Separate dim32 and dim512 embedding results into different folders and update paths

Author: Morten Terhart

complete_triple_patterns.py

@@ -13,8 +13,8 @@ def main():
     wikidata5m_valid = load_wikidata5m_dataset('valid')
     wikidata5m_test = load_wikidata5m_dataset('test')
 
-    model = torch.load('embeddings/ComplEx/trained_model.pkl')
-    model_factory = TriplesFactory.from_path_binary('embeddings/ComplEx/training_triples')
+    model = torch.load('embeddings/dim_32/complex/trained_model.pkl')
+    model_factory = TriplesFactory.from_path_binary('embeddings/dim_32/complex/training_triples')
     print('Training set stats:')
     print(f'  Num Triples:   {model_factory.num_triples}')
     print(f'  Num Entities:  {model_factory.num_entities}')
@@ -32,11 +32,11 @@ def main():
     print(preds_df[preds_df['in_testing'] == True])
 
     predicate_label = wikidata5m_test['P'].iloc[10]
-    arithmetic_mean_rank = get_predicate_metric(predicate_metrics, 'arithmetic_mean_rank', predicate_label, 'ComplEx', 'tail', 'realistic')
-    hits_at_1 = get_predicate_metric(predicate_metrics, 'hits_at_1', predicate_label, 'ComplEx', 'tail', 'realistic')
-    hits_at_3 = get_predicate_metric(predicate_metrics, 'hits_at_3', predicate_label, 'ComplEx', 'tail', 'realistic')
-    hits_at_5 = get_predicate_metric(predicate_metrics, 'hits_at_5', predicate_label, 'ComplEx', 'tail', 'realistic')
-    hits_at_10 = get_predicate_metric(predicate_metrics, 'hits_at_10', predicate_label, 'ComplEx', 'tail', 'realistic')
+    arithmetic_mean_rank = get_predicate_metric(predicate_metrics, 'arithmetic_mean_rank', predicate_label, 'complex', 'tail', 'realistic')
+    hits_at_1 = get_predicate_metric(predicate_metrics, 'hits_at_1', predicate_label, 'complex', 'tail', 'realistic')
+    hits_at_3 = get_predicate_metric(predicate_metrics, 'hits_at_3', predicate_label, 'complex', 'tail', 'realistic')
+    hits_at_5 = get_predicate_metric(predicate_metrics, 'hits_at_5', predicate_label, 'complex', 'tail', 'realistic')
+    hits_at_10 = get_predicate_metric(predicate_metrics, 'hits_at_10', predicate_label, 'complex', 'tail', 'realistic')
     print(f'Arithmetic mean rank: {arithmetic_mean_rank}')
     print(f'Hits at 1:            {hits_at_1}')
     print(f'Hits at 3:            {hits_at_3}')
@@ -45,7 +45,7 @@ def main():
 
 
 def load_wikidata5m_dataset(subset_type: SubsetType):
-    return pd.read_csv(f'dataset/knowledge_graph/wikidata5m_transductive_{subset_type}.txt', sep='\t',
+    return pd.read_csv(f'dataset/wikidata5m/wikidata5m_transductive_{subset_type}.txt', sep='\t',
                        names=['S', 'P', 'O'])
 
 

compute_model_predictions.py

@@ -15,13 +15,13 @@ def main():
     torch.cuda.empty_cache()
 
     print(f'[X] Loading {model_name} model')
-    train_factory = TriplesFactory.from_path_binary(f'embeddings/{model_name}/training_factory')
+    train_factory = TriplesFactory.from_path_binary(f'embeddings/dim_512/{model_name}/training_factory')
 
     model = TransE(
         triples_factory=train_factory,
         embedding_dim=512
     )
-    model.load_state_dict(torch.load(f'embeddings/{model_name}/trained_model_state_dict.pt'))
+    model.load_state_dict(torch.load(f'embeddings/dim_512/{model_name}/trained_model_state_dict.pt'))
     model.to(device).eval()
 
     print(f'[X] Loading Wikidata5M datasets')
@@ -39,7 +39,7 @@ def main():
     scores_df = pack.process(factory=train_factory).df
 
     print(f'[X] Saving predicted scores')
-    scores_df.to_csv(f'embeddings/{model_name}/predicted_scores.csv')
+    scores_df.to_csv(f'embeddings/dim_512/{model_name}/predicted_scores.csv', index=False)
 
 
 if __name__ == '__main__':

compute_predicate_metrics.py

@@ -12,7 +12,7 @@
 
 
 def main():
-    wikidata5m_test_set = pd.read_csv('dataset/knowledge_graph/wikidata5m_transductive_test.txt', sep="\t",
+    wikidata5m_test_set = pd.read_csv('dataset/wikidata5m/wikidata5m_transductive_test.txt', sep="\t",
                                       names=['S', 'P', 'O'], header=None)
     trained_models = get_trained_models()
     print(
@@ -32,7 +32,7 @@ def main():
 
     print(f'[X] Finished evaluation in {timedelta(seconds=timer() - start)}')
 
-    predicate_metrics.to_csv('metrics/predicate_metrics.csv')
+    predicate_metrics.to_csv('metrics/predicate_metrics.csv', index=False)
 
 
 def get_number_of_predicates(dataset_df):
@@ -46,10 +46,10 @@ def get_test_set_per_predicate(test_set_file):
 
 def get_trained_models():
     return {
-        'ComplEx': _load_trained_model('embeddings/ComplEx'),
-        'DistMult': _load_trained_model('embeddings/DistMult'),
-        'SimplE': _load_trained_model('embeddings/SimplE'),
-        'TransE': _load_trained_model('embeddings/TransE')
+        'complex': _load_trained_model('embeddings/dim_32/complex'),
+        'distmult': _load_trained_model('embeddings/dim_32/distmult'),
+        'simple': _load_trained_model('embeddings/dim_32/simple'),
+        'transe': _load_trained_model('embeddings/dim_32/transe')
     }
 
 

dataset/convert_csv_to_turtle.py

@@ -4,12 +4,12 @@
 
 
 def main():
-    df = pd.read_csv('./knowledge_graph/wikidata5m_transductive_train.txt', sep='\t', names=['S', 'P', 'O'])
+    df = pd.read_csv('./wikidata5m/wikidata5m_transductive_train.txt', sep='\t', names=['S', 'P', 'O'])
 
     # Transform triples to Turtle format in the dataframe
     turtle_df = df.apply(row_to_turtle, axis=1)
 
-    turtle_file = './knowledge_graph/wikidata5m_transductive_train.ttl'
+    turtle_file = './wikidata5m/wikidata5m_transductive_train.ttl'
     with open(turtle_file, 'w') as f:
         f.write(f'@prefix wd: <{wikidata_prefix}>\n\n')