def test_shape_of_output_of_6d_rnn(self):
units = 7
last_dim_size = 12
rnn = MDRNN(units=units, input_shape=(None, None, None, None, None, None, last_dim_size),
return_sequences=True,
activation='tanh')
x = tf.zeros(shape=(2, 3, 1, 2, 2, 1, 5, last_dim_size))
result = rnn.call(x)
self.assertEqual((2, 3, 1, 2, 2, 1, 5, units), result.shape)
def test_shape(self):
rnn2d = MDRNN(units=5, input_shape=(None, None, 1),
kernel_initializer=initializers.Constant(1),
recurrent_initializer=initializers.Constant(1),
bias_initializer=initializers.Constant(-1),
return_sequences=True,
activation='tanh')
x = np.arange(6).reshape((1, 2, 3, 1))
res = rnn2d.call(x)
self.assertEqual((1, 2, 3, 5), res.shape)
def test_for_two_step_sequence(self):
kernel_initializer = initializers.Zeros()
recurrent_initializer = kernel_initializer
mdrnn = MDRNN(units=2,
input_shape=(None, 1),
kernel_initializer=kernel_initializer,
recurrent_initializer=recurrent_initializer,
bias_initializer=initializers.Constant(5),
return_sequences=True)
x = np.zeros((1, 3, 1))
a = mdrnn.call(x)
expected_result = np.ones((1, 3, 2)) * 0.9999
np.testing.assert_almost_equal(expected_result, a.numpy(), 4)
def test_result(self):
rnn2d = MDRNN(units=1, input_shape=(None, None, 1),
kernel_initializer=initializers.Identity(),
recurrent_initializer=initializers.Identity(1),
bias_initializer=initializers.Constant(-1),
return_sequences=True,
activation=None)
x = np.arange(6).reshape((1, 2, 3, 1))
actual = rnn2d.call(x)
desired = np.array([
[-1, -1, 0],
[1, 3, 7]
]).reshape((1, 2, 3, 1))
np.testing.assert_almost_equal(desired, actual.numpy(), 6)
def test_2drnn_output_when_providing_initial_state(self):
rnn2d = MDRNN(units=1, input_shape=(None, None, 1),
kernel_initializer=initializers.Identity(),
recurrent_initializer=initializers.Identity(1),
bias_initializer=initializers.Constant(-1),
return_sequences=True,
activation=None)
x = np.arange(6).reshape((1, 2, 3, 1))
initial_state = [tf.ones(shape=(1, 1)), tf.ones(shape=(1, 1))]
actual = rnn2d.call(x, initial_state=initial_state)
desired = np.array([
[1, 1, 2],
[3, 7, 13]
]).reshape((1, 2, 3, 1))
np.testing.assert_almost_equal(desired, actual.numpy(), 6)
def test_result_after_running_rnn_on_3d_input(self):
rnn3d = MDRNN(units=1,
input_shape=(None, None, None, 1),
kernel_initializer=initializers.Identity(),
recurrent_initializer=initializers.Identity(1),
bias_initializer=initializers.Constant(1),
return_sequences=True,
return_state=True,
activation=None)
x = np.arange(2 * 2 * 2).reshape((1, 2, 2, 2, 1))
outputs, state = rnn3d.call(x)
desired = np.array([[[1, 3], [4, 11]], [[6, 15], [17, 51]]]).reshape(
(1, 2, 2, 2, 1))
np.testing.assert_almost_equal(desired, outputs.numpy(), 6)
desired_state = desired[:, -1, -1, -1]
np.testing.assert_almost_equal(desired_state, state.numpy(), 6)
def test_1d_rnn_produces_correct_output_for_2_steps(self):
kernel_initializer = initializers.identity()
recurrent_initializer = kernel_initializer
bias = 3
bias_initializer = initializers.Constant(bias)
mdrnn = MDRNN(units=3,
input_shape=(None, 3),
kernel_initializer=kernel_initializer,
recurrent_initializer=recurrent_initializer,
bias_initializer=bias_initializer,
activation=None,
return_sequences=True)
x1 = np.array([1, 2, 4])
x2 = np.array([9, 8, 6])
x = np.array([x1, x2])
a = mdrnn.call(x.reshape((1, 2, -1)))
expected_result = np.array([[x1 + bias, x1 + x2 + 2 * bias]])
np.testing.assert_almost_equal(expected_result, a.numpy(), 8)
def test_feeding_layer_created_with_default_initializer(self):
mdrnn = MDRNN(units=2, input_shape=(None, 1))
x = np.zeros((1, 1, 1)) * 0.5
mdrnn.call(x)