def setUp(self):
auto_np.random.seed(1234)
self.nn = Dnn()
self.input_layer = DenseLayer(2, 3, activation=sigmoid)
self.middle_layer = DenseLayer(3, 4, activation=relu)
self.output_layer = DenseLayer(4, 1)
self.nn.add_layer(self.input_layer)
self.nn.add_layer(self.middle_layer)
self.nn.add_layer(self.output_layer)
self.W1 = self.nn.layers[0].weights
self.b1 = self.nn.layers[0].biases
self.W2 = self.nn.layers[1].weights
self.b2 = self.nn.layers[1].biases
self.W3 = self.nn.layers[2].weights
self.b3 = self.nn.layers[2].biases
self.params = [self.W1, self.b1, self.W2, self.b2, self.W3, self.b3]
def f(x, w1, b1, w2, b2, w3, b3):
z1 = x @ w1 + b1
z2 = sigmoid_np(z1) @ w2 + b2
z3 = relu_np(z2) @ w3 + b3
return z3
self.f_np = f
def test_symmetry(P, D):
dnn = Dnn()
layer = DenseLayer(P * D, 1)
dnn.add_layer(layer)
sdnn = InputSorter(dnn)
# Random system, sorted by distance to origin.
s = np.random.randn(P, D)
s = s[np.argsort([row.dot(row) for row in s])]
# Limit the number of permutations, for computational feasibility.
for _, p in zip(range(100), permutations(s)):
p = np.asarray(p)
# Permutation invariant:
assert sdnn(s) == sdnn(p)
# Values should be equal to dnn applied to the pre-sorted system.
assert sdnn(p) == dnn(s)
assert sdnn.laplacian(p) == dnn.laplacian(s)
assert all(sdnn.gradient(p) == dnn.gradient(s))
# Drift force should have equal values, but permuted according to the permutation p.
# This because the drift force depends on which positions are which. Values should
# still be the same, so we only allow for sequence particle-wise permutations.
sdnn_drift, dnn_drift = (
sdnn.drift_force(p).reshape(s.shape),
dnn.drift_force(s).reshape(s.shape),
)
sdnn_drift.sort(axis=0)
dnn_drift.sort(axis=0)
assert (sdnn_drift == dnn_drift).all()
P, D = 2, 2 # Particles, dimensions
system = np.empty((P, D))
H = CoulombHarmonicOscillator()
# Wave functions:
simple_gaussian = SimpleGaussian(alpha=0.5)
jastrow = JastrowPade(alpha=1, beta=1)
simple_and_jastrow = WavefunctionProduct(simple_gaussian, jastrow)
layers = [
DenseLayer(P * D, 32, activation=tanh, scale_factor=0.001),
DenseLayer(32, 16, activation=tanh),
DenseLayer(16, 1, activation=exponential),
]
dnn = Dnn()
for l in layers:
dnn.add_layer(l)
psi = WavefunctionProduct(simple_and_jastrow, dnn)
psi_sampler = ImportanceSampler(system, psi, step_size=0.1)
# Sorted
simple_gaussian2 = SimpleGaussian(alpha=0.5)
jastrow2 = JastrowPade(alpha=1, beta=1)
simple_and_jastrow2 = WavefunctionProduct(simple_gaussian2, jastrow2)
layers2 = [
DenseLayer(P * D, 32, activation=tanh, scale_factor=0.001),
DenseLayer(32, 16, activation=tanh),
DenseLayer(16, 1, activation=exponential),
]
class TestDnn(unittest.TestCase):
def setUp(self):
auto_np.random.seed(1234)
self.nn = Dnn()
self.input_layer = DenseLayer(2, 3, activation=sigmoid)
self.middle_layer = DenseLayer(3, 4, activation=relu)
self.output_layer = DenseLayer(4, 1)
self.nn.add_layer(self.input_layer)
self.nn.add_layer(self.middle_layer)
self.nn.add_layer(self.output_layer)
self.W1 = self.nn.layers[0].weights
self.b1 = self.nn.layers[0].biases
self.W2 = self.nn.layers[1].weights
self.b2 = self.nn.layers[1].biases
self.W3 = self.nn.layers[2].weights
self.b3 = self.nn.layers[2].biases
self.params = [self.W1, self.b1, self.W2, self.b2, self.W3, self.b3]
def f(x, w1, b1, w2, b2, w3, b3):
z1 = x @ w1 + b1
z2 = sigmoid_np(z1) @ w2 + b2
z3 = relu_np(z2) @ w3 + b3
return z3
self.f_np = f
def test_evaluate(self):
for _ in range(10):
for x in auto_np.random.randn(500, 2):
np.testing.assert_almost_equal(self.f_np(x, *self.params),
self.nn(x))
def test_parameter_gradients(self):
for _ in range(10):
x = auto_np.random.randn(1, 2)
output = self.nn(x)
grads = self.nn.gradient(x)
# Compute gradients using autograd
auto_grads = []
for i in range(len(self.nn.layers)):
W_grad = elementwise_grad(self.f_np, 1 + 2 * i)(x,
*self.params)
b_grad = elementwise_grad(self.f_np,
1 + 2 * i + 1)(x, *self.params)
auto_grads.extend(W_grad.ravel())
auto_grads.extend(b_grad.ravel())
# Test each indivudual layer
np.testing.assert_almost_equal(
W_grad, self.nn.layers[i].weights_gradient)
np.testing.assert_almost_equal(b_grad,
self.nn.layers[i].bias_gradient)
# Test extraction of full gradient vector
np.testing.assert_almost_equal(auto_grads, grads * output)
def test_drift_force(self):
for _ in range(10):
x = auto_np.random.randn(1, 2)
# Autograd computes gradient per row of x. Want the sum over rows.
auto_gradient = np.sum(elementwise_grad(self.f_np,
0)(x, *self.params),
axis=0)
np.testing.assert_almost_equal(
auto_gradient,
self.nn.drift_force(x) * self.nn(x) / 2)
def test_laplace(self):
# Autograd makes some warnings about code that is not ours. Ignore them here.
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=FutureWarning)
hess = hessian(self.f_np)
for _ in range(10):
x = auto_np.random.randn(
1, 2) # Autograd hessian slow, less testing.
output = self.nn(x)
# Need to feed autograd hessian one row at a time and sum results.
expected = sum(
np.trace(hess(x[i], *self.params)[0])
for i in range(x.shape[0]))
self.assertAlmostEqual(expected, self.nn.laplacian(x) * output)