def test_siren_initialization_with_w0():
tf.random.set_seed(0)
w0 = 2.0
c = 6.0
samples = 10
dim = 128
activation = tf_siren.Sine(w0=w0, dtype=tf.float32)
initializer = tf_siren.SIRENInitializer(w0=w0, c=c, seed=0)
x = tf.random.uniform(shape=[samples, dim], minval=-1.0, maxval=1.)
out = x
# weight matrix
w = initializer(shape=[dim, dim], dtype=tf.float32)
for i in range(100): # Simulate 100 layer SIREN forward pass
# for w0 > 1, we multiply w0 by weight matrix before forward pass
out = tf.matmul(
out, w0 *
w) # The paper suggests output should have zero mean, unit std.
out_np = out.numpy()[0] # select just one sample to test later
# print(i + 1, out_np.mean(), out_np.std())
out = activation(out)
# Simulate new layer with new weights for next pass
w = initializer(shape=[dim, dim], dtype=tf.float32)
out = out_np
check_dist = tf.random.normal(out.shape, stddev=1.0).numpy()
# Compute 2 tailed z statistic
means = out.mean() - check_dist.mean()
stds = np.sqrt((np.square(out.std()) + np.square(check_dist.std())))
z_statistic = abs(means / (stds))
# multiply by 2.0 as it is a 2 tailed test
probability_same_dist = 2 * (1. - normal_dist.cdf(z_statistic))
assert probability_same_dist >= 0.90 # more than 90% confident the two distributions have same mean and std
def test_sine_activation():
x = tf.constant([0.0, np.pi / 2.0, np.pi])
activation = tf_siren.Sine(w0=1.0)
y = activation(x)
assert (tf.reduce_sum(y - [0.0, 1.0, 0.0])**2) <= 1e-6
def test_sine_activation_with_w0():
x = tf.constant([0.0, np.pi / 2.0, np.pi])
activation = tf_siren.Sine(w0=1.5)
y = activation(x)
assert (tf.reduce_sum(y - [0.0, 1. / np.sqrt(2.0), -1.0])**2) <= 1e-6