def test_high_entropy_bounds(self): # For high entropy distributions, the training bound should be very tight, # and the overhead of range coding manageable. noisy = uniform_noise.NoisyNormal(loc=0., scale=100.) em = ContinuousBatchedEntropyModel(noisy, 1, compression=True) x = noisy.base.sample([10000]) bits_eval = em.bits(x, training=False) bits_training = em.bits(x, training=True) bits_compressed = 8 * len(em.compress(x).numpy()) self.assertAllClose(bits_training, bits_eval, atol=0, rtol=5e-5) self.assertAllClose(bits_compressed, bits_eval, atol=0, rtol=5e-3)
def test_information_bounds(self):
# `bits(training=True)` should be greater than `bits(training=False)`
# because it is defined as an upper bound (albeit for infinite data). The
# actual length of the bit string should always be greater than
# `bits(training=False)` because range coding is only asymptotically
# optimal, and because it operates on quantized probabilities.
for scale in 2 ** tf.linspace(-2., 7., 10):
noisy = uniform_noise.NoisyNormal(loc=0., scale=scale)
em = ContinuousBatchedEntropyModel(noisy, 1, compression=True)
x = noisy.base.sample([10000])
bits_eval = em.bits(x, training=False)
bits_training = em.bits(x, training=True)
bits_compressed = 8 * len(em.compress(x).numpy())
self.assertGreater(bits_training, .9975 * bits_eval)
self.assertGreater(bits_compressed, bits_eval)