Fix lru cache multiprocessing memory leak issue

Author: basnetsoyuj

src/deep_impact/evaluation/ranker.py

@@ -1,4 +1,5 @@
 from itertools import product
+from multiprocessing import Pool
 from pathlib import Path
 from typing import Union
 
@@ -9,6 +10,11 @@
 from src.utils.datasets import QueryRelevanceDataset, Queries, RunFile
 
 
+def rank(args):
+    index, qid, query_terms = args
+    return qid, index.score(query_terms=query_terms)
+
+
 class Ranker:
     def __init__(
             self,
@@ -29,9 +35,11 @@ def __init__(
         self.model_cls = DeepPairwiseImpact if pairwise else DeepImpact
 
     def run(self):
-        for qid in tqdm(self.qrels.keys()):
-            scores = self.rank(qid=qid)
-            self.run_file.writelines(qid, scores)
+        with Pool(self.num_workers) as p, tqdm(total=len(self.qrels)) as pbar:
+            for qid, scores in p.imap_unordered(rank, [(self.index, qid, self.get_query_terms(qid)) for qid in
+                                                       self.qrels.keys()]):
+                self.run_file.writelines(qid, scores)
+                pbar.update(1)
 
     def get_query_terms(self, qid):
         query_terms = self.model_cls.process_query(query=self.queries[qid])
@@ -42,6 +50,3 @@ def get_query_terms(self, qid):
                     query_terms.add(f'{term1}|{term2}')
 
         return query_terms
-
-    def rank(self, qid):
-        return self.index.score(query_terms=self.get_query_terms(qid=qid))

src/deep_impact/inverted_index/inverted_index.py

@@ -37,7 +37,7 @@ def term_location(self, term):
             start, end = struct.unpack(LOC_BLOCK_FORMAT, bf.read(LOC_BLOCK_BYTES))
         return term_id, start, end
 
-    @lru_cache(maxsize=10000)
+    # @lru_cache(maxsize=10000)
     def term_docs(self, term):
         term_id, start, end = self.term_location(term)
         if term_id is None: