def test_actnorm_3d(self):
with tf.Graph().as_default():
x_t = tf.random_normal((16, 5, 32, 32, 3), mean=50.0, stddev=2.0)
ops = [glow_ops.actnorm, glow_ops.get_variable_ddi]
with arg_scope(ops, init=True):
x_act, _ = glow_ops.actnorm_3d("actnorm", x_t)
with tf.Session() as session:
x_act_np = session.run(x_act)
# Mean and standard deviation per time-step equals zero and one.
for time_step in range(5):
x_act_curr = x_act_np[:, time_step, :, :, :]
channel_mean = np.mean(x_act_curr, axis=(0, 1, 2))
channel_var = np.var(x_act_curr, axis=(0, 1, 2))
self.assertTrue(np.allclose(channel_mean, 0.0, atol=1e-3))
self.assertTrue(np.allclose(channel_var, 1.0, atol=1e-3))
def test_actnorm_3d(self):
with tf.Graph().as_default():
x_t = tf.random_normal((16, 5, 32, 32, 3), mean=50.0, stddev=2.0)
ops = [glow_ops.actnorm, glow_ops.get_variable_ddi]
with arg_scope(ops, init=True):
x_act, _ = glow_ops.actnorm_3d("actnorm", x_t)
with tf.Session() as session:
x_act_np = session.run(x_act)
# Mean and standard deviation per time-step equals zero and one.
for time_step in range(5):
x_act_curr = x_act_np[:, time_step, :, :, :]
channel_mean = np.mean(x_act_curr, axis=(0, 1, 2))
channel_var = np.var(x_act_curr, axis=(0, 1, 2))
self.assertTrue(np.allclose(channel_mean, 0.0, atol=1e-3))
self.assertTrue(np.allclose(channel_var, 1.0, atol=1e-3))
