def test_evaluate_performance():
X = load_wn18()
model = ComplEx(batches_count=10,
seed=0,
epochs=10,
k=150,
eta=10,
loss='pairwise',
loss_params={'margin': 5},
regularizer=None,
optimizer='adagrad',
optimizer_params={'lr': 0.1},
verbose=True)
model.fit(np.concatenate((X['train'], X['valid'])))
filter = np.concatenate((X['train'], X['valid'], X['test']))
ranks = evaluate_performance(X['test'][:200],
model=model,
filter_triples=filter,
verbose=True)
mrr = mrr_score(ranks)
hits_10 = hits_at_n_score(ranks, n=10)
print("ranks: %s" % ranks)
print("MRR: %f" % mrr)
print("Hits@10: %f" % hits_10)
def test_evaluate_with_ent_subset_large_graph():
set_entity_threshold(1)
X = load_wn18()
model = ComplEx(batches_count=10, seed=0, epochs=2, k=10, eta=1,
optimizer='sgd', optimizer_params={'lr': 1e-5},
loss='pairwise', loss_params={'margin': 0.5},
regularizer='LP', regularizer_params={'p': 2, 'lambda': 1e-5},
verbose=True)
model.fit(X['train'])
X_filter = np.concatenate((X['train'], X['valid'], X['test']))
all_nodes = set(X_filter[:, 0]).union(X_filter[:, 2])
entities_subset = np.random.choice(list(all_nodes), 100, replace=False)
ranks = evaluate_performance(X['test'][::10],
model=model,
filter_triples=X_filter,
corrupt_side='o',
use_default_protocol=False,
entities_subset=list(entities_subset),
verbose=True)
assert np.sum(ranks > (100 + 1)) == 0, "No ranks must be greater than 101"
reset_entity_threshold()
def generate_model(X):
X_train, X_test = train_test_split_no_unseen(X, test_size=100)
print('Train set size: ', X_train.shape)
print('Test set size: ', X_test.shape)
model = ComplEx(batches_count=100,
seed=0,
epochs=10,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr': 1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={
'p': 3,
'lambda': 1e-5
},
verbose=True)
#positives_filter = X
tf.logging.set_verbosity(tf.logging.ERROR)
model.fit(X_train, early_stopping=False)
print("created the model")
save_model(model, './best_model.pkl')
return X_test
def test_evaluate_performance_so_side_corruptions_without_filter():
X = load_wn18()
model = ComplEx(batches_count=10,
seed=0,
epochs=5,
k=200,
eta=10,
loss='nll',
regularizer=None,
optimizer='adam',
optimizer_params={'lr': 0.01},
verbose=True)
model.fit(X['train'])
X_filter = np.concatenate((X['train'], X['valid'], X['test']))
ranks = evaluate_performance(X['test'][::20],
model,
X_filter,
verbose=True,
use_default_protocol=False,
corrupt_side='s+o')
mrr = mrr_score(ranks)
hits_10 = hits_at_n_score(ranks, n=10)
print("ranks: %s" % ranks)
print("MRR: %f" % mrr)
print("Hits@10: %f" % hits_10)
assert (mrr is not np.Inf)
def test_evaluate_performance_ranking_against_specified_entities():
X = load_wn18()
model = ComplEx(batches_count=10,
seed=0,
epochs=1,
k=20,
eta=10,
loss='nll',
regularizer=None,
optimizer='adam',
optimizer_params={'lr': 0.01},
verbose=True)
model.fit(X['train'])
X_filter = np.concatenate((X['train'], X['valid'], X['test']))
entities_subset = np.concatenate(
[X['test'][::1000, 0], X['test'][::1000, 2]], 0)
from ampligraph.evaluation import hits_at_n_score, mrr_score, mr_score
ranks = evaluate_performance(X['test'][::1000],
model,
X_filter,
verbose=True,
corrupt_side='s+o',
use_default_protocol=True,
entities_subset=entities_subset)
ranks = ranks.reshape(-1)
assert (np.sum(ranks > len(entities_subset)) == 0)
def test_evaluate_performance_ranking_against_shuffled_all_entities():
""" Compares mrr of test set by using default protocol against all entities vs
mrr of corruptions generated by corrupting using entities_subset = all entities shuffled
"""
import random
X = load_wn18()
model = ComplEx(batches_count=10,
seed=0,
epochs=1,
k=20,
eta=10,
loss='nll',
regularizer=None,
optimizer='adam',
optimizer_params={'lr': 0.01},
verbose=True)
model.fit(X['train'])
X_filter = np.concatenate((X['train'], X['valid'], X['test']))
entities_subset = random.shuffle(list(model.ent_to_idx.keys()))
from ampligraph.evaluation import hits_at_n_score, mrr_score, mr_score
ranks_all = evaluate_performance(X['test'][::1000],
model,
X_filter,
verbose=True,
corrupt_side='s,o')
ranks_suffled_ent = evaluate_performance(X['test'][::1000],
model,
X_filter,
verbose=True,
corrupt_side='s,o',
entities_subset=entities_subset)
assert (mrr_score(ranks_all) == mrr_score(ranks_suffled_ent))
def test_find_clusters():
X = np.array([['a', 'y', 'b'],
['b', 'y', 'a'],
['a', 'y', 'c'],
['c', 'y', 'a'],
['a', 'y', 'd'],
['c', 'x', 'd'],
['b', 'y', 'c'],
['f', 'y', 'e']])
model = ComplEx(k=2, batches_count=2)
model.fit(X)
clustering_algorithm = DBSCAN(eps=1e-3, min_samples=1)
labels = find_clusters(X, model, clustering_algorithm, mode='triple')
assert np.array_equal(labels, np.array([0, 1, 2, 3, 4, 5, 6, 7]))
labels = find_clusters(np.unique(X[:, 0]), model, clustering_algorithm, mode='entity')
assert np.array_equal(labels, np.array([0, 1, 2, 3]))
labels = find_clusters(np.unique(X[:, 1]), model, clustering_algorithm, mode='relation')
assert np.array_equal(labels, np.array([0, 1]))
labels = find_clusters(np.unique(X[:, 2]), model, clustering_algorithm, mode='entity')
assert np.array_equal(labels, np.array([0, 1, 2, 3, 4]))
with pytest.raises(ValueError):
find_clusters(X, model, clustering_algorithm, mode='hah')
with pytest.raises(ValueError):
find_clusters(X, model, clustering_algorithm, mode='entity')
with pytest.raises(ValueError):
find_clusters(X, model, clustering_algorithm, mode='relation')
with pytest.raises(ValueError):
find_clusters(np.unique(X[:, 0]), model, clustering_algorithm, mode='triple')
def test_fit_predict_wn18_ComplEx():
X = load_wn18()
model = ComplEx(batches_count=1, seed=555, epochs=5, k=100,
loss='pairwise', loss_params={'margin': 1}, regularizer='LP',
regularizer_params={'lambda': 0.1, 'p': 2},
optimizer='adagrad', optimizer_params={'lr':0.1})
model.fit(X['train'])
y = model.predict(X['test'][:1], get_ranks=True)
print(y)
def test_find_duplicates():
X = np.array([['a', 'y', 'b'], ['b', 'y', 'a'], ['a', 'y', 'c'],
['c', 'y', 'a'], ['a', 'y', 'd'], ['c', 'x', 'd'],
['b', 'y', 'c'], ['f', 'y', 'e']])
model = ComplEx(k=2, batches_count=2)
model.fit(X)
entities = set('a b c d e f'.split())
relations = set('x y'.split())
def asserts(tol, dups, ent_rel, subspace):
assert tol > 0.0
assert len(dups) <= len(ent_rel)
assert all(len(d) <= len(ent_rel) for d in dups)
assert all(d.issubset(subspace) for d in dups)
dups, tol = find_duplicates(X,
model,
mode='triple',
tolerance='auto',
expected_fraction_duplicates=0.5)
asserts(tol, dups, X, {tuple(x) for x in X})
dups, tol = find_duplicates(X, model, mode='triple', tolerance=1.0)
assert tol == 1.0
asserts(tol, dups, X, {tuple(x) for x in X})
dups, tol = find_duplicates(np.unique(X[:, 0]),
model,
mode='entity',
tolerance='auto',
expected_fraction_duplicates=0.5)
asserts(tol, dups, entities, entities)
dups, tol = find_duplicates(np.unique(X[:, 2]),
model,
mode='entity',
tolerance='auto',
expected_fraction_duplicates=0.5)
asserts(tol, dups, entities, entities)
dups, tol = find_duplicates(np.unique(X[:, 1]),
model,
mode='relation',
tolerance='auto',
expected_fraction_duplicates=0.5)
asserts(tol, dups, relations, relations)
with pytest.raises(ValueError):
find_duplicates(X, model, mode='hah')
with pytest.raises(ValueError):
find_duplicates(X, model, mode='entity')
with pytest.raises(ValueError):
find_duplicates(X, model, mode='relation')
with pytest.raises(ValueError):
find_duplicates(np.unique(X[:, 0]), model, mode='triple')
def test_evaluate_performance_filter_without_xtest():
X = load_wn18()
model = ComplEx(batches_count=10, seed=0, epochs=1, k=20, eta=10, loss='nll',
regularizer=None, optimizer='adam', optimizer_params={'lr': 0.01}, verbose=True)
model.fit(X['train'])
X_filter = np.concatenate((X['train'], X['valid'])) # filter does not contain X_test
from ampligraph.evaluation import hits_at_n_score, mrr_score, mr_score
ranks = evaluate_performance(X['test'][::1000], model, X_filter, verbose=True, corrupt_side='s,o')
assert(mrr_score(ranks)>0)
def kge(triples, kge_name, verbose):
# Train test split
t_size = math.ceil(len(triples)*0.2)
X_train, X_test = train_test_split_no_unseen(triples, test_size=t_size)
# Select kge_name
if kge_name == 'complex':
# ComplEx model
model = ComplEx(batches_count=50,
epochs=300,
k=100,
eta=20,
optimizer='adam',
optimizer_params={'lr':1e-4},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={'p':3, 'lambda':1e-5},
seed=0,
verbose=verbose)
else:
sys.exit('Given kge_name is not valid.')
model.fit(X_train)
#Embedding evaluation
if verbose:
filter_triples = np.concatenate((X_train, X_test))
ranks = evaluate_performance(X_test,
model=model,
filter_triples=filter_triples,
use_default_protocol=True,
verbose=True)
mrr = mrr_score(ranks)
print("MRR: %.2f" % (mrr))
mr = mr_score(ranks)
print("MR: %.2f" % (mr))
hits_10 = hits_at_n_score(ranks, n=10)
print("Hits@10: %.2f" % (hits_10))
hits_3 = hits_at_n_score(ranks, n=3)
print("Hits@3: %.2f" % (hits_3))
hits_1 = hits_at_n_score(ranks, n=1)
print("Hits@1: %.2f" % (hits_1))
print('''
- Ampligraph example -
MRR: 0.25
MR: 4927.33
Hits@10: 0.35
Hits@3: 0.28
Hits@1: 0.19
''')
return model
def test_missing_entity_ComplEx():
X = np.array([['a', 'y', 'b'], ['b', 'y', 'a'], ['a', 'y', 'c'],
['c', 'y', 'a'], ['a', 'y', 'd'], ['c', 'y', 'd'],
['b', 'y', 'c'], ['f', 'y', 'e']])
model = ComplEx(batches_count=1, seed=555, epochs=2, k=5)
model.fit(X)
with pytest.raises(ValueError):
model.predict(['a', 'y', 'zzzzzzzzzzz'])
with pytest.raises(ValueError):
model.predict(['a', 'xxxxxxxxxx', 'e'])
with pytest.raises(ValueError):
model.predict(['zzzzzzzz', 'y', 'e'])
def train_complex(train_samples: iter):
model = ComplEx(batches_count=100,
seed=0,
epochs=200,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr': 1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={
'p': 3,
'lambda': 1e-5
},
verbose=True)
model.fit(train_samples, early_stopping=False)
return model
def test_fit_predict_CompleEx():
model = ComplEx(batches_count=1, seed=555, epochs=20, k=10,
loss='pairwise', loss_params={'margin': 1}, regularizer='LP',
regularizer_params={'lambda': 0.1, 'p': 2},
optimizer='adagrad', optimizer_params={'lr':0.1})
X = np.array([['a', 'y', 'b'],
['b', 'y', 'a'],
['a', 'y', 'c'],
['c', 'y', 'a'],
['a', 'y', 'd'],
['c', 'y', 'd'],
['b', 'y', 'c'],
['f', 'y', 'e']])
model.fit(X)
y_pred, _ = model.predict(np.array([['f', 'y', 'e'], ['b', 'y', 'd']]), get_ranks=True)
print(y_pred)
assert y_pred[0] > y_pred[1]
def test_discover_facts():
X = np.array([['a', 'y', 'b'], ['b', 'y', 'a'], ['a', 'y', 'c'],
['c', 'y', 'a'], ['a', 'y', 'd'], ['c', 'y', 'd'],
['b', 'y', 'c'], ['f', 'y', 'e']])
model = ComplEx(batches_count=1, seed=555, epochs=2, k=5)
with pytest.raises(ValueError):
discover_facts(X, model)
model.fit(X)
with pytest.raises(ValueError):
discover_facts(X, model, strategy='error')
with pytest.raises(ValueError):
discover_facts(X, model, strategy='random_uniform', target_rel='error')
def test_large_graph_mode_adam():
set_entity_threshold(10)
X = load_wn18()
model = ComplEx(batches_count=100,
seed=555,
epochs=1,
k=50,
loss='multiclass_nll',
loss_params={'margin': 5},
verbose=True,
optimizer='adam',
optimizer_params={'lr': 0.001})
try:
model.fit(X['train'])
except Exception as e:
print(str(e))
reset_entity_threshold()
def test_large_graph_mode():
set_entity_threshold(10)
X = load_wn18()
model = ComplEx(batches_count=100,
seed=555,
epochs=1,
k=50,
loss='multiclass_nll',
loss_params={'margin': 5},
verbose=True,
optimizer='sgd',
optimizer_params={'lr': 0.001})
model.fit(X['train'])
X_filter = np.concatenate((X['train'], X['valid'], X['test']), axis=0)
evaluate_performance(X['test'][::1000],
model,
X_filter,
verbose=True,
corrupt_side='s,o')
y = model.predict(X['test'][:1])
print(y)
reset_entity_threshold()
def test_retrain():
model = ComplEx(batches_count=1, seed=555, epochs=20, k=10,
loss='pairwise', loss_params={'margin': 1}, regularizer='LP',
regularizer_params={'lambda': 0.1, 'p': 2},
optimizer='adagrad', optimizer_params={'lr':0.1})
X = np.array([['a', 'y', 'b'],
['b', 'y', 'a'],
['a', 'y', 'c'],
['c', 'y', 'a'],
['a', 'y', 'd'],
['c', 'y', 'd'],
['b', 'y', 'c'],
['f', 'y', 'e']])
model.fit(X)
y_pred_1st, _ = model.predict(np.array([['f', 'y', 'e'], ['b', 'y', 'd']]), get_ranks=True)
model.fit(X)
y_pred_2nd, _ = model.predict(np.array([['f', 'y', 'e'], ['b', 'y', 'd']]), get_ranks=True)
np.testing.assert_array_equal(y_pred_1st, y_pred_2nd)
ignore_index=True)
positives_filter = pd.concat(
[train_X_temp.iloc[:, :], test_X_temp.iloc[:, :]],
axis='index',
ignore_index=True)
train_y, test_y, positives_filter = to_categorical(
train_y, dtype=np.int32), to_categorical(
test_y, dtype=np.int32), positives_filter.to_numpy(dtype=np.int32)
print(
"Shape of train_y: %s; Shape of test_y: %s; Shape of positives_filter: %s"
% (train_y.shape, test_y.shape, positives_filter.shape))
# Feature Scaling: Normalize dataset via Generation of Embeddings
print("\nFeature Scaling: Embeddings Generation")
embed_dim = 100
embeds_model = ComplEx(k=embed_dim, verbose=True)
tf.compat.v1.logging.set_verbosity(
tf.compat.v1.logging.ERROR
) # TensorFlow will tell you all messages that have the label ERROR
embeds_model.fit(positives_filter)
embeds_source = embeds_model.get_embeddings(positives_filter[:, 0],
embedding_type='entity')
embeds_dest = embeds_model.get_embeddings(positives_filter[:, 2],
embedding_type='entity')
embeds = np.concatenate((embeds_source, embeds_dest), axis=1)
train_sz = train_X_temp.shape[0]
train_X, test_X = embeds[:train_sz, :], embeds[train_sz:, :]
train_X = train_X.reshape(
train_X.shape[0], 4, embed_dim
def select_kge(kge_name, batch_size, epochs, seed, verbose):
model = ''
# Select kge_name
if kge_name == 'complex':
# ComplEx model
model = ComplEx(
batches_count=batch_size,
epochs=epochs,
k=150,
eta=20,
optimizer='adam',
optimizer_params={'margin':
5}, #,'lr':learning_rate}, # default lr:0.1
loss='multiclass_nll',
loss_params={},
regularizer='LP',
regularizer_params={
'p': 2,
'lambda': 1e-4
},
seed=seed,
verbose=verbose)
elif kge_name == 'hole':
# HolE model
model = HolE(batches_count=batch_size,
epochs=epochs,
k=100,
eta=20,
optimizer='adam',
optimizer_params={'lr': learning_rate},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={
'p': 3,
'lambda': 1e-5
},
seed=seed,
verbose=verbose)
elif kge_name == 'transe':
# TransE model
model = TransE(
batches_count=batch_size,
epochs=epochs,
k=350,
eta=20,
optimizer='adam',
optimizer_params={'margin':
5}, #,'lr':learning_rate}, # default lr:0.1
loss='multiclass_nll', #loss='pairwise',
loss_params={}, #loss_params={'margin:5'},
regularizer='LP',
regularizer_params={
'p': 2,
'lambda': 1e-4
},
seed=seed,
verbose=verbose)
else:
sys.exit('Given kge_name is not valid.')
return model
# Client and the SPARQL Endpoint
endpoint = 'http://10.161.202.101:8890/sparql/'
port = 8890
output_format = HttpClientDataFormat.PANDAS_DF
client = HttpClient(endpoint_url=endpoint, port=port, return_format=output_format, timeout=timeout,
default_graph_uri=default_graph_url, max_rows=max_rows)
# Get all triples where the object is a URI
dataset = graph.feature_domain_range(s, p, o).filter({o: ['isURI']})
# execute
df = dataset.execute(client, return_format=output_format)
# Train/test split and create ComplEx model from ampligraph library
triples = df.to_numpy()
X_train, X_test = train_test_split_no_unseen(triples, test_size=10000)
# use ComplEx model to build the embedding
model = ComplEx(batches_count=50,epochs=300,k=100,eta=20, optimizer='adam',optimizer_params={'lr':1e-4},
loss='multiclass_nll',regularizer='LP', regularizer_params={'p':3, 'lambda':1e-5}, seed=0,verbose=True)
model.fit(X_train)
# Evaluate the embedding model
filter_triples = np.concatenate((X_train, X_test))
ranks = evaluate_performance(X_test, model=model, filter_triples=filter_triples,
use_default_protocol=True, verbose=True)
mr = mr_score(ranks)
mrr = mrr_score(ranks)
from ampligraph.datasets import load_wn18
from ampligraph.latent_features import ComplEx
from ampligraph.evaluation import evaluate_performance, hits_at_n_score, mrr_score
X = load_wn18()
model = ComplEx(batches_count=10,
seed=0,
epochs=20,
k=50,
eta=2,
loss="nll",
optimizer="adam",
optimizer_params={"lr": 0.01})
model.fit(X['train'])
y_pred = model.predict(X['test'][:5, ])
from scipy.special import expit
print(expit(y_pred))
ranks = evaluate_performance(X['test'][:10], model=model)
print(ranks)
mrr = mrr_score(ranks)
hits_10 = hits_at_n_score(ranks, n=10)
print("MRR: %f, Hits@10: %f" % (mrr, hits_10))
import matplotlib.pyplot as plt
"""---
# 3. Training ComplEx model
"""
import tensorflow
print(tensorflow.__version__)
from ampligraph.latent_features import ComplEx
model = ComplEx(batches_count=100,
seed=0,
epochs=200,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr':1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={'p':3, 'lambda':1e-5},
verbose=True)
positives_filter = X
import tensorflow as tf
tf.logging.set_verbosity(tf.logging.ERROR)
model.fit(data['train'], early_stopping = False)
"""---
# 4. Saving and restoring a model
])
X['valid'] = np.array([
list((tok2id[r[0]], r[1], tok2id[r[2]])) for r in X['valid']
if r[0] in known_entities and r[2] in known_entities
])
X['test'] = np.array([
list((tok2id[r[0]], r[1], tok2id[r[2]])) for r in X['test']
if r[0] in known_entities and r[2] in known_entities
])
X_train, X_valid = X['train'], X['valid']
print('Train set size: ', X_train.shape)
print('Test set size: ', X_valid.shape)
ke_kwargs = {"verbose": True, "k": 70, "epochs": 100}
# ComplEx brings double dimensions because of the twofold nature of complex numbers
model = ComplEx(**ke_kwargs)
print("Training...")
model.fit(X_train)
save_model(model, model_name_path=ke_model_path)
# If we don't transpose the multidimensionality of the embeddings to 3D but take just 3-D-embeddings,
# This can't be with ComplEX because, it will be an even number and 3 is not
ke_kwargs['k'] = 3
model2 = TransE(**ke_kwargs)
model2.fit(X_train)
save_model(model2, model_name_path=ke_model_path + '2')
else:
model = restore_model(model_name_path=ke_model_path)
model2 = restore_model(model_name_path=ke_model_path + '2')
with open(ke_wnkeys_path, 'rb') as handle:
tok2id, id2tok = pickle.load(handle)
print("Shape of test_X: %s; Shape of test_y: %s" % (test_X.shape, test_y.shape))
print("Shape of positives_filter: " + str(positives_filter.shape))
for j in range(len(mdl)):
# Fit & Train model via ampliGraph library
log_key = mdl[j]+": "+graph_data[i]
log_file = open("eval_log.txt", "a")
print("\n\n----"+log_key+"----", file=log_file)
print("------------------------------------------------")
print("%d) Implementation Model: %s" % (1, mdl[j]))
print("------------------------------------------------")
start_time = time.time() # START: Training Time Tracker
K.clear_session() # Kills current TF comp-graph & creates a new one
if (mdl[j] == "ComplEx"):
model = ComplEx(verbose=True)
elif (mdl[j] == "ConvKB"):
model = ConvKB(verbose=True)
elif (mdl[j] == "DistMult"):
model = DistMult(verbose=True)
elif (mdl[j] == "HolE"):
model = HolE(verbose=True)
elif (mdl[j] == "TransE"):
model = TransE(verbose=True)
elif (mdl[j] == "RandomBaseline"):
model = RandomBaseline(verbose=True)
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR) # TensorFlow will tell you all messages that have the label ERROR
model.fit(train_X)
# Save model at its best-performance point
save_model(model, 'best_ampliGraph_model.pkl')
num_epochs = 300
X = load_from_rdf('.', graph_file, rdf_format='nt', data_home='.')
test_size = int(0.05 * X.shape[0])
X_train, X_test = train_test_split_no_unseen(X, test_size=test_size)
print('Train set size: ', X_train.shape)
print('Test set size: ', X_test.shape)
model = ComplEx(batches_count=100,
seed=0,
epochs=num_epochs,
k=embedding_dims,
eta=5,
optimizer='adam',
optimizer_params={'lr': 1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={
'p': 3,
'lambda': 1e-5
},
verbose=True)
positives_filter = X
tf.logging.set_verbosity(tf.logging.ERROR)
print("Model training started...")
model.fit(X_train, early_stopping=False)
print("Save the model...")
save_model(model, model_name_path=out_embeddings_file)
X_train_valid, X_test = train_test_split_no_unseen(x, test_size=3000)
X_train, X_valid = train_test_split_no_unseen(X_train_valid, test_size=3000)
# X_train, X_test = train_test_split_no_unseen(x, test_size=3000)
print('Train set size: ', X_train.shape)
print('Test set size: ', X_test.shape)
print('Validation set size: ', X_valid.shape)
from ampligraph.latent_features import ComplEx
model = ComplEx(batches_count=100,
seed=0,
epochs=300,
k=150,
eta=5,
optimizer='adam',
optimizer_params={'lr':1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={'p':3, 'lambda':1e-5},
verbose=True)
positives_filter = x
# tf.logging.set_verbosity(tf.logging.ERROR)
# # model.fit(X_train, ,early_stopping = True)
# model.fit(X_train, early_stopping = True,early_stopping_params = \
# {
# 'x_valid': X_valid, # validation set
# 'criteria':'hits1', # Uses hits10 criteria for early stopping
# 'burn_in': 100, # early stopping kicks in after 100 epochs
import os
from ampligraph.utils import save_model, restore_model
import tensorflow as tf
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
from ampligraph.evaluation import evaluate_performance
ke_model_path = "./models/football_ke.amplimodel"
if not os.path.isfile(ke_model_path):
model = ComplEx(batches_count=50,
epochs=1,
k=1,
eta=20,
optimizer='adam',
optimizer_params={'lr': 1e-3},
loss='multiclass_nll',
regularizer='LP',
regularizer_params={
'p': 3,
'lambda': 1e-5
},
seed=0,
verbose=True)
print("Training...")
model.fit(X_train)
save_model(model, model_name_path=ke_model_path)
filter_triples = np.concatenate((X_train, X_valid))
else:
model = restore_model(model_name_path=ke_model_path)
def test_query_topn():
X = np.array([['a', 'y', 'b'],
['b', 'y', 'a'],
['a', 'y', 'c'],
['c', 'y', 'a'],
['a', 'y', 'd'],
['c', 'x', 'd'],
['b', 'y', 'c'],
['f', 'y', 'e'],
['a', 'z', 'f'],
['c', 'z', 'f'],
['b', 'z', 'f'],
])
model = ComplEx(k=2, batches_count=2)
with pytest.raises(ValueError): # Model not fitted
query_topn(model, top_n=2)
model.fit(X)
with pytest.raises(ValueError):
query_topn(model, top_n=2)
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a')
with pytest.raises(ValueError):
query_topn(model, top_n=2, relation='y')
with pytest.raises(ValueError):
query_topn(model, top_n=2, tail='e')
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', relation='y', tail='e')
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='xx', relation='y')
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', relation='yakkety')
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', tail='sax')
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', relation='x', rels_to_consider=['y', 'z'])
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', tail='f', rels_to_consider=['y', 'z', 'error'])
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', tail='e', rels_to_consider='y')
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', relation='x', ents_to_consider=['zz', 'top'])
with pytest.raises(ValueError):
query_topn(model, top_n=2, head='a', tail='e', ents_to_consider=['a', 'b'])
subj, pred, obj, top_n = 'a', 'x', 'e', 3
Y, S = query_topn(model, top_n=top_n, head=subj, relation=pred)
assert len(Y) == len(S)
assert len(Y) == top_n
assert np.all(Y[:, 0] == subj)
assert np.all(Y[:, 1] == pred)
Y, S = query_topn(model, top_n=top_n, relation=pred, tail=obj)
assert np.all(Y[:, 1] == pred)
assert np.all(Y[:, 2] == obj)
ents_to_con = ['a', 'b', 'c', 'd']
Y, S = query_topn(model, top_n=top_n, relation=pred, tail=obj, ents_to_consider=ents_to_con)
assert np.all([x in ents_to_con for x in Y[:, 0]])
rels_to_con = ['y', 'x']
Y, S = query_topn(model, top_n=10, head=subj, tail=obj, rels_to_consider=rels_to_con)
assert np.all([x in rels_to_con for x in Y[:, 1]])
Y, S = query_topn(model, top_n=10, relation=pred, tail=obj)
assert all(S[i] >= S[i + 1] for i in range(len(S) - 1))
def main():
# load Wordnet18 dataset:
# X = load_wn18()
X = load_fb15k_237()
modify_flag = False
# Initialize a ComplEx neural embedding model with pairwise loss function:
# The model will be trained for 300 epochs.
model = ComplEx(
batches_count=10,
seed=0,
epochs=30,
k=150,
eta=10,
# Use adam optimizer with learning rate 1e-3
optimizer='adam',
optimizer_params={'lr': 1e-3},
# Use pairwise loss with margin 0.5
loss='pairwise',
loss_params={'margin': 0.5},
# Use L2 regularizer with regularizer weight 1e-5
regularizer='LP',
regularizer_params={
'p': 2,
'lambda': 1e-5
},
# Enable stdout messages (set to false if you don't want to display)
verbose=True) #, modify_flag = modify_flag)
if False:
# ground truth params (have not tried yet)
# k: 350; epochs: 4000; eta: 30; loss: self_adversarial; loss_params: alpha: 1; margin: 0.5; optimizer: adam; optimizer_params: lr: 0.0001; seed: 0; batches_count: 50
model = ComplEx(
batches_count=50,
seed=0,
epochs=4000,
k=350,
eta=30,
# Use adam optimizer with learning rate 1e-3
optimizer='adam',
optimizer_params={'lr': 1e-4},
# Use pairwise loss with margin 0.5
loss='self_adversarial',
loss_params={
'margin': 0.5,
'alpha': 1
},
# Use L2 regularizer with regularizer weight 1e-5
regularizer='LP',
regularizer_params={
'p': 2,
'lambda': 1e-5
},
# Enable stdout messages (set to false if you don't want to display)
verbose=True,
modify_flag=modify_flag)
# For evaluation, we can use a filter which would be used to filter out
# positives statements created by the corruption procedure.
# Here we define the filter set by concatenating all the positives
filter = np.concatenate((X['train'], X['valid'], X['test']))
# Fit the model on training and validation set
model.fit(X['train'],
early_stopping=True,
early_stopping_params= \
{
'x_valid': X['valid'], # validation set
'criteria': 'hits10', # Uses hits10 criteria for early stopping
'burn_in': 100, # early stopping kicks in after 100 epochs
'check_interval': 20, # validates every 20th epoch
'stop_interval': 5, # stops if 5 successive validation checks are bad.
'x_filter': filter, # Use filter for filtering out positives
'corruption_entities': 'all', # corrupt using all entities
'corrupt_side': 's+o' # corrupt subject and object (but not at once)
}
)
# Run the evaluation procedure on the test set (with filtering).
# To disable filtering: filter_triples=None
# Usually, we corrupt subject and object sides separately and compute ranks
# restore model
sys.exit()
# import the inspect_checkpoint library
sys.exit()
ranks = evaluate_performance(
X['test'],
model=model,
filter_triples=filter,
use_default_protocol=
True, # corrupt subj and obj separately while evaluating
verbose=True)
# compute and print metrics:
mrr = mrr_score(ranks)
hits_10 = hits_at_n_score(ranks, n=10)
print("MRR: %f, Hits@10: %f" % (mrr, hits_10))