def testCustomNoiseShape(self):
inputs = array_ops.ones((5, 3, 2))
noise_shape = [5, 1, 2]
dp = core_layers.Dropout(0.5, noise_shape=noise_shape, seed=1)
dropped = dp.apply(inputs, training=True)
self.evaluate(variables.global_variables_initializer())
np_output = self.evaluate(dropped)
self.assertAlmostEqual(0., np_output.min())
self.assertAllClose(np_output[:, 0, :], np_output[:, 1, :])
def testDynamicRate(self):
with self.cached_session() as sess:
rate = array_ops.placeholder(dtype='float32', name='rate')
dp = core_layers.Dropout(rate, name='dropout')
inputs = array_ops.ones((5, 5))
dropped = dp.apply(inputs, training=True)
self.evaluate(variables.global_variables_initializer())
np_output = sess.run(dropped, feed_dict={rate: 0.5})
self.assertAlmostEqual(0., np_output.min())
np_output = sess.run(dropped, feed_dict={rate: 0.0})
self.assertAllClose(np.ones((5, 5)), np_output)
def testBooleanLearningPhase(self):
dp = core_layers.Dropout(0.5)
inputs = array_ops.ones((5, 3))
dropped = dp.apply(inputs, training=True)
if not context.executing_eagerly():
self.evaluate(variables.global_variables_initializer())
np_output = self.evaluate(dropped)
self.assertAlmostEqual(0., np_output.min())
dropped = dp.apply(inputs, training=False)
np_output = self.evaluate(dropped)
self.assertAllClose(np.ones((5, 3)), np_output)
def testDynamicLearningPhase(self):
with self.cached_session() as sess:
dp = core_layers.Dropout(0.5, seed=1)
inputs = array_ops.ones((5, 5))
training = array_ops.placeholder(dtype='bool')
dropped = dp.apply(inputs, training=training)
self.evaluate(variables.global_variables_initializer())
np_output = sess.run(dropped, feed_dict={training: True})
self.assertAlmostEqual(0., np_output.min())
np_output = sess.run(dropped, feed_dict={training: False})
self.assertAllClose(np.ones((5, 5)), np_output)
def testDropoutProperties(self):
dp = core_layers.Dropout(0.5, name='dropout')
self.assertEqual(dp.rate, 0.5)
self.assertEqual(dp.noise_shape, None)
dp.apply(array_ops.ones(()))
self.assertEqual(dp.name, 'dropout')
