fix len calculation for new data loader

allennlp/data/data_loaders/multi_process_data_loader.py

@@ -203,7 +203,9 @@ def __init__(
             deque(self.iter_instances(), maxlen=0)
 
     def __len__(self) -> int:
-        if self.max_instances_in_memory is None:
+        if self.batches_per_epoch is not None:
+            return self.batches_per_epoch
+        elif self.max_instances_in_memory is None:
             # We haven't read the instances yet, so we do so now, caching them as we go.
             if not self._instances:
                 deque(self.iter_instances(), maxlen=0)
@@ -218,8 +220,6 @@ def __len__(self) -> int:
                 return num_instances // batch_size
             else:
                 return 1 + num_instances // batch_size
-        elif self.batches_per_epoch is not None:
-            return self.batches_per_epoch
         else:
             # We can't know the number of batches for a lazy loader when batches_per_epoch
             # is not specified.

tests/data/data_loaders/multi_process_data_loader_test.py

@@ -156,3 +156,14 @@ def test_drop_last():
     for batch in batches:
         assert len(batch["index"]) == 16
     assert len(batches) == 6
+
+
+def test_batches_per_epoch():
+    loader = MultiProcessDataLoader(
+        MockDatasetReader(), "some path", batch_size=4, batches_per_epoch=10
+    )
+    vocab = Vocabulary.from_instances(loader.iter_instances())
+    loader.index_with(vocab)
+
+    assert len(loader) == 10
+    assert len(list(loader)) == 10