UserWarning: Legacy tensor constructor is deprecated.

[rafi-cohen]

🐛 Bug

Using the ToTensor transform results in the following warning:

/path/to/env/lib/python3.8/site-packages/torchvision/transforms/functional.py:135: UserWarning: Legacy tensor constructor is deprecated. Use: torch.tensor(...) for creating tensors from tensor-like objects; or torch.empty(...) for creating an uninitialized tensor with specific sizes. (Triggered internally at  /opt/conda/conda-bld/pytorch_1618124769378/work/torch/csrc/utils/tensor_new.cpp:477.)

According to the message, the following line triggers the warning:

vision/torchvision/transforms/functional.py

Line 135 in eee4249

img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))

Note that there might be more places where the legacy constructor is used.

To Reproduce

Steps to reproduce the behavior:

1. Simply apply the ToTensor transform

Below is a pure torchvision script generating the warning:

from torchvision.transforms import ToTensor
from torchvision.datasets import FakeData

data = FakeData(transform=ToTensor())
data[0]

Expected behavior

No warning

Environment

PyTorch version: 1.9.0.dev20210411
Is debug build: False
CUDA used to build PyTorch: 11.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 18.04.5 LTS (x86_64)
GCC version: (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
Clang version: Could not collect
CMake version: Could not collect

Python version: 3.8 (64-bit runtime)
Is CUDA available: True
CUDA runtime version: 9.1.85
GPU models and configuration: 
GPU 0: GeForce GTX 1080 Ti
GPU 1: GeForce GTX 1080 Ti

Nvidia driver version: 450.102.04
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A

Versions of relevant libraries:
[pip3] numpy==1.19.2
[pip3] torch==1.9.0.dev20210411
[pip3] torchaudio==0.9.0a0+e8c5e45
[pip3] torchvision==0.10.0.dev20210411
[conda] blas                      1.0                         mkl  
[conda] cudatoolkit               11.1.1               h6406543_8    conda-forge
[conda] mkl                       2020.4             h726a3e6_304    conda-forge
[conda] mkl-service               2.3.0            py38h1e0a361_2    conda-forge
[conda] mkl_fft                   1.3.0            py38h5c078b8_1    conda-forge
[conda] mkl_random                1.2.0            py38hc5bc63f_1    conda-forge
[conda] numpy                     1.19.2           py38h54aff64_0  
[conda] numpy-base                1.19.2           py38hfa32c7d_0  
[conda] pytorch                   1.9.0.dev20210411 py3.8_cuda11.1_cudnn8.0.5_0    pytorch-nightly
[conda] torchaudio                0.9.0.dev20210411            py38    pytorch-nightly
[conda] torchvision               0.10.0.dev20210411      py38_cu111    pytorch-nightly

Additional context

pytorch/pytorch#54414
However, this change is about to be reverted, here: pytorch/pytorch#55831
But I'm guessing that once pytorch/pytorch#55780 is resolved, the legacy constructors will be deprecated again.
Perhaps something like pytorch/pytorch#54142 should be done in this repository as well.

[NicolasHug]

It seems that the warning is raised in pytorch rather than in torchvision:

Triggered internally at /opt/conda/conda-bld/pytorch_1618124769378/work/torch/csrc/utils/tensor_new.cpp:477

So possibly, there's nothing to do on torchvision's side and we can just wait for pytorch to sort this out?

Perhaps something like pytorch/pytorch#54142 should be done in this repository as well.

I don't thing there are calls to torch.ToTensor or torch.Tensor in torchvision (apart in the tests)

[rafi-cohen]

Correct me if I'm wrong, but my understanding is that

vision/torchvision/transforms/functional.py

Line 135 in eee4249

img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))

is a legacy constructor call, which is to be deprecated.
The message indicates that this line, which is on torchvision's side, triggers the warning.

[NicolasHug]

It's possible, but then I don't understand why you mention ToTensor? Either ToTensor is called somewhere during torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes())) and then this is a pytorch issue, or the call to ByteStorage or ByteTensor is indeed deprecated, but that's not directly related to ToTensor as far as torchvision is concerned.

[rafi-cohen]

The offending line torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes())) is part of the implementation of torchvision.transforms.functional.to_tensor, which is called by ToTensor. Therefore, using ToTensor triggers a warning.

[NicolasHug]

Sorry I misread, thanks

[fmassa]

I've re-opened pytorch/pytorch#47112 (comment) to try to track this down, thanks for the discussion @rafi-cohen !

[fmassa]

If the PyTorch team doesn't plan on adding torch.frombuffer, then I propose that we replace the construct with numpy.frombuffer or just the np.asarray(pic), as I believe this is going to share memory and thus be efficient

[fmassa]

Self-assigning to me as someone in the team will be working on fixing this

[fmassa]

Fixed in #3813