add compression rates

modeling_provence.py

@@ -158,6 +158,9 @@ class Provence(DebertaV2PreTrainedModel):
         reranking_scores = [
             [None for j in range(len(contexts[i]))] for i in range(len(queries))
         ]
+        compressions = [
+            [0 for j in range(len(contexts[i]))] for i in range(len(queries))
+        ]
         with torch.no_grad():
             for batch_start in tqdm(
                 range(0, len(dataset), batch_size), desc="Pruning contexts..."
@@ -225,18 +228,24 @@ class Provence(DebertaV2PreTrainedModel):
                     )
                 else:
                     selected_contexts[i][j] = selected_contexts[i][j][0]
+                len_original = len(contexts[i][j])
+                len_compressed = len(selected_contexts[i][j])
+                compressions[i][j] = (len_original-len_compressed)/len_original * 100
             if reorder:
                 idxs = np.argsort(reranking_scores[i])[::-1][:top_k]
                 selected_contexts[i] = [selected_contexts[i][j] for j in idxs]
                 reranking_scores[i] = [reranking_scores[i][j] for j in idxs]
+                compressions[i] = [compressions[i][j] for j in idxs]
 
         if type(context) == str:
             selected_contexts = selected_contexts[0][0]
             reranking_scores = reranking_scores[0][0]
+            compressions = compressions[0][0]
                 
         return {
             "pruned_context": selected_contexts,
-            "reranking_score": reranking_scores
+            "reranking_score": reranking_scores,
+            "compression_rate": compressions,
         }