:, :], axis=1)
np.testing.assert_almost_equal(bond_lengths,
np.array([1.0] * 4),
decimal=4)
cj = np.take(x, ha.angle_idxs[:, 0], axis=0)
ci = np.take(x, ha.angle_idxs[:, 1], axis=0)
ck = np.take(x, ha.angle_idxs[:, 2], axis=0)
vij = cj - ci
vik = ck - ci
top = np.sum(vij * vik, -1)
bot = np.linalg.norm(vij, axis=-1) * np.linalg.norm(vik, axis=-1)
angles = np.arccos(top / bot)
np.testing.assert_almost_equal(angles,
np.array([ideal_angle] * 6),
decimal=2)
def test_ten_steps(self):
"""
Testing against reference implementation.
"""
friction = 10.0
dt = 0.003
temp = 0.0
num_atoms = len(self.masses)
x_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
hb = self.hb
ha = self.ha
ref_intg = ReferenceLangevinIntegrator(self.masses, dt, friction, temp)
num_steps = 10
x = x_ph
for step in range(num_steps):
all_grads = []
for force in [self.hb, self.ha]:
all_grads.append(force.gradients(x))
all_grads = tf.stack(all_grads, axis=0)
grads = tf.reduce_sum(all_grads, axis=0)
dx = ref_intg.step(grads)
x += dx
ref_x_final_op = x
# verify correctness of jacobians through time
ref_dxdp_hb_op = jacobian(x, hb.get_params(), use_pfor=False)
ref_dxdp_ha_op = jacobian(x, ha.get_params(), use_pfor=False)
test_intg = integrator.LangevinIntegrator(self.masses, x_ph, [hb, ha],
dt, friction, temp)
dx_op, dxdps_op = test_intg.step_op()
# dxdps_op = tf.reduce_sum(dxdps_op, axis=[1,2])
sess = tf.Session()
sess.run(tf.initializers.global_variables())
ref_x_final, ref_dxdp_hb, ref_dxdp_ha = sess.run(
[ref_x_final_op, ref_dxdp_hb_op, ref_dxdp_ha_op],
feed_dict={x_ph: self.x0})
x = np.copy(
self.x0
) # this copy is super important else it just modifies everything in place
for step in range(num_steps):
dx_val, dxdp_val = sess.run([dx_op, dxdps_op], feed_dict={x_ph: x})
x += dx_val
test_dxdp = dxdp_val
test_x_final_val = x
np.testing.assert_array_almost_equal(ref_x_final,
test_x_final_val,
decimal=14)
np.testing.assert_array_almost_equal(np.concatenate(
[ref_dxdp_hb, ref_dxdp_ha]),
test_dxdp,
decimal=14) # BAD, restore to 13
# test grads_and_vars and computation of higher derivatives
x_opt = np.array([
[-0.0070, -0.0100, 0.0000],
[-0.1604, 0.4921, 0.0000],
[0.5175, 0.0128, 0.0000],
],
dtype=np.float64) # idealized geometry
def loss(pred_x):
# Compute pairwise distances
def dij(x):
v01 = x[0] - x[1]
v02 = x[0] - x[2]
v12 = x[1] - x[2]
return tf.stack([tf.norm(v01), tf.norm(v02), tf.norm(v12)])
return tf.norm(dij(x_opt) - dij(pred_x))
x_final_ph = tf.placeholder(dtype=tf.float64, shape=(num_atoms, 3))
l0 = loss(ref_x_final_op)
l1 = loss(x_final_ph)
ref_dLdp_op = tf.gradients(
l0, self.hb.params +
self.ha.params) # goes through reference integrator
test_dLdx_op = tf.gradients(l1, x_final_ph)
test_dLdp_op_gvs = test_intg.grads_and_vars(
test_dLdx_op[0]) # multiply with dxdp
# need to fix this test.
ref_dLdp = sess.run(ref_dLdp_op, feed_dict={x_ph: self.x0})
test_dLdp = sess.run([a[0] for a in test_dLdp_op_gvs],
feed_dict={x_final_ph: test_x_final_val})
np.testing.assert_array_almost_equal(ref_dLdp, test_dLdp)
