python - Numpy as_strided() 分段错误

Question

我正在尝试“修补”步幅小于补丁大小的图像（即它们将重叠）。我这样做的方法是读取数组并使用as_strided()，然后将值复制到新数组。它适用于 uint8 格式的图像，但是，在 float32 上使用时会引发分段错误。MWE：

import numpy as np

def patch_sample(arr, patch_size=(128, 128), stride=64):
    arr = np.ascontiguousarray(arr)
    H, W, C = arr.shape
    h, w = patch_size
    patches_h = int((H - h) / stride + 1)
    patches_w = int((W - w)/ stride + 1)
    new_shape = (patches_h, patches_w, h, w, C)
    new_strides = arr.itemsize * np.array(
        [
            W * arr.strides[1] * stride,
            arr.strides[1] * stride,
            W * arr.strides[1],
            arr.strides[1],
            1
        ]
    )
    patches_out = np.lib.stride_tricks.as_strided(
        arr,
        shape=new_shape,
        strides=new_strides,
        writeable=False
    )

    # Here segfault occurs
    patches_out_cp = np.ascontiguousarray(patches_out)

    # Segfault as well:
    # patches_out_cp = patches_out.copy()

    # Same for:
    # patches_out_cp = np.empty_like(patches_out)
    # np.copyto(patches_out_cp, patches_out)

    patches_out_cp = patches_out_cp.reshape((-1, h, w, C))

    return patches_out_cp

# No segfault
test_array = np.random.randint(0, 100, (1000, 1500, 3), dtype=np.uint8)
print(patch_sample(test_array).shape)

# Segfault
test_array = np.random.rand(1000, 1500, 3).astype(np.float32)
print(patch_sample(test_array).shape)

不知道这可能是什么原因。任何帮助都会很棒！

标签： pythonarraysnumpy