update code to work with sentence-transformers instead of simcse

app.py

@@ -9,6 +9,7 @@ import huggingface_hub
 import json
 from simcse import SimCSE # use for gpt
 from evaluate_data import store_sample_data, get_metrics_trf
+from sentence_transformers import SentenceTransformer
 
 # store_sample_data()
 
@@ -108,8 +109,9 @@ def find_orgs(uploaded_file):
         true_orgs.append(sent['orgs'])
         
 
-    sim_model = SimCSE('sentence-transformers/all-MiniLM-L6-v2')
-    all_metrics['gpt'] = calc_metrics(true_orgs, gpt_orgs, sim_model)        
+    # sim_model = SimCSE('sentence-transformers/all-MiniLM-L6-v2')
+    sim_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
+    all_metrics['gpt'] = calc_metrics(true_orgs, gpt_orgs, sim_model, threshold=0.85)        
 
     return all_metrics
 # radio_btn = gr.Radio(choices=['GPT', 'iSemantics'], value='iSemantics', label='Available models', show_label=True)

metrics.py

@@ -1,3 +1,5 @@
+from sentence_transformers import util
+
 def calc_recall(true_pos, false_neg, eps=1e-8):
     return true_pos / (true_pos + false_neg + eps)
 
@@ -44,8 +46,12 @@ def calc_metrics(true, predicted, model, threshold=0.95, eps=1e-8):
                 false_neg_ids.append(j)
             
             continue
-            
-        similarities = model.similarity(true_ents, pred_ents, device='cuda')
+        
+        embed_true = model.encode(true_ents, convert_to_tensor=True)
+        embed_pred = model.encode(pred_ents, convert_to_tensor=True)
+
+        similarities = util.pytorch_cos_sim(embed_true, embed_pred)
+        # similarities = model.similarity(true_ents, pred_ents, device='cuda')
         
         for row in similarities:
             if (row >= threshold).any():

requirements.txt

@@ -5,4 +5,4 @@ datasets==2.18.0
 evaluate
 seqeval
 rich
-simcse
+sentence-transformers