def test_masking_v1():
nb_entities = 10
nb_predicates = 5
embedding_size = 10
init_size = 1.0
rs = np.random.RandomState(0)
for _ in range(1):
for position in [0, 1]:
for st in ['min', 'concat']:
with torch.no_grad():
triples = [
('a', 'p', 'b'),
('c', 'q', 'd')
]
entity_to_index = {'a': 0, 'b': 1, 'c': 2, 'd': 3}
predicate_to_index = {'p': 0, 'q': 1}
kernel = GaussianKernel()
entity_embeddings = nn.Embedding(nb_entities, embedding_size * 2, sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates, embedding_size * 2, sparse=True)
entity_embeddings.weight.data *= init_size
predicate_embeddings.weight.data *= init_size
fact_rel = torch.LongTensor(np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.LongTensor(np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.LongTensor(np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
model = NeuralKB(entity_embeddings=entity_embeddings, predicate_embeddings=predicate_embeddings,
kernel=kernel, facts=facts, scoring_type=st)
xs_np = rs.randint(nb_entities, size=32)
xp_np = rs.randint(nb_predicates, size=32)
xo_np = rs.randint(nb_entities, size=32)
xi_np = np.array([position] * xs_np.shape[0])
xs_np[0] = 0
xp_np[0] = 0
xo_np[0] = 1
xs_np[1] = 2
xp_np[1] = 1
xo_np[1] = 3
xs = torch.LongTensor(xs_np)
xp = torch.LongTensor(xp_np)
xo = torch.LongTensor(xo_np)
xi = torch.LongTensor(xi_np)
xs_emb = entity_embeddings(xs)
xp_emb = predicate_embeddings(xp)
xo_emb = entity_embeddings(xo)
model.mask_indices = xi
scores = model.forward(xp_emb, xs_emb, xo_emb)
inf = model.score(xp_emb, xs_emb, xo_emb)
if position == 0:
assert inf[0] < 0.5
assert inf[1] > 0.9
elif position == 1:
assert inf[0] > 0.9
assert inf[1] < 0.5
scores_sp, scores_po = scores
inf = inf.cpu().numpy()
scores_sp = scores_sp.cpu().numpy()
scores_po = scores_po.cpu().numpy()
for i in range(xs.shape[0]):
np.testing.assert_allclose(inf[i], scores_sp[i, xo[i]], rtol=1e-5, atol=1e-5)
np.testing.assert_allclose(inf[i], scores_po[i, xs[i]], rtol=1e-5, atol=1e-5)
def test_masking_v2():
nb_entities = 10
nb_predicates = 5
embedding_size = 10
rs = np.random.RandomState(0)
for _ in range(1):
for position in [0, 1, 2]:
for st in ['min', 'concat']:
with torch.no_grad():
triples = [
('a', 'p', 'b'),
('b', 'q', 'c'),
('a', 'p', 'c')
]
entity_to_index = {'a': 0, 'b': 1, 'c': 2, 'd': 3}
predicate_to_index = {'p': 0, 'q': 1}
kernel = GaussianKernel()
entity_emb = nn.Embedding(nb_entities, embedding_size * 2, sparse=True)
predicate_emb = nn.Embedding(nb_predicates, embedding_size * 2, sparse=True)
fact_rel = torch.LongTensor(np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.LongTensor(np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.LongTensor(np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
base = NeuralKB(entity_embeddings=entity_emb, predicate_embeddings=predicate_emb,
kernel=kernel, facts=facts, scoring_type=st)
indices = torch.LongTensor(np.array([predicate_to_index['p'], predicate_to_index['q']]))
reformulator = SymbolicReformulator(predicate_emb, indices)
model = SimpleHoppy(base, entity_emb, hops=reformulator)
xs_np = rs.randint(nb_entities, size=32)
xp_np = rs.randint(nb_predicates, size=32)
xo_np = rs.randint(nb_entities, size=32)
xi_np = np.array([position] * xs_np.shape[0])
xs_np[0] = 0
xp_np[0] = 0
xo_np[0] = 1
xs_np[1] = 1
xp_np[1] = 1
xo_np[1] = 2
xs_np[2] = 0
xp_np[2] = 0
xo_np[2] = 2
xs = torch.LongTensor(xs_np)
xp = torch.LongTensor(xp_np)
xo = torch.LongTensor(xo_np)
xi = torch.LongTensor(xi_np)
xs_emb = entity_emb(xs)
xp_emb = predicate_emb(xp)
xo_emb = entity_emb(xo)
# xi = None
base.mask_indices = xi
scores = model.forward(xp_emb, xs_emb, xo_emb)
inf = model.score(xp_emb, xs_emb, xo_emb)
if position in {0, 1}:
assert inf[2] < 0.5
else:
assert inf[2] > 0.9
scores_sp, scores_po = scores
inf = inf.cpu().numpy()
scores_sp = scores_sp.cpu().numpy()
scores_po = scores_po.cpu().numpy()
for i in range(xs.shape[0]):
np.testing.assert_allclose(inf[i], scores_sp[i, xo[i]], rtol=1e-5, atol=1e-5)
np.testing.assert_allclose(inf[i], scores_po[i, xs[i]], rtol=1e-5, atol=1e-5)
def main(argv):
parser = argparse.ArgumentParser('KBC Research', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--train', action='store', required=True, type=str)
parser.add_argument('--dev', action='store', type=str, default=None)
parser.add_argument('--test', action='store', type=str, default=None)
parser.add_argument('--test-i', action='store', type=str, default=None)
parser.add_argument('--test-ii', action='store', type=str, default=None)
# model params
parser.add_argument('--model', '-m', action='store', type=str, default='distmult',
choices=['distmult', 'complex'])
parser.add_argument('--embedding-size', '-k', action='store', type=int, default=100)
parser.add_argument('--batch-size', '-b', action='store', type=int, default=100)
parser.add_argument('--eval-batch-size', '-B', action='store', type=int, default=None)
# training params
parser.add_argument('--epochs', '-e', action='store', type=int, default=100)
parser.add_argument('--learning-rate', '-l', action='store', type=float, default=0.1)
parser.add_argument('--optimizer', '-o', action='store', type=str, default='adagrad',
choices=['adagrad', 'adam', 'sgd'])
parser.add_argument('--N2', action='store', type=float, default=None)
parser.add_argument('--N3', action='store', type=float, default=None)
parser.add_argument('--seed', action='store', type=int, default=0)
parser.add_argument('--validate-every', '-V', action='store', type=int, default=None)
parser.add_argument('--input-type', '-I', action='store', type=str, default='standard',
choices=['standard', 'reciprocal'])
parser.add_argument('--gradient-accumulation-steps', '--gas', action='store', type=int, default=1)
parser.add_argument('--load', action='store', type=str, default=None)
parser.add_argument('--save', action='store', type=str, default=None)
parser.add_argument('--quiet', '-q', action='store_true', default=False)
args = parser.parse_args(argv)
import pprint
pprint.pprint(vars(args))
train_path = args.train
dev_path = args.dev
test_path = args.test
test_i_path = args.test_i
test_ii_path = args.test_ii
model_name = args.model
optimizer_name = args.optimizer
embedding_size = args.embedding_size
batch_size = args.batch_size
eval_batch_size = batch_size if args.eval_batch_size is None else args.eval_batch_size
nb_epochs = args.epochs
seed = args.seed
learning_rate = args.learning_rate
N2_weight = args.N2
N3_weight = args.N3
validate_every = args.validate_every
input_type = args.input_type
gradient_accumulation_steps = args.gradient_accumulation_steps
load_path = args.load
save_path = args.save
is_quiet = args.quiet
# set the seeds
np.random.seed(seed)
random_state = np.random.RandomState(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
logger.info(f'Device: {device}')
data = Data(train_path=train_path, dev_path=dev_path, test_path=test_path,
test_i_path=test_i_path, test_ii_path=test_ii_path, input_type=input_type)
triples_name_pairs = [
(data.dev_triples, 'dev'),
(data.test_triples, 'test'),
(data.test_i_triples, 'test-I'),
(data.test_ii_triples, 'test-II'),
]
rank = embedding_size * 2 if model_name in {'complex'} else embedding_size
init_size = 1e-3
entity_embeddings = nn.Embedding(data.nb_entities, rank, sparse=True)
predicate_embeddings = nn.Embedding(data.nb_predicates, rank, sparse=True)
entity_embeddings.weight.data *= init_size
predicate_embeddings.weight.data *= init_size
parameters_lst = nn.ModuleDict({
'entities': entity_embeddings,
'predicates': predicate_embeddings
})
parameters_lst.to(device)
if load_path is not None:
parameters_lst.load_state_dict(torch.load(load_path))
kernel = facts = None
if model_name in {'ntpzero'}:
kernel = GaussianKernel(slope=None)
fact_rel = torch.from_numpy(np.array([data.predicate_to_idx[p] for (_, p, _) in data.train_triples])).to(device)
fact_arg1 = torch.from_numpy(np.array([data.entity_to_idx[s] for (s, _, _) in data.train_triples])).to(device)
fact_arg2 = torch.from_numpy(np.array([data.entity_to_idx[o] for (_, _, o) in data.train_triples])).to(device)
facts = [fact_rel, fact_arg1, fact_arg2]
model_factory = {
'distmult': lambda: DistMult(entity_embeddings=entity_embeddings),
'complex': lambda: ComplEx(entity_embeddings=entity_embeddings),
'ntpzero': lambda: NeuralKB(entity_embeddings=entity_embeddings, predicate_embeddings=predicate_embeddings,
kernel=kernel, facts=facts)
}
assert model_name in model_factory
model = model_factory[model_name]()
model.to(device)
logger.info('Model state:')
for param_tensor in parameters_lst.state_dict():
logger.info(f'\t{param_tensor}\t{parameters_lst.state_dict()[param_tensor].size()}')
optimizer_factory = {
'adagrad': lambda: optim.Adagrad(parameters_lst.parameters(), lr=learning_rate),
'adam': lambda: optim.Adam(parameters_lst.parameters(), lr=learning_rate),
'sgd': lambda: optim.SGD(parameters_lst.parameters(), lr=learning_rate)
}
assert optimizer_name in optimizer_factory
optimizer = optimizer_factory[optimizer_name]()
loss_function = nn.CrossEntropyLoss(reduction='mean')
N2_reg = N2() if N2_weight is not None else None
N3_reg = N3() if N3_weight is not None else None
for epoch_no in range(1, nb_epochs + 1):
batcher = Batcher(data, batch_size, 1, random_state)
nb_batches = len(batcher.batches)
epoch_loss_values = []
for batch_no, (batch_start, batch_end) in enumerate(batcher.batches, 1):
xp_batch, xs_batch, xo_batch, xi_batch = batcher.get_batch(batch_start, batch_end)
xp_batch = torch.from_numpy(xp_batch.astype('int64')).to(device)
xs_batch = torch.from_numpy(xs_batch.astype('int64')).to(device)
xo_batch = torch.from_numpy(xo_batch.astype('int64')).to(device)
xi_batch = torch.from_numpy(xi_batch.astype('int64')).to(device)
xp_batch_emb = predicate_embeddings(xp_batch)
xs_batch_emb = entity_embeddings(xs_batch)
xo_batch_emb = entity_embeddings(xo_batch)
sp_scores, po_scores = model.forward(xp_batch_emb, xs_batch_emb, xo_batch_emb)
factors = [model.factor(e) for e in [xp_batch_emb, xs_batch_emb, xo_batch_emb]]
s_loss = loss_function(sp_scores, xo_batch)
o_loss = loss_function(po_scores, xs_batch)
loss = s_loss + o_loss
loss += N2_weight * N2_reg(factors) if N2_weight is not None else 0.0
loss += N3_weight * N3_reg(factors) if N3_weight is not None else 0.0
if gradient_accumulation_steps > 1:
loss = loss / gradient_accumulation_steps
loss.backward()
if batch_no % gradient_accumulation_steps == 0 or batch_no == nb_batches:
optimizer.step()
optimizer.zero_grad()
loss_value = loss.item()
epoch_loss_values += [loss_value]
if not is_quiet:
logger.info(f'Epoch {epoch_no}/{nb_epochs}\tBatch {batch_no}/{nb_batches}\tLoss {loss_value:.6f}')
loss_mean, loss_std = np.mean(epoch_loss_values), np.std(epoch_loss_values)
logger.info(f'Epoch {epoch_no}/{nb_epochs}\tLoss {loss_mean:.4f} ± {loss_std:.4f}')
if validate_every is not None and epoch_no % validate_every == 0:
for triples, name in [(t, n) for t, n in triples_name_pairs if len(t) > 0]:
metrics = evaluate(entity_embeddings=entity_embeddings, predicate_embeddings=predicate_embeddings,
test_triples=triples, all_triples=data.all_triples,
entity_to_index=data.entity_to_idx, predicate_to_index=data.predicate_to_idx,
model=model, batch_size=eval_batch_size, device=device)
logger.info(f'Epoch {epoch_no}/{nb_epochs}\t{name} results\t{metrics_to_str(metrics)}')
for triples, name in [(t, n) for t, n in triples_name_pairs if len(t) > 0]:
metrics = evaluate(entity_embeddings=entity_embeddings, predicate_embeddings=predicate_embeddings,
test_triples=triples, all_triples=data.all_triples,
entity_to_index=data.entity_to_idx, predicate_to_index=data.predicate_to_idx,
model=model, batch_size=eval_batch_size, device=device)
logger.info(f'Final \t{name} results\t{metrics_to_str(metrics)}')
if save_path is not None:
torch.save(parameters_lst.state_dict(), save_path)
logger.info("Training finished")
def main(argv):
parser = argparse.ArgumentParser('KBC Research', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--train', action='store', required=True, type=str)
parser.add_argument('--dev', action='store', type=str, default=None)
parser.add_argument('--test', action='store', type=str, default=None)
parser.add_argument('--test-i', action='store', type=str, default=None)
parser.add_argument('--test-ii', action='store', type=str, default=None)
parser.add_argument('--embedding-size', '-k', action='store', type=int, default=20)
parser.add_argument('--k-max', '-K', action='store', type=int, default=1)
parser.add_argument('--max-depth', '-d', action='store', type=int, default=1)
parser.add_argument('--hops', nargs='+', type=str, default=['1', '2'])
# training params
parser.add_argument('--epochs', '-e', action='store', type=int, default=100)
parser.add_argument('--learning-rate', '-l', action='store', type=float, default=0.1)
parser.add_argument('--batch-size', '-b', action='store', type=int, default=8)
parser.add_argument('--N2', action='store', type=float, default=None)
parser.add_argument('--N3', action='store', type=float, default=None)
parser.add_argument('--reformulator', '-r', action='store', type=str, default='linear',
choices=['static', 'linear', 'attentive', 'memory'])
parser.add_argument('--nb-rules', '-R', action='store', type=int, default=4)
parser.add_argument('--seed', action='store', type=int, default=0)
parser.add_argument('--validate-every', '-V', action='store', type=int, default=None)
parser.add_argument('--input-type', '-I', action='store', type=str, default='standard',
choices=['standard', 'reciprocal'])
parser.add_argument('--init-size', '-i', action='store', type=float, default=1.0)
parser.add_argument('--load', action='store', type=str, default=None)
parser.add_argument('--save', action='store', type=str, default=None)
parser.add_argument('--quiet', '-q', action='store_true', default=False)
args = parser.parse_args(argv)
import pprint
pprint.pprint(vars(args))
train_path = args.train
dev_path = args.dev
test_path = args.test
test_i_path = args.test_i
test_ii_path = args.test_ii
embedding_size = args.embedding_size
hops_str = args.hops
nb_epochs = args.epochs
learning_rate = args.learning_rate
batch_size = args.batch_size
N2_weight = args.N2
N3_weight = args.N3
reformulator_type = args.reformulator
nb_rules = args.nb_rules
eval_batch_size = batch_size
seed = args.seed
validate_every = args.validate_every
input_type = args.input_type
init_size = args.init_size
load_path = args.load
save_path = args.save
is_quiet = args.quiet
# set the seeds
np.random.seed(seed)
random_state = np.random.RandomState(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
device = 'cpu'
logger.info(f'Device: {device}')
if torch.cuda.is_available():
torch.set_default_tensor_type(torch.cuda.FloatTensor)
data = Data(train_path=train_path, dev_path=dev_path, test_path=test_path,
test_i_path=test_i_path, test_ii_path=test_ii_path, input_type=input_type)
triples_name_pairs = [
(data.dev_triples, 'dev'),
(data.test_triples, 'test'),
(data.test_i_triples, 'test-I'),
(data.test_ii_triples, 'test-II'),
]
rank = embedding_size
init_size = init_size
entity_embeddings = nn.Embedding(data.nb_entities, rank, sparse=True)
predicate_embeddings = nn.Embedding(data.nb_predicates, rank, sparse=True)
entity_embeddings.weight.data *= init_size
predicate_embeddings.weight.data *= init_size
kernel = GaussianKernel(slope=1.0)
fact_rel = torch.from_numpy(np.array([data.predicate_to_idx[p] for (_, p, _) in data.train_triples])).to(device)
fact_arg1 = torch.from_numpy(np.array([data.entity_to_idx[s] for (s, _, _) in data.train_triples])).to(device)
fact_arg2 = torch.from_numpy(np.array([data.entity_to_idx[o] for (_, _, o) in data.train_triples])).to(device)
facts = [fact_rel, fact_arg1, fact_arg2]
base_model = NeuralKB(entity_embeddings=entity_embeddings,
predicate_embeddings=predicate_embeddings,
facts=facts,
kernel=kernel).to(device)
memory = None
def make_hop(s: str) -> Tuple[BaseReformulator, bool]:
nonlocal memory
if s.isdigit():
nb_hops, is_reversed = int(s), False
else:
nb_hops, is_reversed = int(s[:-1]), True
res = None
if reformulator_type in {'static'}:
res = StaticReformulator(nb_hops, rank)
elif reformulator_type in {'linear'}:
res = LinearReformulator(nb_hops, rank)
elif reformulator_type in {'attentive'}:
res = AttentiveReformulator(nb_hops, predicate_embeddings)
elif reformulator_type in {'memory'}:
memory = MemoryReformulator.Memory(nb_hops, nb_rules, rank) if memory is None else memory
res = MemoryReformulator(memory)
elif reformulator_type in {'ntp'}:
res = NTPReformulator(nb_hops=nb_hops, embedding_size=embedding_size, kernel=kernel)
assert res is not None
return res, is_reversed
hops_lst = [make_hop(s) for s in hops_str]
model = MultiHoppy(model=base_model, entity_embeddings=entity_embeddings, hops_lst=hops_lst).to(device)
params_lst = [p for p in model.parameters()] + [predicate_embeddings.weight]
params = nn.ParameterList(params_lst).to(device)
if load_path is not None:
model.load_state_dict(torch.load(load_path))
for tensor in params_lst:
logger.info(f'\t{tensor.size()}\t{tensor.device}')
optimizer = optim.Adagrad(params, lr=learning_rate)
loss_function = nn.BCELoss()
N2_reg = N2() if N2_weight is not None else None
N3_reg = N3() if N3_weight is not None else None
for epoch_no in range(1, nb_epochs + 1):
batcher = Batcher(data, batch_size, 1, random_state)
nb_batches = len(batcher.batches)
epoch_loss_values = []
for batch_no, (batch_start, batch_end) in enumerate(batcher.batches, 1):
xp_batch_np, xs_batch_np, xo_batch_np, xi_batch_np = batcher.get_batch(batch_start, batch_end)
xp_batch = torch.from_numpy(xp_batch_np.astype('int64')).to(device)
xs_batch = torch.from_numpy(xs_batch_np.astype('int64')).to(device)
xo_batch = torch.from_numpy(xo_batch_np.astype('int64')).to(device)
xi_batch = torch.from_numpy(xi_batch_np.astype('int64')).to(device)
xp_batch_emb = predicate_embeddings(xp_batch)
xs_batch_emb = entity_embeddings(xs_batch)
xo_batch_emb = entity_embeddings(xo_batch)
sp_scores, po_scores = model.forward(xp_batch_emb, xs_batch_emb, xo_batch_emb, mask_indices=xi_batch)
factors = [model.factor(e) for e in [xp_batch_emb, xs_batch_emb, xo_batch_emb]]
sp_objects = [data.sp_to_o_lst.get((xs, xp), None) for xs, xp in zip(xs_batch_np, xp_batch_np)]
po_subjects = [data.po_to_s_lst.get((xp, xo), None) for xp, xo in zip(xp_batch_np, xo_batch_np)]
sp_targets = compute_bce_targets(xp_batch.shape[0], data.nb_entities, sp_objects, device=device)
po_targets = compute_bce_targets(xp_batch.shape[0], data.nb_entities, po_subjects, device=device)
s_loss = loss_function(sp_scores, sp_targets)
o_loss = loss_function(po_scores, po_targets)
loss = s_loss + o_loss
loss += N2_weight * N2_reg(factors) if N2_weight is not None else 0.0
loss += N3_weight * N3_reg(factors) if N3_weight is not None else 0.0
loss.backward()
optimizer.step()
optimizer.zero_grad()
loss_value = loss.item()
epoch_loss_values += [loss_value]
if not is_quiet:
logger.info(f'Epoch {epoch_no}/{nb_epochs}\tBatch {batch_no}/{nb_batches}\tLoss {loss_value:.6f}')
loss_mean, loss_std = np.mean(epoch_loss_values), np.std(epoch_loss_values)
logger.info(f'Epoch {epoch_no}/{nb_epochs}\tLoss {loss_mean:.4f} ± {loss_std:.4f}')
if validate_every is not None and epoch_no % validate_every == 0:
for triples, name in [(t, n) for t, n in triples_name_pairs if len(t) > 0]:
metrics = evaluate(entity_embeddings=entity_embeddings, predicate_embeddings=predicate_embeddings,
test_triples=triples, all_triples=data.all_triples,
entity_to_index=data.entity_to_idx, predicate_to_index=data.predicate_to_idx,
model=model, batch_size=eval_batch_size, device=device)
logger.info(f'Epoch {epoch_no}/{nb_epochs}\t{name} results\t{metrics_to_str(metrics)}')
for triples, name in [(t, n) for t, n in triples_name_pairs if len(t) > 0]:
metrics = evaluate(entity_embeddings=entity_embeddings, predicate_embeddings=predicate_embeddings,
test_triples=triples, all_triples=data.all_triples,
entity_to_index=data.entity_to_idx, predicate_to_index=data.predicate_to_idx,
model=model, batch_size=eval_batch_size, device=device)
logger.info(f'Final \t{name} results\t{metrics_to_str(metrics)}')
if save_path is not None:
torch.save(model.state_dict(), save_path)
logger.info("Training finished")
def test_learning_v2():
embedding_size = 100
torch.manual_seed(0)
triples, hops = [], []
for i in range(16):
triples += [(f'a{i}', 'p', f'b{i}'), (f'b{i}', 'q', f'c{i}')]
hops += [(f'a{i}', 'r', f'c{i}')]
entity_lst = sorted({e
for (e, _, _) in triples + hops}
| {e
for (e, _, e) in triples + hops})
predicate_lst = sorted({p for (_, p, _) in triples + hops})
nb_entities, nb_predicates = len(entity_lst), len(predicate_lst)
entity_to_index = {e: i for i, e in enumerate(entity_lst)}
predicate_to_index = {p: i for i, p in enumerate(predicate_lst)}
kernel = GaussianKernel()
entity_embeddings = nn.Embedding(nb_entities,
embedding_size * 2,
sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates,
embedding_size * 2,
sparse=True)
fact_rel = torch.LongTensor(
np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.LongTensor(
np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.LongTensor(
np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
model = NeuralKB(entity_embeddings=entity_embeddings,
predicate_embeddings=predicate_embeddings,
kernel=kernel,
facts=facts)
reformulator = AttentiveReformulator(2, predicate_embeddings)
hoppy = SimpleHoppy(model, entity_embeddings, hops=reformulator)
for s, p, o in hops:
xs_np = np.array([entity_to_index[s]])
xp_np = np.array([predicate_to_index[p]])
xo_np = np.array([entity_to_index[o]])
with torch.no_grad():
xs = torch.LongTensor(xs_np)
xp = torch.LongTensor(xp_np)
xo = torch.LongTensor(xo_np)
xs_emb = entity_embeddings(xs)
xp_emb = predicate_embeddings(xp)
xo_emb = entity_embeddings(xo)
inf = hoppy.score(xp_emb, xs_emb, xo_emb)
inf_np = inf.cpu().numpy()
assert inf_np < 0.5
def test_learning_v1():
embedding_size = 50
triples, hops = [], []
for i in range(16):
triples += [(f'a{i}', 'p', f'b{i}'), (f'b{i}', 'q', f'c{i}')]
hops += [(f'a{i}', 'r', f'c{i}')]
entity_lst = sorted({e
for (e, _, _) in triples + hops}
| {e
for (e, _, e) in triples + hops})
predicate_lst = sorted({p for (_, p, _) in triples + hops})
nb_entities, nb_predicates = len(entity_lst), len(predicate_lst)
entity_to_index = {e: i for i, e in enumerate(entity_lst)}
predicate_to_index = {p: i for i, p in enumerate(predicate_lst)}
kernel = GaussianKernel()
entity_embeddings = nn.Embedding(nb_entities,
embedding_size * 2,
sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates,
embedding_size * 2,
sparse=True)
fact_rel = torch.LongTensor(
np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.LongTensor(
np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.LongTensor(
np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
for st in ['min', 'concat']:
model = NeuralKB(entity_embeddings=entity_embeddings,
predicate_embeddings=predicate_embeddings,
kernel=kernel,
facts=facts,
scoring_type=st)
for s in entity_lst:
for p in predicate_lst:
for o in entity_lst:
xs_np = np.array([entity_to_index[s]])
xp_np = np.array([predicate_to_index[p]])
xo_np = np.array([entity_to_index[o]])
with torch.no_grad():
xs = torch.LongTensor(xs_np)
xp = torch.LongTensor(xp_np)
xo = torch.LongTensor(xo_np)
xs_emb = entity_embeddings(xs)
xp_emb = predicate_embeddings(xp)
xo_emb = entity_embeddings(xo)
inf = model.score(xp_emb, xs_emb, xo_emb)
inf_np = inf.cpu().numpy()
if (s, p, o) in triples:
assert inf_np[0] > 0.95
else:
assert inf_np[0] < 0.01
def test_learning_v3():
embedding_size = 10
batch_size = 16
triples, hops = [], []
for i in range(16):
triples += [(f'a{i}', 'p', f'b{i}'), (f'b{i}', 'q', f'c{i}')]
hops += [(f'a{i}', 'r', f'c{i}')]
entity_lst = sorted({e
for (e, _, _) in triples + hops}
| {e
for (e, _, e) in triples + hops})
predicate_lst = sorted({p for (_, p, _) in triples + hops})
nb_entities, nb_predicates = len(entity_lst), len(predicate_lst)
entity_to_index = {e: i for i, e in enumerate(entity_lst)}
predicate_to_index = {p: i for i, p in enumerate(predicate_lst)}
torch.manual_seed(0)
kernel = GaussianKernel()
entity_embeddings = nn.Embedding(nb_entities,
embedding_size * 2,
sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates,
embedding_size * 2,
sparse=True)
fact_rel = torch.LongTensor(
np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.LongTensor(
np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.LongTensor(
np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
model = NeuralKB(entity_embeddings=entity_embeddings,
predicate_embeddings=predicate_embeddings,
kernel=kernel,
facts=facts)
reformulator = AttentiveReformulator(2, predicate_embeddings)
hoppy = SimpleHoppy(model, entity_embeddings, hops=reformulator)
N3_reg = N3()
params = [
p for p in hoppy.parameters()
if not torch.equal(p, entity_embeddings.weight)
and not torch.equal(p, predicate_embeddings.weight)
]
loss_function = nn.CrossEntropyLoss(reduction='mean')
p_emb = predicate_embeddings(
torch.LongTensor(np.array([predicate_to_index['p']])))
q_emb = predicate_embeddings(
torch.LongTensor(np.array([predicate_to_index['q']])))
# r_emb = predicate_embeddings(torch.LongTensor(np.array([predicate_to_index['r']])))
optimizer = optim.Adagrad(params, lr=0.1)
hops_data = []
for i in range(128):
hops_data += hops
batches = make_batches(len(hops_data), batch_size)
c, d = 0.0, 0.0
for batch_start, batch_end in batches:
hops_batch = hops_data[batch_start:batch_end]
s_lst = [s for (s, _, _) in hops_batch]
p_lst = [p for (_, p, _) in hops_batch]
o_lst = [o for (_, _, o) in hops_batch]
xs_np = np.array([entity_to_index[s] for s in s_lst])
xp_np = np.array([predicate_to_index[p] for p in p_lst])
xo_np = np.array([entity_to_index[o] for o in o_lst])
xs = torch.LongTensor(xs_np)
xp = torch.LongTensor(xp_np)
xo = torch.LongTensor(xo_np)
xs_emb = entity_embeddings(xs)
xp_emb = predicate_embeddings(xp)
xo_emb = entity_embeddings(xo)
sp_scores, po_scores = hoppy.forward(xp_emb, xs_emb, xo_emb)
loss = loss_function(sp_scores, xo) + loss_function(po_scores, xs)
factors = [hoppy.factor(e) for e in [xp_emb, xs_emb, xo_emb]]
loss += 0.1 * N3_reg(factors)
tmp = hoppy.hops(xp_emb)
hop_1_emb = tmp[0]
hop_2_emb = tmp[1]
c = kernel.pairwise(p_emb, hop_1_emb).mean().cpu().detach().numpy()
d = kernel.pairwise(q_emb, hop_2_emb).mean().cpu().detach().numpy()
loss.backward()
optimizer.step()
optimizer.zero_grad()
assert c > 0.95
assert d > 0.95
def test_reasoning_v6():
torch.set_num_threads(multiprocessing.cpu_count())
embedding_size = 50
torch.manual_seed(0)
rs = np.random.RandomState(0)
triples = [('a', 'p', 'b'), ('b', 'q', 'c'), ('c', 'p', 'd'),
('d', 'q', 'e'), ('e', 'p', 'f'), ('f', 'q', 'g'),
('g', 'p', 'h'), ('h', 'q', 'i'), ('i', 'p', 'l'),
('l', 'q', 'm'), ('m', 'p', 'n'), ('n', 'q', 'o'),
('o', 'p', 'p'), ('p', 'q', 'q'), ('q', 'p', 'r'),
('r', 'q', 's'), ('s', 'p', 't'), ('t', 'q', 'u'),
('u', 'p', 'v'), ('v', 'q', 'w'), ('x', 'r', 'y'),
('x', 's', 'y')]
entity_lst = sorted({e
for (e, _, _) in triples}
| {e
for (_, _, e) in triples})
predicate_lst = sorted({p for (_, p, _) in triples})
nb_entities = len(entity_lst)
nb_predicates = len(predicate_lst)
entity_to_index = {e: i for i, e in enumerate(entity_lst)}
predicate_to_index = {p: i for i, p in enumerate(predicate_lst)}
for st in ['min', 'concat']:
with torch.no_grad():
kernel = GaussianKernel()
entity_embeddings = nn.Embedding(nb_entities,
embedding_size * 2,
sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates,
embedding_size * 2,
sparse=True)
fact_rel = torch.from_numpy(
np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.from_numpy(
np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.from_numpy(
np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
model = NeuralKB(entity_embeddings=entity_embeddings,
predicate_embeddings=predicate_embeddings,
kernel=kernel,
facts=facts,
scoring_type=st)
indices = torch.from_numpy(
np.array([predicate_to_index['p'], predicate_to_index['q']]))
reformulator = SymbolicReformulator(predicate_embeddings, indices)
k = 5
rhoppy0 = RecursiveHoppy(model,
entity_embeddings,
hops=reformulator,
depth=0,
k=k)
rhoppy1 = RecursiveHoppy(model,
entity_embeddings,
hops=reformulator,
depth=1,
k=k)
rhoppy2 = RecursiveHoppy(model,
entity_embeddings,
hops=reformulator,
depth=2,
k=k)
rhoppy3 = RecursiveHoppy(model,
entity_embeddings,
hops=reformulator,
depth=3,
k=k)
rhoppy4 = RecursiveHoppy(model,
entity_embeddings,
hops=reformulator,
depth=4,
k=k)
xs_np = rs.randint(nb_entities, size=12)
xp_np = rs.randint(nb_predicates, size=12)
xo_np = rs.randint(nb_entities, size=12)
xs_np[0] = entity_to_index['a']
xp_np[0] = predicate_to_index['r']
xo_np[0] = entity_to_index['c']
xs_np[1] = entity_to_index['a']
xp_np[1] = predicate_to_index['r']
xo_np[1] = entity_to_index['e']
xs_np[2] = entity_to_index['a']
xp_np[2] = predicate_to_index['r']
xo_np[2] = entity_to_index['g']
xs_np[3] = entity_to_index['a']
xp_np[3] = predicate_to_index['r']
xo_np[3] = entity_to_index['i']
xs_np[4] = entity_to_index['a']
xp_np[4] = predicate_to_index['r']
xo_np[4] = entity_to_index['m']
xs_np[5] = entity_to_index['a']
xp_np[5] = predicate_to_index['r']
xo_np[5] = entity_to_index['o']
xs_np[6] = entity_to_index['a']
xp_np[6] = predicate_to_index['r']
xo_np[6] = entity_to_index['q']
xs_np[7] = entity_to_index['a']
xp_np[7] = predicate_to_index['r']
xo_np[7] = entity_to_index['s']
xs_np[8] = entity_to_index['a']
xp_np[8] = predicate_to_index['r']
xo_np[8] = entity_to_index['u']
# xs_np[9] = entity_to_index['a']
# xp_np[9] = predicate_to_index['r']
# xo_np[9] = entity_to_index['w']
xs = torch.from_numpy(xs_np)
xp = torch.from_numpy(xp_np)
xo = torch.from_numpy(xo_np)
xs_emb = entity_embeddings(xs)
xp_emb = predicate_embeddings(xp)
xo_emb = entity_embeddings(xo)
scores0 = rhoppy0.forward(xp_emb, xs_emb, xo_emb)
inf0 = rhoppy0.score(xp_emb, xs_emb, xo_emb)
for i in range(xs.shape[0]):
scores_sp, scores_po = scores0
inf_np = inf0.cpu().numpy()
scores_sp_np = scores_sp.cpu().numpy()
scores_po_np = scores_po.cpu().numpy()
np.testing.assert_allclose(inf_np[i],
scores_sp_np[i, xo[i]],
rtol=1e-5,
atol=1e-5)
np.testing.assert_allclose(inf_np[i],
scores_po_np[i, xs[i]],
rtol=1e-5,
atol=1e-5)
scores1 = rhoppy1.forward(xp_emb, xs_emb, xo_emb)
inf1 = rhoppy1.score(xp_emb, xs_emb, xo_emb)
for i in range(xs.shape[0]):
scores_sp, scores_po = scores1
inf_np = inf1.cpu().numpy()
scores_sp_np = scores_sp.cpu().numpy()
scores_po_np = scores_po.cpu().numpy()
np.testing.assert_allclose(inf_np[i],
scores_sp_np[i, xo[i]],
rtol=1e-5,
atol=1e-5)
np.testing.assert_allclose(inf_np[i],
scores_po_np[i, xs[i]],
rtol=1e-5,
atol=1e-5)
scores2 = rhoppy2.forward(xp_emb, xs_emb, xo_emb)
inf2 = rhoppy2.score(xp_emb, xs_emb, xo_emb)
for i in range(xs.shape[0]):
scores_sp, scores_po = scores2
inf_np = inf2.cpu().numpy()
scores_sp_np = scores_sp.cpu().numpy()
scores_po_np = scores_po.cpu().numpy()
np.testing.assert_allclose(inf_np[i],
scores_sp_np[i, xo[i]],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf_np[i],
scores_po_np[i, xs[i]],
rtol=1e-1,
atol=1e-1)
scores3 = rhoppy3.forward(xp_emb, xs_emb, xo_emb)
inf3 = rhoppy3.score(xp_emb, xs_emb, xo_emb)
for i in range(xs.shape[0]):
scores_sp, scores_po = scores3
inf_np = inf3.cpu().numpy()
scores_sp_np = scores_sp.cpu().numpy()
scores_po_np = scores_po.cpu().numpy()
np.testing.assert_allclose(inf_np[i],
scores_sp_np[i, xo[i]],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf_np[i],
scores_po_np[i, xs[i]],
rtol=1e-1,
atol=1e-1)
scores4 = rhoppy4.forward(xp_emb, xs_emb, xo_emb)
inf4 = rhoppy4.score(xp_emb, xs_emb, xo_emb)
for i in range(xs.shape[0]):
scores_sp, scores_po = scores4
inf_np = inf4.cpu().numpy()
scores_sp_np = scores_sp.cpu().numpy()
scores_po_np = scores_po.cpu().numpy()
np.testing.assert_allclose(inf_np[i],
scores_sp_np[i, xo[i]],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf_np[i],
scores_po_np[i, xs[i]],
rtol=1e-1,
atol=1e-1)
print(inf0)
print(inf1)
print(inf2)
print(inf3)
print(inf4)
inf0_np = inf0.cpu().numpy()
inf1_np = inf1.cpu().numpy()
inf2_np = inf2.cpu().numpy()
inf3_np = inf3.cpu().numpy()
inf4_np = inf4.cpu().numpy()
np.testing.assert_allclose(inf0_np,
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf1_np,
[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf2_np,
[1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf3_np,
[1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0],
rtol=1e-1,
atol=1e-1)
np.testing.assert_allclose(inf4_np,
[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0],
rtol=1e-1,
atol=1e-1)
def test_reasoning_v5():
torch.set_num_threads(multiprocessing.cpu_count())
nb_entities = 10
nb_predicates = 5
embedding_size = 10
rs = np.random.RandomState(0)
triples = [('a', 'p', 'b'), ('b', 'q', 'c'), ('c', 'r', 'd'),
('d', 's', 'e')]
entity_to_index = {'a': 0, 'b': 1, 'c': 2, 'd': 3, 'e': 4}
predicate_to_index = {'p': 0, 'q': 1, 'r': 2, 's': 3}
for st in ['min', 'concat']:
with torch.no_grad():
kernel = GaussianKernel()
entity_embeddings = nn.Embedding(nb_entities,
embedding_size * 2,
sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates,
embedding_size * 2,
sparse=True)
fact_rel = torch.from_numpy(
np.array([predicate_to_index[p] for (_, p, _) in triples]))
fact_arg1 = torch.from_numpy(
np.array([entity_to_index[s] for (s, _, _) in triples]))
fact_arg2 = torch.from_numpy(
np.array([entity_to_index[o] for (_, _, o) in triples]))
facts = [fact_rel, fact_arg1, fact_arg2]
model = NeuralKB(entity_embeddings=entity_embeddings,
predicate_embeddings=predicate_embeddings,
kernel=kernel,
facts=facts,
scoring_type=st)
indices = torch.from_numpy(
np.array([
predicate_to_index['p'], predicate_to_index['q'],
predicate_to_index['r'], predicate_to_index['s']
]))
reformulator = SymbolicReformulator(predicate_embeddings, indices)
hoppy = SimpleHoppy(model, entity_embeddings, hops=reformulator)
rhoppy = RecursiveHoppy(model,
entity_embeddings,
hops=reformulator,
depth=1)
xs_np = rs.randint(nb_entities, size=32)
xp_np = rs.randint(nb_predicates, size=32)
xo_np = rs.randint(nb_entities, size=32)
xs_np[0] = 0
xp_np[0] = 0
xo_np[0] = 1
xs_np[1] = 1
xp_np[1] = 1
xo_np[1] = 2
xs_np[2] = 0
xp_np[2] = 3
xo_np[2] = 4
xs = torch.from_numpy(xs_np)
xp = torch.from_numpy(xp_np)
xo = torch.from_numpy(xo_np)
xs_emb = entity_embeddings(xs)
xp_emb = predicate_embeddings(xp)
xo_emb = entity_embeddings(xo)
scores = hoppy.forward(xp_emb, xs_emb, xo_emb)
inf = hoppy.score(xp_emb, xs_emb, xo_emb)
scores_h = rhoppy.depth_r_forward(xp_emb, xs_emb, xo_emb, depth=1)
inf_h = rhoppy.depth_r_score(xp_emb, xs_emb, xo_emb, depth=1)
print(inf)
print(inf_h)
assert inf[2] > 0.95
scores_sp, scores_po = scores
scores_h_sp, scores_h_po = scores_h
inf = inf.cpu().numpy()
scores_sp = scores_sp.cpu().numpy()
scores_po = scores_po.cpu().numpy()
inf_h = inf_h.cpu().numpy()
scores_h_sp = scores_h_sp.cpu().numpy()
scores_h_po = scores_h_po.cpu().numpy()
np.testing.assert_allclose(inf, inf_h)
np.testing.assert_allclose(scores_sp, scores_h_sp)
np.testing.assert_allclose(scores_po, scores_h_po)
for i in range(xs.shape[0]):
np.testing.assert_allclose(inf[i],
scores_sp[i, xo[i]],
rtol=1e-5,
atol=1e-5)
np.testing.assert_allclose(inf[i],
scores_po[i, xs[i]],
rtol=1e-5,
atol=1e-5)
np.testing.assert_allclose(inf_h[i],
scores_h_sp[i, xo[i]],
rtol=1e-5,
atol=1e-5)
np.testing.assert_allclose(inf_h[i],
scores_h_po[i, xs[i]],
rtol=1e-5,
atol=1e-5)
