wip

eval.py

@@ -19,7 +19,6 @@ model_name = 'sentence-transformers/paraphrase-albert-base-v2'
 #86% so far
 model_name = 'output/training_OnlineConstrativeLoss-2023-03-17_16-10-39'
 
-
 model_sbert = SentenceTransformer(model_name)
 
 dev_sentences1 = []

quick_evaluate.py

@@ -14,6 +14,7 @@ model_name = 'output/training_OnlineConstrativeLoss-2023-03-11_00-24-35'
 model_name = 'output/training_OnlineConstrativeLoss-2023-03-11_01-00-19'
 model_name='output/training_OnlineConstrativeLoss-2023-03-17_16-10-39'
 model_name='output/training_OnlineConstrativeLoss-2023-03-17_23-15-52'
+#model_name='output/training_OnlineConstrativeLoss-2023-03-14_00-40-03'
 model_sbert = SentenceTransformer(model_name)
 
 

train.py

@@ -24,8 +24,8 @@ logger = logging.getLogger(__name__)
 
 
 #As base model, we use DistilBERT-base that was pre-trained on NLI and STSb data
-model_name ='sentence-transformers/paraphrase-albert-base-v2'
 model_name = 'sentence-transformers/all-mpnet-base-v1'
+model_name ='sentence-transformers/paraphrase-albert-base-v2'
 model = SentenceTransformer(model_name)
 num_epochs = 12
 # Smaller is generally better more accurate results.
@@ -35,7 +35,7 @@ train_batch_size = 10
 distance_metric = losses.SiameseDistanceMetric.COSINE_DISTANCE
 
 #Negative pairs should have a distance of at least 0.5
-margin = 0.5
+margin = 0.4
 
 dataset_path = "data_set_training.csv"
 model_save_path = 'output/training_OnlineConstrativeLoss-'+datetime.now().strftime("%Y-%m-%d_%H-%M-%S")

view_all_evals.py

@@ -0,0 +1,2 @@
+# This will take a model and display the cosine similarity
+# for all the dev set.