Fix off by one bug (thanks @tom-andersson!)

Author: wesselb

neuralprocesses/mask.py

@@ -40,6 +40,15 @@ def _pad_zeros(x: B.Numeric, up_to: int, axis: int):
         return B.concat(x, B.zeros(B.dtype(x), *shape), axis=axis)
 
 
+def _ceil_to_closest_multiple(n, m):
+    d, r = divmod(n, m)
+    # If `n` is zero, then we must also round up.
+    if n == 0 or r > 0:
+        return (d + 1) * m
+    else:
+        return d * m
+
+
 @_dispatch
 def _determine_ns(xc: tuple, multiple: Union[int, tuple]):
     ns = [B.shape(xci, 2) for xci in xc]
@@ -48,7 +57,7 @@ def _determine_ns(xc: tuple, multiple: Union[int, tuple]):
         multiple = (multiple,) * len(ns)
 
     # Ceil to the closest multiple of `multiple`.
-    ns = [((n - 1) // m + 1) * m for n, m in zip(ns, multiple)]
+    ns = [_ceil_to_closest_multiple(n, m) for n, m in zip(ns, multiple)]
 
     return ns
 

tests/test_mask.py

@@ -2,7 +2,7 @@
 import pytest
 
 from .test_architectures import generate_data
-from .util import nps, approx  # noqa
+from .util import approx, nps  # noqa
 
 
 @pytest.mark.flaky(reruns=3)
@@ -31,3 +31,23 @@ def test_convgnp_mask(nps):
     # Check that the two ways of doing it coincide.
     approx(pred.mean, pred_masked.mean)
     approx(pred.var, pred_masked.var)
+
+
+@pytest.mark.parametrize("ns", [(10,), (0,), (10, 5), (10, 0), (0, 10), (15, 5, 10)])
+@pytest.mark.parametrize("multiple", [1, 2, 3, 5])
+def test_mask_contexts(nps, ns, multiple):
+    x, y = nps.merge_contexts(
+        *((B.randn(nps.dtype, 2, 3, n), B.randn(nps.dtype, 2, 4, n)) for n in ns),
+        multiple=multiple
+    )
+
+    # Test that the output is of the right shape.
+    if max(ns) == 0:
+        assert B.shape(y.y, 2) == multiple
+    else:
+        assert B.shape(y.y, 2) == ((max(ns) - 1) // multiple + 1) * multiple
+
+    # Test that the mask is right.
+    mask = y.mask == 1  # Convert mask to booleans.
+    assert B.all(B.take(B.flatten(y.y), B.flatten(mask)) != 0)
+    assert B.all(B.take(B.flatten(y.y), B.flatten(~mask)) == 0)