pytorch - 💡(How to fix) Fix Expose internal `warn_if_not_writeable` on Python `torch.from_numpy` [10 comments, 3 participants]

Pitch

I’d like to propose exposing the existing internal warn_if_not_writeable flag on torch.from_numpy as a keyword argument, defaulting to True:

torch.from_numpy(ndarray, *, warn_if_not_writeable=True)

I have a use case where I intentionally want zero-copy aliasing from a read-only NumPy array, and would like to suppress the warning.

A per-call opt-out seems to me like the right thing here, especially since this is already exposed internally. If this is acceptable, may I open a PR for it (It would be my first contribution to PyTorch, I'll read all documentation etc.)?

Alternatives

Right now the alternatives are:

1. local warnings.catch_warnings() + filterwarnings(...), which is awkward
2. global filter, which is too broad and can hide a real bug at other callsites where the warning would be generated.

Example of warning ITW

import numpy as np
import torch
from PIL import Image

img = Image.new("L", (4, 1), color=7)
arr = np.asarray(img)
assert arr.flags.writeable is False
torch.from_numpy(arr) # warn_if_not_writeable=True) > this is what I'm proposing, to prevent the warning

Prints: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:213.)

cc @mruberry @rgommers @albanD