def test_cell_shapes(self):
"""Check that all the NeuralStackCell tensor shapes are correct.
"""
batch_size = 5
embedding_size = 4
memory_size = 12
num_units = 8
deque = neural_stack.NeuralDequeCell(num_units, memory_size,
embedding_size)
deque.build(None)
self.assertEqual([1, 1, memory_size, memory_size],
deque.get_read_mask(0).shape)
self.assertEqual([1, 1, memory_size, memory_size],
deque.get_read_mask(1).shape)
deque_input = tf.zeros([batch_size, 1, embedding_size],
dtype=tf.float32)
zero_state = deque.zero_state(batch_size, tf.float32)
(outputs, (deque_next_state)) = deque.call(deque_input, zero_state)
# Make sure that deque output shapes match deque input shapes
self.assertEqual(outputs.shape, deque_input.shape)
assert_cell_shapes(self, deque_next_state, zero_state)
def test_enqueue_dequeue(self):
"""Test enqueueing a dequeueing from a NeuralDequeCell.
The sequence of operations is:
enqueue_bottom([1.0, 0.0, 0.0, 0.0])
enqueue_bottom([0.0, 1.0, 0.0, 0.0])
enqueue_bottom([0.0, 0.0, 1.0, 0.0])
enqueue_top([0.0, 0.0, 0.0, 1.0])
dequeue_top()
dequeue_top()
"""
input_values = np.array([[[[1.0, 0.0, 0.0, 0.0]],
[[0.0, 1.0, 0.0, 0.0]],
[[0.0, 0.0, 1.0, 0.0]],
[[0.0, 0.0, 0.0, 1.0]],
[[0.0, 0.0, 0.0, 0.0]],
[[0.0, 0.0, 0.0, 0.0]]]])
expected_values = np.array([[[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 1.0, 0.0],
[0.0, 1.0, 0.0, 0.0],
[1.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 1.0],
[0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0]]])
expected_read_strengths = np.array([[[[0.0], [0.0], [0.0], [1.0], [1.0],
[0.0], [0.0], [0.0], [0.0], [0.0],
[0.0], [0.0]]]])
expected_write_strengths = np.array([[[[0.0], [0.0], [0.0], [0.0], [0.0],
[0.0], [0.0], [0.0], [0.0], [0.0],
[0.0], [1.0]],
[[1.0], [0.0], [0.0], [0.0], [0.0],
[0.0], [0.0], [0.0], [0.0], [0.0],
[0.0], [0.0]]]])
expected_read_values = np.array([[[0.0, 0.0, 1.0, 0.0],
[0.0, 1.0, 0.0, 0.0]]])
batch_size = input_values.shape[0]
memory_size = input_values.shape[1] * 2
embedding_size = input_values.shape[3]
num_units = 8
deque = neural_stack.NeuralDequeCell(num_units, memory_size, embedding_size)
rnn_input = tf.constant(input_values, dtype=tf.float32)
deque_zero_state = tf.zeros([batch_size, num_units])
controller_outputs = deque.call_controller(None, None,
deque_zero_state,
batch_size)
assert_controller_shapes(self, controller_outputs,
deque.get_controller_shape(batch_size))
(outputs, state) = tf.nn.dynamic_rnn(cell=deque,
inputs=rnn_input,
time_major=False,
dtype=tf.float32)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
_, state_vals = sess.run([outputs, state])
(_, read_values,
memory_values,
read_strengths,
write_strengths) = state_vals
print(read_values)
self.assertAllClose(expected_values, memory_values)
self.assertAllClose(expected_write_strengths, write_strengths)
self.assertAllClose(expected_read_strengths, read_strengths)
self.assertAllClose(expected_read_values, read_values)