Minor updates to Baleen

Author: okhat

baleen/condenser/condense.py

@@ -22,9 +22,9 @@ def __init__(self, collectionX_path, checkpointL1, checkpointL2, deviceL1='cuda'
 
     def condense(self, query, backs, ranking):
         stage1_preds = self._stage1(query, backs, ranking)
-        stage2_preds = self._stage2(query, stage1_preds)
+        stage2_preds, stage2_preds_L3x = self._stage2(query, stage1_preds)
 
-        return stage1_preds, stage2_preds
+        return stage1_preds, stage2_preds, stage2_preds_L3x
 
     def _load_model(self, path, device):
         model = torch.load(path, map_location='cpu')
@@ -128,8 +128,14 @@ def _stage2(self, query, preds):
 
             preds = [(score, (pid, sid)) for (pid, sid), score in zip(preds, scores)]
             preds = sorted(preds, reverse=True)[:5]
+
+            preds_L3x = [x for score, x in preds if score > min(0, preds[1][0] - 1e-10)] # Take at least 2!
             preds = [x for score, x in preds if score > 0]
 
-        # TODO: Apply L3x for final stage.
-        
-        return preds
+            earliest_pids = f7([pid for pid, _ in preds_L3x])[:4]  # Take at most 4 docs.
+            preds_L3x = [(pid, sid) for pid, sid in preds_L3x if pid in earliest_pids]
+
+            assert len(preds_L3x) >= 2
+            assert len(f7([pid for pid, _ in preds_L3x])) <= 4
+
+        return preds, preds_L3x

baleen/engine.py

@@ -36,12 +36,12 @@ def search(self, query, num_hops, depth=100, verbose=False):
                 if len(pids_bag) < k * (hop_idx+1):
                     pids_bag.add(pid)
 
-            stage1_preds, facts = condenser.condense(query, backs=facts, ranking=ranking_)
+            stage1_preds, facts, stage2_L3x = condenser.condense(query, backs=facts, ranking=ranking_)
             context = ' [SEP] '.join([collectionX.get((pid, sid), '') for pid, sid in facts])
 
         assert len(pids_bag) == depth
 
-        return facts, pids_bag, stage1_preds
+        return stage2_L3x, pids_bag, stage1_preds
 
 
 

colbert/indexing/collection_indexer.py

@@ -299,26 +299,11 @@ def _collect_embedding_id_offset(self):
         assert len(self.embedding_offsets) == self.num_chunks
 
     def _build_ivf(self):
-        # TODO: If this is slow or memory intensive, it can be done as a torch.sort, torch.add offset, torch.unique
-        # operations over the concatenated codes.
-        # On MS MARCO, this seems to take 10 minutes! I can imagine that's 40 minutes on Wikipedia.
-
-
-        """
-        codes = ResidualCodec.Embeddings.load_all_codes(index_path)
-
-        codes.sort().{values, indices}
-
-        values.unique_consecutive -> counts -> cumsum -> offsets
-        (indices.int32(), offsets)
-
-
         # Maybe we should several small IVFs? Every 250M embeddings, so that's every 1 GB.
         # It would save *memory* here and *disk space* regarding the int64.
         # But we'd have to decide how many IVFs to use during retrieval: many (loop) or one?
         # A loop seems nice if we can find a size that's large enough for speed yet small enough to fit on GPU!
         # Then it would help nicely for batching later: 1GB.
-        """
 
         codes = torch.empty(self.num_embeddings,)
         print_memory_stats(f'RANK:{self.rank}')

colbert/utilities/annotate_em.py

@@ -106,10 +106,8 @@ def save(self, new_path):
 
 
 if __name__ == '__main__':
-    r = '/future/u/okhattab/root/unit/experiments/2021.08/retrieve.py/2021-09-04_15.50.02/ranking.tsv'
-    r = '/future/u/okhattab/root/unit/experiments/2021.08/retrieve.py/2021-09-04_15.59.37/ranking.tsv'
-    r = sys.argv[1]
+    r = sys.argv[2]
 
-    a = AnnotateEM(collection='/future/u/okhattab/root/unit/data/NQ-mini/collection.tsv',
-                   qas='/future/u/okhattab/root/unit/data/NQ-mini/dev/qas.json')
+    a = AnnotateEM(collection='/dfs/scratch0/okhattab/OpenQA/collection.tsv',
+                   qas=sys.argv[1])
     a.annotate(ranking=r)

colbert/utils/amp.py

@@ -3,13 +3,9 @@
 from contextlib import contextmanager
 from colbert.utils.utils import NullContextManager
 
-PyTorch_over_1_6 = True # float('.'.join(torch.__version__.split('.')[0:2])) >= 1.6
-
 
 class MixedPrecisionManager():
     def __init__(self, activated):
-        assert (not activated) or PyTorch_over_1_6, "Cannot use AMP for PyTorch version < 1.6"
-
         self.activated = activated
 
         if self.activated:
@@ -22,7 +18,6 @@ def backward(self, loss):
         if self.activated:
             self.scaler.scale(loss).backward()
         else:
-            assert False, "for now"
             loss.backward()
 
     def step(self, colbert, optimizer, scheduler=None):