def test_smart_clutrr_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, sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates, embedding_size, sparse=True)
for scoring_type in ['concat']:
model = NeuralKB(kernel=kernel)
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.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)
rel_emb = encode_relation(facts=triples, relation_embeddings=predicate_embeddings,
relation_to_idx=predicate_to_index)
arg1_emb, arg2_emb = encode_arguments(facts=triples, entity_embeddings=entity_embeddings,
entity_to_idx=entity_to_index)
facts = [rel_emb, arg1_emb, arg2_emb]
inf = model.score(xp_emb, xs_emb, xo_emb, facts=facts)
inf_np = inf.cpu().numpy()
assert inf_np[0] > 0.95 if (s, p, o) in triples else inf_np[0] < 0.01
def test_smart_clutrr_v2():
embedding_size = 20
triples, hops = [], []
xxx = []
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}')]
xxx += [(f'a{i}', 'p', f'c{i}'), (f'a{i}', 'q', f'c{i}'), (f'a{i}', 'r', f'c{i}')]
entity_lst = sorted({s for (s, _, _) in triples + hops} | {o for (_, _, o) 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, sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates, embedding_size, sparse=True)
for scoring_type in ['concat']:
model = NeuralKB(kernel=kernel)
indices = torch.from_numpy(np.array([predicate_to_index['p'], predicate_to_index['q']]))
_hops = SymbolicReformulator(predicate_embeddings, indices)
hoppy = Hoppy(model, hops_lst=[(_hops, False)], depth=1)
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.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)
rel_emb = encode_relation(facts=triples, relation_embeddings=predicate_embeddings,
relation_to_idx=predicate_to_index)
arg1_emb, arg2_emb = encode_arguments(facts=triples, entity_embeddings=entity_embeddings,
entity_to_idx=entity_to_index)
facts = [rel_emb, arg1_emb, arg2_emb]
inf = hoppy.score(xp_emb, xs_emb, xo_emb, facts=facts)
inf_np = inf.cpu().numpy()
print(s, p, o, inf_np)
assert inf_np[0] > 0.9 if (s, p, o) in (triples + xxx) else inf_np[0] < 0.1
def test_smart_clutrr_v7():
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)}
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)
rel_emb = encode_relation(facts=triples,
relation_embeddings=predicate_embeddings,
relation_to_idx=predicate_to_index)
arg1_emb, arg2_emb = encode_arguments(
facts=triples,
entity_embeddings=entity_embeddings,
entity_to_idx=entity_to_index)
facts = [rel_emb, arg1_emb, arg2_emb]
k = 5
model = NeuralKB(kernel=kernel, k=k)
indices = torch.LongTensor(
np.array([predicate_to_index['p'], predicate_to_index['q']]))
reformulator = SymbolicReformulator(predicate_embeddings, indices)
hoppy0 = Hoppy(model, hops_lst=[(reformulator, False)], depth=0)
hoppy1 = Hoppy(model, hops_lst=[(reformulator, False)], depth=1)
hoppy2 = Hoppy(model, hops_lst=[(reformulator, False)], depth=2)
hoppy3 = Hoppy(model, hops_lst=[(reformulator, False)], depth=3)
# hoppy4 = Hoppy(model, hops_lst=[(reformulator, False)], depth=4)
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 = 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)
res0 = hoppy0.forward(xp_emb, xs_emb, xo_emb, facts=facts)
inf0 = hoppy0.score(xp_emb, xs_emb, xo_emb, facts=facts)
(scores0_sp, subs0_sp), (scores0_po, subs0_po) = res0
res1 = hoppy1.forward(xp_emb, xs_emb, xo_emb, facts=facts)
inf1 = hoppy1.score(xp_emb, xs_emb, xo_emb, facts=facts)
(scores1_sp, subs1_sp), (scores1_po, subs1_po) = res1
res2 = hoppy2.forward(xp_emb, xs_emb, xo_emb, facts=facts)
inf2 = hoppy2.score(xp_emb, xs_emb, xo_emb, facts=facts)
(scores2_sp, subs2_sp), (scores2_po, subs2_po) = res2
res3 = hoppy3.forward(xp_emb, xs_emb, xo_emb, facts=facts)
inf3 = hoppy3.score(xp_emb, xs_emb, xo_emb, facts=facts)
(scores3_sp, subs3_sp), (scores3_po, subs3_po) = res3
# res4 = hoppy4.forward(xp_emb, xs_emb, xo_emb, facts=facts)
# inf4 = hoppy4.score(xp_emb, xs_emb, xo_emb, facts=facts)
# (scores4_sp, subs4_sp), (scores4_po, subs4_po) = res4
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()
scores0_sp_np = scores0_sp.cpu().numpy()
scores1_sp_np = scores1_sp.cpu().numpy()
scores2_sp_np = scores2_sp.cpu().numpy()
scores3_sp_np = scores3_sp.cpu().numpy()
# scores4_sp_np = scores4_sp.cpu().numpy()
subs0_sp_np = subs0_sp.cpu().numpy()
subs1_sp_np = subs1_sp.cpu().numpy()
subs2_sp_np = subs2_sp.cpu().numpy()
subs3_sp_np = subs3_sp.cpu().numpy()
# subs4_sp_np = subs4_sp.cpu().numpy()
scores0_po_np = scores0_po.cpu().numpy()
scores1_po_np = scores1_po.cpu().numpy()
scores2_po_np = scores2_po.cpu().numpy()
scores3_po_np = scores3_po.cpu().numpy()
# scores4_po_np = scores4_po.cpu().numpy()
subs0_po_np = subs0_po.cpu().numpy()
subs1_po_np = subs1_po.cpu().numpy()
subs2_po_np = subs2_po.cpu().numpy()
subs3_po_np = subs3_po.cpu().numpy()
# subs4_po_np = subs4_po.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)
def test_smart_clutrr_v4():
embedding_size = 50
rs = np.random.RandomState(0)
for _ in range(32):
with torch.no_grad():
triples = [('a', 'p', 'b'), ('c', 'q', 'd'), ('e', 'q', 'f'),
('g', 'q', 'h'), ('i', 'q', 'l'), ('m', 'q', 'n'),
('o', 'q', 'p'), ('q', 'q', 'r'), ('s', 'q', 't'),
('u', 'q', 'v')]
entity_lst = sorted({s
for (s, _, _) in triples}
| {o
for (_, _, o) in triples})
predicate_lst = sorted({p for (_, p, _) in triples})
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)
rel_emb = encode_relation(facts=triples,
relation_embeddings=predicate_embeddings,
relation_to_idx=predicate_to_index)
arg1_emb, arg2_emb = encode_arguments(
facts=triples,
entity_embeddings=entity_embeddings,
entity_to_idx=entity_to_index)
facts = [rel_emb, arg1_emb, arg2_emb]
model = NeuralKB(kernel=kernel)
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] = 2
xp_np[1] = 1
xo_np[1] = 3
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)
# print('xp_emb', xp_emb.shape)
res_sp, res_po = model.forward(xp_emb, xs_emb, xo_emb, facts=facts)
inf = model.score(xp_emb, xs_emb, xo_emb, facts=facts)
assert inf[0] > 0.9
assert inf[1] > 0.9
scores_sp, emb_sp = res_sp
scores_po, emb_po = res_po
# print(scores_sp.shape, emb_sp.shape)
# print(scores_po.shape, emb_po.shape)
inf = inf.cpu().numpy()
scores_sp = scores_sp.cpu().numpy()
scores_po = scores_po.cpu().numpy()
def test_smart_clutrr_v3():
embedding_size = 20
batch_size = 8
torch.manual_seed(0)
triples, hops = [], []
for i in range(32):
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({s
for (s, _, _) in triples + hops}
| {o
for (_, _, o) 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(slope=None)
entity_embeddings = nn.Embedding(nb_entities, embedding_size, sparse=True)
predicate_embeddings = nn.Embedding(nb_predicates,
embedding_size,
sparse=True)
# _hops = LinearReformulator(2, embedding_size)
_hops = AttentiveReformulator(2, predicate_embeddings)
model = NeuralKB(kernel=kernel)
hoppy = Hoppy(model, hops_lst=[(_hops, False)], depth=1)
params = [
p for p in hoppy.parameters()
if not torch.equal(p, entity_embeddings.weight)
and not torch.equal(p, predicate_embeddings.weight)
]
# for tensor in params:
# print(f'\t{tensor.size()}\t{tensor.device}')
loss_function = nn.BCELoss()
optimizer = optim.Adagrad(params, lr=0.1)
hops_data = []
for i in range(64):
hops_data += hops
batches = make_batches(len(hops_data), batch_size)
rs = np.random.RandomState()
c, d = 0.0, 0.0
p_emb = predicate_embeddings(
torch.LongTensor(np.array([predicate_to_index['p']])))
q_emb = predicate_embeddings(
torch.LongTensor(np.array([predicate_to_index['q']])))
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]
nb_positives = len(s_lst)
nb_negatives = nb_positives * 3
s_n_lst = rs.permutation(nb_entities)[:nb_negatives].tolist()
nb_negatives = len(s_n_lst)
o_n_lst = rs.permutation(nb_entities)[:nb_negatives].tolist()
p_n_lst = list(islice(cycle(p_lst), nb_negatives))
xs_np = np.array([entity_to_index[s] for s in s_lst] + s_n_lst)
xp_np = np.array([predicate_to_index[p] for p in p_lst + p_n_lst])
xo_np = np.array([entity_to_index[o] for o in o_lst] + o_n_lst)
xs_emb = entity_embeddings(torch.LongTensor(xs_np))
xp_emb = predicate_embeddings(torch.LongTensor(xp_np))
xo_emb = entity_embeddings(torch.LongTensor(xo_np))
rel_emb = encode_relation(facts=triples,
relation_embeddings=predicate_embeddings,
relation_to_idx=predicate_to_index)
arg1_emb, arg2_emb = encode_arguments(
facts=triples,
entity_embeddings=entity_embeddings,
entity_to_idx=entity_to_index)
facts = [rel_emb, arg1_emb, arg2_emb]
scores = hoppy.score(xp_emb, xs_emb, xo_emb, facts=facts)
labels_np = np.zeros(xs_np.shape[0])
labels_np[:nb_positives] = 1
labels = torch.LongTensor(labels_np).float()
loss = loss_function(scores, labels)
hop_1_emb = hoppy.hops_lst[0][0].hops_lst[0](xp_emb)
hop_2_emb = hoppy.hops_lst[0][0].hops_lst[1](xp_emb)
c = kernel.pairwise(p_emb, hop_1_emb).mean().cpu().detach().numpy()
d = kernel.pairwise(q_emb, hop_2_emb).mean().cpu().detach().numpy()
# print(c, d)
loss.backward()
optimizer.step()
optimizer.zero_grad()
assert c > 0.95 and d > 0.95
def main(argv):
argparser = argparse.ArgumentParser(
'CLUTRR', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
train_path = test_path = "data/clutrr-emnlp/data_test/64.csv"
argparser.add_argument('--train',
action='store',
type=str,
default=train_path)
argparser.add_argument('--test', nargs='+', type=str, default=[test_path])
# model params
argparser.add_argument('--embedding-size',
'-k',
action='store',
type=int,
default=20)
argparser.add_argument('--k-max',
'-m',
action='store',
type=int,
default=10)
argparser.add_argument('--max-depth',
'-d',
action='store',
type=int,
default=2)
argparser.add_argument('--hops',
nargs='+',
type=str,
default=['2', '2', '1R'])
# training params
argparser.add_argument('--epochs',
'-e',
action='store',
type=int,
default=100)
argparser.add_argument('--learning-rate',
'-l',
action='store',
type=float,
default=0.1)
argparser.add_argument('--batch-size',
'-b',
action='store',
type=int,
default=8)
argparser.add_argument('--optimizer',
'-o',
action='store',
type=str,
default='adagrad',
choices=['adagrad', 'adam', 'sgd'])
argparser.add_argument('--seed', action='store', type=int, default=0)
argparser.add_argument('--evaluate-every',
'-V',
action='store',
type=int,
default=32)
argparser.add_argument('--N2', action='store', type=float, default=None)
argparser.add_argument('--N3', action='store', type=float, default=None)
argparser.add_argument('--entropy',
'-E',
action='store',
type=float,
default=None)
argparser.add_argument('--reformulator',
'-r',
action='store',
type=str,
default='linear',
choices=['static', 'linear', 'attentive', 'memory'])
argparser.add_argument('--nb-rules',
'-R',
action='store',
type=int,
default=4)
argparser.add_argument('--slope',
'-S',
action='store',
type=float,
default=None)
argparser.add_argument('--init-size',
'-i',
action='store',
type=float,
default=1.0)
argparser.add_argument('--debug', '-D', action='store_true', default=False)
args = argparser.parse_args(argv)
train_path = args.train
test_paths = args.test
embedding_size = args.embedding_size
k_max = args.k_max
max_depth = args.max_depth
hops_str = args.hops
nb_epochs = args.epochs
learning_rate = args.learning_rate
batch_size = args.batch_size
optimizer_name = args.optimizer
seed = args.seed
evaluate_every = args.evaluate_every
N2_weight = args.N2
N3_weight = args.N3
entropy_weight = args.entropy
reformulator_name = args.reformulator
nb_rules = args.nb_rules
slope = args.slope
init_size = args.init_size
is_debug = args.debug
np.random.seed(seed)
random_state = np.random.RandomState(seed)
torch.manual_seed(seed)
device = torch.device('cuda' if torch.cuda.is_available() else '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, test_paths=test_paths)
relation_to_predicate = data.relation_to_predicate
predicate_to_relations = data.predicate_to_relations
entity_lst, predicate_lst, relation_lst = data.entity_lst, data.predicate_lst, data.relation_lst
nb_examples = len(data.train)
nb_entities = len(entity_lst)
nb_predicates = len(predicate_lst)
nb_relations = len(relation_lst)
entity_to_idx = {e: i for i, e in enumerate(entity_lst)}
relation_to_idx = {r: i for i, r in enumerate(relation_lst)}
kernel = GaussianKernel(slope=slope)
entity_embeddings = nn.Embedding(nb_entities, embedding_size,
sparse=False).to(device)
if entropy_weight is None:
relation_embeddings = nn.Embedding(nb_relations,
embedding_size,
sparse=False).to(device)
relation_embeddings.weight.data *= init_size
else:
relation_embeddings = AttentiveEmbedding(nb_predicates=nb_predicates,
nb_relations=nb_relations,
embedding_size=embedding_size,
device=device).to(device)
make_easy(predicate_lst, predicate_to_relations, relation_to_idx,
relation_embeddings)
model = NeuralKB(kernel=kernel, k=k_max).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_name in {'static'}:
res = StaticReformulator(nb_hops, embedding_size)
elif reformulator_name in {'linear'}:
res = LinearReformulator(nb_hops, embedding_size)
elif reformulator_name in {'attentive'}:
res = AttentiveReformulator(nb_hops, relation_embeddings)
elif reformulator_name in {'memory'}:
memory = MemoryReformulator.Memory(
nb_hops, nb_rules,
embedding_size) if memory is None else memory
res = MemoryReformulator(memory)
assert res is not None
return res, is_reversed
hops_lst = [make_hop(s) for s in hops_str]
hoppy = Hoppy(model=model, depth=max_depth, hops_lst=hops_lst).to(device)
def scoring_function(story: List[Fact], targets: List[Fact]) -> Tensor:
story_rel = encode_relation(story, relation_embeddings,
relation_to_idx, device)
story_arg1, story_arg2 = encode_arguments(story, entity_embeddings,
entity_to_idx, device)
targets_rel = encode_relation(targets, relation_embeddings,
relation_to_idx, device)
targets_arg1, targets_arg2 = encode_arguments(targets,
entity_embeddings,
entity_to_idx, device)
facts = [story_rel, story_arg1, story_arg2]
scores = hoppy.score(targets_rel, targets_arg1, targets_arg2, facts)
return scores
def evaluate(instances: List[Instance],
path: str,
sample_size: Optional[int] = None) -> float:
res = 0.0
if len(instances) > 0:
res = accuracy(scoring_function=scoring_function,
instances=instances,
sample_size=sample_size,
relation_to_predicate=relation_to_predicate,
predicate_to_relations=predicate_to_relations)
logger.info(f'Test Accuracy on {path}: {res:.6f}')
return res
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
entropy_reg = Entropy(
use_logits=False) if entropy_weight is not None else None
params_lst = [
p for p in hoppy.parameters()
if not torch.equal(p, entity_embeddings.weight)
]
params_lst += relation_embeddings.parameters()
params = nn.ParameterList(params_lst).to(device)
for tensor in params_lst:
logger.info(f'\t{tensor.size()}\t{tensor.device}')
optimizer_factory = {
'adagrad': lambda arg: optim.Adagrad(arg, lr=learning_rate),
'adam': lambda arg: optim.Adam(arg, lr=learning_rate),
'sgd': lambda arg: optim.SGD(arg, lr=learning_rate)
}
assert optimizer_name in optimizer_factory
optimizer = optimizer_factory[optimizer_name](params)
global_step = 0
hinton = HintonDiagram(max_arr=[0.0, 1.0])
for epoch_no in range(1, nb_epochs + 1):
batcher = Batcher(batch_size=batch_size,
nb_examples=nb_examples,
nb_epochs=1,
random_state=random_state)
nb_batches = len(batcher.batches)
epoch_loss_values = []
for batch_no, (batch_start, batch_end) in enumerate(batcher.batches,
start=1):
global_step += 1
indices_batch = batcher.get_batch(batch_start, batch_end)
instances_batch = [data.train[i] for i in indices_batch]
batch_loss_values = []
for i, instance in enumerate(instances_batch):
story, target = instance.story, instance.target
s, r, o = target
if is_debug is True and i == 0:
# print('STORY', story)
# print('TARGET', target)
r_lst = [
r for p in predicate_lst
for r in predicate_to_relations[p]
]
r_idx_lst = [relation_to_idx[r] for r in r_lst]
with torch.no_grad():
# show_rules(model=hoppy, kernel=kernel, relation_embeddings=relation_embeddings,
# data=data, relation_to_idx=relation_to_idx, device=device)
r_idx_tensor = torch.from_numpy(
np.array(r_idx_lst, dtype=np.int64)).to(device)
r_tensor = relation_embeddings(r_idx_tensor)
k = kernel.pairwise(r_tensor, r_tensor)
# print(r_lst)
print(hinton(k.cpu().numpy()))
story_rel = encode_relation(story, relation_embeddings,
relation_to_idx, device)
story_arg1, story_arg2 = encode_arguments(
story, entity_embeddings, entity_to_idx, device)
facts = [story_rel, story_arg1, story_arg2]
pos_predicate = relation_to_predicate[r]
p_relation_lst = sorted(relation_to_predicate.keys())
target_lst = [(s, x, o) for x in p_relation_lst]
label_lst = [
int(pos_predicate == relation_to_predicate[r])
for r in p_relation_lst
]
rel_emb = encode_relation(target_lst, relation_embeddings,
relation_to_idx, device)
arg1_emb, arg2_emb = encode_arguments(target_lst,
entity_embeddings,
entity_to_idx, device)
scores = hoppy.score(rel_emb, arg1_emb, arg2_emb, facts)
labels = torch.Tensor(label_lst).float()
# if i == 0:
# print(scores)
# print(labels)
loss = loss_function(scores, labels)
factors = [
hoppy.factor(e) for e in [rel_emb, arg1_emb, arg2_emb]
]
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 entropy_weight is not None:
attention = relation_embeddings.attention
if i == 0:
pass
# print(scores.cpu().detach().numpy())
# print(labels.cpu().detach().numpy())
# print(hinton(attention.cpu().detach().numpy()))
# print(attention.cpu().detach().numpy())
loss += entropy_weight * entropy_reg([attention])
loss_value = loss.item()
batch_loss_values += [loss_value]
epoch_loss_values += [loss_value]
loss.backward()
optimizer.step()
optimizer.zero_grad()
loss_mean, loss_std = np.mean(batch_loss_values), np.std(
batch_loss_values)
logger.info(
f'Epoch {epoch_no}/{nb_epochs}\tBatch {batch_no}/{nb_batches}\tLoss {loss_mean:.4f} ± {loss_std:.4f}'
)
if global_step % evaluate_every == 0:
for test_path in test_paths:
instances = data.test[test_path]
evaluate(instances=instances, path=test_path)
if is_debug is True:
for i in range(3):
story, target = instances[i].story, instances[
i].target
# print('INSTANCE', target, story)
if is_debug is True:
r_lst = [
r for p in predicate_lst
for r in predicate_to_relations[p]
]
r_idx_lst = [relation_to_idx[r] for r in r_lst]
with torch.no_grad():
show_rules(model=hoppy,
kernel=kernel,
relation_embeddings=relation_embeddings,
data=data,
relation_to_idx=relation_to_idx,
device=device)
r_idx_tensor = torch.from_numpy(
np.array(r_idx_lst, dtype=np.int64)).to(device)
r_tensor = relation_embeddings(r_idx_tensor)
k = kernel.pairwise(r_tensor, r_tensor)
# print(r_lst)
print(hinton(k.cpu().numpy()))
loss_mean, loss_std = np.mean(epoch_loss_values), np.std(
epoch_loss_values)
slope = kernel.slope.item() if isinstance(kernel.slope,
Tensor) else kernel.slope
logger.info(
f'Epoch {epoch_no}/{nb_epochs}\tLoss {loss_mean:.4f} ± {loss_std:.4f}\tSlope {slope:.4f}'
)
for test_path in test_paths:
evaluate(instances=data.test[test_path], path=test_path)
logger.info("Training finished")