INIT = conf["training"]["init"]
NUM_CORRUPTIONS = 0
if SAMPLING_SCHEME == "all":
NUM_CORRUPTIONS = 4
else:
NUM_CORRUPTIONS = 2
BATCH_SIZE = POS_PER_BATCH + POS_PER_BATCH * NEG_PER_POS * NUM_CORRUPTIONS
kb = load_from_file(conf["data"]["kb"])
print("Batch size: %d, pos: %d, neg: %d, corrupted: %d" %
(BATCH_SIZE, POS_PER_BATCH, NEG_PER_POS, NUM_CORRUPTIONS))
if TEMPLATES_PATH is not None:
rule_templates = load_from_file(TEMPLATES_PATH, rule_template=True)
kb = augment_with_templates(kb, rule_templates)
kb = normalize(kb)
nkb, kb_ids, vocab, emb, predicate_ids, constant_ids = \
kb2nkb(kb, INPUT_SIZE, unit_normalize=UNIT_NORMALIZE,
keep_prob=KEEP_PROB)
known_facts = kb_ids2known_facts(kb_ids)
# using same embedding matrix
# tf.get_variable_scope().reuse_variables()
goal_struct = rule2struct(normalize([[Atom('p1', ["c0", "c1"])]])[0])
if EXPERIMENT == "animals":
goal_struct = rule2struct(normalize([[Atom('p1', ["c0"])]])[0])
def embed(goal, emb, keep_prob=1.0):
def test_learn():
conf = load_conf("./conf/synth/synth_three.conf")
DEBUG = conf["meta"]["debug"]
OUTPUT_PREDICTIONS = conf["meta"]["output_predictions"]
CHECK_NUMERICS = conf["meta"]["check_numerics"]
TFDBG = conf["meta"]["tfdbg"]
TEST_GRAPH_CREATION = conf["meta"]["test_graph_creation"]
TRAIN = conf["meta"]["train"]
TEST_TIME_NEURAL_LINK_PREDICTION = conf["meta"][
"test_time_neural_link_prediction"]
TEST_TIME_BATCHING = conf["meta"]["test_time_batching"]
ENSEMBLE = conf["meta"]["ensemble"]
EXPERIMENT = conf["meta"]["experiment_prefix"]
TEMPLATES_PATH = conf["data"]["templates"]
INPUT_SIZE = conf["model"]["input_size"]
UNIFICATION = conf["model"]["unification"]
L2 = conf["model"]["l2"]
UNIT_NORMALIZE = conf["model"]["unit_normalize"]
K_MAX = conf["model"]["k_max"]
NEURL_LINK_PREDICTOR = conf["model"]["neural_link_predictor"]
TRAIN_0NTP = conf["model"]["train_0ntp"]
KEEP_PROB = conf["model"]["keep_prob"]
MAX_DEPTH = conf["model"]["max_depth"]
TRAIN_NTP = TRAIN_0NTP or TEMPLATES_PATH is not None
if NEURL_LINK_PREDICTOR is None and not TRAIN_0NTP:
raise AttributeError("Can't train non-0NTP without link predictor")
REPORT_INTERVAL = conf["training"]["report_interval"]
NUM_EPOCHS = conf["training"]["num_epochs"]
CLIP = conf["training"]["clip"]
LEARNING_RATE = conf["training"]["learning_rate"]
EPSILON = conf["training"]["epsilon"]
OPTIMIZER = conf["training"]["optimizer"]
POS_PER_BATCH = conf["training"]["pos_per_batch"]
NEG_PER_POS = conf["training"]["neg_per_pos"]
SAMPLING_SCHEME = conf["training"]["sampling_scheme"]
MEAN_LOSS = conf["training"]["mean_loss"]
INIT = conf["training"]["init"]
NUM_CORRUPTIONS = 0
if SAMPLING_SCHEME == "all":
NUM_CORRUPTIONS = 4
else:
NUM_CORRUPTIONS = 2
BATCH_SIZE = POS_PER_BATCH + POS_PER_BATCH * NEG_PER_POS * NUM_CORRUPTIONS
kb = load_from_file(conf["data"]["kb"])
print("Batch size: %d, pos: %d, neg: %d, corrupted: %d" %
(BATCH_SIZE, POS_PER_BATCH, NEG_PER_POS, NUM_CORRUPTIONS))
if TEMPLATES_PATH is not None:
rule_templates = load_from_file(TEMPLATES_PATH, rule_template=True)
kb = augment_with_templates(kb, rule_templates)
kb = normalize(kb)
nkb, kb_ids, vocab, emb, predicate_ids, constant_ids = \
kb2nkb(kb, INPUT_SIZE, unit_normalize=UNIT_NORMALIZE, keep_prob=KEEP_PROB)
known_facts = kb_ids2known_facts(kb_ids)
goal_struct = rule2struct(normalize([[Atom('p1', ["c0", "c1"])]])[0])
if EXPERIMENT == "animals":
goal_struct = rule2struct(normalize([[Atom('p1', ["c0"])]])[0])
def embed(goal, emb, keep_prob=1.0):
return [
embed_symbol(x,
emb,
unit_normalize=UNIT_NORMALIZE,
keep_prob=keep_prob) for x in goal
]
def get_mask_id(kb, goal_struct, goal):
if goal_struct in kb:
facts = kb[goal_struct][0]
num_facts = len(facts[0])
mask_id = None
for i in range(num_facts):
exists = True
for j in range(len(goal)):
exists = exists and goal[j] == facts[j][i]
if exists:
mask_id = i
if mask_id is not None:
return mask_id
return None
mask_indices = tf.placeholder("int32", [POS_PER_BATCH, 2],
name="mask_indices")
goal_placeholder = [
tf.placeholder("int32", [BATCH_SIZE], name="goal_%d" % i)
for i in range(0, len(goal_struct[0]))
]
goal_emb = embed(goal_placeholder, emb, KEEP_PROB)
num_facts = len(kb_ids[goal_struct][0][0])
mask = tf.Variable(np.ones([num_facts, BATCH_SIZE], np.float32),
trainable=False,
name="fact_mask")
mask_set = tf.scatter_nd_update(mask, mask_indices, [0.0] * POS_PER_BATCH)
mask_unset = tf.scatter_nd_update(mask, mask_indices,
[1.0] * POS_PER_BATCH)
target = tf.placeholder("float32", [BATCH_SIZE], name="target")
AGGREGATION_METHOD = conf["model"]["aggregate_fun"]
print(AGGREGATION_METHOD)
if AGGREGATION_METHOD == "Max":
def fun(x):
return tf.reduce_max(x, 1)
aggregation_fun = fun
else:
raise AttributeError("Aggregation function %s unknown" %
AGGREGATION_METHOD)
def corrupt_goal(goal, args=[0], tries=100):
if tries == 0:
print("WARNING: Could not corrupt", goal)
return goal
else:
goal_corrupted = copy.deepcopy(goal)
for arg in args:
corrupt = constant_ids[random.randint(0,
len(constant_ids) - 1)]
goal_corrupted[arg + 1] = corrupt
if tuple(goal_corrupted) in known_facts:
return corrupt_goal(goal, args, tries - 1)
else:
return goal_corrupted
def get_batches():
facts = kb_ids[goal_struct][0]
num_facts = len(facts[0])
fact_ids = list(range(0, num_facts))
assert num_facts >= POS_PER_BATCH
def generator():
random.shuffle(fact_ids)
feed_dicts = []
mask_indices_init = np.zeros([POS_PER_BATCH, 2], dtype=np.int32)
goals_in_batch = [[] for _ in goal_placeholder]
targets_in_batch = []
j = 0
jj = 0
for i, ix in enumerate(fact_ids):
current_goal = [x[ix] for x in facts]
for k in range(len(current_goal)):
goals_in_batch[k].append(current_goal[k])
targets_in_batch += [1] + [0] * (NEG_PER_POS * NUM_CORRUPTIONS)
mask_indices_init[j] = [ix, jj]
j += 1
jj += 1 + (NEG_PER_POS * NUM_CORRUPTIONS)
for _ in range(NEG_PER_POS):
currupt_goal_1 = corrupt_goal(current_goal, [0])
for k in range(len(currupt_goal_1)):
goals_in_batch[k].append(currupt_goal_1[k])
currupt_goal_2 = corrupt_goal(current_goal, [1])
for k in range(len(currupt_goal_2)):
goals_in_batch[k].append(currupt_goal_2[k])
if SAMPLING_SCHEME == "all":
currupt_goal_3 = corrupt_goal(current_goal, [0, 1])
for k in range(len(currupt_goal_3)):
goals_in_batch[k].append(currupt_goal_3[k])
currupt_goal_4 = corrupt_goal(current_goal, [0, 1])
for k in range(len(currupt_goal_4)):
goals_in_batch[k].append(currupt_goal_4[k])
if j % POS_PER_BATCH == 0:
feed_dict = {
mask_indices: mask_indices_init,
target: targets_in_batch,
}
for k in range(len(goal_placeholder)):
feed_dict[goal_placeholder[k]] = goals_in_batch[k]
feed_dicts.append(feed_dict)
mask_indices_init = np.zeros([POS_PER_BATCH, 2],
dtype=np.int32)
goals_in_batch = [[] for _ in goal_placeholder]
targets_in_batch = []
j = 0
jj = 0
for f in feed_dicts:
yield f
return GeneratorWithRestart(generator)
train_feed_dicts = get_batches()
prove_success = prove(nkb,
goal_emb,
goal_struct,
mask,
trace=True,
aggregation_fun=aggregation_fun,
k_max=K_MAX,
train_0ntp=TRAIN_0NTP,
max_depth=MAX_DEPTH)
print("Graph creation complete.")
print("Variables")
for v in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES):
print("\t", v)
if TEST_GRAPH_CREATION:
exit(1)
if DEBUG and TRAIN_NTP:
prove_success = tfprint(prove_success, "NTP success:\n")
def caculate_loss(success, target):
if AGGREGATION_METHOD == "LogSumExp":
return -(target * 2 - 1) * prove_success
else:
x = success
z = target
return - z * tf.log(tf.clip_by_value(x, EPSILON, 1.0)) - \
(1 - z) * tf.log(tf.clip_by_value(1 - x, EPSILON, 1.0))
prover_loss = caculate_loss(prove_success, target)
if DEBUG:
prover_loss = tfprint(prover_loss, "NTP loss:\n")
if NEURL_LINK_PREDICTOR is not None:
neural_link_prediction_success = \
tf.squeeze(neural_link_predict(goal_emb, model=NEURL_LINK_PREDICTOR))
if DEBUG:
neural_link_prediction_success = \
tfprint(neural_link_prediction_success, "link predict:\n")
neural_link_prediction_loss = \
caculate_loss(neural_link_prediction_success, target)
if TRAIN_NTP:
loss = neural_link_prediction_loss + prover_loss
else:
loss = neural_link_prediction_loss
if TEST_TIME_NEURAL_LINK_PREDICTION:
test_time_prediction = \
tf.maximum(neural_link_prediction_success, prove_success)
# fixme: refactor!
if ENSEMBLE:
prove_success = \
tf.maximum(neural_link_prediction_success, prove_success)
loss = caculate_loss(prove_success, target)
else:
loss = prover_loss
test_time_prediction = prove_success
if DEBUG:
loss = tfprint(loss, "loss:\n")
if MEAN_LOSS:
loss = tf.reduce_mean(loss)
else:
loss = tf.reduce_sum(loss)
def pre_run(sess, epoch, feed_dict, loss, predict):
results = sess.run(mask_set, {mask_indices: feed_dict[mask_indices]})
def post_run(sess, epoch, feed_dict, loss, predict):
results = sess.run(mask_unset, {mask_indices: feed_dict[mask_indices]})
optim = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE,
epsilon=EPSILON)
gradients = optim.compute_gradients(loss)
variables = [x[1] for x in gradients]
gradients = [x[0] for x in gradients]
hooks = []
session_config = tf.ConfigProto()
session_config.gpu_options.allow_growth = True
sess = tf.Session(config=session_config)
if TFDBG:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
sess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)
if TRAIN:
train(loss,
optim,
train_feed_dicts,
max_epochs=NUM_EPOCHS,
hooks=hooks,
pre_run=pre_run,
post_run=post_run,
sess=sess,
l2=L2,
clip=CLIP,
check_numerics=CHECK_NUMERICS)
else:
sess.run(tf.global_variables_initializer())
predicate_ids_with_placholders = copy.deepcopy(predicate_ids)
predicate_ids = []
for id in predicate_ids_with_placholders:
if not is_parameter(vocab.id2sym[id]):
predicate_ids.append(id)
if OUTPUT_PREDICTIONS:
goal_placeholder = [
tf.placeholder("int32", [1], name="goal_%d" % i)
for i in range(0, len(goal_struct[0]))
]
goal_emb = embed(goal_placeholder, emb, keep_prob=1.0)
if TEST_TIME_BATCHING:
copies = BATCH_SIZE
for i, x in enumerate(goal_emb):
goal_emb[i] = tf.tile(x, [copies, 1])
prove_success_test_time = \
prove(nkb, goal_emb, goal_struct, mask_var=None, trace=True,
aggregation_fun=aggregation_fun, k_max=K_MAX,
train_0ntp=TRAIN_0NTP, max_depth=MAX_DEPTH)
if NEURL_LINK_PREDICTOR is not None:
neural_link_prediction_success_test_time = \
neural_link_predict(goal_emb, model=NEURL_LINK_PREDICTOR)
if TEST_TIME_NEURAL_LINK_PREDICTION:
prove_success_test_time = \
tf.maximum(prove_success_test_time,
neural_link_prediction_success_test_time)
table = []
for sym in vocab.sym2id:
id = vocab.sym2id[sym]
vec = sess.run(emb[id])
table.append([sym, id, vec])
def predict(predicate, arg1, arg2):
feed_dict = {}
goal = [vocab(predicate), vocab(arg1), vocab(arg2)]
for k, d in zip(goal_placeholder, goal):
feed_dict[k] = [d]
success = prove_success_test_time
if AGGREGATION_METHOD == "LogSumExp":
success = tf.sigmoid(success)
if TEST_TIME_NEURAL_LINK_PREDICTION:
success = tf.squeeze(success)
success_val = sess.run(success, feed_dict=feed_dict)
if TEST_TIME_BATCHING:
if not all([x == success_val[0] for x in success_val]):
print("WARNING! Numerical instability?", success_val)
return success_val
table = []
headers = [vocab.id2sym[rid] for rid in predicate_ids]
for i, e1id in enumerate(constant_ids):
for j, e2id in enumerate(constant_ids):
e1 = vocab.id2sym[e1id]
e2 = vocab.id2sym[e2id]
row = [e1, e2]
for r in headers:
score = predict(r, e1, e2)
if TEST_TIME_BATCHING:
score = score[0]
row.append(score)
table.append(row)
print(tabulate(table, headers=["e1", "e2"] + headers))
def decode(x, emb, vocab, valid_ids, sess):
valid_ids = set(valid_ids)
num_rules = int(x.get_shape()[0])
num_symbols = int(emb.get_shape()[0])
mask = np.ones([num_symbols], dtype=np.float32)
for i in range(len(vocab)):
if i not in valid_ids:
mask[i] = 0
# -- num_rules x num_symbols
mask = tf.tile(tf.expand_dims(mask, 0), [num_rules, 1])
# -- num_rules x num_symbols
match = representation_match(x, emb)
success_masked = match * mask
success, ix = tf.nn.top_k(success_masked, 1)
success_val, ix_val = sess.run([success, ix], {})
syms = []
for i, row in enumerate(ix_val):
sym_id = row[0]
sym_success = success_val[i][0]
sym = vocab.id2sym[sym_id]
syms.append((sym, sym_success))
return syms
def unstack_rules(rule):
rules = []
num_rules = len(rule[0].predicate)
for i in range(num_rules):
current_rule = []
confidence = 1.0
for atom in rule:
predicate = atom.predicate
if isinstance(predicate, list):
predicate, success = predicate[i]
confidence = min(confidence, success)
arguments = []
for argument in atom.arguments:
if isinstance(argument, list):
argument, success = argument[i]
arguments.append(argument)
confidence = min(confidence, success)
else:
arguments.append(argument)
current_rule.append(Atom(predicate, arguments))
rules.append((current_rule, confidence))
return rules
for struct in nkb:
if len(struct) > 1:
rule = nkb[struct]
rule_sym = []
for atom in rule:
atom_sym = []
for i, sym in enumerate(atom):
if is_tensor(sym):
valid_ids = predicate_ids if i == 0 else constant_ids
atom_sym.append(
decode(sym, emb, vocab, valid_ids, sess))
else:
atom_sym.append(sym[0])
rule_sym.append(Atom(atom_sym[0], atom_sym[1:]))
rules = unstack_rules(rule_sym)
rules.sort(key=lambda x: -x[1])
max_a, max_b = 0.0, 0.0
for rule, confidence in rules:
head, body = rule[0], rule[1]
if head.predicate == 'p' and body.predicate == 'q' and head.arguments == body.arguments:
max_a = max(max_a, confidence)
elif head.predicate == 'q' and body.predicate == 'p' and head.arguments == body.arguments:
max_b = max(max_b, confidence)
assert max_a > 0.9
assert max_b > 0.9