Implement clip(..., out=)

src/mygrad/math/misc/funcs.py

@@ -439,7 +439,12 @@ def minimum(
 
 @implements_numpy_override()
 def clip(
-    a: ArrayLike, a_min: ArrayLike, a_max: ArrayLike, *, constant: Optional[bool] = None
+    a: ArrayLike,
+    a_min: ArrayLike,
+    a_max: ArrayLike,
+    out: Optional[Union[np.ndarray, Tensor]] = None,
+    *,
+    constant: Optional[bool] = None,
 ) -> Tensor:
     """Clip (limit) the values in an array.
 
@@ -470,6 +475,11 @@ def clip(
         `None`. If `a_min` or `a_max` are ArrayLike, then the three
         arrays will be broadcasted to match their shapes.
 
+    out : Optional[Union[ndarray, Tensor]]
+        A location into which the result is stored. If provided, it must have
+        a shape that the inputs broadcast to. If not provided or None, a
+        freshly-allocated tensor is returned.
+
     constant : bool, optional(default=False)
         If ``True``, the returned tensor is a constant (it
         does not backpropagate a gradient)
@@ -495,11 +505,10 @@ def clip(
         raise ValueError("`a_min` and `a_max` cannot both be set to `None`")
 
     if a_min is not None:
-        a = maximum(a_min, a, constant=constant)
+        a = maximum(a_min, a, out=out, constant=constant)
 
     if a_max is not None:
-        a = minimum(a_max, a, constant=constant)
-
+        a = minimum(a_max, a, out=out, constant=constant)
     return a
 
 

src/mygrad/tensor_base.py

@@ -3130,7 +3130,12 @@ def any(
         return np.any(self.data, axis=axis, out=out, keepdims=keepdims)
 
     def clip(
-        self, a_min: ArrayLike, a_max: ArrayLike, *, constant: Optional[bool] = None
+        self,
+        a_min: ArrayLike,
+        a_max: ArrayLike,
+        out: Optional[Union[np.ndarray, "Tensor"]] = None,
+        *,
+        constant: Optional[bool] = None,
     ) -> "Tensor":  # pragma: no cover
         """Clip (limit) the values in an array.
 
@@ -3158,6 +3163,11 @@ def clip(
             `None`. If `a_min` or `a_max` are ArrayLike, then the three
             arrays will be broadcasted to match their shapes.
 
+        out : Optional[Union[ndarray, Tensor]]
+            A location into which the result is stored. If provided, it must have
+            a shape that the inputs broadcast to. If not provided or None, a
+            freshly-allocated tensor is returned.
+
         constant : bool, optional(default=False)
             If ``True``, the returned tensor is a constant (it
             does not backpropagate a gradient)

tests/math/test_misc.py

@@ -106,13 +106,6 @@ def amax_clip_only(clip_func, a, b, constant=False):
     )
 
 
-skip_if_lower_than_numpy_1p17 = pytest.mark.skipif(
-    np.__version__ < "1.17",
-    reason="numpy.clip behavior was made consistent in numpy-1.17; "
-    "test must by run on numpy 1.17 or later",
-)
-
-
 @pytest.mark.parametrize(
     ("mygrad_clip", "numpy_clip", "num_arrays"),
     [
@@ -129,7 +122,6 @@ def amax_clip_only(clip_func, a, b, constant=False):
         (clip, np.clip, 3),
     ],
 )
-@skip_if_lower_than_numpy_1p17
 def test_clip_fwd(mygrad_clip: Callable, numpy_clip: Callable, num_arrays: int):
     @fwdprop_test_factory(
         num_arrays=num_arrays, mygrad_func=mygrad_clip, true_func=numpy_clip
@@ -162,7 +154,6 @@ def is_not_close_clip(a: Tensor, a_min=None, a_max=None) -> bool:
         (clip, np.clip, 3),
     ],
 )
-@skip_if_lower_than_numpy_1p17
 def test_clip_bkwd(mygrad_clip: Callable, numpy_clip: Callable, num_arrays: int):
     @backprop_test_factory(
         num_arrays=num_arrays,
@@ -193,7 +184,6 @@ def wrapped_test():
         dtype=float,
     ),
 )
-@skip_if_lower_than_numpy_1p17
 @pytest.mark.filterwarnings("ignore: invalid value")
 def test_clip_input_validation(a, a_min, a_max):
     try:
@@ -218,3 +208,18 @@ def test_clip_method_bkwd():
     x = mg.tensor([1.0, 5.0, 10.0])
     x.clip(2, 7).backward()
     assert_allclose(x.grad, [0.0, 1.0, 0.0])
+
+
+@pytest.mark.parametrize("as_method", [False, True])
+def test_inplace_clip(as_method: bool):
+    x = mg.arange(4.0)
+    y = mg.tensor([-1.0, 2.0, 3.0])
+
+    clipper = y.clip if as_method else partial(mg.clip, y)
+    out = clipper(0, 2.1, out=x[1:])
+    out.backward()
+
+    assert_allclose(out, mg.tensor([0.0, 2.0, 2.1]))
+    assert_allclose(x, mg.tensor([0.0, 0.0, 2.0, 2.1]))
+    assert_allclose(x.grad, np.array([0.0, 1.0, 1.0, 1.0]))
+    assert_allclose(y.grad, np.array([0.0, 1.0, 0.0]))