def test_construction(self):
interface = DNC.interface(
read_keys=None,
read_strengths=None,
write_key=np.random.uniform(0, 1, (3, 9, 1)).astype(np.float32),
write_strength=np.random.uniform(0, 1, (3, 1)).astype(np.float32),
erase_vector=tf.convert_to_tensor(
np.zeros((3, 9)).astype(np.float32)),
write_vector=tf.convert_to_tensor(
np.random.uniform(0, 1, (3, 9)).astype(np.float32)),
free_gates=np.random.uniform(0, 1, (3, 5)).astype(np.float32),
allocation_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32),
write_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32),
read_modes=None,
)
memory = Memory(13, 9, 5)
memory_state = memory.get_initial_state(batch_size=3)
usage, write_weighting, memory, link_matrix, precedence = memory.write(
memory_state, interface)
self.assertEqual(usage.shape, (3, 13))
self.assertEqual(write_weighting.shape, (3, 13))
self.assertEqual(memory.shape, (3, 13, 9))
self.assertEqual(link_matrix.shape, (3, 13, 13))
self.assertEqual(precedence.shape, (3, 13))
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)
memory_matrix = tf.convert_to_tensor(memory_matrix)
op = mem.update_memory(memory_matrix, write_weighting,
write_vector, erase_vector)
M = session.run(op)
self.assertEqual(M.shape, (2, 4, 5))
self.assertTrue(np.allclose(M, predicted))
def test_write(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 1)
M, u, p, L, ww, rw, r = session.run(mem.init_memory())
key = np.random.uniform(0, 1, (1, 5, 1)).astype(np.float32)
strength = np.random.uniform(0, 1, (1, 1)).astype(np.float32)
free_gates = np.random.uniform(0, 1, (1, 2)).astype(np.float32)
write_gate = np.random.uniform(0, 1, (1, 1)).astype(np.float32)
allocation_gate = np.random.uniform(0, 1,
(1, 1)).astype(np.float32)
write_vector = np.random.uniform(0, 1,
(1, 5)).astype(np.float32)
erase_vector = np.zeros((1, 5)).astype(np.float32)
u_op, ww_op, M_op, L_op, p_op = mem.write(
M, u, rw, ww, p, L, key, strength, free_gates,
allocation_gate, write_gate, write_vector, erase_vector)
session.run(tf.initialize_all_variables())
u, ww, M, L, p = session.run([u_op, ww_op, M_op, L_op, p_op])
self.assertEqual(u.shape, (1, 4))
self.assertEqual(ww.shape, (1, 4))
self.assertEqual(M.shape, (1, 4, 5))
self.assertEqual(L.shape, (1, 4, 4))
self.assertEqual(p.shape, (1, 4))
def test_construction(self):
interface = DNC.interface(
read_keys=None,
read_strengths=None,
write_key=np.random.uniform(0, 1, (3, 9, 1)).astype(np.float32),
write_strength=np.random.uniform(0, 1, (3, 1)).astype(np.float32),
erase_vector=tf.convert_to_tensor(
np.zeros((3, 9)).astype(np.float32)),
write_vector=tf.convert_to_tensor(
np.random.uniform(0, 1, (3, 9)).astype(np.float32)),
free_gates=np.random.uniform(0, 1, (3, 5)).astype(np.float32),
allocation_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32),
write_gate=np.random.uniform(0, 1, (3, 1)).astype(np.float32),
read_modes=None,
)
memory = Memory(13, 9, 5)
memory_state = memory.initial_state(3)
write_op = memory.write(memory_state, interface)
init_op = tf.global_variables_initializer()
with self.test_session() as session:
init_op.run()
usage, write_weighting, memory, link_matrix, precedence = session.run(
write_op)
self.assertEqual(usage.shape, (3, 13))
self.assertEqual(write_weighting.shape, (3, 13))
self.assertEqual(memory.shape, (3, 13, 9))
self.assertEqual(link_matrix.shape, (3, 13, 13))
self.assertEqual(precedence.shape, (3, 13))
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)
memory_matrix = tf.convert_to_tensor(initial_mem)
op = mem.get_lookup_weighting(memory_matrix, keys, strengths)
c = session.run(op)
self.assertEqual(c.shape, (2, 4, 2))
self.assertTrue(np.allclose(c, predicted_wieghts))
def test_update_usage_vector(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
free_gates = np.random.uniform(0, 1, (2, 2)).astype(np.float32)
init_read_weightings = random_softmax((2, 4, 2), axis=1)
init_write_weightings = random_softmax((2, 4), axis=1)
init_usage = np.random.uniform(0, 1, (2, 4)).astype(np.float32)
psi = np.product(
1 - init_read_weightings * free_gates[:, np.newaxis, :],
axis=2)
predicted_usage = (init_usage + init_write_weightings -
init_usage * init_write_weightings) * psi
read_weightings = tf.convert_to_tensor(init_read_weightings)
write_weighting = tf.convert_to_tensor(init_write_weightings)
usage_vector = tf.convert_to_tensor(init_usage)
op = mem.update_usage_vector(usage_vector, read_weightings,
write_weighting, free_gates)
u = session.run(op)
self.assertEqual(u.shape, (2, 4))
self.assertTrue(np.array_equal(u, predicted_usage))
def test_update_link_matrix(self):
graph = tf.Graph()
with graph.as_default():
with tf.compat.v1.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
_write_weighting = random_softmax((2, 4), axis=1)
_precedence_vector = random_softmax((2, 4), axis=1)
initial_link = np.random.uniform(0, 1,
(2, 4, 4)).astype(np.float32)
np.fill_diagonal(initial_link[0, :], 0)
np.fill_diagonal(initial_link[1, :], 0)
# calculate the updated link iteratively as in paper
# to check the correctness of the vectorized implementation
predicted = np.zeros((2, 4, 4), dtype=np.float32)
for i in range(4):
for j in range(4):
if i != j:
reset_factor = (1 - _write_weighting[:, i] -
_write_weighting[:, j])
predicted[:, i, j] = reset_factor * initial_link[:, i, j] + _write_weighting[:, i] * \
_precedence_vector[:, j]
link_matrix = tf.convert_to_tensor(value=initial_link)
precedence_vector = tf.convert_to_tensor(
value=_precedence_vector)
write_weighting = tf.constant(_write_weighting)
op = mem.update_link_matrix(precedence_vector, link_matrix,
write_weighting)
L = session.run(op)
self.assertTrue(np.allclose(L, predicted))
def test_get_allocation_weighting(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
mock_usage = np.random.uniform(0.01, 1,
(2, 4)).astype(np.float32)
sorted_usage = np.sort(mock_usage, axis=1)
free_list = np.argsort(mock_usage, axis=1)
predicted_weights = np.zeros((2, 4)).astype(np.float32)
for i in range(2):
for j in range(4):
product_list = [
mock_usage[i, free_list[i, k]] for k in range(j)
]
predicted_weights[i, free_list[
i, j]] = (1 - mock_usage[i, free_list[i, j]]
) * np.product(product_list)
op = mem.get_allocation_weighting(sorted_usage, free_list)
a = session.run(op)
self.assertEqual(a.shape, (2, 4))
self.assertTrue(np.allclose(a, predicted_weights))
def __init__(self,
controller_class,
input_size,
output_size,
max_sequence_length,
memory_words_num=256,
memory_word_size=64,
memory_read_heads=4,
batch_size=128):
"""
constructs a complete DNC architecture as described in the DNC paper
http://www.nature.com/nature/journal/vaop/ncurrent/full/nature20101.html
Parameters:
-----------
controller_class: BaseController
a concrete implementation of the BaseController class
input_size: int
the size of the input vector
output_size: int
the size of the output vector
max_sequence_length: int
the maximum length of an input sequence
memory_words_num: int
the number of words that can be stored in memory
memory_word_size: int
the size of an individual word in memory
memory_read_heads: int
the number of read heads in the memory
batch_size: int
the size of the data batch
"""
self.input_size = input_size
self.output_size = output_size
self.max_sequence_length = max_sequence_length
self.words_num = memory_words_num
self.word_size = memory_word_size
self.read_heads = memory_read_heads
self.batch_size = batch_size
self.memory = Memory(self.words_num, self.word_size, self.read_heads,
self.batch_size)
self.controller = controller_class(self.input_size, self.output_size,
self.read_heads, self.word_size,
self.batch_size)
# input data placeholders
self.input_data = tf.placeholder(tf.float32, [None, None, chunk_size],
name='input')
self.target_output = tf.placeholder(tf.float32,
[None, None, output_size],
name='targets')
#self.input_data = tf.placeholder(tf.float32, [batch_size, None, input_size], name='input')
#self.target_output = tf.placeholder(tf.float32, [batch_size, None, output_size], name='targets')
self.sequence_length = tf.placeholder(tf.int32, name='sequence_length')
self.build_graph()
def test_init_memory(self):
memory = Memory(words_num=13, word_size=7, read_heads_num=2)
state = memory.get_initial_state(batch_size=9)
self.assertEqual(state.memory_matrix.shape, (9, 13, 7))
self.assertEqual(state.usage_vector.shape, (9, 13))
self.assertEqual(state.link_matrix.shape, (9, 13, 13))
self.assertEqual(state.precedence_vector.shape, (9, 13))
self.assertEqual(state.write_weighting.shape, (9, 13))
self.assertEqual(state.read_weightings.shape, (9, 13, 2))
def test_init_memory(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
M, u, p, L, ww, rw, r = session.run(mem.init_memory())
self.assertEqual(M.shape, (2, 4, 5))
self.assertEqual(u.shape, (2, 4))
self.assertEqual(L.shape, (2, 4, 4))
self.assertEqual(ww.shape, (2, 4))
self.assertEqual(rw.shape, (2, 4, 2))
self.assertEqual(r.shape, (2, 5, 2))
self.assertEqual(p.shape, (2, 4))
def test_construction(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
session.run(tf.initialize_all_variables())
self.assertEqual(mem.words_num, 4)
self.assertEqual(mem.word_size, 5)
self.assertEqual(mem.read_heads, 2)
self.assertEqual(mem.batch_size, 2)
self.assertEqual(mem.memory_matrix.get_shape().as_list(),
[2, 4, 5])
self.assertEqual(mem.usage_vector.get_shape().as_list(),
[2, 4])
self.assertEqual(mem.link_matrix.get_shape().as_list(),
[2, 4, 4])
self.assertEqual(mem.write_weighting.get_shape().as_list(),
[2, 4])
self.assertEqual(mem.read_weightings.get_shape().as_list(),
[2, 4, 2])
self.assertEqual(mem.read_vectors.get_shape().as_list(),
[2, 5, 2])
def test_update_read_vectors(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph = graph) as session:
mem = Memory(4, 5, 2, 4)
memory_matrix = np.random.uniform(-1, 1, (4, 4, 5)).astype(np.float32)
read_weightings = random_softmax((4, 4, 2), axis=1)
predicted = np.matmul(np.transpose(memory_matrix, [0, 2, 1]), read_weightings)
op = mem.update_read_vectors(memory_matrix, read_weightings)
session.run(tf.global_variables_initializer())
r = session.run(op)
#updated_read_vectors = session.run(mem.read_vectors.value())
self.assertTrue(np.allclose(r, predicted))
def test_update_precedence_vector(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)
initial_precedence = random_softmax((2, 4), axis=1)
predicted = (1 - write_weighting.sum(axis=1, keepdims=True)) * initial_precedence + write_weighting
precedence_vector = tf.convert_to_tensor(initial_precedence)
op = mem.update_precedence_vector(precedence_vector, write_weighting)
p = session.run(op)
self.assertEqual(p.shape, (2,4))
self.assertTrue(np.allclose(p, predicted))
def test_read(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph = graph) as session:
mem = Memory(4, 5, 2, 1)
M, u, p, L, ww, rw, r = session.run(mem.init_memory())
keys = np.random.uniform(0, 1, (1, 5, 2)).astype(np.float32)
strengths = np.random.uniform(0, 1, (1, 2)).astype(np.float32)
link_matrix = np.random.uniform(0, 1, (1, 4, 4)).astype(np.float32)
read_modes = random_softmax((1, 3, 2), axis=1).astype(np.float32)
memory_matrix = np.random.uniform(-1, 1, (1, 4, 5)).astype(np.float32)
wr_op, r_op = mem.read(memory_matrix, rw, keys, strengths, link_matrix, read_modes)
session.run(tf.global_variables_initializer())
wr, r = session.run([wr_op, r_op])
self.assertEqual(wr.shape, (1, 4, 2))
self.assertEqual(r.shape, (1, 5, 2))
def test_updated_write_weighting(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
write_gate = np.random.uniform(0, 1, (2,1)).astype(np.float32)
allocation_gate = np.random.uniform(0, 1, (2,1)).astype(np.float32)
lookup_weighting = random_softmax((2, 4, 1), axis=1)
allocation_weighting = random_softmax((2, 4), axis=1)
predicted_weights = write_gate * (allocation_gate * allocation_weighting + (1 - allocation_gate) * np.squeeze(lookup_weighting))
op = mem.update_write_weighting(lookup_weighting, allocation_weighting, write_gate, allocation_gate)
w_w = session.run(op)
self.assertEqual(w_w.shape, (2,4))
self.assertTrue(np.allclose(w_w, predicted_weights))
def test_get_directional_weightings(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
_link_matrix = np.random.uniform(0, 1, (2, 4, 4)).astype(np.float32)
_read_weightings = random_softmax((2, 4, 2), axis=1)
predicted_forward = np.matmul(_link_matrix, _read_weightings)
predicted_backward = np.matmul(np.transpose(_link_matrix, [0, 2, 1]), _read_weightings)
read_weightings = tf.convert_to_tensor(_read_weightings)
fop, bop = mem.get_directional_weightings(read_weightings, _link_matrix)
forward_weighting, backward_weighting = session.run([fop, bop])
self.assertTrue(np.allclose(forward_weighting, predicted_forward))
self.assertTrue(np.allclose(backward_weighting, predicted_backward))
def get_addressing_weights(m, i, link_matrix):
lookup_w = ContentAddressing.weighting(m.memory_matrix, i.read_keys,
i.read_strengths)
fwd_w, bkwd_w = TemporalLinkAddressing.weightings(
link_matrix, m.read_weightings)
read_w, read_v = Memory.read(m.memory_matrix, m.read_weightings,
link_matrix, i)
if not tf.executing_eagerly():
lookup_w = lookup_w.eval()
fwd_w, bkwd_w = fwd_w.eval(), bkwd_w.eval()
read_w, read_v = read_w.eval(), read_v.eval()
return lookup_w, fwd_w, bkwd_w, read_w, read_v
def test_read_vectors_and_weightings(self):
m = Memory.state(
memory_matrix=np.random.uniform(-1, 1,
(5, 11, 7)).astype(np.float32),
usage_vector=None,
link_matrix=None,
precedence_vector=None,
write_weighting=None,
read_weightings=DNCMemoryTests.softmax_sample((5, 11, 3), axis=1),
)
# pull out read_modes due to https://github.com/tensorflow/tensorflow/issues/1409
# hack to circumvent tf bug in not doing `convert_to_tensor` in einsum reductions correctly
read_modes = DNCMemoryTests.softmax_sample((5, 3, 3), axis=1)
i = DNC.interface(
read_keys=np.random.uniform(0, 1, (5, 7, 3)).astype(np.float32),
read_strengths=np.random.uniform(0, 1, (5, 3)).astype(np.float32),
write_key=None,
write_strength=None,
erase_vector=None,
write_vector=None,
free_gates=None,
allocation_gate=None,
write_gate=None,
read_modes=tf.convert_to_tensor(read_modes),
)
# read uses the link matrix that is produced after a write operation
new_link_matrix = np.random.uniform(0, 1,
(5, 11, 11)).astype(np.float32)
# assume ContentAddressing and TemporalLinkAddressing are already correct
lookup_weightings, forward_weighting, backward_weighting, \
updated_read_weightings, updated_read_vectors = self.get_addressing_weights(
m, i, new_link_matrix)
self.assertEqual(updated_read_weightings.shape, (5, 11, 3))
self.assertEqual(updated_read_vectors.shape, (5, 7, 3))
expected_read_weightings = np.zeros((5, 11, 3)).astype(np.float32)
for read_head in range(3):
backward_weight = read_modes[:, 0, read_head, np.
newaxis] * backward_weighting[:, :,
read_head]
lookup_weight = read_modes[:, 1, read_head, np.newaxis] * \
lookup_weightings[:, :, read_head]
forward_weight = read_modes[:, 2, read_head, np.newaxis] * \
forward_weighting[:, :, read_head]
expected_read_weightings[:, :, read_head] = backward_weight + \
lookup_weight + forward_weight
expected_read_vectors = np.matmul(
np.transpose(m.memory_matrix, [0, 2, 1]), updated_read_weightings)
self.assertAllClose(updated_read_weightings, expected_read_weightings)
self.assertEqual(updated_read_weightings.shape, (5, 11, 3))
self.assertAllClose(updated_read_vectors, expected_read_vectors)
def test_construction(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
session.run(tf.initialize_all_variables())
self.assertEqual(mem.words_num, 4)
self.assertEqual(mem.word_size, 5)
self.assertEqual(mem.read_heads, 2)
self.assertEqual(mem.batch_size, 2)
def test_update_read_weightings(self):
graph = tf.Graph()
with graph.as_default():
with tf.Session(graph=graph) as session:
mem = Memory(4, 5, 2, 2)
lookup_weightings = random_softmax((2, 4, 2), axis=1)
forward_weighting = random_softmax((2, 4, 2), axis=1)
backward_weighting = random_softmax((2, 4, 2), axis=1)
read_mode = random_softmax((2, 3, 2), axis=1)
predicted_weights = np.zeros((2, 4, 2)).astype(np.float32)
# calculate the predicted weights using iterative method from paper
# to check the correcteness of the vectorized implementation
for i in range(2):
predicted_weights[:, :,
i] = read_mode[:, 0, i, np.
newaxis] * backward_weighting[:, :,
i] + read_mode[:,
1,
i,
np
.
newaxis] * lookup_weightings[:, :,
i] + read_mode[:,
2,
i,
np
.
newaxis] * forward_weighting[:, :,
i]
op = mem.update_read_weightings(lookup_weightings,
forward_weighting,
backward_weighting, read_mode)
session.run(tf.initialize_all_variables())
w_r = session.run(op)
#updated_read_weightings = session.run(mem.read_weightings.value())
self.assertTrue(np.allclose(w_r, predicted_weights))
def test_read_vectors_and_weightings(self):
m = Memory.state(
memory_matrix=np.random.uniform(-1, 1,
(5, 11, 7)).astype(np.float32),
usage_vector=None,
link_matrix=None,
precedence_vector=None,
write_weighting=None,
read_weightings=DNCMemoryTests.softmax_sample((5, 11, 3), axis=1),
)
i = DNC.interface(
read_keys=np.random.uniform(0, 1, (5, 7, 3)).astype(np.float32),
read_strengths=np.random.uniform(0, 1, (5, 3)).astype(np.float32),
write_key=None,
write_strength=None,
erase_vector=None,
write_vector=None,
free_gates=None,
allocation_gate=None,
write_gate=None,
read_modes=tf.convert_to_tensor(
DNCMemoryTests.softmax_sample((5, 3, 3), axis=1)),
)
# read uses the link matrix that is produced after a write operation
new_link_matrix = np.random.uniform(0, 1,
(5, 11, 11)).astype(np.float32)
# assume ContentAddressing and TemporalLinkAddressing are already correct
op_ca = ContentAddressing.weighting(m.memory_matrix, i.read_keys,
i.read_strengths)
op_f, op_b = TemporalLinkAddressing.weightings(new_link_matrix,
m.read_weightings)
read_op = Memory.read(m.memory_matrix, m.read_weightings,
new_link_matrix, i)
with self.test_session() as session:
lookup_weightings = session.run(op_ca)
forward_weighting, backward_weighting = session.run([op_f, op_b])
updated_read_weightings, updated_read_vectors = session.run(
read_op)
# hack to circumvent tf bug in not doing `convert_to_tensor` in einsum reductions correctly
read_modes_numpy = tf.Session().run(i.read_modes)
self.assertEqual(updated_read_weightings.shape, (5, 11, 3))
self.assertEqual(updated_read_vectors.shape, (5, 7, 3))
expected_read_weightings = np.zeros((5, 11, 3)).astype(np.float32)
for read_head in range(3):
backward_weight = read_modes_numpy[:, 0, read_head, np.
newaxis] * backward_weighting[:, :,
read_head]
lookup_weight = read_modes_numpy[:, 1, read_head, np.
newaxis] * lookup_weightings[:, :,
read_head]
forward_weight = read_modes_numpy[:, 2, read_head, np.
newaxis] * forward_weighting[:, :,
read_head]
expected_read_weightings[:, :,
read_head] = backward_weight + lookup_weight + forward_weight
expected_read_vectors = np.matmul(
np.transpose(m.memory_matrix, [0, 2, 1]), updated_read_weightings)
self.assertAllClose(updated_read_weightings, expected_read_weightings)
self.assertEqual(updated_read_weightings.shape, (5, 11, 3))
self.assertAllClose(updated_read_vectors, expected_read_vectors)
parser.add_argument("--no-dnc", action='store_true')
parser.add_argument("--savedir", type=str, default="model")
parser.add_argument("--logdir", type=str, default="logs")
parser.add_argument("--learningrate", type=float, default=1e-4)
parser.add_argument("--no-mask", action='store_true')
args = parser.parse_args()
BATCH_SIZE = args.batch_size
task = eval(args.task)
if args.test_params:
test_params = eval(args.test_params)
else:
test_params = tuple(np.max(p) for p in task.default_params)
memory = Memory(args.msize, args.mwidth, init_state=args.minit)
memory.add_head(NTMReadHead, shifts=[-1, 0, 1])
memory.add_head(NTMWriteHead, shifts=[-1, 0, 1])
input = tf.placeholder(tf.float32, shape=(None, None, task.input_size))
#
if args.controller == 'lstm':
controller = LSTMCell(args.controller_size)
elif args.controller == 'multilstm':
controller = tf.nn.rnn_cell.MultiRNNCell(
[LSTMCell(args.controller_size) for i in range(3)])
elif args.controller == 'ff':
controller = dnc.ff.FFWrapper(
dnc.ff.simple_feedforward(hidden=[args.controller_size] * 2))
if not args.no_dnc:
class DNC:
def __init__(self,
controller_class,
input_size,
output_size,
max_sequence_length,
memory_words_num=256,
memory_word_size=64,
memory_read_heads=4,
batch_size=128):
"""
constructs a complete DNC architecture as described in the DNC paper
http://www.nature.com/nature/journal/vaop/ncurrent/full/nature20101.html
Parameters:
-----------
controller_class: BaseController
a concrete implementation of the BaseController class
input_size: int
the size of the input vector
output_size: int
the size of the output vector
max_sequence_length: int
the maximum length of an input sequence
memory_words_num: int
the number of words that can be stored in memory
memory_word_size: int
the size of an individual word in memory
memory_read_heads: int
the number of read heads in the memory
batch_size: int
the size of the data batch
"""
self.input_size = input_size
self.output_size = output_size
self.max_sequence_length = max_sequence_length
self.words_num = memory_words_num
self.word_size = memory_word_size
self.read_heads = memory_read_heads
self.batch_size = batch_size
self.memory = Memory(self.words_num, self.word_size, self.read_heads,
self.batch_size)
self.controller = controller_class(self.input_size, self.output_size,
self.read_heads, self.word_size,
self.batch_size)
# input data placeholders
self.input_data = tf.placeholder(tf.float32, [None, None, chunk_size],
name='input')
self.target_output = tf.placeholder(tf.float32,
[None, None, output_size],
name='targets')
#self.input_data = tf.placeholder(tf.float32, [batch_size, None, input_size], name='input')
#self.target_output = tf.placeholder(tf.float32, [batch_size, None, output_size], name='targets')
self.sequence_length = tf.placeholder(tf.int32, name='sequence_length')
self.build_graph()
def _step_op(self, step, memory_state, controller_state=None):
"""
performs a step operation on the input step data
Parameters:
----------
step: Tensor (batch_size, input_size)
memory_state: Tuple
a tuple of current memory parameters
controller_state: Tuple
the state of the controller if it's recurrent
Returns: Tuple
output: Tensor (batch_size, output_size)
memory_view: dict
"""
last_read_vectors = memory_state[6]
pre_output, interface, nn_state = None, None, None
if self.controller.has_recurrent_nn:
pre_output, interface, nn_state = self.controller.process_input(
step, last_read_vectors, controller_state)
else:
pre_output, interface = self.controller.process_input(
step, last_read_vectors)
usage_vector, write_weighting, memory_matrix, link_matrix, precedence_vector = self.memory.write(
memory_state[0], memory_state[1], memory_state[5], memory_state[4],
memory_state[2], memory_state[3], interface['write_key'],
interface['write_strength'], interface['free_gates'],
interface['allocation_gate'], interface['write_gate'],
interface['write_vector'], interface['erase_vector'])
read_weightings, read_vectors = self.memory.read(
memory_matrix,
memory_state[5],
interface['read_keys'],
interface['read_strengths'],
link_matrix,
interface['read_modes'],
)
return [
# report new memory state to be updated outside the condition branch
memory_matrix,
usage_vector,
precedence_vector,
link_matrix,
write_weighting,
read_weightings,
read_vectors,
self.controller.final_output(pre_output, read_vectors),
interface['free_gates'],
interface['allocation_gate'],
interface['write_gate'],
# report new state of RNN if exists
nn_state[0] if nn_state is not None else tf.zeros(1),
nn_state[1] if nn_state is not None else tf.zeros(1)
]
def _loop_body(self, time, memory_state, outputs, free_gates,
allocation_gates, write_gates, read_weightings,
write_weightings, usage_vectors, controller_state):
"""
the body of the DNC sequence processing loop
Parameters:
----------
time: Tensor
outputs: TensorArray
memory_state: Tuple
free_gates: TensorArray
allocation_gates: TensorArray
write_gates: TensorArray
read_weightings: TensorArray,
write_weightings: TensorArray,
usage_vectors: TensorArray,
controller_state: Tuple
Returns: Tuple containing all updated arguments
"""
step_input = self.unpacked_input_data.read(time)
output_list = self._step_op(step_input, memory_state, controller_state)
# update memory parameters
new_controller_state = tf.zeros(1)
new_memory_state = tuple(output_list[0:7])
new_controller_state = LSTMStateTuple(output_list[11], output_list[12])
outputs = outputs.write(time, output_list[7])
# collecting memory view for the current step
free_gates = free_gates.write(time, output_list[8])
allocation_gates = allocation_gates.write(time, output_list[9])
write_gates = write_gates.write(time, output_list[10])
read_weightings = read_weightings.write(time, output_list[5])
write_weightings = write_weightings.write(time, output_list[4])
usage_vectors = usage_vectors.write(time, output_list[1])
return (time + 1, new_memory_state, outputs, free_gates,
allocation_gates, write_gates, read_weightings,
write_weightings, usage_vectors, new_controller_state)
def build_graph(self):
"""
builds the computational graph that performs a step-by-step evaluation
of the input data batches
"""
self.unpacked_input_data = dnc.utility.unpack_into_tensorarray(
self.input_data, 1, self.sequence_length)
outputs = tf.TensorArray(tf.float32,
self.sequence_length,
name='outputs')
free_gates = tf.TensorArray(tf.float32,
self.sequence_length,
name='free_gates')
allocation_gates = tf.TensorArray(tf.float32,
self.sequence_length,
name='allocation_gates')
write_gates = tf.TensorArray(tf.float32,
self.sequence_length,
name='write_gates')
read_weightings = tf.TensorArray(tf.float32,
self.sequence_length,
name='read_weightings')
write_weightings = tf.TensorArray(tf.float32,
self.sequence_length,
name='write_weightings')
usage_vectors = tf.TensorArray(tf.float32,
self.sequence_length,
name='usage_vectors')
controller_state = self.controller.get_state(
) if self.controller.has_recurrent_nn else (tf.zeros(1), tf.zeros(1))
memory_state = self.memory.init_memory()
if not isinstance(controller_state, LSTMStateTuple):
controller_state = LSTMStateTuple(controller_state[0],
controller_state[1])
final_results = None
with tf.variable_scope("sequence_loop") as scope:
time = tf.placeholder(dtype=tf.int32, name='time')
final_results = tf.while_loop(
cond=lambda time, *_: time < self.sequence_length,
body=self._loop_body,
loop_vars=(time, memory_state, outputs, free_gates,
allocation_gates, write_gates, read_weightings,
write_weightings, usage_vectors, controller_state),
parallel_iterations=32,
swap_memory=False)
dependencies = []
if self.controller.has_recurrent_nn:
dependencies.append(self.controller.update_state(final_results[9]))
with tf.control_dependencies(dependencies):
self.packed_output = dnc.utility.pack_into_tensor(final_results[2],
axis=1)
self.packed_memory_view = {
'free_gates':
dnc.utility.pack_into_tensor(final_results[3], axis=1),
'allocation_gates':
dnc.utility.pack_into_tensor(final_results[4], axis=1),
'write_gates':
dnc.utility.pack_into_tensor(final_results[5], axis=1),
'read_weightings':
dnc.utility.pack_into_tensor(final_results[6], axis=1),
'write_weightings':
dnc.utility.pack_into_tensor(final_results[7], axis=1),
'usage_vectors':
dnc.utility.pack_into_tensor(final_results[8], axis=1)
}
def get_outputs(self):
"""
returns the graph nodes for the output and memory view
Returns: Tuple
outputs: Tensor (batch_size, time_steps, output_size)
memory_view: dict
"""
return self.packed_output, self.packed_memory_view
def save(self, session, ckpts_dir, name):
"""
saves the current values of the model's parameters to a checkpoint
Parameters:
----------
session: tf.Session
the tensorflow session to save
ckpts_dir: string
the path to the checkpoints directories
name: string
the name of the checkpoint subdirectory
"""
checkpoint_dir = os.path.join(ckpts_dir, name)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
tf.train.Saver(tf.trainable_variables()).save(
session, os.path.join(checkpoint_dir, 'model.ckpt'))
def restore(self, session, ckpts_dir, name):
"""
session: tf.Session
the tensorflow session to restore into
ckpts_dir: string
the path to the checkpoints directories
name: string
the name of the checkpoint subdirectory
"""
tf.train.Saver(tf.trainable_variables()).restore(
session, os.path.join(ckpts_dir, name, 'model.ckpt'))
import tensorflow as tf import numpy as np from dnc import DNC, LSTMCell from dnc.memory import Memory, NTMReadHead, NTMWriteHead from tasks import CopyTask, RepeatCopyTask, AndTask, XorTask, MergeTask from utils import * INPUT_SIZE = 8 BATCH_SIZE = 32 memory = Memory(25, 6) memory.add_head(NTMReadHead, shifts=[-1, 0, 1]) memory.add_head(NTMReadHead, shifts=[-1, 0, 1]) memory.add_head(NTMWriteHead, shifts=[-1, 0, 1]) input = tf.placeholder(tf.float32, shape=(None, None, INPUT_SIZE+2)) #lstm = tf.nn.rnn_cell.MultiRNNCell([LSTMCell(256) for i in range(3)]) lstm = LSTMCell(100) net = DNC(input, memory, INPUT_SIZE+2, controller = lstm, log_memory=True) targets = tf.placeholder(dtype=tf.float32, shape=[None, None, INPUT_SIZE+2]) mask = tf.placeholder(dtype=tf.float32, shape=[None, None, INPUT_SIZE+2]) output = net[0] loss = tf.losses.sigmoid_cross_entropy(logits=output, weights=mask, multi_class_labels=targets) cost = tf.reduce_sum( mask*((1 - targets * (1 - tf.exp(-output))) * tf.sigmoid(output)) ) / BATCH_SIZE opt = tf.train.RMSPropOptimizer(1e-4, momentum=0.9) train = minimize_and_clip(opt, loss)
