Spurious "NumPy array is not writeable" warning on torch.tensor(np_array)

[jpetkau]

🐛 Bug

torch.tensor(a), for non-writeable numpy array a, produces this warning:

UserWarning: The given NumPy array is not writeable, and PyTorch does not support non-writeable tensors. This means you can write to the underlying (supposedly non-writeable) NumPy array using the tensor. You may want to copy the array to protect its data or make it writeable before converting it to a tensor.

But since torch.tensor(a) always makes a copy, the warning makes no sense. Avoiding the warning requires making a copy of the numpy array before converting to torch.

A partial workaround is to do torch.from_numpy(a.copy()).

To Reproduce

Steps to reproduce the behavior:

import torch, numpy
a=numpy.arange(5.)
a.flags.writeable=False
t=torch.tensor(a)

Expected behavior

The warning should only be shown when actually aliasing a read-only buffer with a mutable Tensor, not when copying from a read-only buffer.

Environment

Verified on Windows, pytorch 1.5.1, git_version = '3c31d73c875d9a4a6ea8a843b9a0d1b19fbe36f3' (pip version)
And on linux at pytorch 1.8.0a0fb

Additional context

Warning was introduced in pull 33615. #44027 (read-only tensors) would make this irrelevant.

This warning is important for from_numpy, because you can easily accidentally do things like create a mutable view onto a Python str or bytes object and overwrite it. Having the warning also give trivial false alarms like this can cause real problems to be overlooked.

Probably related: torch.tensor(nparray, pin_memory=True) doesn't work, probably for similar reasons: it tries to construct a temporary view before making the copy. If you want the torch tensor in pinned memory, there appears to be no way to avoid the warning without introducing a spurious copy.

cc @mruberry @rgommers @heitorschueroff

[ngimel]

Thank you for your bug report, warning is indeed spurious in this case.

[ArtistBanda]

Hello! I am new to open source development and would like to help to resolve this issue. It would help a lot with a bit of guidance here.

[mruberry]

You probably want to suppress the warning from here:

pytorch/torch/csrc/utils/tensor_new.cpp

Line 273 in 8d5256e

auto tensor = tensor_from_numpy(data);

when copy_numpy is true. Then add tests for the callsites to verify the behavior is correct.

[ArtistBanda]

Sure, I'll try to do that.

[desaixie]

Can I ignore this warning if it is coming from torch.from_numpy(batch).to('cuda'), where batch is a slice of a numpy ndarray and therefore not writeable? I guess it shouldn't matter since to('cuda') is making a copy, but I want to make sure.

[mruberry]

Can I ignore this warning if it is coming from torch.from_numpy(batch).to('cuda'), where batch is a slice of a numpy ndarray and therefore not writeable? I guess it shouldn't matter since to('cuda') is making a copy, but I want to make sure.

Yes, I think you can ignore it