Handle completely empty veclists

miplearn/features/sample.py

@@ -238,7 +238,8 @@ class Hdf5Sample(Sample):
             else:
                 data = np.array(padded)
             break
-        assert data is not None
+        if data is None:
+            data = np.array(padded)
         ds = self._put(key, data)
         ds.attrs["lengths"] = lens
 
@@ -254,7 +255,7 @@ def _pad(veclist: VectorList) -> Tuple[VectorList, List[int]]:
     maxlen = max(lens)
 
     # Find appropriate constant to pad the vectors
-    constant: Union[int, float, str, None] = None
+    constant: Union[int, float, str] = 0
     for v in veclist:
         if v is None or len(v) == 0:
             continue
@@ -266,7 +267,6 @@ def _pad(veclist: VectorList) -> Tuple[VectorList, List[int]]:
             constant = ""
         else:
             assert False, f"Unsupported data type: {v[0]}"
-    assert constant is not None, "veclist must not be completely empty"
 
     # Pad vectors
     for (i, vi) in enumerate(veclist):

tests/features/test_sample.py

@@ -34,6 +34,7 @@ def _test_sample(sample: Sample) -> None:
     _assert_roundtrip_vector_list(sample, [[True], [False, False], None])
     _assert_roundtrip_vector_list(sample, [[1], None, [2, 2], [3, 3, 3]])
     _assert_roundtrip_vector_list(sample, [[1.0], None, [2.0, 2.0], [3.0, 3.0, 3.0]])
+    _assert_roundtrip_vector_list(sample, [None, None])
 
     # Bytes
     _assert_roundtrip_bytes(sample, b"\x00\x01\x02\x03\x04\x05")
@@ -68,6 +69,7 @@ def _assert_roundtrip_vector_list(sample: Sample, expected: Any) -> None:
     actual = sample.get_vector_list("key")
     assert actual == expected
     assert actual is not None
+    if actual[0] is not None:
         _assert_same_type(actual[0][0], expected[0][0])