def test_lookup_weighting(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
initial_mem = np.random.uniform(0, 1,
(2, 4, 5)).astype(np.float32)
keys = np.random.uniform(0, 1, (2, 5, 2)).astype(np.float32)
strengths = np.random.uniform(0, 1, (2, 2)).astype(np.float32)
norm_mem = initial_mem / np.sqrt(
np.sum(initial_mem**2, axis=2, keepdims=True))
norm_keys = keys / np.sqrt(
np.sum(keys**2, axis=1, keepdims=True))
sim = np.matmul(norm_mem, norm_keys)
sim = sim * strengths[:, np.newaxis, :]
predicted_wieghts = np.exp(sim) / np.sum(
np.exp(sim), axis=1, keepdims=True)
mem.memory_matrix = tf.convert_to_tensor(initial_mem)
op = mem.get_lookup_weighting(keys, strengths)
c = session.run(op)
self.assertEqual(c.shape, (2, 4, 2))
self.assertTrue(np.allclose(c, predicted_wieghts))
def test_update_memory(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
write_weighting = random_softmax((2, 4), axis=1)
write_vector = np.random.uniform(0, 1,
(2, 5)).astype(np.float32)
erase_vector = np.random.uniform(0, 1,
(2, 5)).astype(np.float32)
memory_matrix = np.random.uniform(-1, 1,
(2, 4, 5)).astype(np.float32)
ww = write_weighting[:, :, np.newaxis]
v, e = write_vector[:,
np.newaxis, :], erase_vector[:,
np.newaxis, :]
predicted = memory_matrix * (1 - np.matmul(ww, e)) + np.matmul(
ww, v)
mem.memory_matrix = tf.convert_to_tensor(memory_matrix)
op = mem.update_memory(write_weighting, write_vector,
erase_vector)
M = session.run(op)
#updated_memory = session.run(mem.memory_matrix.value())
self.assertEqual(M.shape, (2, 4, 5))
self.assertTrue(np.allclose(M, predicted))