implement more numpy functions #104
Conversation
|
|
||
| @pytest.mark.xfail(reason="how to find if someone is refencing an array") |
There was a problem hiding this comment.
This we can just do from C++. I don't think we can do this from Python.
| @normalizer | ||
| def resize(a: ArrayLike, new_shape=None): | ||
|
|
||
| if new_shape is None: | ||
| return a | ||
|
|
||
| if a.numel() == 0: | ||
| a = torch.zeros(1, dtype=a.dtype) | ||
|
|
||
| numel = math.prod(new_shape) | ||
| quot, rem = divmod(numel, a.numel()) | ||
| repeats = quot * a.numel() | ||
|
|
||
| result = torch.empty(numel, dtype=a.dtype) | ||
| result[:repeats] = torch.tile(a.ravel(), (quot,)) | ||
| if rem > 0: | ||
| result[-rem:] = a.ravel()[:rem] | ||
| return result.reshape(new_shape) |
There was a problem hiding this comment.
This one is implemented in Python here. Why not vendor its implementation?
| @@ -4549,6 +4550,7 @@ def test_none_shape(self): | |||
| x.resize() | |||
| assert_array_equal(x, np.eye(3)) | |||
|
|
|||
| @pytest.mark.xfail(reason="why would one do it") | |||
There was a problem hiding this comment.
nb. These and other corner-cases naturally appear when writing generic code, where you have slices of tensors and then you perform operations on them.
At any rate, these should be easy to fix by simply using the NumPy implementation of this function.
There was a problem hiding this comment.
Turns out ndarray.resize is documented to differ from np.resize, so many of these corner cases are in fact not corner cases but documented differences:
If the new array is larger than the original array, then the new array is filled with repeated copies of a. Note that this behavior is different from a.resize(new_shape) which fills with zeros instead of repeated copies of a.
https://numpy.org/doc/stable/reference/generated/numpy.resize.html
One other annoyance is the support of both resize(tuple) and resize(*tuple). Anyhow, this is all done in gh-106.
work through the lists in gh-87.