错误 At least one stride in the given numpy array is negative, and tensors with negative strides are not currently supported. 解决方案

在训练 Pytorch 网络时遇到错误 At least one stride in the given numpy array is negative, and tensors with negative strides are not currently supported., 本文记录原因与解决方案。

问题复现

• 在numpy 图像数据转为 torch.tensor 之前使用 numpy 执行内部形变的操作，常见的有：

  • 通道转换

    image = image[:,:,::-1]

  • 图像翻转

    image = np.fliplr(image)
    # or
    image = np.flipud(image)

• 此类操作之后经过 Pytorch 的 Dataloader，读取tensor 时会报错

ValueError: At least one stride in the given numpy array is negative, and tensors with negative strides are not currently supported. (You can probably work around this by making a copy of your array with array.copy().)

问题原因

• 输入网络的 Tensor 需要是内存连续的
• 但是 numpy 上述变换后为了速度考虑不会改变数据内存，这就导致拿到的数据在内存中不连续，导致错误

解决方案

• 可以按照报错中建议的方式

image = image.copy()

• 也可以正经解决 numpy 内存连续的问题

image = np.ascontiguousarray(image)

参考资料

• https://www.zywvvd.com/notes/study/deep-learning/numpy-tensor-contiguous/numpy-tensor-contiguous/

文章链接：
https://www.zywvvd.com/notes/study/deep-learning/bug-fix/numpy-negative/numpy-negative/